# Public development log ## Unreleased — Recover terminal startup after working-directory removal - Launch tmux control processes from a stable directory instead of inheriting a removable checkout. For new terminal sessions, windows and splits, enter the selected directory before login-shell initialization. This also works around an already-running tmux 3.7b daemon ignoring `-c` after its own working directory disappears. Existing panes and foreground commands are not reset. - Verification: 167 focused server tests, 251 desktop terminal/transport tests, desktop typecheck and production compilation, using development sources based on `c8f45fe4` plus this fix (not a release package). Isolated native Electron on macOS exercised production IPC, authenticated HTTP/WebSocket, PTY and tmux: new sessions, both split directions, new windows after directory removal, stdout/stderr, 4,000 output lines followed by stderr, and reconnect. - Boundaries: isolated credential-storage adapter; no OS Keychain acceptance, package installation, production-server replacement or release. The fix is server-side and does not silently repair the cwd of an existing running shell. This does not establish that every reported blank-terminal symptom has the same cause. Portable process regressions ran on macOS; live Linux acceptance remains untested. ## Unreleased — Account-scoped model inventories and explicit checks - Stop presenting historical Claude/Codex version fallbacks and saved chat IDs as current account inventory. Keep Claude discovery passive and let real SDK metadata share its cache with the picker despite chat-only scheduling limits. Follow the native account's default selector when reported, unless the server has an explicit configured override; do not assume the default is Sonnet. - Reconcile desktop catalog caches on reconnect; refresh custom Codex inventories with credential-scoped expiry and explicitly refresh OpenCode's CLI registry. Retain same-connection rows on transient failure with an unconfirmed warning, accept authoritative empty lists, and preserve native/default model semantics. - Raise Cursor API inventory capacity from 256 to 512 with partial-list reporting. Exclude known Gemini audio-only products from the text-agent catalog. - If OpenCode's explicit registry refresh fails, read the same CLI's native registry within the existing deadline and label freshness unconfirmed. Do not substitute another account's cached rows or retry an authoritative empty list. - Add a native-only, revision-fenced optional text check for Claude/OpenCode/Gemini API models, with English/Chinese UI and usage disclosure. It does not change defaults, credentials or chats. Quota/auth/network errors do not prove retirement; short-lived rejection observations are not a permanent model blacklist. - Verification: 307 focused server tests; desktop typecheck, production compile and 580 focused tests including locale terminology checks. Isolated native offscreen Electron exercised real settings reads, default selection, reload/reopen and explicit Gemini/OpenCode requests through production IPC/native HTTP/server authorization. A Gemini inventory entry rejected the request while another succeeded; connection readiness stayed unchanged. English/Chinese, dark/light and narrow layout were inspected. Repeated 24-read concurrency batches returned complete backend catalogs. Live Cursor inventory exceeded the previous cap. Account-switch/failure races used fixtures. - Full chat acceptance: isolated native offscreen Electron sent a disposable prompt through each real Claude, Codex, Cursor, OpenCode and Gemini runtime; all five returned the expected answer and completed successfully. Fresh Claude SDK initialization populated the picker with native current and older models. These checks used an accessible, clean workspace without project overrides. - Limits: not every model was invoked; text success does not establish full agent compatibility. No second live account was available for an account-switch test. No mobile changes, packaging acceptance, public release or production-server deployment is included. ## 2026-11-09 - Publish coordinated 1.0.12-beta.6 - Publish desktop build `1275`, the signed server runtime, short CLI and legacy bridge at product source `0e46b3f81e98596da3d91c73137f89efeee05686`. Beta.6 includes admitted-update restart-gap recovery, credential-free update checks and legacy-server file-descriptor headroom. Stable remains `1.0.11`; the unpublished beta.5 candidate is unchanged. See [release notes](RELEASE_1.0.12-beta.6.md). - Native preparation and independent verification passed on all four desktop platforms. Matched CLI lifecycle acceptance passed on macOS and Linux. [Mac upgrade/recovery checks](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/42534945814) passed all eleven cases across retained attempts, including public beta.4 and stable `1.0.11` with both legacy root modes. Both beta.4 cases needed explicit recovery after server completion, not a seamless first attempt. - [Linux rollback/retry](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/42534947832) passed in both root modes using the same signed packages. Original artifact digests, source pins and scoped observation flags were verified and retained. - Separate isolated provider checks retained Codex/Claude context, queued work and a persistent goal across the signed server update. Packaged desktop replacement with active providers, complete history rendering, reboot, natural OAuth renewal and the native Windows updater are not established. [Claude history duplication #166](https://github.com/ZhengyiLuo/AgentsDock/issues/166) remains a documented beta limitation. Windows previews remain unsigned. - [Matched npm publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/42537649524) verified both exact archives. A fresh scripts-disabled registry installation verified `agentsdock@beta`, its exact runtime dependency and CLI help without operating services. Both npm `beta` tags select `1.0.12-beta.6`; both `latest` tags remain `1.0.11`. The matching signed legacy bridge is public and verified. - [Desktop publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/42538264142) reverified all four platforms and published both download repositories. Each public copy has the same 16 asset digests, signed metadata, update feeds and known-limit notes. All stable defaults remain `1.0.11`. Public feed verification is separate from the exact-artifact native rehearsals; no additional installed public-feed upgrade is claimed. No mobile build or upload is included. Later documentation commits are not new artifact sources. ## Unreleased — Preserve both legacy update-recovery receipt fields - The separate legacy-recovery rehearsal accepts a completed schedule receipt when a lost restart response left no operation receipt. Require at least one valid receipt, and preserve both identifiers and their namespaces exactly across native replacement. No identifiers are included in public diagnostics. - Keep the exact failure, identity, current/unpaused server, candidate bytes and no-additional-restart gates. This corrects a harness false refusal; it is not a new product retry or a claim of seamless upgrade or release acceptance. ## Unreleased — Match beta.4's native coordinated-update acceptance contract - Keep the signed beta.5 candidate unchanged while testing beta.4's shipped server-first update behavior: observe offline refusal, reconnect, explicitly retry, and require the paired server ready before native app replacement. - Preserve the older baselines' app-first offline recovery and shared-operation checks. The beta.4 case labels its later two-client state as convergence rather than claiming both clients observed a live update operation. - Add focused refusal, version/order, component-health and client-convergence regressions. Hosted native execution is still required; this entry does not establish successful acceptance or publication. - Retain bounded, read-only native updater-state comparisons during replacement failures so server preparation and desktop quit failures are distinguishable. Raw messages, process details and private logs are not included in evidence. - Keep a transient-503 legacy recovery experiment separate from seamless-upgrade acceptance: require the exact current server and unchanged candidate before one additional explicit UI retry, and prove no second server restart occurs. - Derive the exact displayed-failure fixture from the native client's message composition, without inventing punctuation. Count the additional recovery click in both the detailed checks and console summary; display classification does not establish the unobserved HTTP response body or content type. ## Unreleased — Keep admitted server updates alive across gateway restart gaps - Keep an app update waiting through a generic HTTP 503 only after a verified, exact-target server update has been admitted. Require an explicit active receipt; completed, canceled, failed and paused history does not authorize this recovery. Identity, API, signature and structured-error checks remain. - Preserve the existing bounded wait and restart request identity. Do not reinterpret an initial admission failure or repeat a destructive restart. - Regression: 168 focused coordinator/updater tests, type checks and Electron production build pass. An isolated offscreen native UI exercised the actual Update button, production IPC/coordinator, authenticated managed server admission, signed package verification, real server runtime replacement and reconnection. The app remained installing across a gateway-shaped 503 and continued only after the original server identity returned at the target version. - The first local attempt correctly reported a test download-environment failure; an explicit retry passed after correcting that isolated fixture. Release discovery/download and the final native app-replacement handoff were simulated. Signed desktop replacement and live-provider work are not claimed by this test; release acceptance remains separate. Source change only. ## Unreleased — Retry transient notarization-ticket validation failures - Extend release orchestration's bounded Apple ticket-download recovery to ticket validation and final package verification. Retry only explicit CloudKit connection-loss or timeout errors after accepted outer-DMG notarization. - Keep the frozen product builder, signed payloads and complete native verifier unchanged. Invalid signatures, tickets, metadata and smoke checks still fail; early app-stapling failures are not treated as completed-package validation. - All 433 release-tooling/CLI checks pass, including 14 focused regressions for recovery boundaries, bounded retries, immutable assets and non-transport failure rejection. Hosted native acceptance remains required; this entry does not establish a successful package or publication. ## Unreleased — Prepare exact beta.5 upgrade rehearsals - Add independently verified public beta.4/build 1272 and retained stable 1.0.11 baselines for the exact beta.5 candidate. Exercise both legacy root permission modes; stable users must explicitly opt into Beta, and existing Beta subscriptions are retained. Earlier candidate pins remain unchanged. - Keep the complete reviewed installer hash and all source, signature, native runner and non-publication guards. This changes test configuration only, not application/runtime behavior, release versions or published packages. - Three new admission regressions fail before the change. All 113 Node and 98 Python harness tests pass, including wrong-version, substituted-input and cross-scope rejection. Independently downloaded beta.4 desktop and signed official-registry runtime bytes pass the new baseline verifier. - This is preparation, not native beta.5 acceptance. Exact signed-package install/relaunch, recovery and real active-work checks remain pending. No service deployment or release publication is established by this entry. ## Unreleased — Reconcile beta updater fixes with main - Bring the already-shipped beta.4 stale-update reconciliation fix and bounded recovery diagnostics into main, retaining the later isolated update-check session, pinned CI download fallback and legacy-service resource-limit fixes. This does not change release versions, feeds or published artifacts. - Verification on integrated source `218b7151`: 243 focused server tests, 97 acceptance-harness tests and 91 replay/candidate/download tests pass; TypeScript checking, Electron production compilation and output validation pass. The added regression methods fail against the pre-backport server. - Two isolated servers preserved newer durable updater progress through real authenticated status requests and rejected each other's credentials. Only the failed updater-liveness probe was synthetic; no installer was started. Actual mouse input in an isolated offscreen Electron app exercised Settings, both release channels and repeated public update checks. Stable and Beta returned their published versions without freezing, and both test server versions remained visible. The development app correctly did not self-update. - Acceptance limits: this verifies source integration and status reconciliation, not native update installation, restart or active-work recovery. Those flows still require an exact signed candidate and managed-service rehearsal before release acceptance. No deployment or publication is part of this backport. ## Unreleased — Isolate update checks and legacy-service resource limits - Public release checks and paired-server metadata fetches use a non-persistent, credential-free Electron session. They no longer open the browser cookie encryption store, which can stall behind macOS Keychain access. Existing deadlines, size bounds, channel rules and signed-descriptor checks remain. - The legacy managed-server entrypoint now applies the same file-descriptor headroom as the split worker/gateway. Provider children inherit the raised soft limit without changing the hard limit or lowering an existing limit. - Verification: the new regressions fail before their fixes. Eight additional cases cover HTTP 504/rate limiting, offline retry, invalid UTF-8, and failed, oversized or incorrectly signed paired metadata without starting installation. Desktop updater, coordinator, destination and shutdown suites pass (180 tests, two skipped). The full desktop suite passes (5,957 tests, six skipped), plus eight build/license checks, all 13 execution-service tests, TypeScript checking and production compilation. An ad-hoc packaged native app exercised update checking and cancellation. A managed test instance completed an authenticated published-package upgrade, retained its state, then ran the patched startup and a real Claude reply in an isolated working directory. Repeated real desktop update checks report the HTTP failure without freezing and keep manual retry available. These source changes are not part of a published release. - Source CI can recover an upstream Electron download 5xx/transport failure through the official GitHub asset API, keeping the exact npm-pinned SHA-256, original installer and complete validation. It does not change product feeds, publish artifacts, or accept unknown versions, invalid hashes or API failures. Real macOS and Linux archives passed pinned-hash verification; the Linux probe exercised the API fallback. All 419 release-tooling/CLI checks passed. - Acceptance limits: public release download URLs returned HTTP 504 independently of Electron, so signed desktop download/install/relaunch remains unaccepted. The ad-hoc package still encountered a separate browser-store Keychain wait during quit; isolating release requests does not establish native restart safety. An existing external-volume working directory also failed in the managed service context; the successful provider check does not resolve that boundary. Active-turn continuation across an update was not exercised in this run. ## Unreleased — Preserve updater progress during stale status checks - Recheck the full update receipt under its file lock before finalizing an apparently abandoned updater. New heartbeats, process/phase changes and recovery ownership now preserve the live operation and its maintenance hold. - Update check/start requests keep a refreshed active operation instead of replacing its status or launching another updater. Unchanged dead-runner recovery and retained activation-journal safeguards remain intact. - Verification: 227 focused update, endpoint, runner and recovery tests pass, including real receipt writes interleaved with simulated failed liveness probes. Python compilation and diff checks pass. Native app/service update acceptance and exact release-package validation are still pending; this is source-only, not a published or deployed fix. ## 2026-11-08 - Clarify failed update rehearsal diagnostics - Preserve fixed failure reasons for an exited updater, execution-hold cleanup and activation recovery in bounded release-test reports. Keep raw messages and private logs out of uploaded evidence. - This changes test diagnostics only. Signed app/server packages, updater behavior and pending release acceptance remain unchanged. - Validation: 97 harness tests and 68 candidate/replay guard tests pass. A fresh exact-package hosted replay is still required; no native pass is claimed here. ## 2026-11-08 - Prepare coordinated 1.0.12-beta.3 - Integrate main through `5de30b23` into the published beta.2 release line, retaining previously shipped runtime, update, mailbox and history fixes. Prepare one desktop/server/CLI beta version; stable remains `1.0.11`. - Combine native goal lifecycle filtering with update-continuation provenance, retaining imported and explicitly user-authored messages. Preserve the resumed goal's run-bound tools and durable update recovery boundary. - Isolate update endpoint tests from a developer's managed-install environment; package preparation retains its separate tests and real native acceptance. - Preserve keyword-only defaults when rebinding the isolated route-authentication test functions, so mailbox authorization exercises the production defaults. - Add exact beta.2/stable 1.0.11 upgrade profiles and the reviewed beta.3 installer pin without changing historical artifact identities or relaxing runner, signature, ownership or rollback guards. - Desktop type checking/build and 5,948 tests pass (six skipped), plus eight packaging/license checks. The 117 update endpoint tests and 277 focused timeline/coordinator tests pass. Native release acceptance remains pending; see [candidate notes](RELEASE_1.0.12-beta.3.md). No publication or deployment is established by this preparation entry. ## 2026-11-07 — Complete isolated coordinated-update source acceptance - Source `75da4096` passed actual signed managed-service replacement with two isolated services, real Claude/Cursor work and native Codex goals. The review fixed a first-turn Cursor recovery hole: updates now wait for its native session to be durable rather than restarting into an empty conversation. - Rechecked invalid signatures, corrupt archives before takeover, partial success and retry, and failure after takeover. The latter restored the old healthy service and cleared the failed stage/fence; the same package then installed successfully on retry. Existing identities/history were retained, including installations with legacy 0755/0750 root permissions. - Through actual native app input, two private packaged update cycles exited, installed and reopened automatically. The last cycle updated both saved services before replacing the app, resumed Cursor once, and waited safely for Claude's independent background command. Draft, connection, language and channel settings survived. Rechecking did not reinstall current targets. Offline preflight kept the old app and a persistent actionable error. Detached terminal work survived rollback/retry and reconnected afterward. - Verification: all 11 hosted checks passed on the code head; 96 focused restart/recovery and 254 update/installer/handoff tests passed. Desktop verification also passed 310 focused tests, type checking and production compilation. Final native history showed one interrupted Cursor turn and one successful continuation, with no internal continuation user bubble. - Test surface: isolated, focus-emulated native Electron with real IPC, authenticated HTTP, providers, service manager and native updater. Private QA feeds/packages used ephemeral server signing and an Apple Development- signed shell containing the branch's compiled desktop code. Production signing/notarization, platform-specific release installation and supported old-client feed migration remain release gates; they are not established by this local acceptance. No production server was changed and nothing was released. See `docs/COORDINATED_UPDATES.md` for the exact boundaries. ## Unreleased — Wait for Cursor's first durable native session before updating - A Cursor thread identifier is persisted only after a successful native reply. Its first running turn now waits for that safe boundary instead of being interrupted into a new thread with no original task context. Later turns with a durable native session remain eligible for update continuation. - Verification: 96 restart/recovery tests pass, including a live-but-not-yet- persisted Cursor identifier and the subsequent durable-session case. Managed-package and native desktop acceptance remains in progress below. ## Unreleased — Avoid double-counting a resumable native goal during update - The live managed-install replay exposed a native goal's consumer being counted both as an active chat and as independent background work. Deduplicate only the exact, still-owned active goal consumer. Unrelated operations, changed ownership, prompts and child agents remain restart blockers. - Verification: 77 restart/admission tests pass, including the actual native operation registry in the goal-checkpoint test and stale-owner regressions. - An isolated two-service managed-package replay exercised legacy installation roots with modes 0755 and 0750, invalid signatures, a corrupt archive before takeover, partial success and retry. The failed service stayed healthy; retry did not restart the already-updated service. A real Claude task recovered and completed before its retained queued follow-up, each once. These packages use a private QA signing key and loopback feed, not production release signing. - The subsequent native Codex goal replay passed a real managed-package update: one pause and one recovery, unchanged native goal identity, retained usage, and completion of the existing file checkpoint. It exposed an internal goal resume notice rendered as a user message; the desktop now hides only the server-marked update recovery boundary, retaining user text and the reply. Verified that correction in the isolated native app against the real events. - Follow-up checks pass 269 server tests and 310 desktop tests, TypeScript and production compilation. Native desktop automatic exit/install/relaunch is still pending; private QA signing is not production release acceptance. ## Unreleased — Preserve native goals through coordinated server restarts - Extend update-continuation capability to native Codex goals. Save the exact thread/goal identity, pause before provider teardown, and resume the same goal without resetting its token or time usage. Completion, replacement, exhausted budgets and a later manual pause/clear take precedence over recovery. - Preserve the original interactive transport for ordinary continuations. Claude SDK runs must confirm native Stop before recovery; closing the transport alone is not a safe stop. Bind Claude recovery to its current native goal, excluding earlier replaced goals, and show a recoverable error if stopping cannot be confirmed. Internal continuation prompts stay out of user bubbles. - Pause independent chat owners concurrently within the bounded shutdown phase. A prepared update with a durable continuation receipt receives the cooperative shutdown budget, not the emergency three-second kill deadline. Keep emergency restart behavior and the installer's bounded wait aligned with shutdown. - Verification on review base `2604af6a`: 380 focused server regression tests pass, including lifecycle/queue, native goal identity and cancellation, concurrent stop, transport retention and both restart signal paths. A separate 110-test goal-admission/history-repair run passes. Desktop type checking, production compilation and 306 focused updater/timeline tests pass. - Real acceptance used an isolated native Electron app (offscreen/focus-emulated) with authenticated local servers and real Codex and Claude providers. Goals were created through the UI, interrupted, resumed after server restart, and completed their file checkpoints. Codex retained native usage; a manual pause prevented recovery. Claude resumed only the current goal, leaving old goal checkpoints untouched; another restart produced no second continuation. Interactive approvals were still required and usable after recovery. A further Claude run exercised the production restart signal/watchdog code using a private fixture trigger and completed the current goal after relaunch. - Test boundary: installation receipts were fixture-generated; source servers were restarted, not upgraded through a signed package/service-manager install. Current-source signed multi-server replacement, offline/partial-install failure and native desktop quit/install/relaunch remain pre-merge gates. Unsupported active jobs, standalone tasks and terminals still wait for a safe boundary. No production server was restarted and nothing was released. ## 2026-11-06 — Refresh the coordinated-update review against current main - Integrate main through `eac24fca`, preserving both native background-wake and server-update continuation filtering in the desktop timeline. - Add two- and twelve-profile coordinator regressions: every exact server reservation receives one continuation-enabled restart request, and desktop installation remains gated until the last server is verified current. - Verification: 443 focused desktop tests, 207 update/restart/recovery server tests and 74 provider-history repair tests pass. Desktop type checking, production compilation and compiled-entry verification pass. The server suites run with isolated configuration and state, without inherited managed-installation settings. Two real disposable servers were reached through a native development app; its Settings listed both, and separate Claude and Codex conversations each received their reply. Local installation discovery is isolated in this private harness; native IPC, authenticated HTTP, providers and chat persistence are real. - Remaining gates: current-source signed multi-server replacement, busy-turn recovery, offline/partial-failure retry, and automatic native app quit and relaunch. This review does not establish universal task checkpointing or a global pause across servers. Source/local development only; no release or production app/server replacement. ## Unreleased — Use one product update control in desktop Settings - Remove the separate manual AgentsServer update controls from Settings. The Updates page keeps one Stable/Beta choice and one AgentsDock installation action while listing each server's version and paired-update progress. A failed paired update is retried from one product-level action; server restart remains available on the Server page. - Verification: 236 focused desktop UI/updater tests, TypeScript checking, production compilation and the compiled-entry audit pass. An isolated native Electron development app used its actual Settings controls and authenticated client connection to a disposable local server: the server version appeared, the product channel was shown once and no separate server update panel appeared. A separately identified local QA app containing this compiled UI was also opened and operated through Settings: Updates showed one product channel and action, the Server page had no manual update panel, and saved server versions remained visible. This QA bundle reuses a test shell and is not a release-grade package. The final UI has not been exercised through a packaged update installation. Availability: source and local QA only; no public release or production app/server replacement. ## 2026-10-20 — Sync the coordinated updater with current main - Merge main `b21ef7ea` into the unreleased coordinated-update branch, retaining both the server-first update work and intervening desktop/server fixes. - After integration, 311 focused desktop tests and 206 focused server tests pass with isolated test state. Desktop TypeScript, production compilation, Python compilation and diff checks pass. No native update package or public release was produced by this source integration. ## Unreleased — Prepare paired servers before replacing the desktop app - A downloaded app update now verifies its matching signed server descriptor, checks saved server identities, and waits for every paired server to reach the target before invoking the native installer. The app blocks chat controls while updating and reports an incomplete update on failure. - A capable server can activate its exact queued update immediately. Ordinary Codex, Claude and Cursor chats interrupted by that update receive one durable internal continuation after restart. This does not checkpoint the exact in-flight provider operation; already-completed side effects can be repeated. Goals, other internal work and unsupported runtimes still wait for a safe boundary. - A live forced-update check found that graceful shutdown can write a stopped terminal event before the replacement process scans abandoned turns. Save the active ordinary run identities in the restart receipt and restore their internal continuation after a matching signed update, including when the update has already completed before startup recovery. Completed turns stay complete; a second startup does not queue another continuation. Real-provider testing confirmed one continuation was queued and launched, but the first continuation only reported an interrupted command instead of finishing its remaining work. The internal instruction now explicitly asks the agent to complete unfinished steps, including a necessary retry after checking effects. A later disposable, signed beta.10-to-beta.11 update interrupted a real Codex command. The replacement server retained identity, queued and completed exactly one internal continuation, and the file contained one start and one finish marker. The local app displayed the resumed answer after reload, and a follow-up sent through that app completed on the same chat. - Validation: 151 focused server tests, 71 focused desktop tests, TypeScript and the desktop production build pass. A disposable named server was upgraded with a signed unpublished candidate and completed an ordinary Codex turn. No public release was changed. - A compiled local Electron app from this branch connected to an existing test server through the real client/server transport. A disposable Codex chat and follow-up both completed, and the Updates page resolved the published Stable and Beta feeds. Equal-version and local-suffix preflight behavior has focused regression coverage. The development app cannot install unpublished updates through its public feed; the separate signed package test exercised active-chat continuation through the real server. - An isolated native update replay exposed an early forced-restart request while the server package was still staging. The coordinator now waits for an explicit ready state before requesting activation; a regression test covers staging followed by readiness. - A later isolated replay activated the matching server but did not complete native app replacement. The coordinator now keeps the same user update action alive across the expected server restart connection gap, including an ambiguous restart response. A later isolated macOS native replay with signed, notarized, unpublished beta.6 and beta.7 packages passed the actual one-click journey: the server reached beta.7 before app replacement, the app relaunched at beta.7, and server identity and a saved chat survived. See [native replay](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/39105335755). The replay used a private exact-package test feed and a disposable server; it did not exercise a real provider turn during the update, busy goals, multiple saved servers, or a production release. ## 2026-10-21 — Replay a private busy-chat update - An explicit retry after server preparation failed could reattach to the old failed schedule and fail again without attempting a new preparation. The main Update action now clears that operation ownership on retry while retaining the exact signed target. A regression failed before the fix; 135 focused desktop tests, TypeScript checking and the production compile pass. - An unpublished, isolated macOS QA app and signed server package replayed a beta.10-to-beta.11 update through the real Update button. The server completed first; the app then replaced itself, relaunched and reopened the same chat. A live Claude turn interrupted during the server restart received one durable continuation, produced a final answer on the new server and remained visible in the app. The continuation was marked system-generated and not user-authored. A separate failed-preparation replay verified that the same main button could retry successfully. - These QA app packages use a separate bundle identity, private app data and loopback-only feeds; they were signed for local development, not notarized or published. Busy Codex/Cursor turns, non-chat blockers, multiple clients and rollback still need native acceptance before a product release. ## 2026-11-06 - Improve terminal feedback and bound bulk output - Feed received output directly into xterm's existing input-aware parser instead of imposing another animation-frame delay. Keep port discovery frame-batched with a bounded tail. Remove the redundant tmux existence check before exiting application-owned history scrolling on the next keystroke. - Add conservative visual-only typing prediction after distinct ordinary characters have echoed correctly in the current burst. Predictions occupy blank cells only and retire on mismatches, idle expiry, command/edit keys, paste, composition, scrolling, selection, resizing, navigation and reconnect. They never change terminal history, clipboard contents or transmitted input. First/unconfirmed input, Unicode, line wrapping and unsupported input modes continue to use authoritative output; this is not a universal shell-mode or password detector. - Negotiate optional ack-v1 output credit with matching desktop/server builds. ACK original PTY byte offsets only after xterm parses them, fence callbacks by server generation and socket identity, and bound unconsumed output to a 64 KiB window plus at most one read. Input remains independent of output credit. Older clients/servers retain the existing streaming path. - Source base `69eb0f8a`; development desktop 0.2.0/build 86, Electron 43.1.1, server metadata 1.0.7-beta.11. All 280 selected desktop checks and 54 server regressions, desktop typecheck and production compilation pass. Native hidden-window Electron acceptance uses real mouse/keyboard, IPC, authenticated WebSocket, PTY and tmux: ordinary typing, exact echo reconciliation, a non-echoing test password, Unicode paste, backspace/arrows, Vim entry/exit, 10,000-line scrollback, bulk output/Ctrl-C, reopen/reconnect and light/narrow presentation. An additional 80 ms renderer-ingress delay exercises predictions without changing transport authentication or provider behavior; it is not a WAN or paint benchmark. - A further visible native-app pass repeats five fast-typing bursts and three scroll-then-input cycles, edits pending input, and interrupts bulk output. Confirmed output remains exact, with no duplicate characters or stale prediction overlay after settling. No synthetic ingress delay is used in this pass. Repeated bulk-output interruption checks also guide the smaller credit window, reducing queued parsing before the next command is visible. - Real transport acceptance verifies output stops at its credit limit, ignores invalid ACKs and resumes; Ctrl-C remains accepted while credit is exhausted. No provider request is needed. No Windows/Linux native, real WAN, or exhaustive shell/password-program acceptance is claimed. Source and local test runtime only; no release, mobile change or production-server deployment. ## 2026-11-06 - Clarify Claude context usage and recover failed reads - Replace the ambiguous context-meter “Unavailable” label with an explicit missing-measurement state. Localize the usage details, loading, next-response guidance and retry hints in English and Simplified Chinese. No token counts or context limits are estimated when Claude supplies no valid measurement. - Surface runtime-read errors in the meter instead of silently displaying an empty state. Retry failed reads through the observational runtime endpoint, including when the cached state says active; never sample the SDK alongside a running turn. Successful runtime responses clear obsolete read errors. Keep last-known usage on manual sampling failures and prevent duplicate clicks during initial loading or refresh. - Source: desktop-only correction on main `81e6e75c`; development package version 0.2.0, Electron 43.1.1, unchanged test-server version 1.0.7-beta.11. Five new focused regressions fail against the previous implementation. All 96 selected Claude-control, context-usage, localization and design tests pass; TypeScript and production Electron compilation/output verification pass. - Exercise an isolated native offscreen Electron app with mouse and keyboard: unloaded-chat guidance, real Claude messages and follow-ups, automatic usage after completion, active-turn refresh blocking, manual refresh, navigation, Chinese copy, light/dark themes and a narrow viewport. Keyboard refresh advances the server's usage generation through real IPC, authenticated transport and the native Claude SDK. Inject a runtime-read failure only in the test harness, then remove it and recover with the real retry action. Credential storage and the failure injection are fixtures; normal provider responses and context measurements are not mocked. Analytics is disabled. - The reporter's custom endpoint and app/server versions remain unknown, so persistent missing telemetry specific to that endpoint is not reproduced. No server/mobile code changes, production restart, release or deployment. Available as local source/build only. ## 2026-11-04 - Keep delegated worker tasks separate from creator orchestration - Follow up PR #145 integration on `e435f079`: clarify the provider policy, idle mailbox wake and busy-run notice. The original user instruction bounds authorization; the prepared message identifies the recipient's subtask. A worker asked to calculate and reply must not repeat its creator's setup or create another worker. Requested replies use Chats, not a local final answer. - Clarify that accepted-send fields are snapshots, not live execution status. An empty inbox is not evidence that the peer has not started or read its task. No delegation is promoted from body text, copied to outgoing peer messages, or expanded into new permissions. Receipt formats and authorization checks are unchanged. - Add policy regressions that fail before the clarification, plus a real-ledger check preserving the recipient task separately from its exact user scope. All 598 focused server tests and 493 desktop service tests pass; TypeScript, production Electron compilation and compiled-output verification pass. - Exercise the native offscreen Electron app through actual mouse/keyboard, production IPC and authenticated test-server transport with real Codex and Claude. Each provider completes initial and repeat creation/calculation/reply workflows: four children total, one task and reply per child, no descendant creations, correct settings inheritance and parent result/read receipts. Verify both rendered mailbox cards and replies, including return navigation and Jump to latest. Only isolated credential storage is a fixture; provider generation and mailbox transport are real. - Development preview metadata remains desktop 0.2.0 / server 1.0.7-beta.11. The local test server runs the correction; no production restart, deployment, release or mobile change. Provider instructions guide task interpretation; they are not a new enforcement or transitive-delegation mechanism. Other providers and packaged Windows/Linux acceptance are not retested here. ## 2026-11-04 - Integrate agent-created chats for local acceptance - Integrate PR #145 on main `413e6fa7`, retaining the original contributor commit `b8bf3b0` and both development-log histories. Independent chats inherit applicable settings and receive an exact creator/child mailbox pair without copying native conversation identity or history. Creation alone stays idle. - Correct a permission-ceiling regression found by the existing Claude native continuation tests: capability issuance no longer adds chat creation to explicitly restricted actions. Normal admission and steering opt in; a server without authentication cannot grant creation. Keep the original strict continuation assertions and add explicit empty/restricted-ceiling regressions. - Verify failed acceptance recovery, cancellation settlement, same-key retries, different-setting rejection, denied creation without mutations, and revoked pairs not being silently restored. All 472 focused server tests pass; 493 desktop service tests, TypeScript, production compilation and compiled-output verification pass. The two original continuation failures reproduce before the correction and pass afterward without weakening their assertions. - Exercise actual native offscreen Electron, production preload/IPC and the authenticated test server with real Codex and Claude. Create independent children, confirm inherited folder/runtime settings and no initial turn, send one task through each new route, observe each idle child wake and return exactly one reply, then observe each creator wake and read its reply. Mouse navigation confirms the new chats and mailbox cards in the existing UI. - After an idle test-server restart, both providers' fresh same-key creation calls return the original chat and route with `created: false` and `duplicate: true`. An unsent draft survives the server restart and renderer reload. Identical same-run Claude mutations replay their original receipt through the existing tool replay guard, rather than creating another chat. - Restart the visible local preview on the integrated source and preserve its existing draft. Development desktop metadata remains 0.2.0; server metadata remains 1.0.7-beta.11. These identify local source previews, not new releases. Only credential storage is a fixture in the isolated client. Custom endpoint inheritance, denied requests, cancellation and revoke failure cases have automated coverage, not all forced live-provider scenarios. Other providers and packaged Windows/Linux acceptance remain untested. Local only; no remote push, PR merge, production deployment, mobile change or release publication. ## 2026-11-04 - Keep catalog concurrency checks independent of runner pauses - The first fixture synchronization fix did not resolve full CI by itself. Subsequent logs show the first HTTP test task stalling its event loop for 6.389 seconds, beyond the old five-second test watchdog. Injecting a matching 6.5-second stall locally reproduces that failure. - Use the existing neighboring concurrency test's 30-second bounded watchdog for entry, join and synthetic-worker release. This test checks shared-work ownership, not request latency: retain all 24 overlapping requests, one probe per provider, identical responses, cleanup and single-worker regression guards. Production discovery budgets and transport timeouts are unchanged. - The injected-stall case now passes; another 20 repetitions plus the complete module pass (34 tests). Re-run the full shard and CI before merge. No running app, provider behavior or release artifact changes are involved. ## 2026-11-04 - Avoid worker starvation in the runtime-catalog concurrency test - The connection/sync integration's full CI repeatedly timed out in an unchanged catalog test. Its blocking entry observer used the same default executor that scope reads and catalog discovery needed. A one-worker executor reproduces the fixture deadlock deterministically, even though ordinary local runs pass. - Signal an asyncio event from the synthetic provider workers instead of using another worker to observe them. Deliberately constrain the test's default pool to one worker. Preserve the five-second entry watchdog, all 24 overlapping requests, exact per-provider call counts, response equality and cleanup checks. - The constrained case fails before the change and passes afterward. Twenty repetitions plus the entire catalog-concurrency module pass (34 tests). This changes test synchronization only, not provider discovery, authentication, release artifacts or the running local app. ## 2026-11-03 - Extend connection and background-sync acceptance - Revalidate integration `d73e5748`: 772 focused desktop tests across service, transport, persistence, settings, server selection and locale contracts pass; 47 focused server tests and TypeScript pass again. No functional source change is needed after these checks; the previous production compilation still applies. - Use an isolated native offscreen Electron profile, production IPC and real authenticated Claude and Codex turns. Confirm hidden completed replies reach the semantic cache before opening their chats, repeated navigation retains chat-owned drafts, and replies remain distinct without cross-chat leakage. Scroll the longer Claude history to verify earlier replies remain available. - Interrupt only the isolated client's real TCP connection through a transparent test proxy. Cached replies remain readable and the draft survives offline navigation. Restore connectivity and open a cached chat: the online indicator returns in about 0.3 seconds in this local run, without waiting for the regular poll. This timing is an observation, not a remote-network latency guarantee. - Disconnect again after a real Claude turn is accepted. The server completes that original turn while the client is offline; reconnecting prefetches the complete result without a second execution or duplicate completion. A later real Codex request succeeds. Fully quit and relaunch the isolated app: the selected chat, unsent draft and cached replies persist. Only credential storage is a fixture; the proxy forwards unmodified traffic and injects socket loss. - The user's visible preview and the server stay running throughout these fault tests. Slow startup recovery and failed-readiness presentation remain automated regression coverage; packaged Windows/Linux acceptance is not claimed. Local verification only, without a remote push, merge or release. ## 2026-11-03 - Integrate connection recovery and background reply synchronization - Combine PRs #143 and #151 on main base `897a40e2`, preserving both original contributor histories, including JiaCheng310's background-prefetch change. Code merges cleanly; retain both development-log entries. Correct the Chinese recovery label to use the established term “会话”, fixing the original terminology-contract failure without changing behavior. - On integration commit `6e91eef6` plus that terminology correction, 699 targeted desktop tests, 47 focused server tests, TypeScript, production compilation and compile-output verification pass. Development app metadata remains 0.2.0; this is not a published version or an updater candidate. - An isolated native offscreen Electron app exercises production preload/IPC, authenticated test-server transport and two real Claude turns. Leave the responding chat, observe each completed reply in the local semantic cache before reopening it, then return and reload. Both replies appear exactly once, the contextual follow-up succeeds, and normal server refresh remains online. Only credential storage is a fixture; the provider and server are real. - Restart the visible local preview on the integrated source, retaining the existing profile and checking draft preservation. The updated test server is required for the normal-disconnect handling. Long recovery and failed-recovery presentation, late responses and retry ownership are covered by focused regressions, not claimed as deliberately forced live scenarios in this run. Packaged Windows/Linux and real network-loss acceptance remain untested here. - Local compiled preview only. No remote push, PR merge or release publication. ## 2026-11-03 - Present Claude compact hook output as command diagnostics - Follow up on the Claude reliability integration below: native `/compact` output from PostCompact hooks is command text, not Markdown. Show a concise completion label and a disclosure containing the unchanged native output. Only a recognized successful single-hook envelope with empty output starts collapsed; failures, warnings, non-empty output and unknown envelopes remain expanded. Copying and stored history retain the complete original text. - Apply this presentation only to source-identified Claude output paired with its exact same-session, same-run `/compact` input. Ordinary messages, user quotes and other providers keep their existing rendering. Reconciliation updates the row if the owning input arrives later. - On integration base `503ae3aa`, 378 targeted desktop tests, 483 focused server tests, TypeScript, production compilation and compile-output verification pass. The new rendering regression fails against the previous renderer. The follow-up changes desktop presentation only; it adds no server mutation. - An isolated native offscreen Electron app exercises production IPC and real authenticated server/Claude Code 2.1.304 requests: compact, contextual follow-up, mouse and keyboard disclosure controls, English/dark and Chinese/light narrow layouts, and reload. Native diagnostic bytes remain intact, shell text no longer becomes mathematical notation, and the remembered phrase survives compaction. Only credential storage is a fixture. Hook failures are covered by automated tests rather than changes to the user's hook configuration. - Local compiled preview only. No release, remote push or PR merge. Real print transport, forced connection-race timing, credential rotation and a full 24-hour cache interval are not claimed as live tests in this follow-up. ## 2026-11-03 - Integrate the Claude reliability review batch for local acceptance - Combine PRs #150, #147, #148 and #149 on main base `f1af8090`, retaining their original contributor commits. Code merges cleanly; development-log conflicts retain both entries. No desktop or mobile source change is needed. - On integration commit `5433f8d0`, 483 focused server tests, desktop TypeScript, production compilation and compile-output verification pass. These changes require the updated server; rebuilding the desktop alone is insufficient. - An isolated native offscreen Electron app uses production preload/IPC and authenticated test-server transport with real Claude Code 2.1.304. Exercise message, native-palette `/compact`, contextual follow-up, a second `/compact`, renderer reload and explicit provider-history refresh. Both commands complete without the false empty-response error, the follow-up retains the test phrase, and history refresh adds no duplicate user or assistant content. - Native model names remain present after more than five minutes and a real Settings reload. A structured mention of a disposable target authorizes a real provider-tool inbox request successfully; the same request without a scoped mention is denied. Temporary test authorization is removed afterward. Deliberate connection-race timing and credential rotation are covered by automated regressions, not claimed as forced real-app scenarios in this run. - Existing local PostCompact hooks append lengthy command diagnostics to the native completion output. The output remains verbose in the desktop; this integration does not claim to improve that presentation. Print-transport and a full 24-hour cache interval were not repeated in this local acceptance run. - Local preview only: the visible app is restarted against the integrated code, preserving its profile and unsent draft. No PR merge, remote push or release. ## 2026-11-03 - Revalidate empty-pane chat creation against current workspace behavior - Integrate the empty-pane New chat action with current workspace exit and file-split behavior. Keep unique workspace identities and explicit layout activation. A creation that finishes after exiting workspace mode updates the saved workspace without reopening its panes or replacing the ordinary chat. - The new delayed-create regression fails without the visibility guard and passes with it. On main base `7e590033`, 511 focused desktop tests, TypeScript and production compilation pass; development metadata remains 0.2.0/build 86. - Actual native offscreen Electron input exercises production preload/IPC and authenticated test-server requests: create four distinct persisted chats in their destination panes, create with an empty search result and keyboard, preserve four unsent drafts, exit/reenter workspace mode, and reload. Inspect dark/light and narrow layouts. Only credential storage is a fixture; server creation and responses are real. No provider turn is needed for this layout change; packaged Windows/Linux acceptance is not claimed. - Local compiled preview and test-service validation only; no release publication. ## 2026-11-01 - Create a new chat directly in an empty workspace pane - Add New chat above the existing-chat search in every empty pane. Create a fresh server conversation in the clicked pane while retaining the other chats, drafts and current workspace controls. Keep the action available when the search has no matches, with English and Chinese labels. - Preserve empty workspace focus when background session lists arrive. Keep ordinary chat-mode startup selection, and respect navigation or replacement of the destination while creation is pending. - Validation uses an isolated offscreen native Electron app with production preload/IPC and authenticated requests to a disposable full server. Click New chat in three- and four-pane layouts, type into the new composers, and verify distinct persisted conversations, empty histories and unchanged existing drafts. Exercise dark and light themes and a narrow empty-search layout. Existing histories and provider availability are fixtures; no provider turn, packaged distribution, or native Windows/Linux acceptance is claimed. - Source based on `9f6f1f13`; development app metadata remains 0.2.0 (build 86). Sixty focused workspace/component tests, TypeScript and Electron production compilation pass. Source/build validation only; no server contract change or deployment. ## 2026-11-03 - Reduce terminal input delay after scrolling - Coalesce same-direction desktop wheel input on a fixed 16 ms deadline while preserving the existing 80-line packet limit and total scroll distance. Flush before typing, resizing or reversing direction; discard pending wheel input when detaching, reconnecting, changing profile or switching tmux targets. Ordinary keyboard input is not delayed or retried. - Reduce repeated server history scrolling from four tmux processes to two by using the mode query as the existence check and combining the scroll with its final mode query. Preserve manually owned copy mode, bottom-of-history exit, and the older-tmux hidden-indicator fallback. No terminal protocol change. - Same-machine authenticated WebSocket/PTY/tmux benchmark: 10,000 history lines, 60 five-line wheel steps at requested 8 ms intervals, then an input echo. The final repeated comparison (eight trials per version) reduced median post-scroll echo delay from 390 ms to 18.6 ms; observed maxima were 504 ms and 20.8 ms. Idle echo medians remained 0.84 ms and 0.86 ms. Earlier runs varied with host load but also improved; these are local measurements, not remote network or foreground paint guarantees. Both mixed client/server combinations were additionally exercised against the unchanged wire protocol. - Verification on main base `6dbb6673`, development metadata 0.2.0/build 86: 676 focused desktop tests, 65 server tests, TypeScript and production build pass. New client regressions fail without the fix. Native offscreen Electron exercises real IPC, authenticated transport, PTY and tmux: ordinary typing, repeated scroll/input, direction reversal, Ctrl-C, window creation, split, resize, reconnect, close/reopen with retained contents, manual copy-mode ownership and mouse-capture toggling. Post-scroll input-to-renderer-data median was 31 ms across four native-app trials (not a paint measurement). Only credential storage and the synthetic terminal workload are fixtures; no provider request is required for this terminal change. - Source/local test build only; existing user services and App are unchanged. Client batching works with the previous server; the server process reduction requires the corresponding server update. No release, deployment, mobile change or native Windows/Linux acceptance is included. ## 2026-11-02 - Preserve Claude tool ownership during connection startup - Keep the callback settings used to construct a Claude connection attached to that connection while concurrent command/MCP discovery supplies newer options. Those newer settings apply on the next reconnect. - This prevents a connecting client from retaining unbound AgentsDock tool callbacks while ownership is assigned to a different settings object. - Validation: the delayed-connect race fails before the fix and passes after it. All 13 focused ownership, reconnect, fork, native continuation and Side chat checks pass. An isolated real Claude SDK run performs a concurrent runtime lookup during controlled connection startup, then reads the inbox through the production provider tool and HTTP authorization successfully. Temporary fixture processes are stopped afterward; no production service changes or deployment. ## 2026-11-02 - Treat Claude /compact as a completed command - A native Claude local command such as `/compact` produces no model reply; Claude only marks its result with `local_command`. AgentsDock flagged that as "Claude returned no response…" and failed the turn. Such results now complete normally, and the replayed `` (e.g. "Compacted") is shown as the reply with its native message ID. The output is promoted only after the structured `local_command` field proves it; a human message quoting the markup is not. - Applies to the Agent SDK and `claude -p` transports. The print path no longer crashes on string-valued replayed user content. - Regression cases for SDK and print `/compact` turns fail before the fix; 297 focused server tests pass. An isolated source server (temporary state/config, loopback) took real authenticated HTTP turns on the Agent SDK path, using the desktop's native command selection, against a real Claude (Haiku): message, `/compact`, follow-up, `/compact` again, then reopening the chat. Before the fix this reproduced the empty-turn error and imported wrapper bubbles. With this and the preceding history fix, both compactions show "Compacted" with exit 0 and sync imports nothing. The same sequence on the real `claude -p` transport (no SDK capability) also completes both compactions with "Compacted" and exit 0. - Desktop acceptance: an isolated unpacked Electron 43.1.1 build of main, offscreen in a private headless Wayland compositor with a private session bus and keyring, connected through the real setup dialog and production IPC/authenticated transport to an isolated source server running real Claude Code 2.1.316 (Opus, Agent SDK). `/compact` was chosen from the native command palette twice, around a follow-up, then the renderer was reloaded and the chat reopened. Before the fix: two "Claude returned no response…" errors. After it: both turns end with an assistant "Compacted", no errors, and identical results after reopening. Limits: `--no-sandbox` (the setuid helper needs root), and Linux only. - Source only; no server deploy/restart, desktop release or mobile change. ## 2026-11-02 - Stop importing Claude command rows as user messages - Claude Code records a local slash command such as `/compact` as user-role transcript rows: a metadata caveat, a `` wrapper and its `` output. Provider-history sync imported the wrapper and output as human bubbles and, because they matched nothing on the timeline, re-imported everything since that point. - Rows proven by the native caveat → command → output parent chain now project as the typed command (`/compact`) and its output ("Compacted"), keyed by the native message ID, so they match the live turn instead of duplicating it. Quoted markup without that chain stays ordinary user text; `/goal` keeps its dedicated normalizer. Full import and incremental cursor sync both apply it. Output rows must share the command occurrence's `promptId` (one stdout and one stderr each), so a later human prompt quoting the markup stays human. Incremental seeding recovers complete preceding records within the transcript reader's 32 MiB record bound, so long command arguments keep their provenance across a cursor boundary. - Regression cases use transcript records captured from Claude Code 2.1.316 and fail before the fix; 191 focused server tests pass. - Desktop acceptance (with the follow-up live-turn fix): an isolated unpacked Electron 43.1.1 build of main, offscreen in a private headless Wayland compositor with a private session bus and keyring, connected through the real setup dialog to an isolated source server over authenticated transport. A real Claude Opus chat ran message → `/compact` (from the native command palette) → follow-up → `/compact` → renderer reload and reopen. Before the fix, each compaction added two imported "You" bubbles with raw wrapper markup. After it, the timeline shows four user and four assistant messages, both "Compacted", no markup, and no imported rows on the server. - Rows already imported into existing chats are unchanged. Source only; no server deploy/restart, desktop release or mobile change. ## 2026-11-02 - Keep versioned Claude model labels while chats stay open - The model picker and composer bar showed bare aliases such as `Opus` instead of `Opus 5.5`. Versioned labels come passively from a real chat's Claude initialization, which happens once when its process connects, but were cached for only five minutes. Claude chat processes stay connected for hours and the desktop re-reads the catalog every 15 minutes, so most reads fell back to the alias list. - The passive label cache now lasts 24 hours. Its key still changes with the Claude CLI binary, provider environment, user/managed settings and account identity; each new connection replaces the entry and native authentication failure clears it. The native credential file now counts only by presence (login/logout), so same-account OAuth renewal no longer discards labels. No disposable Claude process, version guessing or static pin was added. Before any chat connects, aliases remain unversioned. - Regression: the long-lived-chat and credential-rotation cases fail before the fix; 22 catalog tests pass. An isolated source server (temporary state/config, loopback) with a real Claude Opus chat returned `Opus 5.5` immediately and, with the chat still open, `Opus` after 320 seconds before the fix versus `Opus 5.5` after it. - Desktop acceptance: an isolated unpacked Electron 43.1.1 build of main, offscreen in a private headless Wayland compositor with a private session bus and keyring, connected through the real setup dialog and production IPC/authenticated transport to an isolated source server with a real Claude Code 2.1.316 Opus chat. The composer read `Opus · high` until the catalog refreshed, then `Opus 5.5 · high` after Settings → My Agents → Reload settings. The picker listed Opus 5.5, Sonnet 5.5 and Haiku 5.5. After 341 seconds with the chat still open, a second reload kept `Opus 5.5 · high` with the fix; on main it fell back to `Opus · high`. Real OAuth renewal was not triggered; same-account rotation is covered by a regression only. Limits: `--no-sandbox` (the setuid helper needs root), and Linux only. - Updated the Claude discovery docs, including marking the removed disposable-probe design as historical. Source only; no server deploy/restart, desktop release or mobile change. ## 2026-11-02 - Prefetch replies of chats that finish out of view - A chat that finished while hidden opened with its cached messages and showed the reply only after the latest page loaded. The background poll now reads that page for unread chats this device has opened before, up to three per poll, so opening the chat shows the reply at once. Each reply is read once; a failed read is retried after a minute. - Opening a chat after a failed health check reconnects at once instead of waiting for the next 30-second poll. - Verification: new service tests; typecheck, desktop tests and build pass. Against a local test server with a stub Cursor CLI, the reply was cached about 40 ms after the poll and still showed in the app with the server stopped. Real Claude/Codex turns and packaged builds were not tested. ## 2026-11-02 - Exit workspace mode into a single chat - Hide workspaces now exits the entire multi-pane view, including empty chat selectors, and restores the ordinary header for the focused conversation. Reopening restores saved pane assignments and layout instead of creating a workspace. Sidebar navigation while closed cannot overwrite that layout. - Retain drafts and server-scoped preferences through toggles, reloads and server switching; release hidden timeline subscriptions without stopping provider runs. Ignore late hidden-pane loads and reconcile deleted chats before reopening. Migrate previously hidden-but-still-split saved layouts. - Regression cases failed before the fix. The beta-based preview passed 5,956 desktop tests (six skipped). The main-based PR passed 5,827 tests (five skipped), eight packaging/license checks, typecheck and production build. Native offscreen Electron with real authenticated test-server transport exercised three-pane selection, focused-chat exit, unsent draft retention, reload, repeated reopening, keyboard activation, and dark/light/narrow views. Credential storage was isolated; no provider message was submitted. - Local user window refreshed to the updated compiled renderer; no server update/restart, mobile change or release. Native Windows/Linux acceptance and packaged distribution remain outside this local preview. ## 2026-11-02 - Restore file split controls inside chat workspaces - Restore Show chat / Full screen, Editor left / Editor right, Split right, and the Files tree's Open to the Side in tabbed chat panes. Retain the workspace model and its focused-pane/profile ownership guards; do not bring back the removed Changes tab. Restore secondary editor selection and layout preferences independently of pane ownership, including unsaved file drafts. - Bound editor grid tracks and wrap narrow toolbars so restored buttons are not clipped in chat-plus-two-file layouts. Reuse existing controls, labels, themes and resizing behavior without changing the provider/server contract. - Regression: the three focused cases fail before the fix; 5,949 desktop tests and eight packaging/license checks pass, with six tests skipped. TypeScript and production compilation pass; final grid sizing is additionally covered by the 19 editor-style/design-system checks. - Actual native offscreen Electron verification uses real authenticated test server file transport: open two fixture files, show chat beside them, switch editor left/right, save only the focused secondary file, retain an unsaved secondary draft and layout after reload, and inspect dark/light and narrow layouts for clipped controls. Test credentials are supplied by an isolated credential-store fixture; file reads/writes, IPC and renderer handlers are real. No provider turn is required for this editor-layout change. - Initially reproduced and verified on beta.2 artifact source `9e816c6b`. Ported only these two UI fixes onto main `9f6f1f13`, without release-line changes. Repeated native workspace exit/restoration and chat/two-file split workflows with the main-based renderer, preload and service transport. Local source and compiled preview only; no server restart, mobile change or publication. Restarted the local user App with its existing profile and confirmed the selected chat, open file and server connection were retained. Exercised Show chat in that window and confirmed the restored layout controls and visible chat/file panes. Packaged distribution and native Windows/Linux acceptance are not part of this change. ## 2026-11-01 - Distinguish normal chat recovery from connection failures - Keep a healthy server connected while its startup queue-restoration task is running. The server already admits each chat when its durable queue is ready; do not label every chat or the connection degraded until the full scan ends. Show “Connected — restoring queued chats” on the active server indicator and menu, then “Online” when recovery completes. Include English and Chinese text. - Continue reporting failed recovery, unavailable tmux, authentication failures and transport failures. Preserve per-chat queue admission, durable messages and coordinated-update readiness requirements. - Treat a validated HTTP/WebSocket client disconnect as a normal gateway close, without synthesizing an interrupted-execution error after closing its IPC transport. Accepted worker cleanup continues; genuine worker interruptions still report failures. - Validation: 490 affected desktop tests, 27 execution-transport tests, TypeScript and production Electron compilation pass. New regressions fail against the previous behavior. Independent review found no blocking issues. - Native app acceptance uses an offscreen Electron window, production renderer, preload and IPC, and authenticated requests to the full isolated server. Hold/fault its real startup recovery task to exercise restoring, ready, failed and recovered status, using actual mouse input for the server menu and refresh button. No provider turns are needed for connection presentation. Actual TCP WebSocket → Uvicorn → Unix IPC acceptance also confirms a normal code-1000 close produces no gateway error and preserves worker cleanup. - Source changes only; no live server restart, deployment or public release. Windows/Linux native app acceptance and packaged distribution are untested. ## 2026-11-01 - Let agents create chats with automatic mailbox connections - Add run-bound `chats create --title ... --folder ...` for independent AgentsDock conversations. Create a permanent bidirectional mailbox route between creator and child automatically and expose it to the creating run immediately. Existing send, inbox, read and reply actions handle communication. Creation itself does not start a provider turn. - Inherit applicable runtime settings, including a custom provider connection, without copying native thread IDs, transcript, goals or other peer routes. The new chat appears in its requested folder through the existing desktop API. There is no new chat-count quota or route-approval dialog. - Persist explicit retry keys across logical runs and restarts. Repeated creation returns the existing chat. A failed route grant preserves the already-visible child and repairs it on retry; disconnects do not duplicate creation. Exact-route sends discover the current mailbox mode even from a shell opened before creation. Keep private creation-admission records out of client timelines. - Validation: 59 helper tests, four creation regressions, 50 existing pair, discovery and provider-input tests, and 13 session/custom-provider tests pass. Server syntax checks and the unchanged Electron production build pass. - Isolated full-server acceptance exercises the production helper over HTTP, actual capability authentication, immediate send, child read/reply, creator read, and explicit-key recovery in a new logical run after restart. Native offscreen Electron uses production IPC and authenticated transport to select the created chat and display its folder and mailbox cards. These transport checks use synthetic provider run bindings and are not generation acceptance. - Real Codex acceptance separately exercises normal turn admission and the run-bound provider tool: the creator creates and messages a child, the child automatically wakes to read and reply, and the creator automatically wakes and reads that reply. Both chats use disposable state; temporary authentication is removed and fixture processes are stopped afterward. Claude and other providers have not received live generation acceptance for this change. - Server/helper change based on `9f6f1f13`; no desktop contract migration, production server restart, installation, release or mobile change. ## 2026-11-02 - Publish coordinated 1.0.12-beta.2 - Publish direct desktop build **1270**, npm runtime/short CLI and the matching signed legacy server bridge from frozen source `9e816c6b3b73508a47764b4d4ea546e41bcd3857`. Preserve stable `1.0.11` across npm and both desktop feeds. This is a beta release, with no mobile submission or deployment to existing services. - [Paired npm publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/41391793355) and [desktop publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/41392454191) pass. Independently verify fresh scripts-disabled npm installation, exact runtime dependency, installed files/modes and CLI help; verify both desktop asset sets, source tag, signed descriptors, beta feeds and bridge publicly. - Exercise real Claude/Codex contextual follow-ups and repeated messages through the exact signed native app and matching runtime in an isolated hidden window. Verify retained history, workspace/editor state and the lifecycle-marker fix. The app was launched directly; local macOS LaunchServices relaunch acceptance remains incomplete, and further restart checks were explicitly deferred for this beta. Do not count that boundary as passed. - All 11 scoped Mac update/recovery cases, both legacy-permission Linux rollback cases and macOS/Linux CLI acceptance pass. Preserve the reports' limits for populated customer history, fresh interaction, busy replacement and reboot. Full native-provider content parity, Windows native updates and idle Codex goal-resume mail authority remain unverified. QA analytics collection was off. - macOS is Developer ID signed and notarized. Windows remains explicitly unsigned. See [release notes](RELEASE_1.0.12-beta.2.md) for changes, download links, validation evidence and limitations. Documentation updates do not move the artifact source or trigger rebuilding accepted bytes. ## 2026-11-01 - Prepare coordinated 1.0.12-beta.2 - Integrate reviewed main through `9f6f1f13704e767717b5131e2df1213306a58abd` with the shipped beta.1 release line. Preserve installer/update protections, signed runtime metadata, custom-provider and Side chat behavior while retaining the new workspace, composer, code-intelligence, sorting and Claude fixes. - Reconcile native Claude goal status with retained continuation-limit details; retain failed evaluation outcomes. Combine current goal/route ownership checks with release-line cleanup, and advance the SDK configuration identity for the combined configuration. Preserve a retired Claude child's exact tool alias when a delayed structured task identity arrives. - Keep internal native Codex goal-resume lifecycle markers out of user bubbles. Preserve genuine input with identical text and the chronological live steer boundary, activity and final answer. Reproduced and verified through the native desktop/server/provider path in an isolated offscreen profile. - Prepare matching desktop/server/CLI version `1.0.12-beta.2`. Keep stable `1.0.11` unchanged. Bind scoped native rehearsal to independently verified stable 1.0.11 and beta.1 bytes, preserving historical replay identities and installer recovery/permission checks. - Desktop type checking, production compilation, 5,947 desktop tests (six skipped), eight packaging/license checks and 393 release-tooling tests pass. Targeted server/recovery checks and hosted native acceptance remain in progress. See [candidate notes](RELEASE_1.0.12-beta.2.md). This entry records preparation only, not signed-package acceptance, publication, deployment or a mobile release. ## 2026-11-01 - Toggle workspace controls without creating a workspace - The chat-header workspace button now toggles the workspace toolbar on/off. Repeated clicks do not create/delete layouts, change the current chat or reload history. Use the toolbar's plus button to create the first or another workspace; existing split panes remain in place while the toolbar is hidden. Update the tooltip and accessible expanded state to match visibility. - Preserve toolbar visibility independently of layouts in server-scoped preferences. Show layout controls only after a workspace exists. Closing the last workspace returns to the ordinary chat view. Keep drafts and existing workspace restoration unchanged. - Validation: 117 focused tests, 5,818 desktop tests and eight packaging/license checks pass; five desktop tests remain skipped. TypeScript and the Electron production build pass. Source base: `7a41121a`; development app metadata 0.2.0 (build 86), test service metadata 1.0.7-beta.11. - Actual-app verification uses an isolated native Electron window with real authenticated server transport and persisted workspace preferences. Reproduce the old header behavior creating a layout on each click. Verify the fixed header alternates visibility with the count unchanged, plus adds one workspace, reload and a full app restart preserve visibility and the unsent draft, workspace switching and split layouts still work, and closing the last workspace restores the ordinary chat. Exercise mouse and keyboard activation and English/dark and Chinese/light layouts, including a narrow viewport. The isolated fixture disables analytics and stubs its credential store only; no server responses or UI action handlers are mocked. No new provider turns are needed for this presentation-only change. - Restart the user's local development app on request with the same profile and credentials. Verify existing layouts, current chats and drafts survive, and the header toggles visibility without adding a workspace in that running app. Close the isolated preview to avoid duplicate test windows. - Local desktop source/build only; no server restart, mobile changes or public release. Windows/Linux native acceptance and packaged distribution were not exercised. ## 2026-10-31 - Isolate Claude native messages from chat routes - Address issue #53: Claude's native ListAgents/SendMessage namespace can resolve a similarly named provider session instead of an AgentsDock @Chat. Route AgentsDock messages through the run-bound Chats helper and exact server-issued route; native success is not an AgentsDock delivery receipt. - Apply PreToolUse guards in both the Claude SDK and print-mode compatibility path. Deny native cross-session discovery/delivery while preserving messages to exact native child IDs observed by SubagentStart and their replies to main. Keep child ownership across SDK reloads, scoped to the chat and provider session. Preserve native permission checks, custom API settings, existing AgentsDock authorization and durable delivery receipts. - Include the new runtime module in server package, install and deployment manifests. No global Claude configuration or desktop/mobile code changes. - Validation: 294 targeted routing, SDK, runtime-settings, terminal-context and packaging tests pass; shell syntax and Python compilation checks pass. Cover permission modes, legacy tool fields, mismatched identities, symlinks, reconnect isolation and private custom-endpoint settings. - Real-use verification: native input in an isolated Electron app on an authenticated test server exercises two Claude-to-Codex @Chat round trips, with one accepted message per send and matching read/reply receipts. A same-name unrelated Claude chat receives neither message. Real child-agent creation, continuation, replies to main and continuation after Reload Claude succeed. Actual Claude print-mode ListAgents and external SendMessage calls are denied even with permission bypass enabled. A subsequent user-driven send is accepted once, read by the intended Codex recipient and acknowledged in that recipient's chat, with no delivery to the same-name Claude chat. A local reply is not a cross-chat return receipt. - Source base: `7f53e589`; native Claude 2.1.304 and SDK 0.2.130. Local server preview only, no public release. Live acceptance ran on macOS; Windows and every permission mode were not separately exercised end to end. ## 2026-10-31 - Keep Claude reports in compact history - Follow up on issue #50: a report emitted before several tool calls and a short final acknowledgement could disappear from the compact history page. Prioritize native Claude assistant text and legacy report-shaped commentary alongside other essential output, using the existing per-turn and page caps. - Preserve original event types, sequence ordering and stored history. The desktop's existing compatibility projection and final-answer deduplication remain responsible for presentation. Actual thinking and other providers' commentary are not promoted to report output. - Validation: 183 history-projection and Claude runner tests, plus 325 desktop timeline/rendering tests pass; desktop type checks and production build pass. Added regressions fail on the unmodified base and cover native text, heading/table/long-form legacy reports, chronological reloads, immutable storage, thinking exclusion and bounded responses. - Real-use verification: native mouse input in an isolated Electron window reproduced missing reports against a real authenticated server process. The fixed process shows both report formats without expansion after reload; genuine thinking still expands and collapses through the run-trace endpoint. A disposable live Claude turn emitted a report, performed five harmless tool calls and completed. Its original server's compact response omitted the report; replaying that generated output through the fixed history API and native app preserved its visible content/order and one final reply. This tests history presentation, not new provider-runtime behavior or a blanket native-client parity claim. No mocked HTTP or provider success. - Source base: `4dd67629`; development app metadata 0.2.0 (build 86), test service metadata 1.0.7-beta.11. Isolated acceptance disables analytics. At the user's request, run the fixed source on the existing local test service and restart the development app against it. A fresh real Claude report/tool/ final-reply turn remains visible after reload with activity collapsed; the legacy history sample also renders directly. Preserve the original service state and app settings, and leave the production service untouched. Additional coverage verifies older-page navigation with a cold index, native/legacy output caps, and report retention after stop, error or an empty final reply. In an isolated native app on the fixed test service, stop a real Claude turn after its report, reload, and send a contextual follow-up; the report remains visible once, and the follow-up uses the previous context. Switching chats and expanding/collapsing activity preserves report placement. Update the source-extracted history test harness to load the new classifier; historical cron, async-message and goal-follow-up projection checks also pass. Server-source fix only; local test preview, no mobile change or public release. ## 2026-10-31 - Desktop Preferences settings - Add Preferences / 偏好设置 directly below General in Settings. Move Show reasoning traces, Unread chats at the top, Include cron jobs in Unread and Show Running section into that page. Preserve their existing handlers, saved values and storage scopes; no migration or reset is needed. - Keep language, theme and version information in General. Updates remains a separate page. Support direct Preferences navigation and accessible focus restoration when reopening Settings. - Validation: 104 focused tests, 5,811 desktop tests and eight packaging/license checks pass; five desktop tests remain skipped. TypeScript and the Electron production build pass. Source base is `0229fb3e`; development metadata remains 0.2.0 (build 86), not a new release. - Exercise an isolated native Electron app connected to an authenticated test service (1.0.7-beta.11), using native Chromium mouse and keyboard input. Toggle all four controls, navigate between pages, close/reopen and reload; verify persistence and restore the original values. Inspect English/dark and Chinese/light layouts at narrow width, including the reachable close button. No test chats or provider turns are created; provider behavior is unchanged. The isolated fixture disables analytics. Desktop-only local preview; no server update, release or mobile changes. ## 2026-10-31 - Desktop sidebar sorting - Add a compact up/down-arrow Sort by button immediately after the chat count. Click or Enter/Space cycles Folders → Latest → Name → Folders; hover or ArrowDown opens the selection menu. Offer concise English and Simplified Chinese labels; do not add an oldest-first option. - Confirm an explicit sort selection with a green label beside the button for two seconds. Use the existing status-green color with 11px regular-weight text, keeping the brief confirmation quieter than the main controls. Repeated clicks replace the text and restart its timeout; keep keyboard/menu selection and screen-reader announcements consistent. Clear feedback when switching servers or disabling the control, not on ordinary pointer movement. Initial rendering and hovering alone show no confirmation. Suppress the sort button's pointer-selection focus halo without blurring it; retain normal focus indication for keyboard navigation and menu dismissal. - Latest follows the server's agent-visible output timestamp, with creation time as the fallback. Opening/reading a chat, sending a prompt or editing metadata does not promote it. Separate pins from the ordinary list beneath them with a quiet theme-aware solid divider and 12px of breathing room on each side, including when pins are collapsed. - Keep existing unread/running, pinned and archived grouping rules above the chosen ordering. Non-default modes hide folder headings, including empty and collapsed folders, without modifying folder membership or manual order. - Persist the choice per server and align next/previous-chat navigation with the visible list. Disable drag reordering in automatic-sort modes. Reject preference changes during server switching and normalize unknown saved modes. - Validation: 5,811 desktop tests pass, five are skipped, and eight packaging/ license checks pass. TypeScript and the Electron production build pass. - Exercise an isolated native Electron app against a real authenticated test server: click and hover the menu, switch all three orders, reload and restart to verify persistence, return to folders, mark a chat unread, and observe an actual provider turn in the running group while changing sort order. Check English/dark and Chinese/light presentations with a narrow window. The system computer-use bridge was unavailable; interaction used native Chromium input events in the actual Electron window, not component-handler invocation. - Source base: main at `0229fb3e`; development package metadata remains 0.2.0 (build 86), not a new release; test service version 1.0.7-beta.11. Server-scope isolation is also regression-tested. Verify two real provider replies reorder the list and reopening the older chat leaves that order intact; check the divider while pins are collapsed and the button/count spacing. Recheck the softened solid divider and spacing in all three sort modes, with pins expanded/collapsed, in both themes through the native app. Verify native repeated clicks cycle exactly once, hover remains independently selectable, Enter/Space cycle and ArrowDown opens the menu, with the chosen mode retained after reloading. Verify the green label through actual clicks, keyboard and menu selection in English/dark and Chinese/light; the latest selection remains visible past the previous timer's deadline, disappears after two seconds, does not overlap the header actions and is absent after reload. Native clicks and hover-menu selection leave no blue halo; Enter, Space and Tab-away/back retain the keyboard focus indicator in both themes. The feedback follow-up creates no test chats. Remove the disposable test chats after acceptance. The local source preview is refreshed on request with the existing user's profile retained. Desktop-only; no server update or mobile changes required. Not published. ## 2026-10-31 - Desktop send connection recovery - Preserve the original transport, server-token or identity error when an explicit send joins a background reconnect. Background refreshes remain quiet without swallowing the error needed by foreground callers. - Distinguish server connection failures from provider sign-in failures in the composer. Use a warning state, retain the reported server version and offer explicit server reconnection without retrying the message submission. - Preserve drafts, uncertain-send receipts and live running/thinking updates. Invalidate HTTP admission after a transport failure without discarding a healthy receive socket. Keep server-token and identity checks intact. - Fence asynchronous reconnection and runtime rechecks to their original server generation and notice context. Do not let stale cached login status show an authentication warning during a successful Codex or Claude run. - Keep native server reconnection out of the restricted shared-browser bridge; shared pages retain their existing stream-recovery controls. - Validation: 5,790 desktop tests and eight packaging/license checks pass; five tests remain skipped. TypeScript and the Electron production build pass. - In an isolated visible native Electron app, reproduce the old green provider-status/network-error mismatch, then exercise the fixed send path through production IPC and an authenticated real test service. Codex and Claude both return actual replies. A real test-service restart preserves the draft; explicit reconnect sends nothing, and manual retry succeeds. Check English/dark and Chinese/light layouts, including a narrow window. Test analytics is disabled and the production service is untouched. - Additional native-input checks cover repeated reconnect clicks, typing while verification is pending, preserving separate drafts after chat navigation, and switching to an unreachable test-server profile before the old health response returns. Returning to the reachable profile keeps the draft; a manual retry submits once and produces an actual provider reply. - Transport fault injection, old-server version compatibility, token failures and identity mismatches also have deterministic coverage. The reporting user's remote installation and its separate installer-permission error have not been reproduced; no server update or security bypass is included. ## 2026-10-31 - Mobile reliability fixes on current main - Extract bug fixes from older mobile review branches without restoring their broad UI redesigns or old release/build metadata. Close superseded PRs #3, #6 and #19; preserve the remaining reviews pending their independent audit. - Keep received HTTP turn acceptance through body decoding, deadlines and connection revalidation. Do not restore accepted text as an unsent draft or automatically retry the submission. Preserve newer typing and attachments; genuine rejected sends still retain a recoverable draft. - Clear the native composer on exact draft consumption, including asynchronous recipient preflight, rather than on a later network completion. - Fence queue snapshots against concurrent live delivery, exact native goal steering and server identity changes. Reconcile sparse membership with bounded reads; preserve newer paused-after-Stop and delivery-barrier behavior. - Cancel obsolete sidebar history navigation when the search changes, without letting an older completion reopen a result or clear a newer loading state. - Run the new accepted-send, queue reconciliation and exact search-target regression suites in source CI. - Real-service search also exposes deduplicated final answers losing their search identity. Retain exact represented event IDs on assistant rows while preserving visible text, row keys, sequence anchors and chronology; do not claim unrelated events in a sequence gap were loaded. Refresh row metadata when a new terminal search identity arrives. - Preserve Android QR camera permission in composed native configuration, scope translated iOS privacy strings to iOS and retain legacy media-save guards. - Validation: 433 rendered/contract tests and 105 library/API/store modules pass, together with mobile TypeScript, iOS JavaScript export, native iOS development builds and Android resource processing (197 Gradle tasks). - Native iPhone simulator against an isolated real service reproduces an accepted message returning to the input before the fix. The same deliberate response-body fault after real provider acceptance no longer restores it or repeats the POST. Actual Claude replies remain visible. - Native Codex checks cover enqueue, Stop preserving the paused item, editing, explicit Run now, a single provider reply and an empty queue. Search and keyboard-open recipient-picker dismissal also pass through the real client. Both Claude and Codex answer searches now open the exact visible reply after the terminal-identity fix; the Claude case failed in the same native app before the change. Queue controls require scrolling/dismissing the keyboard on compact screens; this pass does not claim to redesign that existing constrained layout. - A separate development app was installed and launched on a connected iPhone. The tester confirmed normal send, queue editing/continuation and peer-message workflows; the test-service records and device-local cache agreed on empty queues and consumed drafts. This does not certify every visual interaction. Android physical QR capture and synthetic stale-response race cases on real devices remain unverified. This change does not publish or upload a release. ## 2026-10-31 - Local integration review of goal follow-ups and child lifecycle - Integrate PRs #105, #100, #99, #96 and #24 with their original contributor commits. The editor line-ending implementation from #24 is already equivalent to main; retain it together with newer editor and language-service behavior. - Preserve main's native Claude goal delivery during active SDK turns rather than replacing it with the older end-of-turn queue design. Expose actual durable queued goals separately, and preserve pending-command exclusion. - Combine Codex goal steering of explicitly selected local chat references and skills with current task-provenance, retry and authorization fences. Suspend per-route source instructions during uncertain writes as well as ordinary provenance. Keep Claude's narrower follow-up admission independent. - Reconcile resumed Codex output by exact native item ownership, not equal text. Preserve Claude child completion and retire stale unowned activity into history. - Focused checks: 942 desktop tests and 723 server tests pass; desktop type check and production compilation pass. Tests include exact-authority retries, rejected/uncertain steering, goal history, child lifecycle and editor line endings. - Real native Electron/provider checks: steer a running Codex goal with a local chat mention, deliver and receive one peer exchange without replacing the run, resume the native goal and reopen history without new imported duplicates; run a real Claude child and retain its completed history after reload; edit and save a CRLF file through the real server while preserving its bytes. - The combined review exposed an existing Claude history-sync boundary: an empty native background-wake lifecycle was counted as user input and blocked subsequent occurrence matching. Exclude only proven metadata-only wakes; keep genuine user input, files, external repeats and native goal commands. - Repair already-imported user copies through exact native reply ownership and sentinel-proven goal-command ancestry, including bounded historical pages. Preserve unknown or conflicting provenance and distinct repeated inputs. Repair changes the projection, not either original history file. - Expanded checks: 830 focused server tests and 942 desktop tests pass, along with desktop type checking and production compilation. Both new regressions fail against the pre-fix implementation. A native Electron/Claude run confirms background completion, a goal admitted during the same active run, goal achievement, history reload, old-copy repair and two intentional equal-text submissions remaining distinct. An unrelated split-pane draft remains intact. - Final pre-merge desktop checks pass: 5,757 tests (five skipped), eight packaging/license guards, TypeScript and production compilation. - Additional native Electron checks exercise a catalog-selected skill during a Codex goal, Stop with a paused unsent queue, reload without automatic resend, and explicit Send now delivering once on the original provider thread. Claude checks cover cancelling a pending goal, replacing an active goal, Clear & stop, child retirement, and completed/cleared state after reconnect. - This pass also catches normalized Claude goal commands being reimported: exact native reply ownership plus command metadata or a goal sentinel now recognizes both XML and canonical slash representations. Unproven commands, quotations and distinct repeated inputs remain visible. Start/clear regressions fail before the change; the 95 focused repair/goal tests pass. Real-app reloading repairs existing copies and preserves fresh commands once. - The final combined server pass after this repair passes 743 tests across 26 goal, steering, history, native transport and child-lifecycle modules. - Hosted CI also exposed a blocking observer in the existing concurrent catalog auth-change test. Replace its executor-based entry wait with thread-safe event-loop notification, keeping a two-worker pool to exercise overlap. The old fixture fails under that constraint; the corrected 14-test module and 30 repetitions pass with unchanged runtime assertions and production code. - A hosted rerun additionally exceeds the test's five-second watchdog during an event-loop scheduling stall. Give this ordering test a bounded 30-second watchdog and release both worker gates even on assertion failure. An isolated 6.3-second stall reproduces the old timeout and passes with the corrected fixture; the full 14-test module remains green. This is not a change to runtime timeouts or an assertion of provider latency. - Native acceptance uses an isolated macOS Electron profile, real IPC/HTTP and live Codex/Claude. Mouse, keyboard and scrolling are driven through Electron's debugging interface; analytics are disabled. No mocked provider result is counted as live acceptance. Existing imported-control grouping is unchanged. - Acceptance is local only. An earlier foreground sleep command was refused by the native Claude harness and was not counted as successful execution; the later background and read-only-file workflows completed through the real provider. Native Windows/Linux and forced real-provider transport failures were not exercised. No release or merge is claimed by this review. ## 2026-10-31 - Local review of activation diagnostics and runtime reliability - Integrate PRs #122, #112 and #111 on the previously accepted #134 preview, retaining their original commits and contributor attribution. The combined review is PR #135, not a published release. The earlier #123 changes are already included in that baseline and were not applied again. - Split-service activation failures identify the failed boundary and unsafe recovery-interpreter requirement without relaxing trust checks. Update the existing rollback regression's block extraction; resolve macOS temporary-path aliases in fixtures. Check foreign ownership, symlinks, exact short-circuit order and cold-handoff signal restoration. - Retain disconnected-IPC cleanup, bounded descriptor headroom and shared runtime discovery from #112. Reproduce its hosted concurrency-test failure with a small executor. Give the deliberately overlapping synchronous calls an explicit pool and observe their entry through event-loop notifications, replacing timing sleeps. Thirty repeated runs pass with a one-worker default executor. This corrects test scheduling, not production concurrency policy. - Verify 204 focused server tests and 193 desktop tests, TypeScript, shell syntax and production compilation. Server checks include real disposable gateway/worker processes, descriptor retirement, recovery controls and cancelled catalog waiters; providers in those fixtures are synthetic. - Native macOS Electron against a real disposable source server reproduces #111's stale banner before the fix when an old child update arrives after a successful turn. The fixed build ignores repeated late updates, retains historical errors and displays genuine current errors, including artifacts. Native Codex and Claude sends complete successfully through production IPC and authenticated transport. Settings reload and concurrent catalog requests succeed when one caller disconnects; unrelated drafts are preserved. - The late-child failure sequence was injected through the production event writer into disposable chats, not reproduced by inducing a provider fault. Native input used Electron's debugging mouse/keyboard interface; QA analytics were disabled. Full installed-version upgrade/rollback, native Linux/Windows UI, and long-duration resource-load acceptance were not repeated. No release publication or production-service update is included. - Additional acceptance: all 5,727 desktop tests (five skipped), eight packaging/license guards, TypeScript and production compilation pass. Correct an older durability regression's script-block extraction and wait for the goal editor's asynchronous hydration in its button assertion; both failures reproduced in hosted CI. The 41 durability/activation tests pass. Native Codex and Claude runs complete concurrently; a busy follow-up appears in the queue and is delivered once. Switching chats and reloading preserves an unsent draft without restoring a stale error banner. ## 2026-10-31 - Inline workspace-file naming and temporary empty buffers - New file in the desktop Files explorer opens an Untitled buffer and focuses an inline filename field in the selected directory. Enter saves through the existing revision-checked create/write path; extensions such as `.py` select the editor language. Existing filenames cannot be overwritten inline. Save As and folder creation remain available. - Empty unnamed buffers are not restored after leaving a chat or relaunching. Nonempty drafts and named files, including empty files, remain intact; this cleanup never deletes workspace files from disk. - Verified on the combined PR #127/#125/#123 source-preview baseline `01dd9b81`: 157 editor tests, all 5,704 desktop tests (five skipped), eight packaging/license guards, TypeScript and production compilation pass. Regression cases cover inline naming, Python detection, duplicate-name rejection, failed-write retry, repeated clicks, and warm/cold restoration of empty versus nonempty drafts. - Exercised the compiled native macOS app against an isolated real test server: inline creation, save, rename, Python definition navigation, folder creation, duplicate-name rejection and chat-switch draft preservation. Files on the server confirmed the results. Dark and light/narrow rendering were inspected; light preference and viewport width were emulated in the native renderer. Input used Electron's keyboard/mouse debugging interface, not React handlers. QA analytics were disabled. No new server implementation is required. - Availability: local development build, not published or merged. Windows/Linux native input and physical Mac function-key mappings were not exercised. ## 2026-10-31 - Combined code intelligence and Claude integration check - Locally integrate PRs #127, #125 and #123 on the post-#129 main baseline. Preserve their original commits. Resolve the workspace-editor conflict by keeping code intelligence and the current navigation layout, without restoring the removed Changes-panel condition. - Source baseline: `c9893949`; combined code: `d0392711`. The source-preview labels remain Electron `0.2.0`/build `86` and server `1.0.7-beta.11`; these are development metadata, not new published release versions. - The compiled native macOS app used its production IPC and authenticated HTTP/WebSocket transport against a matching source-preview test server. Files/Symbols inserted editable references from real Python and TypeScript language services. Both editors showed diagnostics and unsaved-buffer completions without changing files on disk; TypeScript definition and reference navigation also worked. - A real Claude one-shot native cron moved from idle to running with Stop, then returned to idle with its final reply. Reopening preserved chronological output without an extra user bubble for the internal scheduled prompt. During a separate real Claude turn, editing a queued message saved promptly; Stop ended the current turn and the edited follow-up completed successfully. - Focused desktop checks: 1,171 tests pass. Claude runtime, history repair, transcript and mailbox checks: 416 pass. Language-service bridge checks: 15 pass, including real Pyright and TypeScript servers. Eight build-output and license checks, TypeScript and production compilation pass. - This is local integration acceptance, not a release. Windows/Linux packaging, native schedule recovery after provider teardown, forced admission races, and every provider-native thinking/content-parity case were not repeated in this combined pass. Desktop input used Electron's debugging protocol because the computer-use surface was unavailable. QA analytics were disabled. ## 2026-10-31 - Composer history isolation and deliberate queue shortcuts - Integrated prompt-history and empty-composer Enter shortcuts with main `45fe1710`, retaining Escape-to-stop and the opt-in workspace UI. - Holding Enter no longer promotes a queued message or submits a selected command accidentally. Release and press Enter again to use Send now; Shift/Alt+Enter multiline editing and IME composition remain unchanged. - Recall history and its navigation cursor use the existing server-identity and chat scope, with a saved-profile fallback. Aliases of the same known server share history; different servers cannot recall each other's prompts. Legacy chat-only history is not adopted because its server is unknown. Late receipts do not reset another chat's history cursor; failed sends keep their drafts and are not added to history. Storage failures retain old data. - Verification: 284 focused component/helper/shortcut tests, TypeScript, production compilation and compiled-output verification passed. Reproduced held-Enter queue promotion in the original native macOS app, then exercised the fixed build with an isolated, analytics-disabled profile and a disposable service through real Electron IPC/HTTP and Codex/Claude requests. Both agents completed repeated messages; Up/Down recall, draft and multiline preservation, chat navigation, idle Enter and held Enter passed. A deliberate second Enter issued one successful run-now request per provider: ordinary Codex stopped and continued on its thread; Claude admitted the native follow-up without a stopped turn. No server implementation changes were required. - Cross-server identical-ID, storage-full and late-receipt cases were exercised in deterministic fixtures, not on customer servers. Native Windows/Linux and every provider's own CLI keyboard UI were not exercised. Computer-control integration was unavailable; native app input used Electron's debugging keyboard/mouse interface, not direct React handler calls. Availability: local development build only, not published or merged. ## 2026-10-28 - Composer prompt history and Enter to send queued messages - ArrowUp/ArrowDown in the desktop composer recall prompts previously sent from that chat, like the Codex and Claude CLIs. History is stored locally per chat (latest 100), only steps from the first/last line, and never replaces a draft the user typed or edited. - While a chat is running, a plain Enter on an empty composer sends the first user-queued message now, matching the existing Command/Control-Enter and "Send now" path (same confirmation and admission checks). OpenCode, which cannot steer, keeps queueing. The server decides delivery: Claude accepts it as a native follow-up without stopping the run; a plain (non-goal) Codex turn is stopped and the queued message continues on the same thread, so the timeline shows the earlier turn as stopped. - Implemented and accepted on main `9a750c7`: 12 new unit/component checks, TypeScript and production compilation passed; the full run passed 5,496 tests (5 skipped) plus 8 packaging/license checks. Rebased onto `aa1b8c4`; TypeScript, the full suite and production compilation were rerun there. - Test surface: `pnpm dev` Electron app with an isolated user-data directory, connected over its production transport to a loopback AgentsServer from this checkout with isolated HOME/state, using a real Codex login. ArrowUp/Down recalled sent prompts newest-first and returned to an empty draft; an unsent draft was left unchanged; ArrowUp on the second line of a multi-line draft only moved the caret; a second chat did not show the first chat's history; Enter on an empty composer while idle did nothing. Enter while a Codex turn was running issued exactly one `run-now` (200); the server stopped the running turn and the queued prompt completed on the same Codex thread with the earlier context. Claude follow-up insertion was not exercised (no Claude login in the isolated environment). Availability: source only. ## 2026-10-31 - Escape shortcut integration verification - Integrated the Escape-to-stop change with main `827ed452`, retaining the workspace/header updates and peer-message Edit and send behavior. Only the development log required conflict resolution; no server change was needed. - Added focused regressions for modified/repeated Escape, IME composition, per-composer ownership, pending Stop de-duplication and cancelled interruption confirmation. The six affected desktop suites pass (292 tests); TypeScript, production compilation and compiled-output verification pass. - Native macOS Electron acceptance used an isolated analytics-disabled profile and a disposable test service. Confirmed the old source did not stop on composer Escape, then exercised real Codex and Claude turns through native IPC and HTTP. Each Escape stop produced one HTTP 200 response and a server terminal event marked stopped. Drafts survived; slash, chat-mention and team- mention palettes closed first; idle, modified and repeated Escape did not stop work. A second Codex stop and a successful Claude follow-up also passed. - Inspected the Chinese dark Stop tooltip and English shortcut settings at a narrow viewport with simulated light styling. Native Windows/Linux behavior, live subagent-only stopping and live inbound-delivery cancellation were not exercised; existing Stop behavior is reused and focused regressions cover ownership and confirmation. Available as a local development app only; not released or merged. ## 2026-10-29 - Escape stops the running agent from the composer - Escape in the desktop composer stops the running agent, like the Codex and Claude CLIs. It reuses the Stop button's action (including the inbound delivery confirmation and Stop subagents for idle Codex chats with live children). An open command, @Chat or team-mention palette is closed first; the draft is kept; key repeat, modifiers and idle chats are ignored. Escape elsewhere in the app is unchanged. The Stop button tooltip and the shortcut list show the new ⎋/Esc shortcut in English and Chinese. - Based on main `aa1b8c4`. 5 new regression checks (the two composer stop checks fail without the change); TypeScript and production compilation passed; the full run passed 5,518 tests (5 skipped) plus 8 packaging/license checks. - Test surface: `pnpm dev` Electron app with isolated user data, connected over its production transport to a loopback AgentsServer from this checkout with isolated HOME/state and a real Codex login. Escape during a running Codex turn sent exactly one `stop` (200) and the turn ended as stopped; the Stop tooltip showed ⎋; with the command palette open, the first Escape only closed it and the second stopped the turn; an unsent draft survived Escape; Escape while idle did nothing. Claude was not exercised (no Claude login in the isolated environment); the stop request itself is provider-agnostic. Availability: source only. ## 2026-10-31 - Publish coordinated 1.0.12-beta.1 - Publish the [direct desktop beta](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.12-beta.1), build **1269**, with matching `@agentsdock/server` and `agentsdock` npm packages and the signed legacy server bridge. Both npm `beta` tags advance to this version; stable desktop and npm `latest` remain **1.0.11**. - Freeze artifact source at `c52aad2ed807a911000e41ffc0ade24d7edd4778`, combining main through `827ed452140db0a09bfab39905aa133f736871be` with the fixes shipped in stable 1.0.11. Publish identical desktop assets and final notes to the canonical and legacy download repositories; no package is rebuilt during publication. - Add named workspaces with up to four live chat panes, paired-chat user delegation without a fresh mention, and inbox **Edit and send** that preserves the matching draft. Retain released progress and Side chat behavior alongside native child ownership, provider follow-up and composer fixes. - Hosted source/server checks, matched npm/CLI checks, four-platform package verification, scoped macOS upgrade/recovery rehearsals and Linux rollback checks passed. Fresh registry installation matched accepted archive contents and the exact CLI/runtime version. macOS uses Developer ID and notarization; Windows retains the disclosed unsigned-installer policy. - Isolated live Claude/Codex checks exercised conversations, cold context resume, workspace restoration, paired mail, draft review/send and Codex Side chat continuation through real app/server transports. Full provider-content parity, populated customer-history migration, normal local quit/logout/reboot and busy-work replacement remain outside accepted coverage. - The known idle Codex goal-resume mail-authority limitation remains. Mobile builds and store submissions are separately scoped. See the [release notes](RELEASE_1.0.12-beta.1.md) for details and verification links. ## 2026-10-31 - Prepare coordinated 1.0.12-beta.1 - Integrate main `827ed452140db0a09bfab39905aa133f736871be` with stable `1.0.11` release protections for the next beta desktop, npm server, short CLI and signed legacy bridge. Preserve the stable channel and existing immutable release artifacts. - Retain main's native child ownership checks together with the released subagent activity labels, Side chat progress and per-turn runtime settings. Preserve mailbox delivery for older paired-route callers while carrying main's server-attested user delegation and reply identity checks. - Advance Claude's runtime configuration revision so existing release sessions receive the updated native tool policy. Include notification and unread-state modules consistently in both runtime package formats. - Local validation: desktop TypeScript and production build pass, along with the full desktop test suite and release-tooling Node/Python checks. Focused server merge checks pass after isolating inherited service selectors in the local test process. Native package and hosted source checks remain pending. - Hosted source CI and all eight canonical server shards passed. The separate signing gate exposed a read-only preparation test comparing access timestamps; its own file/link reads can change those timestamps. Compare contents and exact inode, ownership, mode, size, link count and nanosecond write timestamps instead. Runtime behavior is unchanged; rerun the signing gate from the new committed test source before accepting any package. - See [candidate notes](RELEASE_1.0.12-beta.1.md) for public changes and the known idle Codex goal-resume limitation. Package and native acceptance remain pending; this entry records preparation, not publication or deployment. ## 2026-10-31 - Workspace toolbar entry and contextual chat headers - Move Create workspace from the chat menu to a matching icon button next to Scheduled jobs and Terminal. Ordinary chats show title and folder together, with their copyable provider session ID underneath. Workspace chats retain a single compact title/ID/action row without the folder chip. - Workspace menus now offer Change folder as a hover/keyboard submenu, listing saved and in-use folders, including folders containing only pinned chats. Mark the current folder and preserve the existing per-chat update path. - Native desktop acceptance on base `aa1b8c47` (development 0.2.0/build 86) used a test server: created a workspace from the toolbar, reloaded, closed back to ordinary mode, hovered the submenu, changed and restored a test chat's folder through production IPC/HTTP, and verified server state and the newly selected checkmark. Copied the full native session ID in both modes. Restored the original test layout. Checked both themes and a narrow simulated viewport; no provider turn, server restart or deployment was needed. - Targeted header/layout/design regressions: 58 passed. Full desktop suite: 5,532 passed, 5 skipped. TypeScript and production build/output verification pass. Native Windows/Linux UI and live cross-server switching were not repeated. Source and local development app only; no release published. - CI follow-up: make the existing Team Hub maintenance-cancellation test use event-loop notifications instead of blocking executor observers. Keep real temporary-store snapshots, repeated cancellation and exact-fence assertions; use explicit release gates and a bounded watchdog independent of filesystem speed. The regression now runs with a single worker. Local before/after probes reproduce the old failure with a constrained executor and a slow snapshot; both pass with the revised synchronization. No server runtime behavior changed. ## 2026-10-29 - User-authorized mailbox tasks and editable drafts - Existing authorized chat pairs now carry the originating user's exact task without requiring another `@` mention. Shared, idle-wake and busy-mail guidance tells the receiving agent to perform work covered by that instruction without requesting the same authorization again. Peer-written bodies, quoted wrappers, reply links and internal wakes remain distinct from user instructions. - Native steering records the incoming user's text without generated attachment scaffolding. Goal steering suspends earlier task attestation at the actual transport write, binds the correction after a valid acknowledgement and keeps the existing provider owner, routes and tools. Previously accepted messages retain their original source instruction on exact retries. Empty or oversized ordinary tasks still deliver peer mail without truncated authorization. - Inbox messages have an **Edit and send** action. It loads the complete message into the matching chat's composer, preserves existing drafts and attachments, and requires normal Send after review. The original received message and its agent-read state are unchanged. Late responses cannot populate another pane, chat or server. English and Simplified Chinese are included. - Validation uses source-built Electron 0.2.0/build 86 in an isolated native offscreen profile, production IPC/authenticated HTTP and a disposable full server with real Codex. Repeated no-mention delegation produced the requested files without asking for authorization again. A quoted authorization lookalike did not authorize an additional file. Native app clicks appended full mail, preserved an existing draft across chat navigation, and sent user-edited text through normal Send to a real provider response. Light/dark and narrow English/ Chinese presentation were inspected. The 229 focused server tests and 343 desktop regressions pass, along with TypeScript and production compilation. - Component/transport regressions cover unread/read messages, attachments, stale responses, forged provenance, size boundaries and retry identity. Those edge cases are not all repeated in the native UI; Claude-provider acceptance is not claimed. A separate pre-existing limitation was reproduced: an idle native goal resume can acknowledge steering but lacks the runtime capability/ metadata needed by Chats, which fails before mailbox dispatch. This change does not claim to repair that distinct goal-resume lifecycle path. Full CI also exposed an invalid object-only goal authority fixture and a peer-panel assertion preceding its effect; use a real capability fixture and await the existing observer without weakening assertions or increasing timeouts. Changes are source-only and need a matching app/server update; no public release or production server restart was performed. - Latest-main revalidation: integrated `d79677ef` without changing the new workspace/header behavior; the only merge conflict was overlapping log entries. On the matching source-preview test server, native Electron input exercised initial pairing, repeated no-mention Codex and Claude delegation, recipient file creation and replies. Reviewed mail appended to an existing draft; navigation preserved it, a manually edited Send reached a real provider, and split-pane drafting left the other pane unchanged. Checked dark English and narrow light Chinese rendering. Focused checks: 212 server tests and 376 desktop tests pass; TypeScript and production compilation pass. Clarified the documented difference between oversized ordinary-turn fallback and explicit-reference delegation. Live goal steering, cross-server stale replies and attachment retention were not repeated in this acceptance pass; their automated coverage and the separate idle-goal limitation above still apply. No production restart, public release or mobile change. ## 2026-10-28 - Optional desktop workspaces and single-row chat headers - Ordinary chats no longer manufacture a default workspace. The chat action menu's Create workspace entry enables the workspace/layout bar on demand. Preserve existing named and multi-pane layouts, while migrating the old implicit single-chat default back to ordinary mode. Closing the last workspace keeps the selected chat and its draft; no chat or directory is deleted. Workspace preferences remain scoped to the connected server. - Keep title, folder, provider session ID/copy and right-side actions in one chat header, including three- and four-pane layouts. Remove the redundant pinned Chat/file row while viewing chat and the Changes entry altogether. File tabs and their Chat return action remain available when viewing a file. - Source base: `aa1b8c47`; local desktop metadata: 0.2.0/build 86. Exercised the visible native Electron app with disposable chats on an authorized AgentsServer 1.0.11-beta.12 test instance: ordinary mode, explicit and repeated workspace creation, three/four panes, closing a workspace, last-workspace exit, and persistence on reload. Opened an existing file through the real preload/IPC/server boundary and returned to chat without the duplicate header. Checked dark and simulated light system appearance and narrow three-column layouts. Primary services were not restarted. No new provider turn was needed for this layout change. - Regression coverage includes legacy preference migration, explicit opt-in, draft retention, per-pane menu ownership, hidden tab navigation and file return. Desktop suite: 5,528 passed, 5 skipped; build/license script tests: 8 passed. TypeScript, production compilation and output verification pass. Native Windows/Linux presentation and live cross-server switching were not exercised; ownership/isolation remain covered by automated regressions. Available in source and the local development app only; not published. ## 2026-10-28 - Cursor compound tool IDs and live analytics acceptance - Cursor CLI 2026.10.22 emits compound tool correlation IDs containing an LF separator. The server previously rejected these and interrupted otherwise valid tool-using turns. Recognize the observed bounded compound format and map its entire value to a deterministic, control-free digest. Preserve ordinary IDs, separate the encoded namespace from literal provider IDs, and continue rejecting empty, oversized and unrelated control-bearing IDs. Tool start and completion retain the same key without conflating distinct calls. - Reproduced the native CLI format before the fix. Added a failing-before/passing- after parser regression and production-runner subprocess coverage. An isolated native offscreen Electron app connected through its production IPC/HTTP boundary to a source-built test server and real Cursor API key: two file reads completed with distinct paired IDs, their results appeared in the UI, and a subsequent tool-using follow-up succeeded. No production service was restarted or deployed. - Repeated acceptance in the user-designated installed test instance, retaining its existing connection identity and configuration and changing only the parser. The visible native desktop app selected Cursor API key and a disposable working directory through the UI. Three real-provider turns and six paired tool calls (including four file reads) completed successfully, and reloading retained the results. The main service was left running unchanged; this test-only patch is not a published update. - Removed the temporary local-launcher analytics cancellation used during automated QA; this was not shipped app code. With normal consent handling unchanged, three real UI submissions produced exactly three Cursor Custom API message events in the actual Mixpanel project's raw export. Verification must distinguish raw ingestion from dashboards that exclude test installations. Earlier intercepted events were not replayed or backfilled. No new payload fields, credentials or message content were added. - Follow-up desktop acceptance found that `/model default` could remain in the composer after a successful API-model reset: the server returned its concrete default rather than null. Clear the command after that confirmed normalization, retaining existing failure/newer-draft/server-switch guards. Added a regression that fails before this correction. No prompt or provider turn is submitted. - Validation: 120 focused Cursor parser/runner/API-key regressions, 577 desktop regressions, TypeScript and Electron production build/output verification pass. Source repair is not an installed server update: existing services need to load the new parser before their tool-using chats benefit. ## 2026-10-28 - Custom API message metrics and local model commands - Desktop accepted-message analytics now includes an allowlisted connection kind (`default` or `custom_api`) when explicitly known at submission. Keep it separate from the five Agent families; omit legacy/unknown values. No model, endpoint, credential, account, chat ID or message content is collected. Update the event catalog and source privacy disclosure accordingly. - Typed `/model` opens the existing model picker; `/model ` updates the session through the same model/effort logic as that picker. Commands never enter prompt composition or count as sent messages. Unknown, ambiguous and locked choices are rejected without starting a turn. Preserve attachments, failed-command drafts and newer drafts; ignore late results after switching servers. Keep existing slash skills and ordinary text intact. - Based on main `9a750c7a`; source-built desktop metadata remains 0.2.0/build 86. An isolated native offscreen Electron app connected to a test AgentsServer 1.0.7-beta.11 reproduced the old typed command starting a Claude turn without changing the selected model. After the fix, real keyboard input changed Claude and Codex models through production preload/IPC/HTTP and server persistence without adding turns. Subsequent real provider messages returned expected replies. Repeated switching, default reset, reload persistence, unknown-model draft retention and the existing picker were also exercised. - Captured and cancelled test analytics before network transmission: default Claude/Codex messages included the new property; model commands did not emit message events. Custom API classification across all five Agents, Custom API accepted sends, queued/failed sends, stale-response ownership and failure rollback were verified with deterministic tests. The isolated test service had no configured Custom API endpoint; see the user-acceptance follow-up below. No production analytics events were sent. - 576 focused regressions, TypeScript and the production build/output check passed. No server/mobile changes, release, deployment, or website publication. - User-acceptance follow-up: at the user's request, gracefully closed the prior local source app and launched this branch with its existing profile and cache. Rebuilt and verified the loaded renderer asset, online connection, composer and model picker through native input. No server restart or release occurred. The user's subsequent live Cursor Custom API submission produced one accepted `message_sent` event with `backend: cursor` and `connection_kind: custom_api`. This verifies client submission acceptance and classification, not provider completion. Test telemetry is captured locally and cancelled before Mixpanel submission. Other live Custom API backends and a live queued-send scenario remain untested; their deterministic coverage is described above. ## 2026-10-28 - Claude native cron activity - Track successful native CronCreate receipts and the native command/prompt hook that identifies their later wakes. Give each observed wake its own run for live activity, tools, output, Stop, and provider file delivery without submitting a replacement prompt or restarting Claude. - Keep the chat idle between occurrences. A clean Stop interrupts the current occurrence and preserves future native schedules. Preserve the schedule's original authority across intervening human messages; successful CronDelete and provider teardown retire the corresponding registration. Native cron inputs stay out of ordinary user bubbles. - Exact schedule IDs are not exposed on SDK wake frames. Identical registered prompts retain their candidate IDs and use the most recent matching original owner, without combining grants. Simultaneous background-task provenance is retained separately. - If a native wake wins before a human message reaches Claude, preserve the original accepted run and its effects, then submit it once after the native work finishes. Stop/delete can cancel that unsent message. Replies and activity stay in execution order while the human input remains at its original position. If Claude already accepted the input and folds it into native work, retain proven progress and native acknowledgment/result boundaries without replay. - Cancellation during native-to-human ownership restoration emits one terminal event and keeps the original accepted metadata until cleanup finishes. The production-supervised regression covers cancellation before, during, and after restoration without sending the canceled query or evicting Claude. - An isolated production-app baseline with real Claude reproduced a native cron command executing while the app remained idle without Stop. The patched app showed the wake as active, streamed its tools and reply, delivered its attachment, and reopened without duplicate output or an internal user bubble. The recurring app check stopped one occurrence, sent a human follow-up, then observed the next occurrence and successful schedule deletion on the same native thread. Scrolling the virtualized timeline confirmed exactly the original user messages remained visible. - A deliberate admission delay through the actual app exercised a native cron wake winning before human submission: the original human query arrived once afterward, the connection survived, and replies appeared in the correct order before and after reopening. A separate real native overlap probe exercised Claude folding already-submitted human input into a scheduled turn, retaining progress before acknowledgment without an extra query or synthetic result. - SDK, runner, admission, queue/recovery, and timeline regressions passed, along with TypeScript, production compilation, and compiled-entry verification. Two earlier app-harness assertions were corrected: a follow-up accidentally prohibited future scheduled work, and a DOM count ignored virtualization. They are not counted as passing acceptance; the corrected full run passed. - Scope: schedules observed in the current native connection. Recovery of durable schedules after provider teardown is not established by these checks. No app installation, release, or live-service change has been performed. ## 2026-10-28 - Per-chat code intelligence - Add Python and TypeScript/JavaScript completion, hover information, diagnostics, definition navigation, and references to the workspace editor. Language analysis runs on the chat's connected server and uses its project configuration. Separate editor connections preserve independent unsaved buffers; navigation opens files in the originating pane. - Add Files/Symbols search beside the composer. Selected paths and optional line ranges become ordinary editable draft text, with removable references. Symbol startup waits for the language server's project index; it does not send workspace contents or hidden instructions to a chat provider. - Discover existing Pyright and TypeScript language servers on demand. Missing tools and connection failures show Setup/Retry while file editing and the composer remain usable. Scope requests to the server, chat, and editor; discard stale responses and reap only the owned language-server processes. - Validation: focused renderer, native HTTP/WebSocket, scope-fencing, real language-server, lifecycle, packaging-manifest, and license checks passed. TypeScript and production Electron compilation passed. Fresh TypeScript independent-file symbol search also passed ten isolated process runs. - Exercised the compiled app in an isolated native offscreen Electron profile against real authenticated AgentsServer instances and actual Pyright and TypeScript language servers: completion, unsaved diagnostics with disk preservation, F12 navigation, references, unopened-file symbols, editable draft references, chat/server switching, missing tools, installation followed by Retry, and language-process failure followed by Retry. Inspected English light/dark, narrow layouts, Chinese labels, and diagnostic tooltips. No renderer errors occurred in accepted runs. Provider catalog fixtures were used only to create disposable chats; no AI-provider request was needed. - This change requires matching desktop/server support and the documented language tools. No release, deployment, user installation, or live-server restart was performed. Packaged Windows/Linux desktop validation remains outside this local macOS acceptance. ## 2026-10-27 - Honest native sign-in status in My Agents - Separate credential discovery from verified native login. A reported authentication failure now overrides retained credential files; detected but unverified credentials show a neutral status rather than a connected check. Native and Custom API cards remain independent. Keep send-time verification and retries available after credentials are renewed outside the app. - Add English/Chinese guidance and regressions for expired credentials, passive detection, latest-diagnostic selection, Recheck, recovery and independent API connectivity. No provider login, token storage or server admission changes. - Based on main `eb165198`; availability is local source/build only. An isolated offscreen Electron app reproduced the previous false Connected label using a real authenticated server catalog, then verified the corrected CLI status, login guidance and reopening My Agents through the production renderer, preload, IPC and client transport. Inspected English light/dark and Chinese narrow-window layouts. No server deployment or user-app replacement. - 86 focused checks, TypeScript and production compilation passed. One full regression run passed 5,484 tests (5 skipped), plus 8 packaging/license checks. A repeat full run hit an unrelated existing SecurePeerPanel timing assertion; all 156 tests in the focused/auth and secure-peer follow-up run passed. That repeat-suite failure is recorded, not counted as a successful full run. - Live credential renewal and an actual post-login provider request remain pending. Recheck failure/recovery uses component fixtures; the live legacy server's forced catalog refresh was deliberately not invoked because it may initiate an unrelated provider auth handoff. Real-app inspection used passive reads and did not modify credentials or run agent tasks. - Follow-up: launched the complete compiled Electron app interactively with an isolated cache and existing connection metadata for user acceptance. Native clicks opened My Agents and expanded Claude. The server now reported a successful native authentication, and the CLI card correctly showed Signed in rather than relying on credential presence. This observes recovery; the login operation and provider request were not initiated by this test. All 86 focused checks, TypeScript and compilation passed again. User chat acceptance is pending; no server restart, deployment, publication or PR was performed. - PR validation: repeated the full suite successfully (5,484 passed, 5 skipped, plus 8 packaging/license checks), TypeScript and production compilation. In an isolated native offscreen app, opened My Agents, inspected the verified CLI status, then typed and sent a short no-tools prompt to an exact-ID-bound disposable Claude chat. The real server/provider returned the expected reply and the renderer displayed it without errors. Screenshots stay local; account metadata is not attached to the public PR. Login renewal itself remains a native-provider operation, and this display fix does not prevent OAuth expiry. ## 2026-10-26 - Claude scheduled-message import recovery - Keep source-proven scheduled inputs out of ordinary user bubbles on first import. Preserve the scheduled-job card, result, genuine messages, and separate occurrences of recurring jobs. - Recover native Claude transcripts containing NUL-only padding records without changing source bytes, checkpoint hashes, or cursor offsets. Malformed JSON and ambiguous ownership retain the existing behavior; server journals remain strict. - Reader, history-repair, and provider-history synchronization regressions pass. Native offscreen desktop acceptance against an isolated production server reproduced the duplicate, then verified its removal with real Claude output. One ordinary turn and two scheduled turns completed successfully; repeated app reopen checks preserved both job results and the genuine user message. - An initial harness assertion incorrectly expected different job revisions to share a card. The corrected assertion passed against the retained fixture. A redundant fresh-provider rerun encountered expired authentication; later history-reopen checks completed without additional provider calls. - Availability: source fix only. Desktop package unchanged; no release, installation, live history rewrite, or server deployment performed. ## 2026-10-27 - Queued-message saves during reconnect - Queued edits can finish as soon as the current server passes its existing identity check, without waiting for an unrelated chat/job list refresh. Reuse the same health request; retain profile-switch and validation failures. Other mutations retain their existing metadata ordering. - Native offscreen production-app acceptance against an isolated full server with real Claude reproduced the old blocked Save by delaying the chat-list response after a simulated health failure. The corrected app persisted the edit and closed the editor while the list remained delayed and Claude kept working. Repeat save, Cancel, an injected save failure, manual retry, and reopening all preserved the expected draft and queued message. - All 460 service tests, TypeScript, production compilation, and compiled-entry verification passed. Regression tests fail against the previous service. The transport delays/errors were controlled test faults; no live incident timing or signed-package acceptance is claimed. - Availability: source fix only; no app installation or live-server change. ## 2026-10-27 - Claude goal activity and recovered timeline order - Distinguish a saved Claude goal from live execution, waiting, idle, and an unloaded session. Remove the wall-clock timer that implied continuous work. Show only native evaluator duration and offer an explicit **Continue goal** action in the same thread when Claude is idle or unloaded. Preserve native failed verdicts and their reasons. - Place recovered provider replies between the appropriate native turns using their original timestamps and identities. Keep separate replies separate, retain genuine identical-text messages, and keep history imports from ending the active turn. Live streaming remains incremental after catch-up. - Capture Claude's actual background-task completion wakes after the parent result. Bind the observed wake to a new owned run before tools execute, without submitting another prompt or keeping background Bash permanently busy. Preserve the existing task authorization and native thread. - Native offscreen desktop checks used production IPC and authenticated HTTP against isolated full servers. A real Claude goal exercised working, Stop, idle, Continue, native completion, chat navigation, narrow/light/dark layouts, and reopening with the same thread. The isolated server restart check reopened the app; automatic reconnection after restart is not an accepted boundary. - Real Claude background tasks exercised idle between turns, live wake output, file attachment through the provider helper, history reopening without a duplicate, stopping a second wake, and a subsequent human message. Separate synthetic journal fixtures through the actual app reproduced the old ordering and verified corrected chronology after navigation and reopening. - Focused SDK, goal, history-repair, native-runner, shared-transcript, renderer, and localization regressions pass, along with TypeScript and desktop production compilation. Timeline coverage includes incremental projection, pagination, a large import batch, and silent native-wake boundaries. - Availability: source changes and local production compilation only. No release, installation, live history rewrite, or server deployment performed. ## 2026-10-26 - Claude scheduled-message import recovery - Keep source-proven scheduled inputs out of ordinary user bubbles on first import. Preserve the scheduled-job card, result, genuine messages, and separate occurrences of recurring jobs. - Recover native Claude transcripts containing NUL-only padding records without changing source bytes, checkpoint hashes, or cursor offsets. Malformed JSON and ambiguous ownership retain the existing behavior; server journals remain strict. - Reader, history-repair, and provider-history synchronization regressions pass. Native offscreen desktop acceptance against an isolated production server reproduced the duplicate, then verified its removal with real Claude output. One ordinary turn and two scheduled turns completed successfully; repeated app reopen checks preserved both job results and the genuine user message. - An initial harness assertion incorrectly expected different job revisions to share a card. The corrected assertion passed against the retained fixture. A redundant fresh-provider rerun encountered expired authentication; later history-reopen checks completed without additional provider calls. - Availability: source fix only. Desktop package unchanged; no release, installation, live history rewrite, or server deployment performed. ## 2026-10-25 - Mobile provider pre-merge validation - Integrated main `3ac069e`; only the development log required conflict resolution, retaining both sets of entries. Updated older source-contract tests and the composer fixture for the additional providers. - All 428 mobile MJS checks passed, with one generated-native-metadata check skipped. Session creation, provider selection, recovery, optional permissions, transport, cross-chat references, notifications and Android discovery tests passed, along with TypeScript and iOS/Android production JavaScript exports. - Maestro operated the real iOS simulator through Settings, update destinations, My agents and Gemini chat navigation. Home creation retained Gemini Custom API and Default permissions, confirmed through authenticated server readback. Initial scripted attempts missed a navigation tap and a keyboard-obscured row; waiting for navigation and bringing the row fully into view completed the flow. These failed attempts are not counted as passing acceptance. - The user confirmed the refreshed physical iPhone Dev build works. Existing cross-agent receipt checks still show no duplicate mail or pending QA work. Store handoff, signed APK installation and Android provider end-to-end checks remain outside this validation. No release or TestFlight upload was performed. ## 2026-10-25 - Mobile Gemini identity and recoverable new-chat defaults - Embed the desktop multicolor Gemini SVG in the mobile bundle, without a network image dependency or monochrome tint. Light and dark iOS simulator screenshots show the multicolor mark. The dark-mode scripted scroll to a specific QA row timed out; direct screenshot inspection verified other Gemini rows instead. - Home new-chat creation keeps an available inherited provider/mode, otherwise randomly chooses from available choices across all five agents. Report no available agent only when the candidate set is empty. Preserve model and permissions after message failures; an earlier local fallback based on turn failure was too broad and has been removed. On confirmed pre-creation API/workspace validation failures, try remaining choices at most once each; never retry ambiguous writes, authentication failures, or explicit provider forms. Reset provider-specific defaults when switching; preserve existing chats. - Confirmed unavailable native diagnostics override stale credential presence; Custom API and native modes retain independent availability. Regression coverage includes every agent-to-agent fallback, all-unavailable cases, same-agent API fallback, and bounded retries across rejected candidates. Ten randomized session-creation runs passed; actual simulator creation also verified fallback to OpenCode after a confirmed invalid-workspace rejection. - Actual iOS simulator testing against the isolated beta.12 server reproduced a deleted Gemini workspace, then verified fallback creation and a real Claude reply. Mobile mention selection granted routes to all five providers. Codex sent mail to Claude, Cursor, and OpenCode and read their explicit replies. - Gemini cross-chat preflight was blocked by native workspace trust, while ordinary replies worked. After explicit approval to trust only the isolated QA directory, actual simulator mention/send flows exercised Gemini against Codex, Claude, Cursor, and OpenCode in both initiating directions, with native one-time tool approvals and explicit return acknowledgments. No permission policy was disabled. Initial Claude initiation used a legacy target index and failed; an explicit route-discovery retry succeeded. This first-attempt provider tool-use failure remains a limitation, not a passing first attempt. Receipt checks found exactly one request and one acknowledgment per initiating direction, with no duplicate messages or pending QA turns/approvals. - Simulator acceptance also verified that after a Gemini communication error, Home new-chat creation retained Gemini and received a real ordinary reply. An intermediate retry at the server default directory exposed an API-auth override; confirmed invalid-directory creation still falls back to another provider rather than silently changing Gemini's authentication. - Session-creation and chat-reference regressions, 51 focused component/source checks, TypeScript, and iOS production JS export passed. Availability is local development only; no release or TestFlight build was published. ## 2026-10-25 - Mobile optional agents and discoverable app updates - Add capability-gated OpenCode and Gemini choices, API-aware model catalogs, saved selections, permissions, and new-chat inheritance. OpenCode free models remain selectable without login credentials; Gemini uses configured APIs and its required client protocol. Preserve bound API catalogs in chat details. - Add run-owned Gemini approvals with native-control authentication, duplicate response protection, foreground/event invalidation, and server-generation fencing. Do not expose unsupported provider reloads or Gemini forks. - Move App updates into the primary Settings categories on both platforms. iOS opens App Store or TestFlight. Android discovery now reads bounded release pages; desktop releases can no longer hide Android updates after page one. Failed checks do not persist a successful up-to-date cache entry. - Actual iOS simulator acceptance against AgentsServer 1.0.11-beta.11 verified OpenCode free-model and Custom API replies, Gemini replies and contextual follow-ups, permission deny/allow-once with file readback, and OpenCode model and Plan-permission inheritance. One default OpenCode model returned an empty response in its native transcript too; an explicitly selected free model and the configured API returned normally. No server code or deployment changed. - Settings/agent/approval/interaction component checks, provider and transport regressions, TypeScript, iOS production JS export, and a local Android debug APK build passed. Real public release discovery found an Android candidate on page two. Android updater Settings, offline failure, and successful retry were exercised on an emulator. iOS Settings retained both agents and update destinations in dark mode. - Availability: local development builds only, based on main `7d94b05`. iOS store installation is not testable on a simulator; physical-device App Store/TestFlight handoff and a signed production APK upgrade remain unverified. Android provider end-to-end acceptance remains pending. ## 2026-10-25 - Mobile new-chat defaults follow the last open chat - Inherit the last open chat's agent, native or Custom API mode, model, effort, and agent-specific permissions when creating a chat from Home. Keep saved preferences as the fallback when no valid chat is selected on that server. - Resolve omitted permissions with a bounded detail read before creation; failed reads remain retryable instead of silently resetting permissions. Preserve duplicate-tap protection and connection-generation checks. - iOS simulator acceptance reproduced Claude reverting to Codex on the previous client, then verified Claude Plan and Codex read-only/on-request inheritance through the actual app with authenticated server readback. Temporary chats were removed without starting provider runs. - Session-creation regressions, 39 interaction tests, TypeScript checks, and iOS/Android production JS exports passed. Custom API selection, Cursor, lookup failure, and server-switch races have automated coverage; Android interaction and physical-device acceptance were not exercised for this change. ## 2026-10-25 — Codex child cancellation acknowledgement (#85) - Keep a localized Stop subagents action available for an idle Codex chat with authoritative, owned live children, without changing Send into Queue. Ignore forked history and unrelated roots; refresh the selected chat's child snapshot once after stopping, fenced by the current server and timeline subscription. - Reconcile children already terminal when Stop reaches the native runtime. Report pending children and interrupt errors instead of treating parent completion as successful child cancellation. Recheck ownership after native reads and preserve newer lifecycle transitions during stop confirmation. - Source: `fix/codex-subagent-stop-85`, based on `7d94b05b`; local development build only. Both desktop and server changes are required; no release published. - Actual visible native Electron acceptance exercised composer submission, native Codex child execution, Stop, reopening the chat, and repeated use. Native provider records confirmed interrupted child turns; the stop control cleared, an independently running control chat continued, and another pane's draft was retained. The actual IPC/authenticated HTTP/native-provider boundary was exercised without a stub provider. - The installed native provider retained the parent turn until child completion or interruption, even after emitting its final reply. The fully idle-parent case, stale terminal state, missing runtime, concurrent state/ownership changes, pending errors and legacy response compatibility have regression coverage; that historical idle-parent lifecycle was not recreated end-to-end. - Pre-merge validation: 1,190 focused desktop tests and 240 focused server tests passed; TypeScript and Electron production compilation passed. - Updated the compact-composer layout contract to recognize the shared Stop label, including the idle-child label; action placement and sizing are unchanged. - Simplified Chinese consistently calls subagents “子 Agent” in the inspector, limits, lifecycle labels and Stop feedback. Localize the inspector heading and unnamed display fallback without changing provider names, identities or chat content. English remains unchanged. Actual native-app Settings language switching verified the Chinese heading, empty state and per-chat limit text, then restored English. The terminology follow-up passed 308 focused tests, TypeScript checks and the Electron production build; no server restart required. ## 2026-10-25 — Remove host-imposed provider restrictions - Allow native background shell commands and Claude scheduling tools. Remove prompt rules forbidding detached work, full-history subagent forks, other integrations and ordinary use of the chat terminal. Native permissions and the exact authenticated owner of AgentsDock helpers remain intact. - Remove default execution silence deadlines and automatic retained-connection eviction. Idle warnings remain informational; Stop and transport failures still settle owned work. Explicit operator timeouts and retention limits remain effective. Available-memory launch floors now default to disabled, matching the existing opt-in concurrency limits. - Remove the internal Claude MCP two-call ceiling while preserving mutation deduplication and checking authorization before replay. Allow Codex Side chat subagents with verified native ancestry; inherited parent agents do not gain side-chat authority. Retain idle side conversations until explicit cleanup, and use the existing request-body bound instead of a separate 8,000-character question limit in both desktop validation and the server. Older servers' advertised limits remain honored by the desktop. - Allow background-terminal inspection and explicit termination during a Codex turn without replacing its owner. Submit confirmed shell commands through that same live owner and notification stream; native command receipts retain the owner until accepted commands finish, including a late native shell turn. Register requests at the actual transport write boundary and do not replay uncertain delivery. A real service-managed source checkout can restart without the official installer directory layout. Remove the normal restart cooldown; in-flight request coalescing and exact service ownership remain in place. - Real offscreen app/IPC/authenticated HTTP acceptance with Claude Code 2.1.303 verified native background Bash and detached workers continuing after the reply, native schedule creation and retention through the next turn, exact schedule deletion and reuse of the same native connection. This verifies schedule retention, not delivery of a future scheduled wake into the app. - Native Codex 0.156.1 acceptance verified a saved Side chat's full-history child, inherited context and file work while the parent completed normally; original parent readback confirmed completion. Active terminal list, terminate and clean used the exact original owner. Approval ancestry and closed-owner rejection have regression coverage; an actual child approval dialog was not required by the selected native permissions. - Actual offscreen app restart acceptance used a disposable named launchd service and source checkout. Two confirmed requests 4.7 seconds apart each returned acceptance and reconnected to a new process with the same durable server identity. The previous source reproduced the disabled Restart button. Test services, processes and credential copies were cleaned up. - Preserve native background task accounting after a Claude reply. Inspect retained Side chat descendants when preparing an idle-update snapshot, and check the same native generation before using that snapshot. Active work delays an idle update; idle retained conversations do not. Cancelling an optional Claude context sample leaves its live connection and tasks intact. - Actual offscreen app/IPC/HTTP acceptance with Codex verified a confirmed shell command during an existing native tool, one normal completion of the original turn and a successful follow-up. A separate Side chat journey verified a full-history child, an immediate reply, a context-preserving follow-up, and the child continuing after the main chat completed. The production idle-update snapshot reported only that child until it finished, then cleared all work labels. Original credentials were unchanged and isolated copies removed. - Focused server suites cover disabled deadlines, positive configured limits, long-lived native connections, ownership, concurrent helper calls, side context retention and restart races. Cursor/OpenCode execution tests use real subprocess pipes with fixture CLIs; live provider acceptance for those adapters is not claimed. No live server deployment or public release. - App source `492360f7`, paired server source `1b07dbbe`; accepted local macOS arm64 0.2.0/build 86. TypeScript, all 5,460 desktop tests (five skipped), production compilation, bundle audit, Developer ID signing and strict signature verification passed. The exact packaged archive repeated the live Codex active-shell and subsequent-message journey through production IPC and authenticated HTTP. All 91 compiled assets matched the archive, with no renderer errors. Archive SHA-256 remained `c71bd2ccee0c7fdc7387804ff075fd85c82f065dfb4ec929fbf042cd6250cc73` before and after acceptance. Packaged code ran through an isolated native Electron bootstrap; signed-executable launch and Gatekeeper were not tested. No installed app or live server was replaced or restarted. ## 2026-10-25 — Set and replace Claude goals during active work - Remove the desktop busy-turn disable and route active goal requests through the owning Claude session. Native queue admission is acknowledged separately from application of the goal. A running tool and its original subagents continue, including when replacing a goal that is keeping Claude running. - Surface later native refusals in the dialog and beside the composer. Match each pending goal to its exact run and native command receipt; stale history and receipts from another request cannot settle it. Goal input remains separate from the existing Force Send lane. - Explicit Stop withdraws the queued goal and uses native task controls for owned commands that become background tasks during interruption. Wait for their actual terminal receipts before retiring that provider connection. Clear & stop then clears the native goal through the normal idle path. Successful user-requested cancellation settles quietly; genuine refusals and unconfirmed cancellation remain visible errors. - App source `e1cd724c`, server source `0392ba4c`; local macOS arm64 0.2.0/build 85. TypeScript, all 5,458 desktop tests (five skipped), production compilation, bundle audit, Developer ID signing and strict signature checks passed. The server SDK, runner, goal-route and goal-projection suites passed 226 tests. Focused regressions cover native refusal, uncertain delivery, concurrent Force Send, stale results and cancellation-control failure. - The exact packaged archive passed native offscreen interaction through production IPC and authenticated HTTP into an isolated server with real Claude Code 2.1.303. Set a goal while parent and child tools were running, replace it while the original child remained active, and observe the same native session reach Goal achieved after both original tasks completed. The archive and signature remained unchanged, with no JavaScript errors. - Native SDK Stop acceptance verified cancellation of the queued goal, original process exit without completion outputs, and a successful subsequent tool turn. Delayed superseded results and cancellation-control errors were checked with focused fixtures after that native run. Actual app Stop and Clear & stop also passed through production IPC/authenticated HTTP with real Claude: original process IDs exited, pending goals disappeared, cleared native state was read back, and each chat accepted a successful subsequent message without reviving the stopped work. The final Clear run verified quiet cancellation. Test profiles and processes were cleaned up. Signed-executable launch and Gatekeeper were not exercised. - This is a paired app/server source fix and local package, not an installed update or published release. Setting a goal before the provider has bound its initial native session remains outside this active-session acceptance. ## 2026-10-23 — Quieter workspace selection - Replace persistent blue stripes on workspace/chat/file tabs and focused grid panes with neutral selected-tab backgrounds and medium-weight titles. The focused pane's selected tab has a slightly stronger neutral background. - Resize handles use a neutral hover treatment; keyboard focus retains a clear outline. Tab order, drag/reorder, pane selection and file state are unchanged. - Source `88372526`, local macOS arm64 package 0.2.0/build 85. TypeScript, production compilation, all 5,455 desktop tests (five skipped), bundle audit, Developer ID signing and deep strict signature checks passed. - Native offscreen checks exercised the exact packaged archive: select chat and file tabs across three/four panes, retain open files while focusing another pane, resize to the desktop minimum, and inspect dark/light appearance. Native keyboard navigation retained a visible focus outline. Production IPC and authenticated file reads used isolated server state; no provider ran. All 91 compiled assets matched; archive/signature unchanged, no JavaScript errors, and test processes/credentials cleaned up. Signed-executable launch and Gatekeeper were not exercised. - Local desktop presentation change only. No server change, installed-app replacement or publication is included. ## 2026-10-23 — Keep queued messages from crowding the split-pane draft - Give queued messages, attachments and notices a separate scrolling area above the draft. Short panes can grow the composer instead of forcing everything into a 75-pixel box. Keep the text field and Send/Stop toolbar visible together; smaller panes retain a full editable line and scroll longer drafts internally. - Allocate the composer from the space remaining below the goal/context controls. Preserve the same draft, focus and selection when queued messages arrive or disappear. Single-pane queue behavior and sending contracts are unchanged. - Source `6175fb76`, local macOS arm64 package 0.2.0/build 85. All 5,455 desktop tests passed (five skipped), with TypeScript, production compilation, bundle audit, Developer ID signing and deep strict signature verification. An existing asynchronous file-picker focus assertion now waits for focus; its targeted rerun also passed after the initial full-suite timing failure. - Native offscreen baseline reproduced a queued message pushing the input fully outside its visible area. The exact packaged archive then passed queue submission, edit/save with authoritative readback, removal and repeat use, retained multiline drafts, goal/alert pressure, three/four panes, divider dragging, light/dark and narrow-window checks. Draft and queue scroll independently; Send/Stop remain inside the tested panes. - These exercised production IPC, authenticated HTTP and WebSockets in isolated server state with synthetic held turns; no provider ran. All 91 compiled assets matched the package, and its archive and signature remained unchanged. The archive ran through stock Electron offscreen; signed-executable launch and Gatekeeper were not exercised. - App-only local trial; this layout fix needs no server update. No live service, installed app or published release is changed. ## 2026-10-23 — Reorder workspace tabs and preserve Claude tasks on Force Send - Replace boxed workspace tabs with a flat rail and an active underline. Drag tabs to reorder, or use Alt-Shift-Left/Right. Remember order per server without changing selection, drafts, files or pane layouts. Keep the active tab visible after resizing and retain the adjacent new-workspace button during overflow. - Claude Force Send uses native queued-input acknowledgement and Send now, preserving the same process, logical run, tool authority and tracked children. It no longer falls through Stop for ordinary user follow-ups. Explicit Stop keeps its cancellation behavior. Unsupported native protocols leave input queued; uncertain delivery is never replayed or treated as permission to terminate children. Proven prewrite cancellation restores the unsent input. - Accepted follow-ups appear once in desktop/shared history, including during scheduled work. Recovery consumes only acknowledged input and retains a truncated delivery record without resending it. New route grants/provider actions still require their own turn; Force Send retains its original owner. - Source `be4b6806`, local macOS arm64 package 0.2.0/build 85. All 5,457 desktop tests passed (five skipped), with TypeScript, compilation, bundle audit, Developer ID signing and deep strict verification. The scoped server SDK, runner, queue, goal, history and projection suites passed 518 tests. - Real Claude Code 2.1.296 and 2.1.303 probes submitted Force Send during a foreground Bash command. An existing native Agent survived, completed and woke the parent; acknowledgement took approximately 0.7 and 1.2 seconds. A separate explicit Stop probe observed killed/stopped child receipts. These exercised the candidate SDK adapter in disposable state; server queue races and workflow receipts used isolated fixtures, not live user chats. - The exact packaged archive passed native offscreen drag/keyboard, rename, close/add, draft/file/layout retention, per-server persistence, full restart, narrow overflow, resize and light/dark checks through real authenticated HTTP/WebSockets and production IPC. All 91 compiled assets matched; archive and signature remained unchanged. This ran stock Electron with the packaged archive, not the signed executable or Gatekeeper. - Local trial only: no app installation, service update/restart or publication. Workspace tabs work with the existing server; the Claude Force Send fix requires installing the paired server source. ## 2026-10-23 — Readable syncing status in compact chat panes - Place the neutral syncing status beside the working-directory control, outside the compact input's scrolling area. Keep the full input line and toolbar visible without increasing the desktop composer's height. Shared chats show the status without exposing the working-directory control. - Source `99d50432`, local macOS arm64 development package 0.2.0/build 85. Placement regressions failed before the fix; all 5,430 desktop tests passed (five skipped), plus TypeScript, production compilation, bundle audit, Developer ID signing, required entitlements and deep strict signature checks. - Native offscreen acceptance reproduced the clipped input in the previous package, then verified a full input line, readable status, accessible toolbar and retained typing through reconnection. Three/four panes, completed goal banners, light/dark themes and narrow windows passed through production IPC and authenticated HTTP/WebSockets in isolated state. All 91 compiled assets matched; archive and signature remained unchanged. - Reconnecting activity was a disposable fixture; no provider ran. The packaged archive ran in stock Electron offscreen, without launching the signed executable or exercising Gatekeeper. This is an app-only local trial; no server update is required. No installed app, live service or published release was changed. ## 2026-10-23 — Keep every visible chat receiving live updates - Remove the obsolete two-stream eviction from the desktop service. Opening a third or fourth pane now preserves every visible chat's history request and live connection. Explicit pane/workspace cleanup still releases hidden chats; late events from retired subscriptions remain fenced to their original owner. - Clear ended-subscription bookkeeping so explicitly selecting a cached chat can reconnect it. Reconnecting streams keep their owner; this adds no polling or automatic retries. - Source `50c89366`, local macOS arm64 development package 0.2.0/build 85. Regression tests reproduced dropped primary history/completion and stale subscription bookkeeping before the fixes. All 5,429 desktop tests passed (five skipped), plus TypeScript, production compilation, bundle audit, Developer ID signing, required entitlements and deep strict signature checks. - Native offscreen acceptance reproduced missing primary replies in the old package, then verified all three/four panes receiving new assistant, reasoning, tool and completion events through production IPC, authenticated HTTP and WebSockets. Working, Running and Stop appear during held fixture turns and clear on completion. Concurrent hydration, workspace switching, hidden-pane cleanup and catch-up after reopening passed. All 91 compiled assets matched; the packaged archive and signature stayed unchanged. - Provider events and active-run ownership were synthetic fixtures in disposable server state; no provider ran. The actual packaged archive ran in stock Electron offscreen; signed-executable/Gatekeeper launch was not exercised. No installed app, running service or published release was changed. This is an app-only local trial using existing contracts; no server update is required. ## 2026-10-23 — Three-column chat workspaces - Place the new-workspace button immediately beside the workspace tabs. Keep it accessible when tabs overflow and reveal the newly selected tab. - Add three full-height, independently resizable columns alongside the existing three-pane split. Remember each workspace's layout and widths separately; retain pane assignments, drafts and file tabs when changing arrangements. Keyboard focus navigation follows the visible columns. Command-Control-Left and Command-Control-Right cycle workspace tabs, including from the composer. - Show each pane's chat title once in its pinned tab. Double-click that tab or use Rename chat in the actions menu to rename it; keep the folder control accessible in the toolbar. Selecting a chat already assigned to a workspace returns to that workspace and focuses its pane instead of opening a duplicate. - Local macOS arm64 package: source `7187ce27`, development version 0.2.0/build 85. All 5,425 desktop tests passed (five skipped), followed by TypeScript, production compilation, bundle audit, Developer ID signing, required entitlements and deep strict signature checks. - Native offscreen acceptance used the packaged archive through production IPC and authenticated HTTP against isolated services. Verified both three-pane arrangements, both draggable dividers, keyboard resizing, workspace shortcuts, sidebar jumps without duplicate placement, title consolidation and real rename, pane-local file opening, retained drafts/layouts after reload, tab overflow and narrow presentation. All 91 compiled assets matched; archive and signature remained unchanged. No provider ran. Signed-executable/Gatekeeper launch was not exercised; this remains a local, unnotarized trial. - Scope: app only, using existing server contracts. No server update is required; no installed app, live service or published release was changed. ## 2026-10-23 — Cron unread grouping on older servers - When cron inclusion is off and an older server omits output provenance, use the existing scheduled-job list to keep registered cron chats in their folders or pinned section. Visible ordering and keyboard navigation use the same rule. No additional requests, transcript scans or polling are needed. - Preserve saved inclusion settings, manual unread overrides and unread dots. Newer servers continue distinguishing ordinary replies from cron output. Older servers cannot distinguish replies within a mixed-use scheduled chat; their fallback applies to the registered chat until inclusion is enabled. - Four new regressions failed before the fix. All 281 focused sidebar, session and store tests pass. The local macOS arm64 package uses source `13193ad2`, development version 0.2.0/build 85. All 5,398 desktop tests passed (five skipped), followed by TypeScript, production compilation, bundle audit, Developer ID signing, required entitlements and deep strict signature checks. - Native offscreen acceptance reproduced the failure in the prior packaged app, then verified the corrected signed archive against an isolated authenticated server returning missing/null legacy provenance. Settings on/off, preserved saved choices after restart, pinned placement, ordinary and manual unread, keyboard order, server switching, offline cached jobs and reconnect passed. Auxiliary job-add/remove fixtures used the existing read-only jobs-list IPC through real HTTP; the existing background polling cadence was not changed. All 91 compiled assets matched and archive/signature remained unchanged. No scheduled job or provider ran. Signed-executable/Gatekeeper launch remains untested; this is a local unnotarized trial, not a publication or installation. - Scope: app compatibility fix; no server update or live-state change required. ## 2026-10-23 — Tabbed chat workspaces and pane-local file tabs - Add named workspace tabs with one, two, three or four chat panes. Three panes use a tall left pane and stacked right panes; four use a resizable grid. Each pane owns its chat and file tabs, so opening a file leaves other chats visible and preserves the conversation in its pinned tab. - Remember workspace layouts, focused panes and divider sizes per server. Retain chat drafts, scroll positions and file edits while switching workspaces, servers and layouts; reducing the pane count keeps hidden chat assignments. Keyboard pane navigation follows the visible grid and file commands target the focused pane. - Focused renderer, editor and store regressions, TypeScript and production compilation pass. Native offscreen desktop acceptance exercised all four layouts, mouse and keyboard dividers, pane-local file reads and saves, dirty edits, drafts, exact scroll restoration, named workspace switching, server isolation and app/server restart persistence through real IPC and authenticated HTTP in disposable state. Provider events were synthetic. Narrow-pane headers retain compact controls without wrapping their provider status. Escape, focus loss and navigation clear the drag highlight immediately. - Local macOS arm64 package: source `fe26f598`, development version 0.2.0/build 85. All 5,393 desktop tests passed (five skipped), followed by production compilation, bundle audit, Developer ID signing, required entitlements and deep strict signature checks. - Packaged native offscreen smoke loaded main, preload and renderer directly from the signed archive; all 91 compiled assets matched. Pane-local file reads/saves, four-pane tabs, per-server state, app restart, retained drafts, native drag cancellation and narrow light layouts passed. The archive and signature were unchanged after testing. Signed-executable/Gatekeeper launch was not exercised; this remains a local, unnotarized trial build. - Scope: desktop only, using existing server contracts. No server update is required for the tabbed grid. This local trial does not publish a release or replace an installed app. ## 2026-10-23 — Local desktop package with sidebar controls - Built macOS arm64 from `29951a90`, combining the sidebar settings with the merged TIFF-preview and split-chat file-opening fixes. Development metadata remains 0.2.0/build 85; this is a local package, not a published release. - The full desktop suite passed: 5,367 tests, five skipped. TypeScript, production compilation, bundle audit, Developer ID signing, entitlements and deep strict signature checks passed. - Packaged native offscreen checks loaded main, preload and renderer directly from the signed archive. All 91 compiled assets match. The three sidebar switches, cron grouping, running-chat placement and restart persistence passed through real IPC and an isolated authenticated server. The archive and signature remained unchanged. Provider output was synthetic; no provider or scheduled job executed. Signed-executable/Gatekeeper launch was not tested. - Keep metadata-free server summaries sparse, fixing the existing summary-size regression without changing its threshold. Known counters and pending-recovery nulls remain explicit; 44 targeted server tests pass. The packaged smoke used this corrected server source. - Availability: local app package only. The grouping switches work with older servers; distinguishing cron output requires the matching server patch. Nothing was installed, published or deployed. ## 2026-10-22 — Configure Unread and Running sidebar sections - Add three General settings remembered locally per server: Unread chats at the top (on by default), Include cron jobs in Unread (off), and Show Running section (on). Hidden sections return chats to their saved folders or pins; unread dots and activity indicators remain visible. Active chats stay out of Unread even when Running is hidden. Keyboard navigation follows the same order. - Distinguish scheduled output from ordinary unread replies on the server. A later cron result cannot hide an unread ordinary reply, and explicit Mark unread still promotes the chat. Imported history does not count as a new ordinary reply. Existing read cursors, folders and pins remain unchanged. - Recover existing provenance once in a background worker and persist a journal checkpoint for subsequent restarts. Unknown provenance and older servers keep the previous unread behavior; catalog, session-list and turn-admission paths do not wait for recovery. Include the helper in server installation and packaging inventories. - Verification: 474 focused desktop tests, 14 unread-provenance regressions, related server projection/provider tests and runtime archive checks pass. TypeScript, Python compilation, production Electron compilation and entry verification pass. Native offscreen Electron exercised the controls through production IPC and authenticated server transport, including active unread placement, manual unread, mixed replies, keyboard navigation, narrow/light layouts, profile switching and app/server restarts. Final server checks also recovered legacy metadata without changing read cursors or event journals. These used synthetic provider events in disposable state; no live provider or scheduled job ran. A packaged release and other native platforms were not exercised. - Availability: source changes for desktop and server; cron filtering requires both. No release, installation or live-server deployment is included. ## 2026-10-22 — TIFF previews and split-chat file opening - Desktop attachment and workspace viewers render TIFF files through a PNG preview of the first page. Portable background workers decode the image; downloads and native Open preserve the original file. Failed previews retain download recovery. Queued previews retain stream backpressure instead of buffering every visible TIFF, and remote transfers gain no new timeout. - Split chat delivers a fresh file-open event: Chromium otherwise drops the replay of the original stopped event. Delivery also waits for saved workspace tabs and mount effects, so restoration cannot select the prior tab over the clicked file. Both conversations and their drafts remain in place. - Validation: the full desktop Vitest suite passed (5,340 tests; five skipped), followed by 42 media tests and 201 split/workspace tests on the final changes. Eight build/license script tests, TypeScript and production compilation passed. - Native macOS offscreen acceptance reproduced both failures before the fix and passed the final production-compiled candidate through real authenticated server routes. Checks covered attachment/workspace TIFF rendering, unchanged original download bytes, split file links and Files browsing, repeat opens, focus, chat/file draft preservation, navigation, light/dark and narrow views. Typing and idle checks observed no workspace requests or long tasks. Test chats were synthetic; no provider was invoked. Native Linux/Windows execution remains untested. - Local macOS arm64 package: source `541f60a17d42573c97048323dd2f9ea7152a9b60`, development version 0.2.0/build 85, Developer ID signed. Bundle, entitlements and deep strict signature checks passed. All 91 packaged compiled assets match the accepted outputs; TIFF worker, dependencies and notices are present. The same native journeys passed loading main/preload/renderer and TIFF worker from the signed app archive under stock Electron 43.1.1 offscreen. Archive bytes and signature remained unchanged. The signed executable's own launch and Gatekeeper acceptance were not exercised; this is not a notarized release. - Availability: local Mac app package. This desktop fix requires no server update; nothing was installed, published or deployed. ## 2026-10-22 — Reduce execution-worker descriptor pressure - Close disconnected gateway sockets without cancelling accepted application work. Retain buffered request bodies and WebSocket disconnect details, and keep closing transports tracked for explicit worker shutdown. - Share overlapping runtime catalog refreshes and provider installation probes instead of launching duplicate subprocess pipelines for each client. Preserve explicit refresh, deadline and authentication-generation behavior; one departing caller does not cancel work shared by the remaining callers. - Raise low inherited worker/gateway file-descriptor soft limits toward 8192 within the existing hard limit, including with older service definitions. Preserve higher limits and continue startup if the OS refuses the increase. - Validation: 220 tests passed across transport/gateway integration, real child process descriptor limits, startup/installation, concurrent runtime catalog requests, sign-in handoffs and custom-provider readiness. One native systemd parser case was skipped on macOS. The installer fixtures were run with their expected umask 022; their failure under umask 077 also reproduces on main. Python compilation and diff checks passed. Tests use isolated state and synthetic providers; no live OAuth or production-load acceptance is claimed. - These changes address reproduced pressure mechanisms; they do not establish which descriptors caused a previously restarted process to exhaust its limit. - Source change only; no user services, installations or credentials changed. ## 2026-10-21 — Recover held mobile messages after a stopped queue handoff - Handle bulk Stop events on mobile, clear obsolete promotion flags and retain delivery-uncertainty warnings. If a promoted row has already left the mobile shelf, recover it through the existing guarded authoritative queue read. Recovery never resends messages or introduces polling. - Preserve a message held by Stop even when its earlier Run now request succeeds; show an informational paused status instead of a misleading send failure. Add reducer and live-store regressions to CI, including stale navigation reads. - Make the server cancellation test's startup barrier tolerate busy CI workers; cancellation and slot-ownership assertions remain unchanged. - Verification: the new mobile regressions fail before the correction and pass afterward. Mobile type checking, iOS JavaScript export, related store suites, 46 interaction contracts and 121 targeted server tests pass. The broader local server shard is not clean in this environment and is not a passing CI result. - Native offscreen Electron against the exact test-server source and real Claude exercised Run now immediately followed by Stop: both messages returned paused in order, one explicit retry started exactly once and the other stayed held. Editing, navigation, renderer reconnect and removal used production IPC/HTTP. - Native iOS Dev with the patched bundle observed the same real cross-client handoff race and restored both held rows. A native Run now tap completed one real Claude request exactly once while the other stayed paused. Editing, removal, navigation and cold relaunch retained the expected queue state; relaunch did not resend the held message. The queue area scrolls to reveal actions when its content exceeds the available height. Android native acceptance was not exercised. No release or deployment to users is included. - Follow-up manual testing in the visible native desktop was accepted by the user against the patched test service. A message already admitted before Stop remains interrupted in history; only an unadmitted promotion returns to the held queue. This distinction prevents automatic duplicate execution. - Pre-review recheck: 280 targeted desktop tests, 126 targeted server tests, mobile queue reducer/store regressions and both client type checks pass. Full hosted CI remains a separate merge gate. ## 2026-10-21 — Restore queued messages after Stop interrupts a handoff - When Stop pauses a message that was about to start, return it to the held queue immediately and clear the obsolete Starting state. Retain its text, attachments and order, and restore Send now, edit and remove controls. - Treat paused snapshots from older servers as paused even if they still carry a promotion flag. Preserve a paused message when a successful Send now response is followed by a newer Stop; do not filter it out of the cache. - Describe pending admission as waiting for the provider instead of claiming that the message was already handed over. Explicit Stop still prevents automatic delivery; retry uses the existing authenticated queue operation. - Verification: five new queue/projection/UI regressions fail before the fix; 190 focused server tests and 964 related desktop tests pass. TypeScript, Python compilation, production Electron compilation and entry verification pass. Native offscreen Electron reproduces the stale Starting row before the fix and restores retry afterward, through production IPC, authenticated HTTP, websocket events and server queue admission. Each of two messages starts exactly once on explicit retry; the other stays paused and ordered. Chat navigation, renderer reconnect and full app reopen retain the queue. This initial race was seeded and Claude replies were simulated; the subsequent real-provider and native-client acceptance is recorded in the follow-up above. No live service was modified during that initial simulated check. - Availability: source only; the complete correction includes app and server changes. No release, installation, deployment or restart is included. ## 2026-10-21 — Keep old subagent updates from reviving chat errors - Select the composer's latest chat error from the parent turn lifecycle. A later subagent status update retains its original run association but cannot make an older failed parent turn current again after a newer reply. - Preserve genuine errors for the latest turn, including failures whose start lies outside the loaded history page. Imported history and background bookkeeping do not clear or resurrect a current failure. Historical error messages remain in the timeline. - Verification: 48 focused component tests and 139 supporting runtime/history tests pass, along with TypeScript and the production Electron build/entry check. An isolated native Electron run reproduced the old warning on unchanged main and verified the fix through authenticated server reads and websocket updates, navigation, reconnection and full app reopening. A new failure remains visible and the original error remains in history. These runs use synthetic historical events through the production event writer; no live provider request was made. - No user installation, live service or goal state was changed. - Availability: app source only; this correction needs an updated app. ## 2026-10-20 — Recover Codex model-switch compaction without replacing history - Address the remaining compaction path in #89; authentication-aware model retirement/default handling remains covered by #92. A compaction failure no longer triggers a bounded-memory replacement thread. - Attempt at most one same-thread continuation only after a full native failed turn proves no execution occurred. If the original input was already accepted, continue with empty native input instead of inserting the user message again. Unknown acceptance, output/tool activity, cancellation and lost ownership all prevent automatic replay. - Close failed/interrupted automatic-compaction status without reporting success. Explain the selected model and failure stage while retaining the native error; do not invent diagnostics for a fallback error the provider did not expose. - Verification: 137 targeted runner, retirement, resume, restart and compaction status tests pass, along with Python source compilation and the dependency lock check. Two new regression tests fail on unmodified main. A wider controls sweep encounters an existing idle-thread-load mock failure, also reproduced on unmodified main; it is not fixed by this change. - The first CI run exposed three isolated lifecycle tests whose AST fixture omitted the runner's compaction helpers. Load those real pure helpers in the fixture; all 60 follow-up, child-continuation and provider-session checks pass locally afterward. No production behavior changes in this follow-up. - Personally exercised the actual native Electron client through mouse/keyboard input and production IPC/HTTP against an authorized test server with Codex CLI 0.156.1. Disposable old-model fixtures and a local fault-injection endpoint cover persistent compaction failure, the one-retry limit, native fallback, visible failure status and Stop. After removing the fault injection and restarting the test server, the same native conversation produced a real OpenAI reply recalling its earlier marker. A separate real native continuation confirmed that empty input adds no duplicate user message. - The retired-model seed and failures are controlled fixtures, not a claim to reproduce an uncontrolled upstream outage. Cancellation during recovery proof and abandoned-run restart cleanup have deterministic test coverage. No new CLI version, mobile change, public release or production deployment is included; this source fix requires a server update. ## 2026-10-26 — Prepare stable 1.0.11 from the published beta.12 source - Start the stable candidate at published beta.12 product source `ab95ad3284f91997c2b249ac413b155f99c09ce6`. Preserve its desktop, server and CLI functionality; only the release version and preparation documentation change. Later main changes are outside this release. - Prepare matching desktop, npm runtime, short CLI and signed legacy bridge packages as version `1.0.11`. Exact new package verification and local native upgrade acceptance remain pending; prior beta checks are not relabeled. - Hold public publication until populated provider-history retention, busy/queued work, restart and recovery observations have been reviewed. No mobile build or publication is included. See [candidate notes](RELEASE_1.0.11.md). ## 2026-10-21 — Prepare 1.0.11-beta.12 source - Integrate main through `b21ef7ea20bd8bc1121708590fa3bc21785c0984` with the published beta.2 release line. Retain the shipped Codex streaming, pending update, CLI PATH and managed service lifecycle protections. - Include CLI pairing QR output, shared-chat analytics, connection recovery, Claude fork timestamps, and the server's Codex goal summary fields used by notification clients. Existing mobile changes come from main; no mobile package is built or uploaded as part of this preparation. - Pin the reviewed beta.12 installer separately from every historical pin. Keep historical release exception tests tied to their exact original source fixtures, and include the pairing modules in the inert npm PATH test package. - Focused desktop checks (1,263 cases), TypeScript and production compilation pass. CLI, installer, packaging, session-summary and release-tooling checks also pass after correcting isolated test setup. Full source CI and exact signed candidate acceptance remain pending. - This is source preparation only. The new candidate still requires native first-install and upgrade acceptance, active-work safety, and notification checks through the actual client/server boundary. No new npm package, desktop release, bridge, mobile build or installed service is changed. ## 2026-10-21 — Verify public bridge and npm server upgrades - On disposable GitHub-hosted Macs, exercise the published signed desktop app against real managed servers: `1.0.3` → legacy bridge `1.0.10` → npm `1.0.11-beta.2`, and npm-capable `1.0.9` → npm `1.0.10` → npm `1.0.11-beta.2`. - Observe real public TLS download hosts without replacing payloads or disabling signature verification. The legacy first hop uses GitHub release assets; npm-capable hops use `registry.npmjs.org`. - Verify automatic recovery after an offline start, a refused npm connection that leaves the incumbent usable, and successful retry through the App's visible **Retry server update** control. Preserve server identity, token, saved test session/event and settings. Compare all 120 stable and 129 beta runtime files with their independently verified public packages. - Both scoped native jobs pass in [run 39557877984](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/39557877984). Live provider history, busy/queued work, activation rollback, logout or reboot, other desktop platforms and desktop binary replacement remain outside this acceptance. This verifies the published versions, not the new beta.12 source. ## 2026-10-18 — Publish 1.0.11-beta.2 - Complete [paired npm publication, attempt 2](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/39103091104) after the accepted runtime becomes publicly available. Verify and skip that unchanged runtime, publish the exact matching CLI through its separate Trusted Publisher, and verify both public archives. Both `beta` tags now resolve to `1.0.11-beta.2`; both stable `latest` tags remain `1.0.10`. - A fresh isolated registry installation with setup scripts disabled resolves the matched CLI/runtime pair and exact dependency, and exercises command help without operating existing services. Native first-install and lifecycle acceptance remains the separately recorded candidate check below. - [Desktop publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/39177990728) passes all four native platform verifiers and publishes the [canonical beta](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.11-beta.2) and [compatibility mirror](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/v1.0.11-beta.2). The signed [legacy server bridge](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.11-beta.2) is public as well. Preserve product source `853dcd285ef442176370632149deba21ef732eda`, native build 1261 and asset-manifest digest `bff0ef6024feb840c7b8e9d1a8c84817b775fc1c23711aa3f29132c96e71ec8c`. No accepted artifact was rebuilt, repacked or replaced. - Keep stable releases, mobile builds and existing installations unchanged. Retain the disclosed clean-quit, unsigned Windows, native Windows/Linux UI, populated-history/busy-upgrade and logout/reboot limits. Gemini is included for desktop/server API connections, not mobile or Google/CLI sign-in. ## 2026-10-18 — Accept the 1.0.11-beta.2 candidate - Freeze product source `853dcd285ef442176370632149deba21ef732eda`, native build 1261. [Desktop preparation](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/39100809951) passes all four native platforms, including Developer ID signing, Apple notarization and the exact asset-set verification. The separate orchestration commit `197f07b9ca949c1a81441d18b46a048764d6dd2d` does not change product code. The sealed asset-manifest digest is `bff0ef6024feb840c7b8e9d1a8c84817b775fc1c23711aa3f29132c96e71ec8c`. - Exercise the unmodified signed macOS app and exact packaged server through real UI, IPC, HTTP and Gemini CLI 0.64.0 against Google inference: connected API settings, direct new chat, a long first prompt, automatic short title, two turns with the same native session, and retained replies/title on reload. Title generation inserts no extra user turn. Temporary credentials and disposable server state are removed afterward; existing installations remain unchanged. Test cleanup required force-closing the owned app, so this is not clean-quit acceptance. - Pass 5,392 desktop tests (six skipped), the full hosted server suite, 18 native Gemini CLI/HTTP cases and a 40-turn, four-chat streaming/reconnect stress check. [Native npm acceptance](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/39099183063) passes on macOS and Linux, including first setup, exact runtime inventory, default/named lifecycle, repeat installation and preserved state on removal. - [Paired npm publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/39103091104) initially accepted the runtime upload, then exhausted 61 read-only visibility checks. At candidate acceptance, the CLI upload and public desktop/bridge release remained pending; the completion receipt above supersedes that state. Candidate acceptance alone was not a completed public release. Mobile Gemini, unsigned Windows warnings, native Windows/Linux UI, populated-history/busy upgrades and logout/reboot remain the disclosed boundaries. ## 2026-10-18 — Bound macOS release ticket-download recovery - Keep product source and signed server/npm candidates frozen independently of the reviewed build orchestrator. Recover only a recognized network failure while attaching an already-accepted outer DMG notarization ticket. - Retry the same artifact with bounded delays; require its ticket, Developer ID signature and every original package/updater check before accepting a build. Compilation, test, signing, notarization rejection and other failures still block release. No artifact is published by the recovery helper. ## 2026-10-18 — Prepare 1.0.11-beta.2 - Integrate current main and the built-in Gemini desktop/server provider with the prior published beta line. Preserve existing native ownership, login handoff, next-turn provider selection, CLI PATH and lifecycle protections. - Include terminal-only first-install connection details, service logs, latest-first history recovery, file-link handling and copied diagnostics. Mobile Gemini and mobile publication remain outside this release. - Combine Gemini's runtime files with the existing Codex runtime inventory. Pin the reviewed beta.2 installer separately; do not alter historical pins, signed artifacts, stable channels or existing installations. - This is source preparation, not a publication or candidate acceptance receipt. Signed-package and real-app acceptance must precede public publication. ## 2026-10-13 — Accept imports and publish 1.0.11-beta.1 - Publish immutable source `e93391294af2cb0c35f5f04fe8e94cce77810458`, native build 1252. No accepted package was rebuilt, repacked or replaced. - Exercise the signed/notarized macOS app and matching packaged server through the real client/server boundary with isolated data. Import a 300 MiB Claude image record and a 64 MiB Codex tool record; preserve native titles, surrounding text and Unicode. Confirm subagent/archive filtering, cancellation, repeat discovery, unchanged provider files and no duplicate incremental imports. - Import and continue real Claude and Codex sessions with their original native IDs and remembered context. Four imported chats survive an isolated relaunch. Excessive retained text fails with the observed amount, limit, record ordinal and byte offset without creating a partial chat. All 103 focused regressions pass. These checks do not establish new image rendering or live Cursor parity. - Hosted source checks and macOS/Linux native CLI lifecycle checks pass. [Paired npm publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/38370285507) verifies both exact public archives through their separate Trusted Publishers. A fresh registry installation with lifecycle scripts disabled verifies both package versions and command help. Both `beta` tags are `1.0.11-beta.1`; both `latest` tags remain `1.0.10`. - [Desktop publication](https://github.com/ZhengyiLuo/AgentsDock/actions/runs/38431444310) passes all four native platform verifiers and publishes the [canonical beta](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.11-beta.1) and [compatibility mirror](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/v1.0.11-beta.1) with the same 16 sealed assets. The signed [legacy server bridge](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.11-beta.1) is also public. The asset-manifest digest remains `dbed888a33160639ba00992bbb5ec1d1f9ba0650c7329c27c7701f55df9579cd`. - Retain the disclosed limits: unsigned Windows installer; no new live Cursor resume or native Windows/Linux app interaction acceptance; no full populated history, active/queued upgrade, logout/reboot or Windows updater acceptance. The isolated relaunch check is not clean-quit lifecycle acceptance. Temporary test credentials/state were removed; existing apps and servers were untouched. ## 2026-10-13 — Prepare the 1.0.11 beta integration - Start `release/1.0.11-beta.1` from the frozen 1.0.10 release line and merge main through `2b666a9f`. Include streaming native-history import, the Files inspector, simplified server settings and connection sharing, recoverable empty chats, and provider/permission inheritance for new chats. - Preserve release-line native writer handoff, pending provider selection, mailbox-only delivery authority, CLI service PATH and stop/start fixes. Combine observational catalog refresh and attributed login-reload diagnostics with those existing ownership protections rather than restoring older code. - Keep the new transcript reader in every runtime manifest and compile check. Pin the new installer separately for this exact beta; historical installer hashes and accepted 1.0.10 artifacts remain unchanged. - Local checks: 5,266 Electron tests passed (six skipped), eight build/license checks passed, TypeScript and production compilation passed; 407 release/CLI checks and 196 Python release-tooling checks passed. Targeted import, provenance, login, cross-chat, cleanup, goal, instance and restart regressions passed after resolving the runtime inventory ordering mismatch. - Availability: source preparation only. New-beta signed packages, hosted native lifecycle/update checks and real-app feature acceptance are pending. These source checks do not establish a public release or native provider resume acceptance. No mobile package is built by this preparation. ## 2026-10-20 — Fix Claude running-chat fork timestamp precision - Fix #94: recognize millisecond native timestamps within a second-truncated completion event when the completed reply has an exact provider message UUID. Keep text matching, provider identity, tool-call exclusion and future-second rejection intact. Legacy replies without UUIDs retain the strict cutoff; imported-history checkpoint handling is unchanged. - The new regressions fail on the original implementation. All 72 focused live-fork and session-fork tests pass with the correction, as do Python source compilation and the dependency lock check. - Personally exercised the visible native Electron app through mouse/keyboard input, production IPC/HTTP and real Claude against an authorized test server running 1.0.11-beta.11 with the narrow source patch. A same-second native completion successfully forked while its parent continued running. The native child preserved earlier tool-result context, excluded the in-flight parent request and completed a contextual follow-up; the parent completed normally. An idle fork was also checked after the parent finished. - Availability: source branch and local test-server patch only. A server update is required; no Electron UI change, public release or deployment to other servers is included. This does not establish broader provider parity. ## 2026-10-20 — Prevent duplicate replies after resumed Codex goals - Native goal replies now retain their provider thread identity. History sync excludes goal lifecycle records and output from text-only matching, then reconciles public replies by their exact native thread, turn and item. Genuine repeated text from another item remains visible. - Existing imported copies can be recognized from an older goal's recorded start and native output identities. Repair is a read-only presentation; original replies and stored history remain unchanged, including after an interrupted goal with multiple native turns. - Source is based on `df15d18b`; Electron UI code is unchanged. All 172 focused server tests pass, covering import prevention, existing-history repair, repeated reads, conflicting ownership, interruption and both history matching stages. The added ownership regression fails against the original code. - An isolated offscreen native Electron app exercised production IPC and authenticated HTTP against the actual server with synthetic provider history. Duplicate replies changed from two copies to one; two intentional equal-text replies stayed visible. History refresh, repeated chat navigation, full app reopen and scrolling passed. Source and event-log hashes stayed unchanged. Desktop type checking, production compilation and compiled-entry verification also passed. No live provider generation was exercised. - Availability: source only. Requires a server update; no installed-app replacement, live-service restart, deployment or public release was performed. ## 2026-10-20 — Verify history repair in the local desktop - Integrate the existing repair-budget fix without changing Electron UI code. The combined branch passes 67 focused history/paging tests, desktop type checking, and the production Electron build. - After an explicitly authorized local restart, AgentsServer 1.0.11-beta.2 with the narrow source patch returns ordinary prompts and replies that its unpatched history page omitted. No conversation history was deleted. - Reopened the actual desktop through its production transport, inspected restored historical user messages, and clicked the latest-message control. Recent user messages were visible directly, without opening thinking traces. - The user also reports recovery in a second, unchanged client after the server restart. That report is separate from the directly exercised desktop checks. - This verifies the targeted history failure, not full cross-machine/provider parity. No public release was published. ## 2026-10-17 — Keep chat messages visible in repair-heavy history pages - Fix #91: reserve semantic-page slots for ordinary user prompts and assistant output before emitting same-ID history repair receipts. Give repairs a separate bounded allowance while retaining the absolute response cap, existing replay suppression, trace access and older-page cursors. - Five focused regressions cover default and small pages, older-page reads, active/interrupted turns, dense corrections and the absolute response limit. The original allocation fails the new message-retention checks. All 159 targeted paging, job-history, async-message, Claude-repair and Codex model retirement checks pass after integrating the latest main branch. The wider Codex native-proof suite has one existing malformed-provider-line failure, reproduced with the unmodified allocation; it is not fixed here. - An isolated offscreen native Electron development app used production IPC/HTTP against the changed server source and synthetic Codex history with real native-replay proof. The 48-item response retained all 169 corrections and the latest user/final messages. Native input exercised trace expansion, collapse, older-history loading and reload: prompt/final output stayed visible without opening the trace. No real provider turn was submitted. - Python source compilation and the dependency lock check pass. After an explicitly authorized local server patch and restart, the real desktop transport returned ordinary prompts and assistant output in the default history page; the user confirmed that the conversation was visible again. The post-restart UI automation did not establish its visibility assertion, so user confirmation and the earlier isolated native checks are recorded separately. This does not establish full provider-content parity. - Availability: source branch and local server patch only; requires a server update, not a desktop UI change. No public release or cross-machine acceptance claim. Existing conversation history was retained. ## 2026-10-18 - Preview Simplified Chinese on mobile - Add a first-pass device-local System / English / Simplified Chinese choice in mobile Settings, reusing unambiguous desktop translations and a mobile-specific catalog. Localize app-authored navigation, settings, server controls, status text and dates without translating user or provider content. Add Chinese iOS permission prompts and platform language metadata; no server update is needed. - Refine preview terminology for model defaults, pinned chats, version, live status, message entry, agent attribution, server token and fork actions. Localize the native quick-message menu and its sent prompts in Chinese mode. - Keep the mobile My Agents inventory aligned with the desktop Gemini backend. - Verification on an isolated iOS Dev simulator: switch English and Chinese, reopen Settings, cold-launch with Chinese retained, and inspect the Chinese session list in light and dark themes, restoring the System theme afterward. The full mobile Node test suite, TypeScript, iOS and Android JavaScript exports, settings touch and quick-message simulator flows pass. The iOS native Debug compilation/install also passed during this preview. - This is a local preview branch, not a TestFlight upload or App Store listing localization. Advanced actions and server-originated errors may remain English; Android native UI and a physical iPhone are not yet accepted. ## 2026-10-18 - Validate cross-device pairing and mobile reliability changes - Prepare the accumulated connection-sharing, mobile recovery, media action and desktop sidebar changes against main 126c74f7. Update four source-contract assertions for multi-address health validation and active/final profile removal, retaining identity, namespace and durable-write ordering checks. Add the new QR, roaming, removal, media permission and HTTP cancellation regressions to CI. - Verification: 128 mobile UI/configuration checks, rendered transfer checks, focused client/store suites and 940 desktop tests pass. Sync stress completes 40 cycles with 36,000 events plus duplicates and 146 bounded reads. Thirty-one server discovery, CLI lifecycle and packaging checks pass from an isolated clean source fixture. Both TypeScript checks, desktop compilation, iOS/Android production JS exports and iOS native dependency parity pass. - Actual iOS simulator input repeats video share-sheet open/cancel four times, image preview reopen/share-sheet open/cancel four times, then six cycles of Servers -> Settings -> Appearance. Download and Pin remain the only preview overflow actions; theme selection, navigation and dismissal keep working. Media downloads use the authenticated test server. Restore the System theme. - The user separately confirmed the final download interaction on their physical iPhone. No physical device was operated during this verification round. Android native saving/camera and third-party share delivery remain unverified; source and component checks are not device acceptance. No app-store upload, npm publication or production server deployment is included. ## 2026-10-18 - Consolidate mobile preview download actions - Keep Download and Pin in a single compact row, removing the separate Photos toolbar action. iOS Download now stages authenticated bytes and opens the system share sheet directly, exposing applicable photo/video saving, Files destinations and installed apps without an extra custom menu. Preserve the existing Share shortcut and do not claim that dismissing the sheet saved or sent anything. - Android Download uses a bottom chooser for Save to Files, Save to Photos and Share / Open in. Cancellation starts no transfer; stale chat/menu callbacks cannot operate on a replacement chat. Documents do not offer photo saving. - Verification on main 126c74f7 plus pending changes: rendered transfer checks, 19 transfer unit tests, TypeScript and iOS/Android production JS exports pass. The rebuilt iOS Dev simulator exercised the actual authenticated test server, system share sheet, repeated presentation and real PNG/MP4 photo-library saves. Independently verify saved media dimensions/duration and a 597860-byte PNG written through the system Save to Files destination picker. - Android and tablet routing have component coverage, not new native-device acceptance. External messages and third-party app delivery were not sent as part of testing. This interaction change requires no additional native module, server update, release or installation over the existing rebuilt Dev app. ## 2026-10-18 - Save mobile chat images and videos to Photos - Add Save to Photos to image/video preview actions on phones and tablets, retaining Download and Share. Reuse authenticated, chat-scoped downloads, progress, cancellation, integrity checks and temporary-file cleanup. Report success only after the native library confirms the save; denied permission offers a Settings action. No server change is required. - Add source-built Expo MediaLibrary with iOS add-only permission. Android 10+ uses scoped MediaStore insertion without library reads; older storage writes stop at API 28. A composed-config regression guards plugin ordering and keeps broad media read permissions and legacy storage mode disabled. - Validation on main 126c74f7 plus pending branch changes: 19 transfer unit tests, rendered file-transfer behavioral checks, nine platform/config checks, TypeScript, iOS/Android production JS exports, both iOS native Debug builds, native dependency parity and packaged framework ABI checks pass. - Actual iPhone-layout simulator actions downloaded synthetic attachments from an authenticated test server and saved a 1000x4800 PNG and a 720x1280 six-second MP4. Repeat both saves using only the system's add-only grant, with no Photos read grant; verify dimensions and duration independently in the photo library. Light/dark previews, the native permission prompt and denied-access feedback were exercised. Denied access added no asset and never claimed success. - The Settings action opens system Settings, but this simulator lands on the root page rather than the app permission page; direct permission navigation still needs physical-device acceptance. Android native saving and iPad native layout are not yet accepted; their control/permission paths use component fixtures. The separate iPhone Dev app was rebuilt, installed and launched successfully, without writing test media to the user's real photo library. No release, upload, server restart or publication. ## 2026-10-18 - Integrate current main into the connection and mobile fixes - Update the working branch to main 126c74f7 while preserving the pending mobile, desktop and QR changes. Resolve the development-log and installer manifest conflicts by retaining both sets of additions, including the shared analytics module and terminal QR runtime files. - Verification: 335 focused desktop tests, 81 mobile checks, both TypeScript checks, desktop production compilation and compiled-entry verification pass. Eleven installer/archive checks pass from an isolated clean source snapshot. The initial live-source archive check correctly rejected generated bytecode; the running service and its cache were not modified to bypass that guard. - The merged native macOS desktop build passes repeated Pinned folding, keyboard expansion, preference reload and light/dark/narrow UI checks in an isolated offscreen profile. Correct the local test setup to reload its reset preference before repeating the workflow. Mobile source and native dependencies were unchanged by upstream; no new physical-device acceptance was performed. - Source integration only: no commit, push, publication or server restart. ## 2026-10-18 — Queue Claude goals during running work - Allow a completion condition to be submitted while Claude is running. The desktop offers Queue goal or Queue replacement and shows the saved request separately from the provider's active goal. The existing message queue supplies cancellation and recovery; the current run is not interrupted. - Retain the exact native command selection through queue admission and drain. Clear remains the explicit clear-and-stop operation. A queued replacement takes effect after the current run finishes, including its native goal loop; this is not an immediate change to a running goal. - Queue events refresh the selected chat's goal display without polling or requests while typing. Native clear-command aliases remain separate from completion conditions. Both English and Chinese controls describe timing. - Validation: 129 focused server checks and 239 desktop checks passed, followed by final goal/header checks, TypeScript and production desktop compilation. An isolated offscreen native desktop reproduced the disabled baseline. The corrected desktop sent a real Claude SDK turn, queued a goal through authenticated native IPC/HTTP while its shell tool ran, observed the original task finish successfully, and then observed the goal become active and achieved. The provider session ID stayed unchanged and no stop occurred. A separate native-desktop fixture passed queue cancellation, failed-request retry, chat switching/reopening, and replacement while preserving the original task. English/dark and Chinese/light dialogs and keyboard focus were checked at the supported minimum width. This fixture used synthetic provider state; the live-provider result above was a separate test. - Availability: source change and local compiled desktop only. App and server updates are both required; no installed app, live service or release changed. ## 2026-10-18 — Allow ordinary chat mentions while steering Codex goals - Send selected native skills and local `@Chat` follow-ups into the current Codex goal turn. Preserve its native thread, running tools, goal and original chat access; add only the explicitly selected live chat routes after delivery. Ordinary text with a local chat mention is a supported follow-up; it does not need a skill or command. Saved chat routes also remain compatible with steering. - Bind provider tools from native goal activation to their exact live owner when Codex omits client metadata. Keep ordinary-turn proofs and rejection of stale, partial or ambiguous ownership. Preserve selected skills in durable queue recovery and retain per-route user instructions across mailbox retries. - Show actionable Send now failures and deferred outcomes without Electron's IPC wrapper. If input was accepted but chat access subsequently disappeared, explain that outcome without replaying the message. - Exclude promptless goal-control starts from native history matching so a later sync does not import the accepted follow-up and reply a second time. - Validation on main base `126c74f7`: 434 targeted server tests, 516 desktop tests, Python compilation, TypeScript and production Electron compilation. Native offscreen Electron acceptance used disposable state and real Codex: selected a skill and a local chat through the actual pickers, queued each follow-up and clicked Send now while the original shell remained running. Both goals received the input and completed without changing their thread or interrupting work. The chat case reused a thread with prior conversation, verified the selected route through the real provider tool, and returned to idle. History was checked again after switching away and reopening the chat. Team/remote mentions and legacy/internal deliveries remain separate-turn operations. Live mailbox sends, cross-device and custom-endpoint provider acceptance are outside these checks. Source only; no release or deployment. - Additional regression: an ordinary local mention with five saved routes and no skill or special-purpose metadata fails the old admission guard and passes the updated guard. Plain-text controls pass both. The full admission suite passes 35 tests; native-turn projection passes 14 tests after completing its isolated cleanup fixture dependency and asserting capability revocation. ## 2026-10-17 — Refresh privacy-limited desktop and sharing metrics - Integrate current main into the existing sharing-analytics work, preserving Gemini support, native report forms, and updated server packaging. Retain the two share-mode selection events and three narrowly scoped shared-web events. - Add confirmed Agent/model/permission changes, imported Gemini API binding, archive/restore, Files-panel entry, native file-open outcomes, explicit server restart request/reconnection, language/font changes, and diagnostic-copy and issue-form request outcomes. Existing provider usage/configuration events can carry only a fixed built-in provider family, never a model/endpoint/name. - Keep Team Network instrumentation unchanged. Do not count initialization, restored settings, automatic model migration, stale profile responses, or rejected session patches as successful user changes. Document accepted messages separately from completed provider work and restart requests separately from verified reconnection. Update the catalog and privacy source. - Validation base: main `4ef55959`; development Electron build, not a release package. Type checking and desktop/shared-web builds pass. Targeted Electron regression suites: 939 passed; sharing and packaging Python suites: 116 passed. - Hands-on surface: isolated native offscreen Electron, actual authenticated local AgentsServer routes, separate browser partition. Exercised new chat, Agent selection, Codex/Claude permission changes, model selection, archive/restore, Files opening, unavailable-file feedback, language selection, error copying, blocked external issue navigation, and both share modes. Redeemed tokens in the browser and verified reload/reconnection deduplication, no shared-browser install identifier, and an interactive Codex prompt/reply. Captured event envelopes locally; production analytics delivery was blocked. - Remaining acceptance: actual Mixpanel ingestion, Gemini API-binding with credentials, managed restart success, native font-menu input, and successful external-file/application launch were not exercised end-to-end. Provider resume across sharing modes is not an acceptance claim of this change. These are source changes only: no deployment, release, or privacy-site publish. ## 2026-10-03 — Point issue reporting to the maintained repository - Correct the server README's issue-report link to ZhengyiLuo/AgentsDock. The desktop `/feedback` command already opens that repository's new-issue form; preserve its existing behavior and add focused coverage for finding it with `/feedback`, `/issue`, and `/bug` without sending a chat message. - All three component cases, TypeScript and the production build pass; the component cases mock the external-browser bridge. Actual app-to-browser interaction remains unverified because computer use is disabled. This documentation/test change does not require a server update. No release or deployment. ## 2026-09-30 — Share selection and recipient usage events - Distinguish choosing View only/Interactive from successful share creation. Preserve the existing share-dialog opening, creation and revocation events. Add three narrowly scoped shared-browser events: authenticated read-only and interactive activation, and server-accepted interactive message submission. See `docs/ANALYTICS_EVENTS.md` for exact event names and counting semantics. - Keep generic desktop analytics disabled in shared pages, including when the restricted bridge is installed after imports. Remove a legacy web install ID on initialization. Web events have no persistent viewer identifier, content, URLs, tokens, names, resource IDs or arbitrary event properties. Respect DNT/GPC and existing same-origin opt-out; deduplicate refresh/reconnect and receipt replay with local-only sessionStorage markers. Counts do not identify unique viewers or form a cross-device per-user funnel. - Generate both server-served web artifacts from the shared tracker, retain dependency license notices and include the snapshot artifact in installation manifests. Successful snapshot GETs permit only the exact analytics script hash and Mixpanel destination; token forms, errors and HEADs stay scriptless. - Exercised the actual native macOS Electron share menu through IPC/HTTP into a disposable AgentsServer. Separate isolated offscreen browser windows opened both token forms, rejected an invalid token, authenticated and showed the transcript. Read-only reload did not recount activation. Interactive browser submissions completed real Codex replies; after correcting the initialization boundary only shared-browser send events were emitted and no install ID remained. Refresh and simulated network interruption/recovery did not recount activation. Revocation disabled sending and returned the token entry screen on reload. Captured requests locally and blocked production Mixpanel delivery. - Focused client/server tests, type checking and production desktop/shared-web builds pass. Actual Mixpanel ingestion, other browser engines, and live queued acceptance are not validated here (queued/refused receipt semantics are covered by targeted bridge tests). No release or server deployment. ## 2026-10-18 - Restore desktop Pinned group folding - Connect the Pinned header to a dedicated collapse state and right-aligned arrow. Persist the preference in the existing server-scoped desktop cache; folders named Pinned remain independent. Chat pins and folder assignments do not change. Next/previous-chat navigation skips hidden pinned rows, while unread/running pinned chats remain eligible in their visible status groups. Emergency notifications retain the existing automatic section expansion. - Reproduced the unresponsive header in an isolated native Electron window before the fix. The locally compiled app then passed mouse collapse/expand, Enter/Space activation, preference restoration after reload, repeated use, same-name folder independence and light/dark/narrow presentation checks. Pinning a disposable chat crossed the real authenticated test-server boundary; folding persisted locally. Test chats were removed without running a provider. - Validation: 55 focused sidebar/session tests, two emergency-state regressions, TypeScript, production compilation and compiled-entry verification pass. The two initial folding/navigation regressions failed before the implementation. Native testing used an offscreen macOS Electron 43.1.1 development process with source changes over 4ef55959, not an installed release. Windows/Linux native acceptance was not performed. No server update, publication or merge. ## 2026-10-18 - Recover lazy media previews after a Dev transport interruption - Backport Expo's upstream HMR recovery fix to the pinned Expo 57.0.4 package through a scoped pnpm patch. Lazy imports after a lost Metro graph now call the platform reload helper instead of accessing `window.location` on native. This affects development recovery, not production chat synchronization, and does not change the native dependency ABI. - Stage iOS video previews with a foreground file-system download. The same authenticated sample failed in the background URLSession and downloaded successfully in the foreground session. Existing cancellation, authentication, cache ownership and integrity checks remain in place. - Verification: 51 HMR and interaction checks, TypeScript, video-staging and deferred-import tests, and iOS/Android production JavaScript exports pass. The HMR regressions fail before the patch and pass afterward. - Actual iOS simulator UI reproduced the reported undefined `reload` error before the patch. After the same forced Metro graph loss, the app reloaded without that error and reopened the chat. Subsequent automated preview-open assertions did not pass; full preview and video playback acceptance remain incomplete. The user separately confirmed opening works on the physical phone after reloading the patched Dev app. - No Android native acceptance or release-binary validation was performed. Changes remain local source only; no server update, PR or release. ## 2026-10-18 - Decouple selected-chat recovery from slow session lists - Foreground recovery waits for server identity validation, but no longer waits for the session list before fetching the selected chat and reopening its live stream. Joining an already-pending list refresh follows the same path. - Reconnect restores a cached selection immediately after validation. A late list response does not restart that selection or supersede a newer selection; a selected chat removed from the server still falls back to an available chat. - Request the same `summary=true` session list as desktop. Existing merge rules retain omitted detail fields and honor explicit reset values. - Skip a redundant legacy tail request only when the returned total proves all events are present and neither side of the page is omitted. Unknown totals and incomplete pages retain the compatibility fallback. Latest-first bounded paging, cache visibility and authoritative stream cursors remain unchanged. - The new foreground regression fails on the prior implementation and passes with this fix. Lifecycle, reconnect, validation rejection, latest-first paging, compatibility, client cancellation, profile isolation, cache adoption and state-hardening checks pass. Stress validation passes 40 cycles and 36,000 events with 146 bounded reads; TypeScript and iOS/Android production JavaScript exports pass. - Actual iOS simulator UI checks use an isolated authenticated HTTP/WebSocket fixture. After a cold launch, foregrounding with a 5.002-second session-list response starts the selected tail immediately and opens one stream 5 ms later. A legacy one-message response delayed by 1.5 seconds now takes one request and opens its stream in 1.506 seconds, instead of two serial requests. Both checks retain the visible message and reach Live without another screen tap. - These are controlled transport measurements, not production latency promises or acceptance on the affected user's server. Android native/device acceptance and matching the reported TestFlight binary to source remain pending. Changes are local source only; no release, server deployment or PR was created. ## 2026-10-18 - Verify npm connection commands and complete setup QR output - Keep `agentsdock token [NAME]` as the combined URL/token/QR entry point and `agentsdock pair [NAME]` as the QR-only shortcut. Clarify the first-install, setup and new-instance output in command help and package documentation. - Real npm lifecycle tests exposed truncated terminal QR output: nonblocking tty writes stopped under buffer pressure, sometimes splitting UTF-8 symbols. Retrying nonblocking writes also allowed npm's progress indicator to corrupt the QR. Use normal blocking tty writes, preserve partial-write offsets, and bound retries of transient write errors. Credential output still goes only to the controlling terminal, never captured npm output or debug logs. - Verify the complete generated QR, not merely a heading, under ordinary npm installation and `--foreground-scripts`, with an authenticated local HTTP fixture. Verify headless omission, repeat-install safety and log secrecy. Service creation and instance discovery in this lifecycle test are synthetic; no native fresh-default service was created or replaced in this iteration. - Verification: 35 focused Node checks and all five package tests pass. The complete-QR lifecycle check also passes three consecutive runs in both npm output modes. Package tests run from an isolated source snapshot so live Python bytecode does not bypass or weaken the runtime packaging allowlist. - Availability: local source only; no global CLI replacement, server restart, npm publication or mobile build. ## 2026-10-18 - Bound native HTTP reads and recover from canceled sync requests - Reproduce the native fetch cancellation error with a deliberately slow HTTP response in an isolated iOS Dev simulator. This is a transport exception exposed through app error state, not a development-only warning overlay. - Cover response bodies as well as headers with the existing request deadline. Race the operation against cancellation because the installed Expo iOS native text reader can remain pending after its transport has been canceled. A partial JSON response reproduced the indefinite Syncing state before this additional deadline guard; aborting the socket alone was insufficient. - Preserve client/caller cancellation ownership, normalize timeouts and unexpected native interruptions, and keep confirmed HTTP failures distinct. A stalled 401 body still revokes validation. Late responses cannot publish a second rejection, and mutations are not automatically retried. - Keep transient automatic session-list failures in profile connection state, preserving cached chats and unrelated user-action errors. Selected-chat reads continue through the existing bounded recovery sequence without a screen tap. - Native UI checks against an isolated authenticated HTTP/WebSocket fault fixture pass for stalled headers and partial JSON bodies: the cached message remains, the approximately 12-second deadline closes the request, and automatic recovery returns to Live with no global error banner. A 30-second session-list timeout leaves the healthy timeline Live, marks the profile degraded, and clears that diagnostic on the next successful refresh. This is fault-injection transport validation, not live-provider or affected-user-server acceptance. - Twelve cancellation regressions, client scopes, store integration, bounded timeline recovery, lifecycle and address fallback checks pass. The sync stress check passes 40 cycles with 36,000 events; TypeScript and iOS/Android production JavaScript exports pass. The exact reported TestFlight binary was not rebuilt or matched to a release source manifest. No physical device, working server, release, publication or merge was operated for this investigation. ## 2026-10-18 - Fix intermittent iOS Settings taps after native menus - A physical-device native snapshot found stale `RCTSurfaceTouchHandler`s on ordinary Settings views beneath the modal's own handler. Expo UI 57.0.4's RNHostView attached these handlers but never detached them before Fabric reused the views. This matches upstream Expo #47713, not a disabled theme option or a server synchronization dependency. - Pin Expo UI to 57.0.7, which includes the upstream disappearance/deinit cleanup. Build ExpoUI from source against the installed ExpoModulesCore and verify package/pod/source parity. Do not change Settings actions to fire on touch-down or replace their existing scrolling and modal interactions. - Add a dependency lifecycle regression and a native Maestro workflow that cycles Servers -> Settings -> Appearance, asserts selected themes and returns to System. The old simulator binary reproduced an ignored Appearance tap after closing Servers, remaining on the category page. Simulator-only push entitlement overlays were dismissed separately, not treated as this failure. - The rebuilt simulator passes the same six-cycle drawer/settings workflow, including selected-state assertions for each Light/Dark switch and working category/back/close actions. This is actual native UI input, not handler calls. - Focused checks (20), TypeScript, iOS/Android production JavaScript exports, native dependency parity and both iOS Debug builds pass. The rebuilt Dev app is installed on the independent physical-device and simulator identifiers; the user confirms Settings navigation and theme buttons work on the phone. No production installation, server configuration or chat content is changed. Android device acceptance and longer physical-device testing remain pending. ## 2026-10-18 - Add desktop unread and running sidebar groups - Match mobile's temporary status grouping: Unread first with a blue dot, Running next with a green dot, then ordinary sections. Active work takes precedence over unread state; archived chats stay archived. Each chat appears once, retaining its saved folder and pinned state. Empty status groups hide. - Add independent status collapse controls and right-aligned folder/group arrows. Keep status groups out of folder drop targets and match next/previous-chat navigation to the visible group order. Collapse state resets on workspace adoption and does not write server preferences. - Verification: focused sidebar/grouping/drag/navigation/design checks, the desktop regression suite, TypeScript, production compilation and output checks pass. An isolated native Electron window exercises real menu marking, collapse and keyboard expansion against the authenticated test server. A disposable live Codex turn moves Running -> Unread -> original folder; reading removes empty groups. Light/dark and narrow-window screenshots confirm dot sizing and rightmost arrows. Temporary chats are removed afterward; no working chats or production app are replaced. Other desktop operating systems are not exercised. - Availability: local source and desktop preview only; no server update or public release is required or performed for this grouping change. ## 2026-10-18 - Restore green mobile sent-message bubbles - Set the shared user-message surface to soft green in light mode and muted deep green in dark mode. Keep assistant surfaces, text, layout and behavior unchanged; document both colors in the mobile design tokens. - Verification: eight theme checks, including primary/secondary text contrast, TypeScript and iOS/Android Hermes exports pass. Actual iOS simulator taps exercise theme switching and chat reopening with temporary in-memory message fixtures; both green bubbles render correctly. Restore the prior workspace and system theme afterward. No provider message was sent; this is visual acceptance, not a new server integration or Android-device check. - Availability: local development source/preview only, no public release. ## 2026-10-18 - Keep mobile agent marks independent of development networking - Embed the original Claude/Codex PNG bytes in the shared backend-mark component, preserving the artwork and Cursor's vector mark. Both Debug and Release now render these icons without fetching a separate image from Metro. - Generate the embedded sources during installation and verify byte-for-byte artwork parity, self-contained sources and repeated mounts in four focused component checks. TypeScript and iOS/Android Hermes production exports pass. - On a physical iPhone running the independent Dev app, a native bundler-host correction restores launch over the tailnet. A subsequent cold launch without an explicit host argument shows the connected home screen and both bitmap marks. Simulator taps open, close and reopen My agents with all marks visible. No chat writes, production-app replacement or server changes are involved. - Document the distinction between the Debug bundle host and the chat server. Debug still needs a reachable Metro service; embedding these marks does not make the entire development build standalone. Android device UI and a physical Wi-Fi-to-cellular switch after this change have not been exercised. - Availability: local development source and installed iOS Dev preview only; no public release or store submission. ## 2026-10-17 - Allow removing the current mobile server - Enable Remove for every saved connection, including the active and final profile, with a confirmation explaining that remote chats/tasks are retained. Select the next saved profile when available; otherwise clear the active workspace and show setup. Empty profile settings survive a cold launch. - Persist removal before detaching the local client or deleting its credential. Preserve identity-scoped caches and drafts for re-addition, invalidate old requests, and do not require remote connectivity or completion of active work. A profile-settings write failure leaves the original connection and token intact. - Verification: menu/confirmation regressions, real-HTTP store fixtures for active/inactive/final/offline removal, busy state, late replies, storage failure, credential cleanup, empty startup and re-addition pass. Existing profile, cache, store and multi-address reconnect regressions, TypeScript and iOS/Android Hermes production exports pass. Fixture requests confirm removal does not send remote stop/delete commands; busy state is simulated, not a newly started live provider task. - In the installed iOS simulator, actual menu taps exercised cancellation and active/final removal. A cold restart retained the empty list. Re-adding the same isolated server through Add and refreshing returned the same chats over the production authenticated transport. QR payload injection was test setup, not physical camera acceptance. Physical-phone/Android UI remain unverified. - Availability: local development source only; no server update, public release, or store submission is needed or performed for this change. ## 2026-10-17 - Automatically select shared server addresses on mobile - Add QR format v2 with up to six explicit candidate addresses and a required durable server identity; retain v1 parsing and single-address sharing. - CLI automatic pairing includes authenticated LAN/Tailscale candidates with one identity. Desktop QR sharing asks the selected remote server for its own binding-aware addresses through a native-token-only endpoint. Older/offline servers retain the saved single address; discovery never uses desktop NICs. - Mobile staggers bounded health probes, selects the first identity match and cancels losers. All failures produce one error. Saved alternatives support reconnect without duplicating profiles or changing cache/draft ownership. Explicitly reusing a scanned saved identity can refresh its addresses without replacing credentials. Manual address edits or identity resets clear old routes. - Verification: 30 mobile component checks, nine parser/selection checks, real-HTTP mobile store roaming/persistence and existing profile/client/store regressions pass. Desktop client/service/contract checks pass (645 tests), as do 25 server discovery/binding checks and 33 CLI/lifecycle checks. Both TypeScript checks, desktop compilation and iOS/Android Hermes exports pass. - Actual offscreen desktop UI generated a multi-address QR against the isolated test server; independent native image decoding confirmed the advertised routes and authenticated identity. The CLI also authenticated both actual routes. The installed iOS simulator accepted the decoded payload, saved through an actual Add tap, and reconnected both directions while one route was deliberately rejected by a test-only fetch wrapper. All-route failure and recovery retained the profile; afterward the wrapper/profile were removed and selection restored. - This is not physical camera or off-Wi-Fi phone acceptance. Actual iPhone Tailscale roaming and Android device UI remain unverified. Only the designated isolated local test service was switched to source; no production deployment, global CLI replacement or release was performed. ## 2026-10-17 - Simplify mobile QR connection setup - Add server shows Scan QR code, a subtle "or add manually" caption and all three manual steps immediately, without an expansion action. Scanning shows a compact name/address confirmation; Edit details reveals the masked manual fields. Saving still requires successful authenticated health and the QR's expected identity. Existing profiles are not silently overwritten. - Use a square camera viewport with a separate width constraint, QR icon and terminal/desktop source guidance. Only the scanner expands the native sheet to full height; ordinary server management retains half/full detents. - Document LAN, Tailscale, remote desktop, localhost/tunnel and HTTPS cases. Format v1 still carries one URL, without automatic network failover. Verify the existing interactive token command displays all discovered address categories, token and terminal QR together; update its help accordingly. - Verification: TypeScript, 28 mobile component checks, 22 binding/address checks and 15 CLI checks pass. The camera-enabled iOS simulator native build and iOS/Android production exports pass. After making manual steps immediately visible, TypeScript and the 28 component checks pass again; actual simulator UI verifies the caption, all three fields, scanner back navigation and drawer close/reopen without an expansion action. Earlier checks cover camera permission and compact confirmation. Native camera viewport bounds are 360 by 360 points. - A decoded QR injected into the actual simulator app reaches the real test server, passes identity validation, and saves/activates through an actual Add tap. This is not a physical camera scan. The temporary profile is removed and prior selection restored. A real PTY token-command run verifies terminal QR output and authenticated health without recording credentials. - Physical iPhone/Android camera capture, off-Wi-Fi tailnet routing and public HTTPS/proxy acceptance remain unverified. No public release, global CLI replacement, server restart or network-policy changes in this iteration. - Additional observation outside the QR change: after switching Appearance to dark, the Settings close action did not dismiss the sheet on two simulator attempts. Restarting the QA app was needed to continue; do not treat this as resolved or assume it is only automation timing. No unrelated Settings fix is included in this iteration. After restart, dark-mode QR choice, scanner and dismissal pass actual UI checks; the test theme is restored afterward. ## 2026-10-17 - Stabilize mobile server setup step badges - Replace the three native rounded-border badges with fixed 24-point SVG circles and numbers. A physical iPhone light-theme screenshot showed step two stretched into a horizontal oval; the previous overflow clipping did not prevent it. Preserve theme colors, step accessibility labels and all connection behavior. The underlying native rendering cause is not confirmed. - Verification: TypeScript and 27 focused server form/scanner tests pass. Actual iOS simulator light-theme checks exercised three drawer reopen, expansion, input focus and return cycles; inspected all three circular badges in the expanded sheet. Corrected automation-only close selectors, keyboard dismissal and screenshot output paths during the run. - Restarted the attached iPhone's separate Dev app and observed a fresh Metro bundle connection. Physical-device visual confirmation remains pending: runtime inspection timed out. No production app or server data changed. ## 2026-10-17 — Make Gemini a built-in provider - Remove the Gemini preview environment gate from provider registration, capabilities, native import discovery, skills and cross-chat discovery. Compatible desktop clients now show Gemini in My Agents without an opt-in flag. Installation of Gemini CLI 0.64.0 and a configured API connection remain required for sending; Google/CLI sign-in is not offered. Old-client checks and unsupported-feature capabilities remain unchanged. - Update setup documentation and preserve the Gemini icon attribution in the standalone server export. Repair integration-test fixtures for the current update-session wrapper, archive rejection helper and per-chat queue recovery; no unrelated production behavior is changed by those fixture corrections. - Verification: 169 targeted server tests and 18 native CLI/HTTP tests pass. Desktop TypeScript, 5,246 tests (five skips), eight build/license checks and production compilation pass. A fresh real HTTP server without any feature flag exposes Gemini while correctly rejecting missing runtime/API readiness. Actual offscreen Electron through production IPC/HTTP and the pinned CLI verifies the unconfigured settings card, API setup without an initial model ID, model discovery/default selection, new chat, two turns and reload. Screens inspected; this rerun uses a loopback model, not Google inference. Full server regression and hosted CI remain separate gates. No installed server restart, public release or package publication performed. - Integrate main's Codex retired-model selection fix without dropping either side's development-log entries. The post-integration run passes 141 focused server tests and 5,250 desktop tests (five skips), plus TypeScript, eight build/license checks and production compilation. The superseded local full server run was stopped; full-suite validation belongs to the new hosted CI. ## 2026-10-17 — Prepare the Gemini preview for review against current main - Integrate main's history-responsiveness and desktop recovery updates while retaining Gemini import, title and session-identity coverage. Resolve the additive development-log and session-summary test conflicts without dropping either side's behavior. Gemini remains an opt-in, API-only development preview. - Use the approved Chinese term for imported sessions. Update archive-label assertions for both the section and per-session badge, make the clipboard permission fixture independent of umask, and give native Gemini fixtures a canonical workspace path. Align the old skills test with API-only policy while retaining native skill discovery and unchanged terminal-login checks. - Clarify provider versus adapter limits: Gemini has native subagents and terminal compression; this integration does not expose managed child lifecycle controls or manual compression. Current ACP Fork returns method-not-found, and unregistered slash commands are ordinary prompts. - Verification: 5,246 desktop tests pass with five skips, plus eight build/license checks; TypeScript and production compilation pass. A focused server run passes 246 tests with seven optional native cases skipped. Actual offscreen Electron through production IPC/HTTP and the pinned CLI verifies both native import/API connection entry points, repeated sends, original session IDs and reopening. Chinese/narrow/light and English/dark rendering were inspected. The current integration rerun uses a loopback model fixture, not Google inference; earlier real-API acceptance and remaining release gaps are recorded in the Gemini preview document. No public release or installed-server update. ## 2026-10-17 — Explicit API continuation for native Gemini imports - Import Chat can explicitly bind Gemini history to its saved API; existing history-only imports gain a localized Connect Gemini API action. The API destination is disclosed before selection. Other providers' import and account-switch restrictions remain unchanged. - Preserve the original native UUID, workspace and transcript location. Pinned CLI process-only API auth does not modify the terminal's saved login, copy history or start a replacement session. Busy/queued work, missing history, revoked bindings and managed auth conflicts fail closed; failed persistence rolls back the binding. Forget cannot fall back to Google login. - Verification: focused server and desktop regressions, TypeScript and production compilation pass. Six native CLI/HTTP contracts cover import, continuation, Stop/retry, Forget, title compatibility and retained history. Actual offscreen Electron through production IPC/HTTP and Google Developer API verifies both connection entry points, original contextual recall, repeated turns, unchanged native identity and reopen. Light/dark, narrow and Chinese presentation were inspected. The native fixture confirms exact API key headers and unchanged terminal settings; no copied transcript is created. - Prepared on the Gemini preview branch for the authorized local desktop/test server update. No public release or merge. External native-history syncing, broader organization policies and platform acceptance remain separate. ## 2026-10-17 — Gemini first-turn summarized titles - Prefer the native summary; if absent after a successful ordinary turn, make one bounded, tool-free background title request through that chat's own API connection. Auto resolves from its exact native model metadata. Long prompt excerpts retain the ending question as well as the beginning. - No hidden chat message, native session or tool invocation is created. Manual names, late native titles, opt-out, connection revocation and identity changes reject stale results. Failure keeps the prompt fallback without repeated usage. - Verification: 160 focused server regressions pass, plus a native CLI/HTTP title contract without injected summary. Actual offscreen Electron through production IPC/HTTP and the Google Developer API verifies short English and long Chinese/log naming, visible refresh, repeated turns, stable native identity and reopen. Manual-name and revocation races have focused server coverage. The broader native lifecycle run passed 29 of 30 cases; the legacy Google-login skills case fails identically on the unchanged baseline, independently of this title change. Full provider/release acceptance remains separate. - No desktop source changes or public release. Prepared on the Gemini preview branch for the explicitly authorized local test-server update. ## 2026-10-17 — Require real first-turn provider title acceptance - Tighten the provider checklist: existing native metadata and synthetic title adoption tests do not establish automatic generation. Require an actual App and real-model first-turn check with long pasted input, missing native titles, manual rename protection, repeated turns and reopen. - Record Gemini's missing fallback title generator separately from its tested native-summary reader. Align the preview's outstanding acceptance list with the API-only policy; Google/CLI login is intentionally unsupported. - Documentation only. No title-generation implementation, chat repair or service change is included; checks are documentation diff/link review. ## 2026-10-17 — Gemini API-only connections and clearer chat errors/archive state - Gemini's preview now offers API connections only. The CLI remains its runtime; other providers' login choices are unchanged. Old Google-login chat history remains readable, but sends cannot fall back to that account or silently move to an API account. Settings, discovery and send admission agree on this policy. - Composer errors are selectable and have a localized copy action. Timeline error cards also support copying. Long command text is preserved rather than truncated, and an empty failed-turn summary no longer hides the actual error. - Archived chats have a muted row, archive icon and explicit label, including in search. They remain readable in the collapsible Archived section and are excluded from ordinary chat mention candidates. - When the owning Codex resume RPC explicitly rejects the exact native thread as archived, the server mirrors it through the normal AgentsDock archive lifecycle and does not retry through exec. Jobs/routes follow the existing archive policy; provider history is untouched. Stale runs and other errors cannot cause this transition. This is response-driven, not a background scan of native archives or a migration of old error records. - Verification: 705 focused desktop tests, TypeScript and production compilation passed. The server suite passed 230 cases with one optional native test skipped; that native test then passed separately. These cover credentials, archive lifecycle, routing and stale-run ownership. Actual offscreen Electron through production IPC/HTTP verifies API-only Settings, copying both error surfaces, legacy-chat recovery navigation, Archive persistence, mention exclusion, reopen, narrow layout, light/dark and Chinese. A native Gemini CLI/loopback endpoint fixture covers real failure propagation; separate Google Developer API Default/Plan sends succeeded. A disposable native Codex archive/resume probe reproduced the exact rejection and retained its native history without a model request. - Not published or deployed; existing user services and chats are unchanged. ## 2026-10-17 — Compact Gemini approval actions - Gemini tool requests now use localized Allow / Deny buttons. Full request details remain available, and approval still applies only to that request. The shared card accepts optional labels; other providers retain their existing controls and approval semantics. - Verification: 37 focused desktop tests cover long tool names, English and Chinese, both decisions, ownership, existing Codex controls and Gemini modes. TypeScript and production compilation pass; 39 transport/process regressions pass. Actual offscreen Electron through production IPC/HTTP and the native Gemini CLI verifies Allow executes a disposable write, Deny does not, mode selection, next-turn isolation and five Stop/retry/reopen cycles. Narrow rendering was inspected. These deterministic model replies use a loopback fixture, not Google inference. - Prepared on the same Gemini preview branch for the authorized local test update; no public release, push or merge. Google OAuth acceptance remains separate from these tool-policy and lifecycle checks. ## 2026-10-17 — Gemini early Stop/Delete stress fixes - Reproduced immediate Stop leaving a chat fenced, first at the real process boundary and again through actual offscreen Electron. A process readiness handshake prevents signalling the guard before its handlers exist; startup cancellation retains the process handle for verified cleanup. Stop/Delete now use the Gemini-owned cleanup operation instead of killing its guard through generic tree termination. A failed native exec with no launched provider is not misclassified as surviving tools. Cursor's legacy exit contract is unchanged; unverified cleanup still fails closed. - Added early Stop/cancel/exec-failure regressions and repeatable opt-in native long-message stress. Existing cleanup-failure and restart fences remain in place; no old fence is automatically repaired by this change. - Actual offscreen desktop, production IPC/HTTP and native CLI: five consecutive Stop/retry/reopen cycles, Default approval and rejection, next-turn policy change, automatic edits, YOLO cancel/confirm, Plan write denial, keyboard operation, narrow/light/dark and Chinese presentation pass. Separate actual desktop journeys verify direct first/repeated new chat, inherited settings, custom-provider switching, native built-in skill activation and repeated picker use, and blocked untrusted handoff followed by an ordinary question. These deterministic flows use a local model fixture, not Google inference. - A separate actual desktop Google Developer API run completes Default then Plan with the same native ID. Google OAuth-account acceptance remains open. Four simultaneous native chats complete 40 long-message turns and 40 socket reconnects with exact replies and stable identities. Native history survives deletion. This is bounded regression coverage, not a production load claim. - Verification: 377 focused server regressions pass, followed by 39 transport/ process tests including two additional failed-exec cases; 41 native CLI/HTTP lifecycle, permissions, skills and API-binding cases pass. Desktop: 5,214 tests pass, five skip, plus eight compile/license checks. TypeScript and production compilation pass. Counts overlap and are not a readiness claim. - Availability: same Gemini development branch and isolated local test Apps; no existing service, working desktop, credentials or trust settings replaced. No deployment, public release, push or merge. Remaining provider limitations are tracked in the Gemini preview documentation. ## 2026-10-17 — Gemini native permission-mode controls - Add a localized Gemini permission menu in the desktop composer for both native-login and API chats: Default, Auto Edit, explicit-confirmation YOLO, and Plan once a native workspace handshake reports it enabled. Preserve next-turn semantics and the native session ID; newly created chats inherit the selected Gemini chat's saved mode without copying another provider's sandbox options. Default remains normal native approval policy. - Pass native approval flags and confirm the requested ACP mode before the prompt. Preserve native folder trust, tool denies and administrator YOLO restrictions. Disabled Plan fails closed rather than silently running in Default. API profiles preserve native Plan settings without changing login. Shared chats do not gain this administrative permission editor. - Keep a mode-change acknowledgement out of assistant replies by correlating the exact pinned CLI notice with the outstanding mode-change RPC. Identical text in an actual model response is still retained. - Verification: 85 focused server/protocol/schema/packaging tests pass. Six additional real CLI/HTTP mode cases verify default approvals, auto-edit writes, shell confirmation, YOLO, Plan denial, native tool/admin restrictions, unavailable modes and same-session next-turn changes. The model in those deterministic native tests is a loopback fixture, not Google inference. - Desktop regression: 5,214 tests pass, five skip, plus eight compile/license checks. TypeScript and production compilation pass. Five additional native regression cases cover API reconnect, resume, Auto routing, cross-chat and Stop/Delete isolation; these are separate from the live Google acceptance. - Actual offscreen Electron with production IPC/HTTP and native Gemini verifies selecting/saving modes, pending approval isolation, denial, automatic writes, YOLO confirmation/cancellation, Plan blocking writes, keyboard focus, light/dark/narrow layouts, English/Chinese and reopening. No internal mode notice appears as an answer. A separate real Google Developer API run through this App completes Default then Plan in the same native session. Google OAuth-account mode acceptance and wider platform/policy coverage remain open. - Source only on the Gemini development branch. No existing service or working desktop was replaced; no cleanup fence bypass, existing native trust change, public release or merge. The broader provider acceptance limitations remain. ## 2026-10-16 — Keep Gemini workspace trust separate from sign-in - Report unavailable communication tools as a workspace/MCP issue, not an invitation to sign in again. Include the exact chat workspace and native trust/settings guidance without changing credentials or granting trust. - Do not require a communication tool for every ordinary question merely because a chat has a durable communication permission. Use this turn's validated references and internal delivery purpose. Explicit communication and mailbox wakes still fail before model work when the tool is unavailable; broker authorization, native trust and approval policy are unchanged. - Verification: the new real-CLI regression failed before the fix and passes afterward. 82 focused regressions and seven native CLI/HTTP cases cover ordinary follow-up, untrusted sender/recipient, queued cancellation, native denial, unknown targets and busy-recipient retry/deduplication. - Actual offscreen Electron with production IPC/HTTP and the real native CLI displays the workspace-specific error, then accepts an ordinary question in the same chat without delivering mail; its answer survives reopening. This isolated check uses a loopback model fixture, not Google authentication or inference. Dependency lock and changed-module compilation checks pass. - Availability: source on the Gemini development branch; not deployed or released. Existing development-service cleanup fencing was not bypassed. ## 2026-10-16 — Distinguish Gemini identity, chat access and delivery - Include Gemini's native identity in detailed session responses and the header's ID/copy fallback. If no resumable native session has been saved, show a localized pending explanation instead of an unexplained dash; never invent a provider ID for a failed first turn. - Label a newly granted chat route as communication permission, explicitly not a sent message. Queued user work still waits its turn; native MCP loading failures now explicitly report that no cross-chat message was sent. Folder trust and provider approval policy remain unchanged. - Hide Gemini's server-attested automatic mailbox wake input, like Codex and Claude, without hiding real user inputs, agent progress or replies. This removes an empty "You" bubble exposed during actual cross-chat acceptance. - Focused verification: 85 server regressions, eight actual CLI/HTTP lifecycle cases and 195 desktop component/projection cases pass; TypeScript and production compile pass. Native tests use a loopback model fixture, including queued sender, cancellation, busy recipient, reply, deduplication, revoked/unknown routes, denied tool approval and untrusted workspaces. - Full desktop regression: 5,201 pass, five skip, plus eight compile/license checks. These counts do not substitute for the live-provider acceptance. - Live acceptance: actual offscreen Electron with production IPC/HTTP and the real Google API verified queued permission without premature delivery, cancellation, one message each way, native IDs, purple message cards and reopening. Only the requested native communication tools were approved. A second actual-App review verifies no fake user bubble on the receiving mailbox turn and retained messages after reload. - Remaining limitation: an extremely early startup Stop left a disposable chat behind a truthful cleanup fence. Its cleanup/race investigation remains open; this acceptance does not establish all lifecycle paths are ready. - Availability: Gemini development branch and designated development service (server source `b668e787`), with the corrected development desktop restarted using its existing profile; no public release or merge. Model quota and workspace trust remain distinct. ## 2026-10-16 — Include Gemini native built-in skills - Discover skills with the pinned native `skills list --all`; its default listing omits built-ins. Parse the native built-in marker as scope metadata rather than part of a skill's name, and show it in the existing composer badge. Keep native trust, disabled-skill and same-name override behavior, opaque selection IDs and file revision checks. Other providers' private skill folders are not advertised as Gemini capabilities. - Verification: 43 focused server tests, three real CLI skill contracts, the real HTTP/CLI five-turn resume/Forget/reconnect/Delete contract, and 226 composer tests pass. Native contracts use an isolated loopback model fixture; they cover built-in activation, user-skill approval, denied and approved writes, scope precedence, disabled entries and stale selections. - Actual offscreen Electron using production IPC/HTTP reproduced the missing built-ins, then selected and activated `skill-creator` through the real CLI, rendered its reply and reopened the picker successfully. That activation run used disposable server state and a loopback model API, not Google inference. - Live acceptance: the development service at source `b3759a17` preserved its identity, credentials and existing sessions. Actual offscreen Electron using production IPC/HTTP showed the native local and built-in inventory, selected `skill-creator`, observed successful native activation and rendered the exact requested Google API reply. Reopening the picker retained the built-in skill. - Availability: Gemini development branch and designated development service; no public release, merge or mobile changes. This does not implement permission-mode selection or resolve the separate native-login eligibility issue. No client installation was required for this server-side discovery fix. ## 2026-10-16 — Always open new chats directly - Remove the retired setup modal from the desktop dialog tree and all new-chat fallbacks, including first use and stale/unavailable runtime discovery. Reuse the last-opened chat settings; also retain custom instructions and supported subagent limits. Available agents remain selectable before the first send. Creation errors remain inline; no silent switch from a revoked API to native credentials, and no conversation identity or delegated/job authority copying. - Rebind the selected provider when switching between custom API agents before the first turn. Do not reuse another provider's immutable connection ID; first-turn backend locking and credential revocation remain unchanged. - Actual offscreen Electron with production IPC/HTTP reproduced the old first-chat modal, then verified direct first creation, eight inherited fields, Claude API to Gemini API switching, model selection, a rendered native CLI reply and repeated creation with the newly selected settings. This run used disposable server state and a loopback model API, not Google inference. - Verification: 5,198 desktop tests pass, five skip; eight compile/license checks, TypeScript checks, production compilation and 37 focused server tests pass. - Live acceptance: the same offscreen native App used the designated development service at source `8837b925`, created directly, inherited eight settings from a disposable OpenCode API chat, switched to Gemini API in the composer, and received the exact requested response from Google with Auto. Repeating New chat retained Gemini API and Auto. No endpoint credentials were changed. - Availability: Gemini development branch and development service; the existing development App was restarted without packaging or changing its selected server. No publication, merge or mobile change. Existing provider-wide acceptance limitations remain unchanged. ## 2026-10-16 — Honor Gemini API Auto and saved model defaults - Fix API chats showing Auto while admission rejected an empty model. Google's official Developer API now uses the pinned CLI's native `auto` alias; explicit chat choices and immutable connection defaults take precedence. Third-party gateways still require a selected model. Credential isolation and Forget revocation remain enforced; no global login or model setting is rewritten. - My Agents and chat catalogs expose the effective default without pretending model discovery or inference succeeded. The runner now passes a saved API model to the CLI even when the chat has no separate override. - Verification: 39 settings/runner/admission cases, 72 related model/runtime cases, 39 desktop catalog/settings cases, and two real CLI/loopback-API contracts including native Auto routing pass. The existing real server/CLI five-turn resume/Forget/reconnect/Delete contract also passes. Desktop type checks pass; the unchanged renderer uses its existing production compilation. - Actual-app verification: isolated offscreen Electron using production IPC/HTTP against the development service, runtime source `70ed513c`. My Agents showed Auto without a saved model ID; an API-bound empty-model test chat displayed Auto and completed two consecutive real Google requests with exact requested replies. No model/key selection was required and no native-login fallback was used. The fixture chat was created through HTTP before UI sending; the rapid new-chat/provider-switch workflow was not accepted in this check. - Availability: Gemini development branch and designated development service; no publication. This validates API Auto, not the separate Google-login quota issue or the remaining provider-wide acceptance checklist. ## 2026-10-16 — Bounded Gemini stress acceptance - Runtime source `ffe22869`: a development-instance HTTP/WebSocket soak ran for 60 seconds with 16 HTTP workers and 16 reconnecting event-stream workers. All 8,515 reads and 1,854 socket catch-ups succeeded; existing history and configuration remained unchanged. This was loopback and empty test chats, not a remote-network/large-history benchmark. - A separate disposable server with actual Gemini CLI 0.64.0 and a fake local model API completed four concurrent chats × ten turns, with 134,890-byte streamed Unicode replies and 40 socket reconnects. Text stayed exact, replay produced no duplicate turns, UUIDs stayed isolated/stable, and Delete retained native history. Native repeated-output loop protection remains enabled. - Nine additional stream/runner/approval regressions pass. An offscreen native Electron client used production IPC/HTTP against the updated development instance to inspect My Agents and select/cancel the configured Gemini API option in New Chat, without changing credentials or sending model requests. - Availability: development test service only; no release/publication. Google account generation remains unaccepted, and full CI/platform, long-duration, real-network and resource-pressure acceptance are still pending. ## 2026-10-16 — Harden Gemini lifecycle and complete shared-chat wiring - Keep native Gemini identity before visible activity/tool approval. Add an owned parent-death guard and verified-cleanup contract; an unconfirmed cleanup keeps Stop/Delete fenced instead of erasing the chat or reporting false success. Persist that failure fence across restart; loss of the live cleanup handle requires operator verification, not an automatic unlock. Permit only the native loader's exact leading environment-header refresh. - Fix native-account skill discovery without an API connection and capability admission for scheduled chat/standalone jobs. Extend the jobs helper's provider selectors. Keep run-bound MCP instructions free of shell authority-file paths. - Wire Gemini prompts, approvals and runtime updates through interactive shares. Fix both the shared-state allowlist and the mounted approval shelf refresh. Preserve share-token, CSRF, exact-chat/run and revocation checks. - Verification: 88 Gemini cases using actual CLI 0.64.0 and a loopback fake API; 247 related server regressions; 83 transport/cleanup/shared-control cases; 29 jobs-helper cases; 160 queue/Stop/goal/job regression cases. These batches overlap and are not a unique-test total. Nine release-manifest checks pass. Desktop: 5,193 tests pass, five skip; type checks, production/shared builds and eight package/license checks pass. Native remote Google requests were not retried: real-account eligibility/billing acceptance remains deferred. - Actual offscreen desktop acceptance covers API setup, New Chat, composer model selection, skills/approval, repeated resume, live answers and title import. Actual packaged shared-web acceptance covers token entry, native approval, two turns and revoke. English/dark, Chinese/light/narrow and keyboard focus were checked; native HTML-select popup interaction remains unaccepted. - Crash-mid-tool/descendant cleanup, failed cleanup, queued Delete, unauthorized and revoked recipients, busy mailbox delivery, idempotent retry, WAV transport and three-chat concurrency have dedicated fixtures. Process sampling is not a security sandbox or a low-memory/load guarantee. Native Fork, native goals, managed subagents, side questions, explicit compaction and later external history reconciliation are not claimed. See the Gemini preview matrix. - Availability: source and isolated test builds only; no server deployment, publication or mobile changes. Full repository CI and real-account acceptance remain release gates. - The full local server-suite attempt is not a pass: it was stopped after an installer scenario stalled. A separate probe reproduced 22 IPC setup errors where the local test filesystem could not provide private POSIX file modes; those guards were not weakened. The relevant Gemini/adjacent suites above were run with disposable POSIX-backed state. ## 2026-10-16 — Extend the Gemini development preview - Add live answer snapshots, bounded text/image/audio attachments, native summary adoption and opt-in main-session discovery/initial text import. Preserve manual titles, exact native workspace/identity, ownership exclusion and compatibility with older import clients. Native edits are not reconciled. - Fix retry after an empty failed first generation without resetting existing conversations. Preserve native identity after exposed activity, including history-recorder failures; retain partial output on Stop and native tool diffs. CLI mode no longer assumes a custom API connection object exists. - Exercise native queue/restart/resume, one-time approvals, denied writes, pending-tool Delete, retained native continuation and cross-chat send/read/ reply. Report missing MCP under native folder trust without granting trust. Include new modules in installation and package manifests. - Verification uses Gemini CLI 0.64.0, disposable state and a loopback fake model API: 72 Gemini tests and 247 related server regressions pass. Actual offscreen desktop IPC/HTTP acceptance covers API/model settings, skills, repeated sends, live answer delivery and native-title import. Desktop tests: 5,191 pass, five skip; type checking, production compilation, eight package checks and nine server manifest checks pass. - Still an opt-in preview, not release-ready. Real-account acceptance is deferred; native Fork, later history reconciliation and remaining optional provider-checklist workflows are not presented as completed. No deployment or publication is part of this validation. ## 2026-10-15 — Use the official Gemini product icon - Replace the approximate sparkle with Google's unmodified four-colour asset, bundled locally with attribution. Preserve its transparency without showing the fallback underneath, and include the SVG in shared-chat packaging. - Actual desktop checks verified the sidebar, composer and new-chat provider picker in dark mode and emulated light/narrow layouts. Type checking, production and shared-renderer builds, 5,189 desktop tests and eight package checks passed; five desktop tests were skipped. The packaged shared-chat asset matches the original, but its live web route was not exercised. - Icon-only acceptance does not resolve the Gemini preview's account, billing or provider-checklist limitations. No server deployment or publication. ## 2026-10-15 — Exercise Gemini connections and native skills - Give Gemini a distinct sparkle mark and correctly name its error notice. Preserve custom endpoint model choices through compact session-list refreshes. - Add bounded native skill discovery, opaque per-chat selection/revalidation, and native activation requests with ordinary tool permissions. API profiles reference installed skill assets while keeping login credentials isolated. - Real pinned CLI/local API fixtures cover native activation, approve/deny, continuation and cleanup. A skill-follow-up regression found native load-time history normalization; accept only its exact known text/tool transformation while still rejecting unrelated changes and suppressing historical replay. Targeted desktop tests, type checking and production compilation pass. Real desktop sends reached Google's service but eligibility and API billing responses block successful real-account acceptance. - This remains an opt-in development preview. Cross-chat, complete streaming parity, titles/import, mobile and the other recorded provider-checklist gates are not accepted merely because basic fixtures pass. Nothing is published. ## 2026-10-15 — Begin an opt-in Gemini provider preview - Add an experimental desktop/server Gemini CLI adapter behind a development flag, pinned to CLI 0.64.0. Google CLI login and Gemini Developer API credentials remain isolated; ACP is the transport, not a separate login method. - Wire My Agents, native model discovery, same-session continuation, bounded approvals and run-bound provider tools. Guard asynchronous native history replay so it cannot be appended as a new response. Forget must not fall back to another account; Delete preserves native history. - Actual CLI and HTTP contracts used a loopback fake API and disposable state. Actual offscreen desktop acceptance exercised configuration, displayed model options, manual default selection and two sends through production IPC/HTTP. It caught and fixed a client backend allowlist omission and duplicate replay. - Validation: 252 focused desktop tests, TypeScript and production compilation; server batches of 773 lifecycle/authorization regressions, 129 runtime checks, 72 provider/transport checks and 20 packaging-manifest checks passed. These batches overlap; they are not a full-suite or real-Google acceptance claim. - Not release-ready: real Google account/API, tool/cross-chat/native parity, new-chat dialog and native select-popup acceptance remain pending. Imports, automatic titles, attachments, fork, skills UI and mobile are not included. See the server's Gemini preview record for precise limits and remaining gates. ## 2026-10-17 - Camera-based server pairing - Add a live QR scanner inside the existing mobile Add Server drawer. Reuse the versioned CLI payload, mask the token, authenticate the decoded address, and require any encoded server identity to match before enabling Add. Scanning never saves immediately; cancel, stale tests and duplicate camera events do not replace existing profiles. Manual connection remains available. - Request camera access only on scanner entry, release the preview in the background, support denied-permission recovery, and keep older Dev binaries usable until their native camera update is installed. No microphone access, QR-image import, connection deep-link activation or third-party QR service. - Interactive installation and token display render a locally generated QR after authenticated health. Fresh npm postinstall writes it only to the controlling terminal, not lifecycle logs. QR failures are nonfatal. Include the renderer and its notices in the exact server runtime file inventory. - Build iOS camera code from source, matching the existing video dependency policy. The precompiled camera framework failed the bundled ABI gate; the source-built Dev app passes dependency parity and bundled-framework checks. - Verification: TypeScript, 383 mobile UI/source tests, QR parser cases, 32 CLI tests, 120 installer/CLI-package tests, 33 focused installer/package checks, two process-startup tests, and iOS/Android Hermes exports. Nested QR payload directories reject linked parents in both packaging and installation. A real isolated server installation emitted a terminal QR; an independent Vision decode of its PNG authenticated the exact encoded identity. Simulator UI verification filled a real decoded payload, tested, saved and switched, then restored the original profile and removed the disposable profile. - Integrate current main and update the isolated local test server without replacing the stopped legacy test instance or its data. Install and launch the signed camera-enabled iPhone Dev app under its separate bundle identifier; leave the production app unchanged. - Acceptance boundary: the simulator received an injected decoded barcode event, not a physical camera image. Live phone capture and Android native interaction still require device verification. No release or upload. ## 2026-10-17 - Validate the mobile UI review candidate - Update rendering-test host mocks for the appearance preference and shared corner tokens. Keep permission and model-selection assertions aligned with provider-aware routing without weakening admission or catalog-isolation checks. This resolves the seven test failures recorded in the preceding UI pass; no additional product UI changes were needed. - All 371 mobile UI/source tests and TypeScript passed. Focused provider-mode, client transport, session creation, multi-server store, latest-first sync, preference migration, optimistic mutation, date formatting, runtime-catalog and smart-group regressions passed. Production-mode Hermes bundles and assets exported successfully for both iOS and Android; this is not a signed native release or store submission. - The final installed iPhone simulator smoke passed three settings-navigation rounds, agent-mark checks, search dismissal, and server-editor back/close. Earlier checks cover API selection through the actual server. Earlier repeated light/dark and enlarged-text checks remain applicable. Actual server selection was verified without sending model requests. Live provider inference, Android-device and tablet acceptance remain unverified. Review candidate only; no TestFlight upload, deployment or merge is authorized by this validation. ## 2026-10-17 - Show agent marks in mobile chat selection - Native chat choices now use the agent name without a Native suffix; Custom API choices retain their explicit label. Share a bounded, scrollable picker between the composer and new-chat defaults, with existing agent marks before each name and a trailing selection check. The iOS native menu cannot render these bitmap marks. Preserve provider guards, model reset and reload behavior; dismiss the picker before an action can show an iOS alert. - TypeScript and focused regressions passed, including 120 picker interactions across iOS/Android host fixtures, explicit API routing, stale-scope rejection and locked-chat behavior. The installed iPhone simulator passed five rounds of settings, theme, agent-choice and server-drawer navigation, plus search/keyboard dismissal and unsaved server-editor back navigation. Light and dark screenshots, including enlarged text, were reviewed. A disposable chat selected an existing API connection through the real server and was deleted without sending a message. Test preferences were restored; no server credentials were edited. - Broader UI regression testing still has seven failures: three host mocks lack Appearance, two theme mocks lack radius, and two source contracts still assume the pre-provider-mode expressions. These are recorded for follow-up, not treated as passing tests. No new interaction failure reproduced in the exercised iOS flows. Physical-device, Android device and tablet acceptance remain unverified. No PR, release build, upload or additional UI fixes in this test pass. ## 2026-10-17 - Distinguish mobile account modes and center server onboarding - My agents now separates native credentials from configured Custom API connections using the selected server's safe runtime metadata. Mobile does not edit endpoint addresses or keys. Old servers retain native behavior and show an upgrade hint for custom connections. - Add native/API choices to new-chat defaults and the existing pre-first-turn agent menu for the mobile client's Codex, Claude and Cursor backends. Send explicit per-chat routing, persist the selected default, isolate API model choices from native catalogs, and keep started chats bound to their original connection. Reject unavailable custom defaults or ignored routing responses rather than silently switching to native. OpenCode remains status-only on mobile; this does not implement another chat backend. - Replace the synthetic welcome-chat entry with a centered welcome surface and setup action that opens the existing half-height Add Server drawer. Clear old ephemeral welcome state without deleting real chats or server configuration. - Basic validation: TypeScript, focused UI, provider-selection, client transport, session creation, mutation and preference regressions passed. The installed Debug iPhone simulator exercised My agents, the default agent picker, creating a disposable chat and switching it to an already-configured Cursor API through the actual server. Confirmed the returned custom binding and removed only that test chat; no messages, model requests or credential changes were made. A reversible empty-workspace fixture exercised centered onboarding and the real native Add Server drawer; no connection was saved. Light-mode screenshots were reviewed. Live Codex/Claude API runs, Android, tablet and physical-device acceptance remain unverified. Local preview only; no build release or upload. ## 2026-10-17 - Contain mobile server setup badge backgrounds - Clip the translucent fill of each numbered connection step to its fixed circle. This narrowly guards against a reported light-mode blue rendering artifact without changing inputs, validation or connection behavior. - Basic validation: TypeScript and 15 focused checks passed. The badge clipping assertion failed before the change and passed after it. The Debug iPhone simulator exercised drawer expansion, secure dummy-token entry and editor back/reopen in light mode, with no stray line visible. No connection was tested or saved. The original intermittent artifact was not reproduced in the simulator, so confirmation on the reporting physical device remains pending. The independent phone Dev app was refreshed; no release build, native dependency change, dark-mode or Android acceptance in this iteration. ## 2026-10-17 - Simplify mobile settings navigation and agent status - Open Privacy directly in the browser with an external-link indicator. Keep Android usage analytics in its own settings row. Add a Server shortcut that waits for the iOS settings sheet to dismiss before opening the existing half-height server drawer. - Replace Agents & runtime with My agents: Codex, Claude Code, Cursor and OpenCode, matching this branch's desktop inventory. Show green checks only for ready agents on a connected server, mute unavailable entries and keep one server-machine sign-in hint. Recheck refreshes the existing server runtime catalog, suppresses duplicate taps and scopes pending UI state to the current server. Remove the persistent-goals settings control, not the underlying chat feature. This does not enable additional chat backends. - Show a Tailscale server hostname only for a connected, configured tailnet domain. No phone VPN/account detection or inference from an IP address. - Basic validation: TypeScript, diff whitespace checks and 14 focused tests passed. The Debug iPhone simulator exercised Recheck, four-agent rendering, repeated settings-to-server handoff, browser privacy loading and returning to the app. Light-mode screenshots were reviewed. Refreshed the independent phone Dev app without replacing the release app; physical visual acceptance, Android and tablet checks remain pending. No chats, profiles, provider runs or server settings were changed. Local preview only, no release build. ## 2026-10-17 - Simplify mobile server-editor actions - Remove the redundant editor Cancel button; retain the header back and close actions. Shorten Add & switch to Add without changing successful connection testing, server creation or automatic activation. Keep Save for edits and the existing-server action when a tested connection is already saved. - Basic validation: TypeScript and 26 focused checks passed, including the test/create/switch sequence in a fixture. The Debug iPhone simulator verified the two-button add form, back navigation and closing/reopening. No server profile was created or changed. Android and physical visual checks remain pending; local preview only. ## 2026-10-17 - Move mobile folder indicators to the trailing edge - Move expand/collapse chevrons after the count at the right of folder headers, including Unread and Running. Keep status dots before the title, whole-row toggling and existing folder actions. Expose the expanded accessibility state. - Basic validation: TypeScript and 27 focused checks passed. The Debug iPhone simulator collapsed and expanded a running group and an ordinary folder, restored both, and verified the light-mode layout. Android, tablet and dark-mode/physical visual acceptance remain pending for this small iteration. ## 2026-10-17 - Clarify mobile home buttons and simplify chat dates - Give the settings, new-chat and server-selector buttons a consistent 1pt outline and themed control surface, with clearer separation in light mode. Preserve their touch targets, actions and the transparent footer fade. - Show Today, Yesterday or a date for chat activity; retain the year for dates outside the current year and omit time-of-day. Use a shared, readable tertiary text color. Keep the existing activity timestamp source and chat ordering. - Document five typography sizes, four named font weights and semantic palette roles. Migrate home/server text styles to these tokens and add a focused check against literal text metrics, weights and colors in those adopted surfaces. Other legacy screens remain a gradual migration, not a global font rewrite. - Basic validation: TypeScript and 57 focused checks passed. Date tests passed in American and Asian time zones, including calendar and DST boundaries. The Debug iPhone simulator exercised settings/theme navigation and the server selector, with light/dark home screenshots reviewed for outlines and dates. Yesterday is fixture-tested; no chat or server data was changed. Android, tablet and physical-device visual acceptance remain pending. Local preview only. ## 2026-10-17 - Group mobile settings and add appearance preferences - Replace the expanded settings form with App, Server and About category rows. Keep text size, new-chat defaults, runtime diagnostics, supported Codex goal settings, privacy controls and Android updates accessible in detail pages. Back returns to the list; close dismisses the sheet; reopening starts at the list. Server address and token editing remain in server management only. - Add System, Light and Dark appearance choices, applied immediately to the app palette and native appearance and saved independently of server profiles. Preserve the latest selection during startup hydration and rapid changes; report local persistence failures without blocking theme selection. - Basic validation: TypeScript and 25 focused checks passed. The Debug iPhone simulator exercised category navigation, back/close/reopen, text-size controls, light/dark switching, theme persistence after process restart and following system appearance changes. Light/dark screenshots were reviewed. Android navigation and update availability were fixture-tested only; Android, tablet and physical-device visual acceptance remain pending. No server settings, chats or provider jobs were changed. Local preview only; no release build. ## 2026-10-17 - Prioritize unread and running chats on mobile - Add temporary Unread (blue dot) and Running (green dot) groups before ordinary sidebar folders, in that order. Each hides when empty. Active chats take precedence over unread classification, appear only once, and retain their saved folder and pin metadata. Archived chats and the local welcome guide remain in their existing sections. - Support collapsing each temporary group without persisting it as a folder, adding it as a move destination, or colliding with similarly named real folders. Preserve search and existing chat ordering; manual row reordering remains in ordinary folders. - Basic validation: TypeScript and 33 focused checks passed. The Debug iPhone simulator displayed real server activity, marked an empty test chat unread, verified group ordering and collapse/expand, restored its read state, and verified opening an unread chat returns it to its original folder. No messages, provider runs or folders were created. Run-completion transitions and empty groups were fixture-tested; light-mode screenshots were reviewed. - Scope: existing unread/read-receipt semantics are unchanged. Visibility gating of automatic read receipts for a selected chat hidden behind the home screen remains a separate follow-up. Android, tablet and dark-mode/physical visual acceptance remain pending. Local preview only; no release or server change. ## 2026-10-17 - Number the server connection fields - Give the shared add/edit connection form compact blue outlined step numbers for Name (optional), Server URL and Access token. Separate the three field groups with consistent spacing and use short example placeholders. Preserve optional-name handling, secure token entry and connection validation. - Basic validation: TypeScript and 26 focused checks passed, including numbered labels and creating a profile without a name in a fixture. The Debug iPhone simulator verified half/full form presentation, input with the software keyboard, back/discard and closing. Light-mode screenshots were reviewed; no connection settings were saved. Android, tablet, dark-mode and physical visual acceptance remain pending for this iteration. Local preview only. ## 2026-10-17 - Open server management at half height - Present server management at a half-height native detent, with a grabber to expand to full height or collapse again. Reopening starts at half height; retain the separate editor-back and drawer-close actions. - Reuse the existing React Native Screens dependency. Keep its presenter mounted during dismissal without intercepting home touches while closed. Let the native form sheet handle keyboard resizing instead of applying duplicate scroll adjustments that can hide the focused input. - Basic validation: TypeScript and 25 focused checks passed. The Debug iPhone simulator exercised half/full/collapsed detents, close/reopen, swipe dismissal, current-server selection, unsaved add input with the software keyboard, editor back, the edit menu, folder cancel, and opening from home search. Light/dark screenshots were reviewed. No server settings were saved and no messages or folders were created. - Availability: local development preview only. Android, tablet and physical device visual acceptance remain pending. No dependency upgrade, native release, publication or server change. ## 2026-10-17 - Separate server-editor back and drawer close - Use one fixed header for the server list and its editor. Add and edit show a back arrow with their page title instead of a redundant Servers heading and a second close icon. Back discards the unsaved form and dismisses its keyboard; the single upper-right close action dismisses the entire drawer. - Basic validation: TypeScript and 23 focused checks passed. The Debug iPhone simulator exercised adding with unsaved text, returning to the list, reopening an empty form, closing the drawer from the editor, and editing/back navigation. Light-mode screenshots confirmed one title and one close action. No connection settings were saved and no server request contract changed. - Availability: local development preview only; dark-mode, Android, tablet and physical-device visual acceptance remain pending for this iteration. No native release, publication or server change. ## 2026-10-17 - Shorten server overflow actions - Remove Move up and Move down from each server's menu and shorten its remaining actions to Edit and Remove. Keep current-server removal disabled and preserve the confirmation before removing any saved connection. Stored ordering and server persistence are unchanged. - Basic validation: TypeScript and 21 focused checks passed. The Debug iPhone simulator verified both menu labels, absence of ordering actions, opening and canceling the editor, and opening and canceling removal confirmation for an inactive profile. No server profiles were modified or removed. - Availability: local development preview only; Android and physical-device visual acceptance remain pending. No release or server change. ## 2026-10-17 - Consolidate home actions and open servers directly - Move New folder beside Search in the top row with a wider gap, removing its separate toolbar row. Make the server name an outlined rounded button while retaining truncation, connection status and the reserved refresh indicator. - Open the existing server sheet directly instead of an intermediate popup. Present saved servers as tappable rows with a current-server checkmark and one Add server entry. Keep edit, reorder and confirmed removal in each row's overflow menu; show the connection editor separately from the server list. - Close the drawer after successful selection, without reconnecting when the current server is selected. Keep failures visible, and prevent a completed switch from dismissing a newer drawer after its original view was closed. Preserve identity validation and secure credential handling. - Basic validation: TypeScript and 45 focused checks passed. The Debug iPhone simulator exercised folder opening/cancel, direct drawer presentation, add/edit/cancel, current-server selection, switching between two saved servers and back, long-name truncation, opening from search, and testing the existing connection through the actual server. Screenshots were reviewed in light mode. No messages, folders or new server profiles were created, and connection settings were not saved or removed. - Availability: local development preview only. Failed-switch and stale-view completion paths were fixture-tested. Dark-theme, Android, tablet and physical visual acceptance remain pending for this iteration; no native release, publication or server deployment. ## 2026-10-17 - Close home search after opening a chat - Close and clear home search after a chat row successfully navigates, including history-result and local welcome rows. Cancel pending search debounce and invalidate stale results through the existing clear-search action. Preserve the query when a history result cannot open or its server scope has changed. - Reuse the same cleanup for the close button and empty-search dismissal; ordinary chat navigation still opens immediately without waiting for sync. - Basic validation: TypeScript and 27 focused checks passed, including executed callback tests for cleanup, immediate navigation, history success/failure and stale-server completion. The Debug iPhone simulator reproduced the retained search before the change, then verified title search, direct and submitted result selection, back navigation without the search field, and reopening an empty search against the configured server. No messages were sent. - Availability: local development preview only. History-result failure and stale-server paths were fixture-tested, not exercised against a failing live server. Android, tablet and physical-device visual acceptance remain pending; no native release, publication or server change. ## 2026-10-17 - Show chat-list refresh beside the server name - Show a small loading indicator immediately after the active server name during pull-to-refresh, sharing the list's existing profile-scoped refresh state. Clear it when that refresh settles; do not add a minimum animation duration or tie it to background timeline synchronization. - Reserve the indicator's space to keep the title stable, expose the busy state to accessibility, and leave server switching available during refresh. - Basic validation: TypeScript and 30 focused sidebar, server-profile and profile-safety checks passed. The Debug iPhone simulator exercised a real pull gesture with a one-shot, test-only eight-second delay before the actual session-list request. The indicator remained visible while pending and disappeared after the request returned HTTP 200. Restored the test hook; no production delay, messages or chat mutations were introduced. - Follow-up validation: two unmodified, real pull-to-refresh requests returned HTTP 200 before the header and native refresh-control props became false in the same React commit. Simulator video confirmed the native control's closing animation remains visible afterward. This is list-refresh progress, not an indication that independent timeline synchronization has completed. Restored the local timing hooks; no application behavior was changed by this check. - Availability: local development preview only. Physical-device visual, Android and tablet acceptance remain pending; no release or server change. ## 2026-10-17 - Let home chats scroll behind the footer fade - Remove the list viewport's bottom margin so chat rows remain visible behind the transparent portion of the footer gradient instead of ending at a hard horizontal boundary. Use trailing content padding equal to the fade height so the final row can scroll fully above the overlay. Preserve phone-search layout and the independent circular actions. - Basic validation: TypeScript and 20 focused sidebar/theme checks passed. The Debug iPhone simulator in dark mode exercised scrolling through the fade and opening/closing Settings; screenshots confirmed progressively fading chat rows underneath the overlay. No messages were sent or chats changed. - Availability: local development preview only. No release, server changes or Android/tablet acceptance in this pass. ## 2026-10-17 - Fade home actions and dismiss unused search - Add a touch-transparent vertical footer gradient, fading from transparent into a darker neutral shade in each theme. Keep the circular action sizes, positions and reserved list space unchanged; reuse existing SVG rendering. - Close an empty or whitespace-only home search on blur or an outside tap. Preserve nonempty queries/results, and exclude touches within the search field and its opening button from outside-tap dismissal. Do not capture chat row responders or change session creation/persistence. - Basic validation: TypeScript and 59 focused UI checks passed. The Debug iPhone simulator exercised empty/whitespace dismissal, input touches, nonempty query retention after outside taps and submission, and Settings in both themes. Verified that tapping a chat while empty search is open navigates on the first tap and returns without the search field. No messages were sent. - Availability: local development preview; refreshed the independent physical iPhone Dev app. Physical visual, Android and tablet acceptance remain pending. No native rebuild, release, server change or publication was performed. ## 2026-10-17 - Use independent circular mobile home actions - Move New chat from the upper toolbar to a circular bottom-right action; place Settings in a matching bottom-left action. Only the circles have a surface color, with no full-width footer background or divider. Reserve list space above the buttons so they cannot cover chat text or intercept row taps. - Retain the existing guarded quick-create flow and hide both actions while phone search is open. Preserve the server-switch and setup guards. - Basic validation: TypeScript, 57 focused UI checks and the existing mocked session-creation/fork regressions passed. The Debug iPhone simulator exercised Settings, search opening/closing, direct creation through the configured server, and back navigation without sending a message. Removed only the disposable chat created by this check. Development warnings briefly covered the first UI attempt; the actions were exercised again after dismissal. - Confirmed a remaining mobile gap: creating a chat persists it immediately; the client does not yet use the server's per-client empty-placeholder dismissal contract. Returning without sending still retains the row. This change does not implement automatic cleanup or alter existing chats/drafts. - Availability: local development preview only. Android and tablet UI acceptance remain pending; no release build, publication or server deployment occurred. ## 2026-10-17 - Simplify mobile home surfaces and dock phone search - Use the same background for the home list and top safe area. Remove the footer divider and persistent selected-chat background/rail; retain pressed feedback and running/unread status indicators. Settings remains a transparent icon with a circular touch target. - Place phone search below the results and immediately above the software keyboard. Closing clears the filter and restores Settings; the keyboard Search action submits the current query and dismisses the keyboard. Keep tablet search at the top and preserve server-generation guards and search debounce. - Replace hard-coded component corner radii with shared small, control, sheet and circular tokens, preserving control dimensions and proportional shapes. - Basic validation: TypeScript and 31 focused sidebar, theme, server-profile and profile-safety checks passed. The existing Debug iPhone simulator exercised search opening, typing, title/history results through the configured server, no-match results, closing/reopening and Settings opening/closing in light and dark themes. Verified software-keyboard positioning and Search submission. No chat messages were sent or server settings changed. - Availability: local development preview only. No server update is required. Android, tablet, full secondary-screen and physical-device visual acceptance are not claimed; broader regression remains deferred for this UI iteration. No native release build or publication was performed. ## 2026-10-17 - Prepare an independent mobile development variant - Add an opt-in development configuration named AgentsDock Dev with separate app identifiers and URL scheme, leaving the default shipping configuration unchanged. Reject unknown variant names instead of silently selecting the shipping application. - Document separate device installation and native regeneration when changing variants. Check the effective Info.plist build version for generated Expo projects, including projects that use a literal rather than an Xcode variable. - Validation: eight variant, native metadata and privacy checks plus TypeScript passed. Expo generated the independent iOS project and its distinct URL schemes. Native dependency parity checks and a local iOS Debug build passed. Verified the signed development identifier, bundled JavaScript and device provisioning, then installed and launched the separate app on a physical iPhone. Confirmed that the shipping app remained installed unchanged and the development process stayed running. On-device visual and connected-server acceptance remain pending; no existing app was replaced and no build was uploaded or published. ## 2026-10-17 - Establish mobile visual foundations - Add shared UI typography roles, spacing and corner-radius tokens. Apply them to primary navigation, server management, settings forms, chat headers, composer surfaces and message metadata without changing their actions. - Increase small navigation labels, soften font weights, and give chat rows more breathing room. Use neutral surfaces and message backgrounds while retaining distinct connection, running, unread and error colors. - Preserve app-wide text scaling and measured Markdown, code and terminal typography. This is an incremental style pass, not a component-library migration or a redesign of every secondary panel. - Basic validation: TypeScript, 31 focused UI/theme checks and the typography regression entrypoint passed. The existing Debug iPhone simulator exercised search typing/clearing, server management, chat opening, input focus and return navigation in light/dark themes, plus repeated Settings opening and closing in light mode. No messages were sent or server profiles changed. - Availability: local development preview only. No fresh native release build, Android/device acceptance, full keyboard regression or publication is claimed. Broader testing remains deferred until the UI iteration is complete. ## 2026-10-17 - Simplify the mobile home header - Replace the homepage brand title and footer counters with one footer Settings action. Keep the existing settings contents and per-chat activity indicators. - Move server selection to the top with its original compact typography and connection-status dot. Long names truncate within half the available row; the native server menu remains, without a duplicate URL subtitle. - Add a top-right search icon that opens and focuses the search field; closing search clears the filter and dismisses the keyboard. - Replace the refresh toolbar button with native list pull-to-refresh using the existing generation-scoped refresh action and a request-bound loading state. - Basic validation: TypeScript and 25 focused UI source checks passed. The local iPhone simulator exercised the pull gesture, server menu opening/dismissal, search opening, typing, closing and reopening with the filter cleared, and earlier repeated Settings opening/closing. No server profile was changed by these checks. Full regression and device coverage are deferred until the UI iteration is complete. Availability: local development preview only. ## 2026-10-17 — Reconcile stale Claude subagent status - Reconcile the subagent panel's historical Claude records against the existing runtime owner. Untracked old work moves to **Tracking lost** history instead of accumulating hours under **Active now**. Preserve confirmed completion, genuine tracked work, and separate identities for reused task IDs across runs. This read does not start providers, interrupt work, or rewrite chat history. - Accept corrected server snapshots in the desktop cache, including corrections whose journal sequence is unchanged or older than a stale projection. Ignore delayed running updates for retired children and reconcile launch-tool aliases without duplicating rows or overriding authoritative task state. - Validation covers historical missing completion, owner replacement, live work, delayed events, terminal receipts, cache repair, reload and chat switching. Focused server/desktop tests, TypeScript and production compilation passed. - Personally exercised the compiled desktop offscreen through native IPC and a real authenticated disposable server. The baseline reproduced 39 stale active rows; the corrected app showed them in history, preserved a controlled live owner, and rejected delayed running updates. Task history and the live-owner fixture were synthetic. A separate real disposable Claude request rejected the legacy Agent launch as unavailable and correctly produced a failed task; a real child's successful full lifecycle remains unverified. - Rechecked the existing Team Space stop-hosting/join fix in the same candidate with two isolated real servers, including a lost reply, TLS joining, message delivery, reload and preservation of the original team. No installed app or production service was changed; these changes are source-only. ## 2026-10-17 — Let a Team Space host stop hosting and join another team - Add a visible **Stop hosting…** action that invokes the existing Host-to-Member operation. Replace the old listener-only control, which left the Host role active and prevented joining another team. Confirmation explains that peer access stops while chats and the original team's stored data remain intact. - Retire only the desktop's obsolete managed-host session cache after confirmed demotion. Keep human sign-in records, server-owned teams and peer pairings. Recover a committed stop whose response or follow-up health request failed using explicit authenticated role evidence, retaining the original warning. - Clear the old workspace and expose joining; restore a previously joined remote team's workspace when the server resumes that membership. Preserve pending invitations and ignore results belonging to a different selected server. - Validation: 1,040 focused desktop checks, TypeScript and production compilation passed. Sixteen isolated server checks and a real TLS lifecycle probe passed. An unrelated filesystem fixture requires its expected POSIX umask; its isolated rerun and the final scoped suite passed with that setting. - Native acceptance: reproduced the missing exit in the baseline desktop, then personally exercised the compiled app offscreen with disposable real servers: cancel, failed stop and retry, successful stop, lost response after committed stop, invitation-based joining, and reconnect after renderer reload. Mail sent through native IPC was read on the new host. Reactivating the original host preserved its hub/team identities and saved message. Host approval and original team reactivation used authenticated fixture APIs; provider runtimes were not invoked. No installed apps or production services were changed. - Visual acceptance: checked the actual desktop at 1040 × 780 in light and dark appearance, including English and Chinese confirmation, keyboard focus, Escape/cancel, and the recovered stop's dismissible warning and join form. - Availability: desktop source and isolated development build only. This uses the existing server host-control contract; no new server deployment is required. Signed-package and other desktop-platform acceptance are not established here. ## 2026-10-17 — Retire unavailable Codex model choices without deleting chats - Scope known Codex model retirements to native ChatGPT sign-in. Filter the discovered list, configured defaults and fallback options; publish replacement metadata so the desktop does not reinsert an old chat's retired selection. Native API-key authentication and custom endpoints keep their own availability. - Migrate an old selection at the next admitted turn, retaining the native conversation ID and history. Preserve supported reasoning levels, record a visible model-change event, and leave running work untouched. Standalone scheduled runs migrate their runtime snapshot without rewriting the parent. - Discard the older desktop model-catalog cache on upgrade, rather than retaining its obsolete default until the cache expires. Historical model names remain in history; this change does not rewrite transcripts or hide provider errors. - Validation: 244 targeted server checks and 477 targeted desktop checks passed; TypeScript, production compilation and output verification passed. The desktop filesystem fixture suite required its expected POSIX umask; a first run under a restrictive umask failed an unrelated clipboard-directory fixture. - Native acceptance: reproduced the obsolete default and reinserted selection in the baseline Electron app. An isolated native development app and real authenticated server then hid that selection, migrated a disposable saved selection on send, and completed two real Codex continuations with the same provider thread and remembered context. Reopening retained history and exactly one migration notice; a new chat using Server default also replied successfully. The QA launcher isolated local-server discovery from existing installations; provider calls, Keychain access and client/server transport were real. - API/custom exceptions and scheduled-parent isolation used deterministic fixtures, not live paid endpoints or a long-running production schedule. Native compaction recovery and its error attribution remain separate work tracked in issue #89; this test does not establish recovery of previously failing long contexts. - Availability: source branch and isolated development testing only. Both server and desktop updates are needed; no production instance or public release changed. ## 2026-10-16 — Integrate responsive history maintenance with current main - Move provider-history preparation outside chat admission locks, retaining ownership proofs and source/cursor validation before committing results. Persist explicit send request identities and queue recovery checkpoints so lost responses can be retried without admitting duplicate work. - Preserve current event-stream liveness and recovery while adding bounded refresh notifications when background history projection changes. - Port the history-responsiveness changes without release-branch version bumps or unrelated release-only changes. This requires both updated app and server. - Validation: 614 focused desktop tests passed; backend batches of 280 and 182 checks passed (including overlapping transport coverage and a 330 MiB import fixture). TypeScript, production compilation and output verification passed. - Isolated native Electron exercised authenticated real transports, Codex and Claude replies, same-provider-session continuation after restart, queued work, Stop and subsequent sending, navigation, and large synthetic history views. Eight concurrent retries after discarding a real send response joined one durable admission; retry after server restart retained that receipt. - Full CI exposed incomplete isolated-test dependencies and regressions in native-title refresh, partial-record migration proof and empty session-list metadata. Restore the title lookup outside admission locks, require a fully scanned journal for migration and keep never-started maintenance sparse. Known maintenance failures and their cleared state remain visible. All 230 targeted history, title, mailbox, provider and session-summary checks pass. - Availability: local source integration and development app only; no public release or installed production update. Large fixtures validate display and storage behavior, not provider memory. Cold indexing still depends on transcript size; signed-package, cross-machine and all-provider acceptance are not established by these checks. ## 2026-10-15 - Remove mobile QR connection setup - Removed the mobile camera scanner, QR-image import, connection-link handler and setup prompts. Removed the camera dependency and plugin, camera permission and pairing URL scheme. Manual server URL/token entry, authenticated connection tests, identity checks, secure storage, photo/video attachments and the legacy terminal URL scheme remain. Desktop and server pairing code is unchanged. - Restrict terminal deep-link matching to the outer URL so a discarded connection payload containing a terminal URL cannot open the terminal accidentally. - A clean candidate using the committed dependencies passed TypeScript, seven focused library/store test entrypoints, 310 UI checks and an iOS JavaScript export. One generated-native-metadata check was skipped in the source-only candidate. CI now checks manual connection and the absence of QR entry points. - Actual iPhone simulator interaction verified manual add/test/save/switch against an isolated production server, three repeated management openings, light/dark presentation and a cold launch with saved-token authentication. Installed-bundle checks verified no camera permission or pairing scheme while retaining photo-library permission. Old connection links do not present setup or open a terminal through an embedded URL. No user chat messages were sent. - Native acceptance used an ad-hoc-signed Debug simulator build with Keychain entitlements and the existing local SDK-aligned dependencies. Those unrelated dependency adjustments remain outside this change. No physical-device or Android acceptance, TestFlight upload, release or server deployment is claimed. ## 2026-10-15 - Stress-test mobile sync and connection fixes - Added repeatable CI coverage for latest-first sync, recovery, stable native Markdown nodes, server-name migration and the first-page navigation change. - Store stress passed 40 cycles across eight chats with 36,000 events and 1,200 duplicate deliveries. Checks cover bounded snapshots, one active stream, repeated retry taps, transient failures, backgrounding, late HTTP responses and retired-stream callbacks. Rendering stress retained all 160 paragraph host instances through 300 text/appearance updates using the real Markdown renderer with mocked native hosts. Profile migration preserved 64 identities, custom names and credential versions across 100 round trips. - A clean candidate using the committed lockfile passed all 28 focused TypeScript test entrypoints, type checking, 307 UI checks and an iOS production JavaScript export. One generated-native-metadata check was skipped because the source-only candidate has no generated native project. Updated stale UI source assertions and removed an integration fixture race by waiting for reconnect's optional runtime request before measuring identity rejection; that integration check also passed 20 consecutive runs. - Six production mobile API clients made 60 latest-page reads and underwent 18 forced WebSocket disconnects against the isolated production service. Every client received all 105 expected events in order, without duplicates. The fixture used synthetic Claude and Codex history, not model requests. - Actual iPhone simulator input repeated 20 long-chat openings, three background/foreground cycles and five right-edge server-menu openings. All returned the latest content and remained operable. Adding an unnamed server displayed its host and port. The original profile was retained throughout. - Native checks reused the installed development app and its local SDK-aligned dependencies. Those dependency adjustments are excluded from this repair PR; no native release archive, physical-device or real-provider parity acceptance is claimed. The affected real conversation still needs user confirmation. ## 2026-10-14 - Open mobile long chats latest-first - Opening a chat, returning to the foreground, or recovering a disconnected timeline now fetches the latest bounded page before subscribing. Cached content remains visible until that page arrives; healthy live windows keep their bounded delta path. Older history remains available through paging, and servers without semantic paging retain the legacy tail fallback. - Replacement reads retire the previous stream. Concurrent send results are retained, but the replacement stream starts at the HTTP snapshot boundary so intervening events cannot be skipped. Failed or obsolete reads cannot clear visible content or restart an old chat's stream. - Markdown now memoizes its parsed AST and gives paragraphs and the root stable structural keys. Text updates no longer remount every native text view; links, code, tables, math and range-selection rendering are retained. - Regression tests reproduced old-cursor reads and native text remounts before the fix. Latest-first, fallback, retry, background, concurrent-event and retired-response tests pass, along with store/client/history regressions, 63 rendering/contract checks, TypeScript and an iOS production JS export. - Exercised the installed iPhone simulator against an isolated instance of the production service with 2,000 synthetic turns / 8,000 initial events per Claude and Codex chat: latest opening, repeated reopening, live appends, draft retention, upward pagination, return to latest and foreground recovery. Light/dark rendering was inspected. A Claude live-update recording sampled at 30 fps had no blank transcript frames; this is not a guarantee against every form of flicker. The history automation was adjusted after automatic pagination moved its target, and the resulting older page was checked in-app. - Availability: local mobile branch and simulator only; no server deployment, publication or native archive. Real provider output parity, the user's affected conversation, physical devices and network switching remain unverified. Server/provider history and model context are unchanged. ## 2026-10-14 - Use readable mobile server names - Unnamed mobile connections now use the URL host and port instead of the internal server identity. Loading saved profiles repairs names equal to that identity while retaining other names, profile IDs, credential versions, timestamps and identity-scoped workspaces. - The server selector measures its available row width explicitly so the intrinsic native menu label does not shrink the button to its text. - Removed the first-page Team Network shortcut as requested, retaining its existing workspace and other entry points. - Verified the installed iPhone simulator development app against the existing connected server: the saved name migrated, the selector filled its row, and right-edge taps, menu reopening and server management displayed the same address in light and dark modes. No chat messages were sent. New-profile, legacy migration, credential/draft preservation and pairing paths also passed focused tests; those tests use isolated fixtures, not a new live pairing. - Profile/storage/store regressions, 14 UI source-contract checks, TypeScript and an iOS production JavaScript export passed. Source: local changes on `988edb6e`; no new native archive, physical-device or Android acceptance. Availability: local simulator only; no server update or publication required. Long-conversation sync recovery is separate and is not changed by this fix. ## 2026-10-14 — Explain failed desktop file links - File links and binary attachment links now surface native open/download failures through the existing dismissible error notice. Missing linked attachments explain, in English and Chinese, that the agent must attach the file. A server filesystem path is not automatically a downloadable attachment; registration and authorization requirements are unchanged. - Late errors are discarded after changing chat or server. Successful opens preserve their existing behavior and do not clear unrelated error notices. - Real native offscreen acceptance reproduced the silent failure before the fix, then exercised missing-file HTTP 404, dismissal/retry and an authorized attachment HTTP 200 through the production IPC and test server. Checked dark English and narrow light Chinese layouts. The packaged local candidate also passed actual click-to-error and normal attachment opening against the server. - Focused regression checks (51), TypeScript and production compilation passed. The first full suite hit the existing same-second server-update receipt test; its standalone rerun passed all 44 checks. The serial full-suite rerun passed 5,176 tests (five skipped). Eight package/license checks passed. - Availability: local desktop candidate `1.0.11-beta.1.sync.4`, retaining the latest-first sync changes; not published. No server update is required. User acceptance of the local file-link test passed before PR submission. Cross-machine acceptance remains pending; mobile is outside this change. ## 2026-10-14 — Open stale desktop chats latest-first - Cached content still paints immediately, but opening/reconnecting a timeline now requests the latest bounded semantic page before subscribing at its snapshot boundary. It no longer starts a full offline-gap replay while also fetching that page. Older history remains available through upward paging; server/provider history, thinking content and chronological order are unchanged. - A main-process authoritative tail can replace a disjoint renderer window without a redundant refetch or fabricated adjacency. Overlapping loaded prefixes retain their existing handling. Failed tail reads retain the cache; retired leases and buffered events cannot overwrite a replacement log. - Native offscreen acceptance used the real authenticated test service with 2,000 synthetic turns / 8,000 events, including long reasoning. The App opened at the recent tail, paged older content through real scroll/click input, and accepted a 57-character unsent draft without timeline requests. Six native service clients each reopened a stale cache five times and underwent three forced disconnects; all received one real Codex final reply without duplicates or replaying the old gap. Claude was verified separately with two real replies and an application-frame blackhole; recovery preserved both replies. - Focused desktop checks, TypeScript, production compilation and 28 server transport/reasoning/stress checks passed. Full desktop validation passed 5,164 tests (five skipped), plus eight package/license checks with two workers. Earlier high-concurrency runs encountered timing failures in unchanged secure peer/update settings tests; those tests were not changed by this work. - Availability: local test candidate only; no publication, push or merge. Physical cross-machine/network-switching and OS sleep/wake acceptance remain pending. This change is desktop-side; the existing test server protocol is sufficient, and the working server is not changed. ## 2026-10-14 — Recover stalled desktop timeline streams - Desktop advances its reconnect cursor only after the event consumer accepts a row. Consumer failures close the stream and retry from the previous cursor, with backoff; they cannot silently skip the failed row. - An additive, negotiated server capability confirms completed WebSocket catch-up and sends lightweight idle heartbeats. Desktop reconnects after 45 seconds without progress and renews open timelines after system wake. A progressing catch-up is not replaced by the periodic metadata refresh. Legacy clients/servers keep their existing protocol. Recovery never resends prompts, refreshes provider credentials or changes thinking content. - Real loopback transport stress covered six clients, 6,000 historical rows, 18 forced reconnects, filtered sequence gaps and 40 live appends. Every client received the complete expected sequence without duplicates; connection reservations were released. These synthetic-history tests do not call models. - Desktop suite: 5,162 passed, five skipped; eight packaging/license checks, TypeScript and production compilation passed. Server catch-up/reasoning regressions passed. Native offscreen acceptance uses a real authenticated test service, separately with Codex and Claude, with a fault proxy keeping protocol ping/pong alive while dropping application frames. The legacy connection remained stale after 60 seconds despite a completed reply; the new connection automatically recovered the missing reply by the 40-second observation. Both runs retained exactly two submitted turns and two final replies. This is transport acceptance, not a new claim of thinking parity. - The local arm64 test package passed bundle audit, ad-hoc signature validation and isolated startup. It disables updates and uses separate app data; it is not a notarized public release. Real cross-machine/network switching and OS sleep/wake remain user acceptance items; wake handling has focused regression coverage. The designated test server was updated; no working server restart, public upload, release or merge is included. ## 2026-10-14 — Shorten the README introduction and quickstart - Reduce the feature overview and explain the app/server relationship in one short paragraph. Keep installation to four steps, with the npm command and URL/token retrieval visible. - Add prominent website, setup and feature-guide links; keep detailed prerequisites and troubleshooting in the existing server guide. - Documentation only; no runtime, installer or release behavior changed. ## 2026-10-14 — Use the published npm CLI in installation guides - Make `npm install -g agentsdock` the primary manual installation command in the product and server READMEs. Explain automatic first setup, existing-state preservation, URL/token retrieval and the distinction from managed updates. - Retain source installation as an advanced path and show the short commands for named instances. Remove obsolete unpublished-package wording from the server and CLI introductions. - Organize first-time setup around the server host, prerequisites, npm install, address/token selection and a first provider chat. Clarify that tmux enables managed updates, localhost addresses cannot connect another device, and native Windows is a client platform rather than an npm server host. - Link desktop downloads to the latest stable release instead of a stale version. Add setup troubleshooting and distinguish unreleased CLI output improvements from commands available in the published packages. - Verified public npm tags and the stable package's integrity, setup guards and command parser without creating, restarting or changing a service. Checked shell syntax for all 18 command blocks, resolved 24 relative file links and verified the prerequisite documentation links. Documentation only; no fresh service installation, runtime or release change. ## 2026-10-13 — Show CLI connection details together - Successful first-time global npm setup shows the default server URL and token directly in the controlling terminal, even with npm's normal hidden lifecycle output. Raw installer output and tokens stay out of npm's captured streams/debug logs. Headless installs retain a non-secret summary; existing installations are still left unchanged. - Interactive `agentsdock token [NAME]` shows binding-aware connection URLs before the selected token and optional clipboard prompt. Interactive `agentsdock status [NAME]` offers an opt-in arrow-key server picker afterward. Cancellation shows no token; redirected status never prompts, and redirected explicit token commands retain raw token-only output. - Checked the actual packaged CLI against a running macOS test instance: URL/token agreement, status → consent → arrow selection → token, decline, and redirected behavior. Authenticated health, service process/start time, server identity/version and configuration stayed unchanged. No clipboard write, service restart or global CLI replacement was performed. - Package tests exercise real npm both with and without foreground scripts, checking terminal visibility and absence of credentials from npm logs. That first-install test uses a synthetic installer, not new native service creation. All 171 focused Node/Python checks passed, along with JavaScript syntax and diff checks. Native Linux terminal acceptance remains untested. Source only; no npm publication is included. ## 2026-10-13 — Explain missing uv during npm setup - When automatic server setup cannot find or run uv, explain its Python environment role and show installation, verification and retry instructions. macOS guidance includes `brew install uv`; other platforms link the official installation guide. Retry commands preserve the selected stable/beta version. - Keep installer logs and credentials out of npm diagnostics. Guidance does not execute package managers, download prerequisites, retry setup, or add telemetry. Unrelated failures retain their existing diagnostic path. - All 44 Node CLI checks and five package checks passed, including real npm lifecycle execution with a synthetic missing-uv installer failure and private output assertions. No uv installation/removal, live service changes or new server creation was performed. This improvement is source-only until released. ## 2026-10-13 — Validate recovery fixes against current main - Integrate the new batch emergency dismissal with the bounded-memory journal reader. A 40 MiB regression reproduced more than 80 MiB of allocations with the unprojected reader and passed below 8 MiB after the correction. - Recheck 201 focused server cases, then all 31 emergency cases including the added regression; 907 desktop checks, TypeScript and production compilation passed. The 1,000-chat/1.21 GB process test passed again, retaining all 192 queued messages and original history bytes; chat readiness took 14 seconds. - An isolated native development app used the real authenticated server to dismiss two alerts with a 40 MiB intervening tool record. Both receipts were durable, the active banner cleared, and refresh retained the historical rows without restoring the banner. This fixture used no model provider. - No published release or running user service was replaced by this merge validation. Signed end-to-end release acceptance remains separate. ## 2026-10-14 — Exercise receiving-client timeline synchronization - Added component coverage for sync backfill plus overlapping authoritative corrections, a newer sync window arriving during scheduled empty-range recovery, and a genuinely empty authoritative history. Sync must preserve loaded messages and a stale recovery must not rebuild the newer window. - Validation: 240 targeted synchronization, history, projection, service and viewport checks passed; 483 unrelated checks were excluded by the test-name filter. TypeScript passed. - A native offscreen Electron fixture exercised 31 viewport observations: empty-cache arrival, incremental catch-up, prepend while reading, expanding and shrinking corrected history, disjoint replacement, hidden-tab sync and return, and a narrow window. It used real timeline/store reducers and wheel events with synthetic snapshots, not a remote server or a second physical machine. The unmodified baseline also passed all 31 observations; these checks do not reproduce the reported persistent blank viewport or establish that the candidate recovery fixes its cause. Remote-machine acceptance is still pending. This follow-up changes only tests and documentation. ## 2026-10-13 — Recover a persistently empty virtual timeline range - When a populated timeline emits an empty virtual range after previously displaying messages, allow layout to settle and rebuild its measurements only if the viewport remains visible and empty. Retain the semantic reading anchor rather than reusing a potentially stale pixel offset. Initial, transient and hidden empty ranges do not trigger this recovery. - Validation: 34 focused component/position checks, TypeScript and production compilation passed. An isolated native offscreen Electron fixture exercised wheel scrolling during response growth, completion/collapse and cached-event replay. It used the production timeline and virtualizer, with stubbed IPC; this was not a new live-provider request. - This is defensive recovery for a suspected stale virtual range. The reported intermittent blank viewport was not reproduced, so exact-cause confirmation and acceptance in the affected live conversation remain pending. Visible-app automation was unavailable. Available as a local test package only; no server change or release is included. ## 2026-10-13 — Stress-test recovery and component updates with large histories - Added an opt-in process test with 1,000 saved chats, approximately 1.21 GB of mixed history, 192 queued messages and 32 interrupted turns. Production worker/gateway startup, concurrent HTTP reads, exact queue order, single interrupted-run closure and original history-byte preservation passed. - Separate isolated native acceptance exercised 32 deterministic concurrent runs and eight child processes: busy worker replacement was rejected, gateway replacement preserved work, and a failed candidate rolled back before a successful retry. The native App reconnected, showed pending blockers and cancelled the fixture reservation without stopping active work. - Both component health and chat readiness are required: rollback restored HTTP before background queue reconstruction finished. Continuous work can still defer an ordinary idle update indefinitely under the existing policy. - See [the stress matrix and boundaries](RECOVERY_READINESS_TESTING.md#large-state-and-busy-update-stress-pass--2026-10-13). These are source/runtime and native-client checks, not a signed release installation or platform-wide acceptance. No production code, public release or working service was changed by this stress pass. ## 2026-10-13 — Distinguish failed chat recovery and stream emergency receipts - Expanded acceptance found and fixed a structured-payload regression: emergency journal projection no longer inherits native-import value-count limits or charges discarded object keys as retained metadata. Import defaults remain unchanged. The new regression failed before the correction. - Additional validation passed 195 recovery/alert checks, 203 update/rollback checks, 99 overlapping import/admission checks and 487 desktop service/update checks. Native Electron also exercised Send and Refresh against an isolated failed-recovery HTTP server: the draft survived, the error was explicit, and server selection remained available. The 256 MiB HTTP fixture repeated at 1.58 seconds. Repeating the real provider/UI acknowledgement after deployment, reopening the test app, and continuing the same conversation also passed. Test-instance identity, token and existing chats were preserved; disposable test data was cleaned up. See [the test matrix](RECOVERY_READINESS_TESTING.md) for boundaries. - Completed queue-recovery timeouts now report failure instead of indefinitely claiming recovery is in progress. Admission stays closed on failure, and disconnecting a request cannot cancel shared queue restoration. - Health separately reports chat readiness. Desktop connection status remains usable for recovery while warning about unavailable chat admission; a paired update stays pending or failed until reported chat recovery succeeds. Older servers without the field retain their existing behavior. - Emergency receipt scans now perform linear reads and retain only lifecycle fields. Large unrelated records no longer cause repeated whole-line copying. Projection reads release the global session lock and recheck sequence state before applying, preserving acknowledgement idempotency and ordering. - A real loopback HTTP fixture with a 256 MiB unrelated record acknowledged in about 1.6 seconds; the unchanged stable source timed out at 30 seconds. This is a synthetic regression, not proof of a particular user's remote cause. - Validation: 187 focused Python checks, 5,142 desktop checks (five skipped), TypeScript, production compilation, bundle audit and strict ad-hoc signature verification passed. Permission-sensitive test fixtures used native temporary storage; build outputs remain local. - A separate native Electron profile completed a real Codex request through the production client transport. A synthetic provider alert appeared in the UI; clicking Acknowledge cleared its active controls and committed exactly one durable acknowledgement. An explicit repeated request returned acknowledged false, preserving idempotency. Recovery failures were injected in isolated HTTP/regression fixtures, not into a user's active service. - Available as a local test server and local desktop build. No release is published by this change. ## 2026-10-13 — View local server logs from the CLI - Add `agentsdock logs [NAME]`, defaulting to the default instance, with `--lines` / `-n` and `--follow` / `-f`. Viewing logs does not install, restart, stop, or reconfigure a service. - Select the installed macOS log layout, including gateway and worker files for split default installations. Bound file reads, handle rotation and missing files, reject unsafe paths, and escape terminal control characters. Linux reads the selected user-service journal without a pager. - Verify the actual packaged command against a running macOS named instance: recent-log output, follow, and interruption preserved service PID/start time, authenticated health, identity, version and configuration. Package tests use disposable installations. Targeted checks: 191 passed, with one native systemd parser check skipped on macOS. Native Linux journal acceptance remains untested. - Logs can contain private data and must be reviewed before sharing. This is an unpublished source change, not included in existing CLI releases. ## 2026-10-13 — Repair legacy cross-chat reply imports - Older cross-chat reply/request imports no longer reappear as messages attributed to the user when the matching delivery is outside the loaded timeline page. The previous repair covered asynchronous handoffs but missed this older exchange format. - Recovery matches the native provider input to durable delivery records. Genuine user quotations remain visible; the original journal, provider history, event identities and ordering are preserved. This does not enable legacy communication routes. - All 48 focused history-repair/import-boundary tests passed. Unchanged source reproduced the gap, while a read-only check of an affected large history confirmed the candidate repair without changing its files or service state. - An isolated native offscreen app reproduced the leaked user bubble before the fix. The candidate used real authenticated timeline reads, removed that duplicate, preserved a separate genuine quotation, and remained correct on chat navigation and reload. Fixture journal/provider bytes stayed identical. No model-provider request or new cross-chat delivery was made. - Availability: source only; requires a server update. No production restart or release publication is included. ## 2026-10-13 — Dismiss emergency alerts for a chat - The current chat's emergency banner now has one **Dismiss all** action. It persists acknowledgements through the alert displayed at click time; later alerts and other chats remain active, and historical alert rows stay available. Individual acknowledgements remain available in the timeline. - The existing authenticated acknowledgement API accepts an exact durable alert boundary and commits the batch together. Repeated requests and another device acknowledging the boundary are harmless. Active work is untouched. - An isolated native offscreen Electron run reproduced the old one-at-a-time behavior, then exercised the new UI through production IPC and authenticated HTTP into a real disposable server. It checked persistence, reconnect, another chat, a new alert during the request, failed request/retry, and supported narrow layout. Component coverage checks localized labels. Alert creation was synthetic; no model-provider request was needed or made. - Focused emergency API/CLI, desktop transport/store/component checks, TypeScript checking and production compilation passed. A pre-existing clipboard test assumes umask 022; its isolated rerun passed with that umask. - Availability: source only. This feature requires the paired app/server change; no release or production restart is included. ## 2026-10-13 — Recover server updates and preserve detached jobs - Linux execution workers now restart independently of detached background jobs. Update admission checks the effective native service policy before accepting a worker replacement that preserves those jobs. - Failed-update recovery accepts valid retained rollback releases. Activation also supports an existing hotfix whose Python environment belongs to a retained sibling release, without changing that environment or its links. - Focused update, handoff, durability and native-service suites passed with platform-specific skips. Regression fixtures reproduced both recovery failures before the fixes. A real isolated Linux service exercised automatic restart and explicit stop/start, preserving a detached child's PID and start time across both operations. Its worker was a fixture, not a model provider. - These changes do not publish a desktop or server release. ## 2026-10-12 — Validate large-history imports for merge - Recheck import, incremental synchronization, Claude/Codex provenance, recent history repair, fork boundaries and runtime manifests: 402 of 406 checks passed on the development filesystem. The remaining four also fail on unchanged main because their fixtures require preserved inode identities or pre-1980 timestamps. All four passed using native temporary storage; the full 15-case recent-history suite passed there as well. Production identity checks and test assertions remain unchanged. - All 11 npm/legacy package checks passed, including actual disposable archive creation and offline npm installation. Python compilation, shell syntax and diff checks passed. - Repeat authenticated imports for Claude and Codex confirmed explicit single-record limits, observed byte counts and correct record offsets. Rejected imports created no chats and preserved native source bytes, existing chats, server identity and credentials; temporary fixtures were removed. - Native desktop UI and actual provider resume acceptance remain unverified. No desktop or server release publication is included in this change. ## 2026-10-12 — Explain native import resource limits - Identify single-record text/metadata, field/scalar size, nesting, value-count, scan-record-count and timeout failures separately. Include observed amounts and limits, plus scan-relative record position and source byte offset. - Explain that retained-byte accounting includes overhead and is distinct from total conversation disk size. Observed sizes are lower bounds because reading stops at exhaustion; do not expose transcript content in diagnostics. - Preserve the import API's error contract, failure semantics and source bytes. All 97 focused reader/import/sync checks passed, including real parser errors reaching both providers' bulk-import responses without creating a chat. Changed Python compilation and diff checks passed. Live authenticated Claude and Codex imports of disposable 33 MiB text records returned the specific 32 MiB accounting limit and correct record offsets. Failed imports created no chats, left native fixture bytes unchanged, and preserved pre-existing chats, server identity and credentials. Fixtures were removed after verification. Native desktop acceptance remains pending while computer control is unavailable. ## 2026-10-12 — Stream large native history payloads - Replace whole-record JSON loading for native imports, sync and native proof reads with bounded-memory projection. Opaque image/tool payloads remain in the unchanged provider transcript; message text and provenance are preserved. - Remove the native import file-size cutoff and raw media-record cap. Retain explicit text/metadata, JSON structure, record-count and work budgets. Hash original bytes in chunks and keep file-identity/offset checks; a processing failure cannot become a successful partial import. - Include the shared reader in both server package formats, installation, deployment and staged import/compilation checks. - Isolated-reader measurements: approximately 281 MiB peak RSS with the former parser versus 23 MiB with streaming for the same 64 MiB image record; a 300 MiB record used approximately 24 MiB. These are fixture measurements, not the running server's total memory or a guaranteed upper bound. - Validation: 267 of 269 import/sync/provenance/package checks passed; the two remaining failures reproduce on unmodified main (candidate ordering and malformed-line native proof). Another 93 legacy-proof/deferral/fork checks passed. The memory-growth regression passed in separate reader processes. Changed Python and installer/deployment shell syntax checks passed. - Native desktop acceptance is pending because computer control is unavailable; no native-provider resume or desktop UI acceptance is claimed. - Final focused checks passed all 30 cases after reviewing blank/malformed and non-object JSON distinctions, including bounded-memory growth and native fork checkpoint behavior. Live authenticated import accepted a 300 MiB Claude image record and a 64 MiB Codex tool record that the preceding runtime rejected. Appending 40 MiB records, reopen-triggered synchronization and repeated reads preserved order without duplicates. Excessive retained text still failed explicitly. Original transcript bytes, pre-existing chat IDs/titles, server identity and credentials were preserved; disposable test data was removed. ## 2026-10-12 — Accept larger native transcript records - Raise the Claude/Codex JSONL record bound from 4 MiB to 32 MiB across import, cursor synchronization, fork scans and native history proof readers. Include the actual bound in import errors. Embedded media and tool output no longer fail solely because one record exceeds the former 4 MiB bound. - Preserve total-file, scan, displayed-text and aggregate proof budgets. Separate the existing proof-text cap from the larger raw-record cap; original provider history remains read-only. - Validation: 250 of 252 targeted regression checks passed, including nine new boundary, synchronization and provenance cases. The two existing failures (candidate ordering and malformed-line native proof) also reproduce against unmodified main. Native desktop acceptance remains pending because computer control is unavailable. - Live authenticated server checks reproduced the former rejection for separate 5 MiB Claude image and Codex tool fixtures, then passed import, reopen-triggered incremental synchronization and repeated reads without duplicates after the update. Over-limit input still failed explicitly. Fixture source bytes and pre-existing chat identities/titles, server identity and credentials were preserved; only disposable test chats and transcripts were removed. These HTTP checks do not establish native-provider resume or desktop UI acceptance. ## 2026-10-12 — Verify the consolidated Files inspector for merge - Integrate current main and preserve its release-tooling changes. Update the old Explorer-column style regression to require the editor's full-width layout. - Validation: 230 of 231 focused checks passed initially; the sole failure was the obsolete layout expectation, and all 13 style checks passed after updating it. Both previously timing-sensitive directory checks passed this time. TypeScript, production compilation and eight package/license checks passed. - The first hosted desktop run passed 5,138 checks but hit the second-page reveal fixture's one-second DOM polling limit. Await its mocked page responses and React commits before checking the rendered target; retain the selected-row and exact paging-offset assertions without changing production deadlines. All three focused pagination/reveal checks passed after that fixture change. - The user reported the local app looks good and authorized the PR and merge. Agent-operated native acceptance remains pending because computer control is disabled. The local app remains the previously verified package; no runtime behavior changed during this merge preparation. No server update is required. ## 2026-10-11 — Show Files when opening a file - Opening or selecting a workspace, attachment or image/PDF file tab now opens the right panel and selects Files, including when the panel was closed or showing Details/Review. The choice is remembered for subsequent reopening. - Apply the same behavior in split chats. Reject stale notifications from a previous chat, working directory or server connection. Cancelling the picker or rejecting a file before opening a tab leaves the panel unchanged. - Validation: all 56 focused app/file-opening checks passed. The broader 194-case run passed 191; a separate rerun passed its large-file timeout case. Two remaining large-directory timing failures also reproduced against the preceding source, so broader editor acceptance retains that limitation. TypeScript and production compilation passed. - Native desktop acceptance is pending because computer control is disabled; component fixtures use mocked IPC. No server changes are required. Local package: version 0.2.0/build 85, source `b771f38f`. Bundle audit, strict ad-hoc signature verification and all 25 compiled-code asset comparisons passed; the local app was gracefully restarted. Isolated user data and disabled automatic updates are retained. ## 2026-10-11 — Consolidate workspace browsing in Files - Move the editor Explorer into the right-hand Files tab. Reuse its new file/folder, refresh, open beside, copy path, download, rename and delete operations, retaining dirty-buffer and revision checks. The editor no longer reserves a second tree column or Explorer resize handle. - Remember the last Details/Files choice across panel and app reopening. Review remains transient. Split-chat browsing follows the focused chat without covering either conversation; opening a file activates the same editor. - Preserve paged results on folder errors, allow retry, reject changed roots, and prevent background folder loads from resetting a quick-open selection. - Validation: 223 focused component/style/localization checks passed (the final editor rerun passed all 138 cases), plus eight compilation/license checks. TypeScript and production compilation passed. - Native desktop acceptance remains pending because computer control is disabled. Component tests use mocked IPC and cannot establish live server authorization or actual desktop appearance. No server API change is required. Local package: version 0.2.0/build 85, source `ec572e80`. Bundle audit, strict ad-hoc signature verification and all 25 compiled-code asset comparisons passed; the local app was gracefully restarted. Isolated user data and disabled automatic updates are retained. ## 2026-10-11 — Browse workspace files from chat details - Add Details / Files tabs to the chat inspector, retaining Review when present. Files shows the server-side working directory, expandable folders, explicit pagination and refresh, and opens selected files in the existing editor. - Load only the root and folders requested by the user. Clear directory state when the chat, server generation or working directory changes; ignore stale responses. Preserve existing authenticated workspace access and symlink rules. - Validation: 172 focused checks passed, including component fixtures for browsing, errors/retry, paging, stale responses, English/Chinese labels and the Files-to-editor handoff. TypeScript and production compilation passed. The full desktop run had 5,125 passes and five skips; its only failure was an existing clipboard-permission fixture under umask 077, which passed with its expected umask 022; all 421 service checks passed on that rerun. All eight compilation-output/license checks passed. Native desktop acceptance is pending because computer control is unavailable; mocked IPC checks do not establish the real app/server journey. Availability: local desktop package, version 0.2.0/build 85, packaged source `61265aca`. Bundle audit, strict ad-hoc signature verification and packaged-renderer byte comparison passed. The package uses isolated app data with automatic updates disabled. No server API change or restart is required. ## 2026-10-11 — Reconcile hosted desktop release readiness - Bring the release line's audited native build-number epochs into the maintained workflow, retaining every historical run/rerun reservation and avoiding stale build-number reuse after local builds. - Add pinned canonical-draft verification without publication. The read-only helper reuses all four native verification jobs, rejects changed source or checksum seals, and emits a receipt only for matching successful results. Normal publication keeps its existing protected credential and mirror guards. - Include the desktop release regressions in source CI and document permanent operator setup, one-time short-CLI bootstrap and npm's initial trust-validation deadline. Credential checks remain separate from candidate acceptance. - Validation: 72 focused checks passed on Node 24; all 17 release-script JavaScript test files also passed on the local Node runtime. Workflow YAML and changed script syntax passed. A real read-only check verified the frozen beta.6 draft's metadata, checksum seal and signed server descriptor; no native assets were rebuilt and this does not establish full native or update acceptance. - Availability: source tooling only. Hosted candidate builds, real native acceptance, first CLI publication and OIDC publication remain separate gates. No release, registry package, running server or signing identity changed. ## 2026-10-11 — Integrate connection pairing with matched npm releases - Include the pairing command, bundled QR renderer and third-party notices in the publication verifier's exact file allowlist. Preserve source-byte binding, file-mode checks and rejection of extra files; include all three new files in tamper-rejection fixtures. - Preserve the latest main release-pipeline checks when integrating this branch. Desktop QR sharing stays commented out. No package publication or deployment is part of this source integration. - Update restart regression journeys to expand the active server card before exercising its controls; preserve safe/force confirmation and recovery checks. Keep the server-settings light background explicitly transparent. All 116 related restart and light-theme checks pass. - Allow two asynchronous test fixtures to reach their intended cancellation and private-revision audit checkpoints despite full-suite scheduling pauses. Retain the real SQLite/threaded-probe paths, all assertions and production deadlines; no tests are skipped. ## 2026-10-11 — Temporarily hide desktop QR sharing - Comment out the Show QR code button and its section heading in Server settings. Keep the QR component and underlying sharing implementation for re-enabling later. Token preview/copy and other server controls remain available. - Validation: 31 focused settings/component checks, TypeScript, production compilation, local bundle audits and strict ad-hoc signature checks passed. Retained QR tests exercise its component directly while the entry is hidden. Actual-app visual acceptance remains pending because computer control is blocked. Availability: local desktop package (0.2.0, build 85), automatic updates disabled. No server update is needed. ## 2026-10-11 — Inline token-copy confirmation - Confirm a successful token copy by replacing the button's copy icon with a checkmark, keeping its Copy label and the card height unchanged. Remove the separate success line; failed copies still show a retryable error. Editing or folding the card clears the prior confirmation. - Validation: 31 focused server-settings tests, TypeScript, production compilation, bundle audits and strict ad-hoc signature checks passed. Actual-app visual acceptance remains pending because computer control is blocked. Availability: local desktop package (0.2.0, build 85), automatic updates disabled. No server update is needed. ## 2026-10-11 — Masked token row in server settings - Show the first four token characters followed by masking dots in each expanded server card, with a Copy button at the end that copies the complete credential. Short tokens remain fully masked; missing tokens disable copying. - Mask the preview in the main process before IPC. Read only on expansion or after a token edit, discard stale results, and keep profile listing free of credentials. Preview errors never display credential-store error details. - Validation: 215 targeted component, IPC, preload, credential-lifecycle and localization checks passed, together with TypeScript and production compilation. Fixtures cover stale credential reads, short-token masking, inactive profiles, edits, reopening and missing credentials. Actual-app visual/clipboard acceptance remains pending because computer control is blocked. Local bundle audits and strict ad-hoc signature verification passed (0.2.0, build 85; automatic updates disabled). This desktop-only local package requires no server update. ## 2026-10-11 — Give server settings a clearer layout - Keep the server name and URL together in the collapsed header, with connection status alongside them. Expanded cards show version and identity on separate labeled rows, with wrapping for long addresses and identifiers. - Group Copy token and Show QR code under Connect another device. Separate Use/Edit from the red Remove action, with more spacing and wrapping. Remove the duplicate status/identity block beside the active server's restart control. - Validation: 118 targeted component/localization tests, TypeScript, production compilation, bundle audits and strict ad-hoc signature checks passed. Tests use mocked IPC; actual-app visual acceptance is pending because computer control is blocked. Availability: local desktop test package (0.2.0, build 85), automatic updates disabled; no server update is needed for this layout change. ## 2026-10-11 — Collapsible server settings cards - Match the My Agents settings layout with server cards collapsed by default. Keep each name, connection status and active selection visible; place address, identity/version, edit/remove/use, token copy and QR actions inside the card. The active server's health/restart controls also live inside its card. - Put Add server above the cards. Keep new-connection editing at the top and existing-connection editing inside its owning card. Folding preserves drafts, dismisses credential-bearing QR panels and resets removal confirmation. A newly saved server expands so a failed activation remains actionable. - Validation: 118 related component/localization tests, final focused settings checks, TypeScript, production compilation, local bundle audit and strict ad-hoc signature checks passed. Component IPC is mocked; the actual-app visual/client-server workflow remains pending because computer control is blocked. This is a local desktop change with no service restart or public publication. ## 2026-10-11 — Remove subagent concurrency from Server settings - Remove the Codex subagent-limit form from desktop Server settings. Opening the page no longer reads or offers to change this per-chat default through the server administration API. - Validation: all 163 related settings/goal/server-management component tests, TypeScript, production compilation, local bundle audit and strict ad-hoc signature checks passed. Native visual acceptance remains pending because computer-control policy blocks the desktop app. Availability: local test package only; no server update is required for this UI removal. ## 2026-10-11 — Make persistent goals available without a global toggle - Remove the persistent-goals switch from desktop Server settings. Goal creation, pause, resume and limits remain per-chat controls. - Updated servers ignore the retired environment/config opt-out. Preserve the authenticated legacy status route as non-configurable; enabling is read-only and disabling returns 409 without restarting providers or changing goal state. Existing paused goals stay paused and unrelated shared settings are preserved. - Validation: 151 desktop component tests, TypeScript, production compilation, local bundle audit and strict ad-hoc signature checks passed. Fifty isolated server tests passed, including actual ASGI routes with synthetic credentials, stale disabled settings, rejected legacy writes and unchanged goal/subagent state. Native App visual/goal acceptance remains pending because computer control is blocked. Availability: local desktop package and server source; existing services were not restarted or deployed. ## 2026-10-11 — Simplify saved-server actions - Remove the up/down ordering buttons from desktop Settings → Server and their unused component handler. Keep the saved list order and the remaining connection-management actions. - Validation: all 26 existing server-settings component tests, TypeScript, production compilation, local bundle audit and strict ad-hoc signature checks passed. Actual-app visual acceptance remains pending because computer-control policy blocks the desktop app. Availability: local test package only. ## 2026-10-11 — CLI and mobile QR connection pairing - Add `agentsdock pair [NAME]` / `qrcode` with binding-aware Tailscale/LAN selection, authenticated health and identity checks, terminal QR output, private PNG export and an explicit link-output option. Local-only and ambiguous network selections require explicit choices. Existing services are never restarted. - Bundle the QR renderer with its licenses so short-CLI installation remains offline-capable and keeps the exact same-version server dependency. - Add mobile camera/image scanning and connection deep links. Users review the address, then connect and save through existing identity validation and secure token storage. Matching server identities are reused; failed authentication or an identity mismatch cannot replace a saved server. Preserve the legacy scheme. - Validation: CLI, offline packaging, 97 instance tests, cross-client payload tests, mobile component/profile/setup checks and TypeScript; generated native permission/deep-link configuration; a packaged CLI QR independently decoded and authenticated against the selected live test server. Component native modules are mocked. Simulator UI control is blocked by computer policy; actual native scan/import, persistence/reopen, deep-link presentation, and physical-phone LAN/Tailscale acceptance remain pending. This is local development, not a public npm/mobile release. The iOS simulator Release build and isolated simulator installation passed; the QR fixture was imported for manual image-scan testing. ## 2026-10-11 — Desktop saved-server token copy and connection QR - Add per-server Settings actions for native clipboard token copying and locally generated, saveable QR images, including inactive/offline saved profiles. Copying never returns the token through IPC. Generation/saving checks the selected credential revision; profile metadata stays credential-free. - Define a versioned connection-link contract and document the separate mobile import work. Show local-only address and credential-sharing explanations; generating a QR does not change network binding or grant restricted chat access. - Validation: focused service/IPC/preload/component/format checks, TypeScript and production compilation; native offscreen copy and QR display; real macOS clipboard and local-server authenticated health; independent Vision QR decode. Native save-picker interaction was blocked by computer-control policy. Picker, theme/layout and mobile end-to-end acceptance remain pending. No server source change, service restart or public publication is included. ## 2026-10-11 — Preserve certificate-chain lookup in the hosted signing probe - Include the disposable keychain in the hosted search list only while signing and verifying the probe, matching macOS certificate-chain lookup requirements. - Restore the exact original list on success and failure without replacing the default keychain. Add executable isolated regression checks for that boundary; no credential values or private signing output are exposed. ## 2026-10-11 — Identify hosted signing-preflight failure stages safely - Add fixed public checkpoint labels to the non-publishing Developer ID probe so operators can distinguish certificate import, identity discovery, signing and verification failures without exposing credentials or raw command output. - Preserve the protected hosted-runner guards, disposable-keychain cleanup and all release acceptance requirements. A failed probe remains a failed check. ## 2026-10-11 — Make release operation independent of the operator computer - Reconcile paired npm release tooling with current main while preserving the frozen release branches. Include the native harness implementation as well as its registered workflow, stronger split-default removal/state checks and safe bounded failure identifiers. Keep failed product boundaries failed. - Add a manual protected credential preflight for existing GitHub release access, Developer ID probe signing and read-only Apple API authentication. It does not publish, submit notarization, build a product, install a service, upload private diagnostics or replace signing identities. Hosted execution remains pending; structural tests are not credential validation. - Document operator versus administrator/npm-owner responsibilities, exact source selection, package-specific trust and the limitations of artifact-only or verification-only runs. The scoped beta.6 runtime is public; the short CLI and matched desktop publication remain incomplete. Preserve stable defaults. - Focused pipeline/harness/packaging source checks pass. No new candidate was built and no existing signed artifact, service, credential or public release was changed by this tooling update. See [the operator checklist](NPM_CLI_RELEASES.md#operating-releases-from-another-computer). ## 2026-10-11 — Inherit new-chat settings from the open conversation - Both quick creation and the new-chat chooser prefer the focused conversation's agent, CLI/custom connection, model, reasoning and provider-specific permissions. Saved defaults apply only without an eligible open conversation, remain scoped to the server, and accept older records without permission fields. Instructions, native session identity, credentials and jobs are not copied. - Unavailable inherited agents/connections require an explicit choice instead of silently switching providers. Late preference reads cannot cross server scopes. - Verification: 729 focused desktop tests, typecheck and production compilation passed. Isolated native offscreen Electron reproduced stale model inheritance with the old renderer, then exercised the new renderer via actual mouse/keyboard input and production IPC/HTTP against an authenticated test service. Codex, Claude and Cursor creation preserved server-side settings; a new Codex chat completed a live provider turn. Unavailable OpenCode opened the chooser; cancel/reopen and explicit provider selection worked. The test used isolated in-memory credential storage, not the user's desktop profile. - Pre-merge review also covered a Claude-only availability fixture: initialization cannot silently replace an unavailable inherited provider, including after cancel/reopen. The native offscreen app checked the chooser against a controlled catalog and created the explicitly selected Claude chat through real IPC/HTTP. User testing accepted the primary inheritance workflow. - OpenCode availability and custom-endpoint propagation are covered by regression fixtures, not live authenticated OpenCode/custom-provider runs. This is a local desktop change; no public release, mobile change or primary-server update. ## 2026-10-11 — Separate catalog refresh from Codex sign-in reload - Automatic reconnect/model discovery and Recheck CLIs no longer request an opaque-login handoff. Genuine native login revisions are still detected; custom endpoint managers and background-work ownership remain unchanged. - Add a native-admin-only sign-in reload POST and wire the existing Codex CLI account action to it, with English/Chinese copy and truthful pending status. Account reads and page mounts remain passive; unsupported older servers keep a read-only Recheck. Profile-generation fencing applies to explicit reload. - Log handoff source, reason and coalesced waiting/release outcomes without credentials, account identity, revision digests or command content. - Validation: 67 focused server HTTP/auth/lifecycle tests and 527 desktop transport/service/component/preload tests passed; desktop typecheck and production build passed. Production desktop transport against an isolated installed service passed account read/reload, browser/bearer rejection, repeated catalog refreshes and two real Codex turns on the same native thread. User testing reported no issue with the refresh/reconnect workflow. Live keychain/background-training and implementer App-click acceptance of the new explicit reload action remain pending; computer control was unavailable. No public release or upgrade of existing primary services is included. ## 2026-10-11 — Clean unused chats on desktop startup and explicit refresh - Recognize unused `New chat` rows without a device-local creation marker, including legacy placeholders and blanks whose model/settings changed. Check again after the initial online refresh when startup used an old cache; do not run cleanup on ordinary background refresh or when leaving a chat. - Persist conditional, desktop-scoped dismissals on the server. Keep the empty record recoverable so a different/offline desktop's local draft remains usable. Subsequent changes make the row visible again. Explicit Delete is unchanged. - Flush and protect local drafts/references/uploads before requesting cleanup; the server repeats its history, provider, queue, job and terminal checks under its lifecycle fence. New clients use a distinct endpoint, so an older server cannot interpret recovery-safe cleanup as permanent deletion. - Verification: reproduced the old missing-marker behavior in an isolated offscreen native Electron app, then exercised the compiled app against an authenticated test server through the real IPC/HTTP boundary. Startup and sidebar refresh removed legacy/model-edited blanks; typed drafts and uploaded files survived. Two independent desktop profiles and server/app restarts preserved offline drafts and per-client cleanup. No model inference was used; chats/files were disposable API fixtures and typing/refresh used real UI input. - Verification build: desktop source metadata `0.2.0` / 85 and server source `1.0.7-beta.11`, based on main `4de6e80b`. All 805 focused desktop and 28 server regressions, TypeScript and production compilation pass. Test fixtures/services are isolated; no primary installation changed. Availability: local source and compiled test app only; not published or merged. ## 2026-10-08 — Configure custom APIs before choosing a model - Remove Model ID from Claude/OpenCode connection forms. Initial connection checks are read-only; saved draft model IDs cannot trigger hidden inference. Choose the default afterward from the endpoint catalog and reload it on demand. - When a public or missing catalog cannot verify credentials, save a private `model_required` configuration, not a successful connection. It remains absent from available chat providers until an explicit model check succeeds. Selecting the first model runs that bounded check with a visible possible-cost notice. Already verified defaults remain metadata-only; no arbitrary model is selected. - Preserve native administrator checks, credential redaction, revision fencing, Forget and existing chat bindings. Old servers that cannot retain pending configuration show an update message, not a Model ID field the user cannot find. - Manual model entry and service-specific official catalog links are fallback controls in the later model section, not initial configuration. Include English and Simplified Chinese copy. Unknown gateways are not assigned another service's model list merely because they use that agent or API protocol. - Verification: 62 focused component/native transport tests and 49 server tests pass; TypeScript and production compilation pass. These cover save/list/select/check, failure, refresh, profile switching and no early model call. A real loopback HTTP gateway exercises the production server routes and native authentication with synthetic credentials, not a live provider account. Actual App interaction remains blocked by unavailable computer control; local packaging and user acceptance do not imply release or merge approval. - Full desktop regression: 5,064 pass and five skipped; eight build/license guards pass. Package commit `70af1129` as an audited, ad-hoc signed local App (package metadata `0.2.0`, build 85) with the existing isolated test profile. The designated local test server runs `1.0.10-beta.2.opencode.70af1129` with authenticated health and installed module/source parity verified. The primary server is unchanged. No artifacts are published and merge remains pending. ## 2026-10-07 — Prepare matched npm runtime and short CLI publication - Extend the manual npm release workflow to prepare both `@agentsdock/server` and `agentsdock` from one committed source/version. Validate the accepted CLI receipt and all packaged files against Git, the exact scoped dependency and the existing signed runtime descriptor before publication. Stage CLI bytes in their own immutable two-asset private draft; do not modify runtime/app bundles. - Publish and verify the runtime before the CLI; verify public resolution metadata and downloaded bytes for both. Preserve explicit stable baselines during beta publication. Matching retries skip existing uploads, while conflicting bytes/tags fail closed. Missing first-package ownership/trust is reported as incomplete publication, not a matched-release success. - Add a manually dispatched disposable macOS/Linux native CLI acceptance workflow and bounded reports. It exercises genuine first-global postinstall, repeat-install preservation, commands, service lifecycle and cache retirement. Known split-default service binding refusals remain failed checks, not bypasses. - Validation: 190 JavaScript and 142 Python checks pass, including actual offline npm packaging/install tests, CLI/postinstall behavior, publication guards, instance-manager fixtures and native harness safeguards. Changed workflows pass actionlint. Real fresh native-service, split-default lifecycle/removal, interactive setup and logout/reboot acceptance remain pending; mocked lifecycle checks and safe existing-state no-op installs do not establish those results. - Preparation only: no new beta build, npm upload, release publication, signing identity change or existing-service modification. The short command remains unpublished. See [the matched npm release procedure](NPM_CLI_RELEASES.md). ## 2026-10-07 — Remember custom API drafts and clarify connection fields - Keep non-secret endpoint drafts on this device per server connection and provider, even after failed checks, closing settings or restarting the app. Codex retains its URL too. Keys stay out of drafts; embedded URL credentials, query strings and fragments are not persisted. Typing does not send requests and disk writes are debounced, with a flush on close/page exit. - Show protocol and authentication as separate, visible fields. Mark required inputs, keep provider-fixed protocols read-only, and require a model for known public Zen catalogs. Restored drafts never imply a verified connection. Include English and Simplified Chinese copy. - Verification: 81 focused component, draft, style and native transport tests pass; TypeScript, eight build/license guards and production compilation pass. Full desktop regression: 5,039 pass, five skipped. Filesystem-sensitive tests use a POSIX temporary directory and the CI-compatible file-creation mask; security checks are unchanged. - Package source `3a25c492` as a local ad-hoc test App, audit/sign the bundle and launch it with the existing isolated local-app profile. Packaged main, preload and renderer bytes match the validated build. No release publication or server restart is part of this desktop follow-up. - A separate live Zen model check succeeds, but the reported App authentication rejection has not been reproduced. Computer interaction remains disabled, so actual App acceptance is still pending; this change must not be merged based on component or HTTP checks alone. ## 2026-10-07 — Align OpenCode custom API checks and runtime routes - Resolve Anthropic API prefixes consistently for model discovery, connection checks and OpenCode runtime configuration. Root URLs and explicit `/v1` URLs work without dropping gateway prefixes or changing saved binding identities. Claude's SDK root convention and OpenAI-compatible routing stay unchanged. - Stop unknown/invalid OpenCode completions instead of treating an empty step as success and waiting while the CLI makes repeated requests. Preserve partial timeline text and explicit provider HTTP status diagnostics; do not dump raw gateway bodies or headers. No additional retry or inference probe is added. - Allow a bounded 256-token output budget for explicit Claude/OpenCode model checks. Reasoning models can consume the former 32/64-token budget before answering; empty or reasoning-only responses still cannot verify a connection. Checks remain single-shot and do not alter runtime token limits. - Verification: 198 focused tests pass, including URL parity, credential isolation/revocation, catalog discovery, stream errors, lifecycle and MCP. Regression cases fail on the prior implementation. Lock validation and compilation/package-source checks pass. Existing runner fixtures emit recovered missing-mailbox-table warnings; these are not live-service errors. - Four opt-in tests exercise real OpenCode 1.18.29 through the production process runner and a synthetic loopback gateway: root URL, explicit `/v1`, HTML-200 streaming response and HTTP 401. Disposable home/state and synthetic keys only; no real model inference in these four repeatable smoke tests. Run `tests.test_opencode_custom_api_smoke` with `OPENCODE_TEST_EXECUTABLE` set. - Separate live OpenCode Zen checks pass for Chat Completions, Anthropic Messages and Responses. Real CLI replies also pass for Anthropic root and `/v1` URLs and Responses. These checks do not establish Go subscription or free-tier compatibility; neither was accepted in this validation. - Desktop interaction is blocked by disabled computer access. Full app acceptance remains pending and the PR is held for that check, not merged. This is a local test candidate, not a published release. Per-model protocols and runtime output limits are not included. ## 2026-10-03 — Preserve automatic naming across follow-up messages - Compare effective title runtime settings before and after session updates, rather than cancelling because a provider/model field was included. Normal and queued follow-ups with unchanged settings keep their in-flight title. Missing legacy default-provider fields normalize to the explicit defaults. - Manual renames (including identical text), opt-out, archive and actual runtime/credential-revision changes still cancel. Stale-result checks and the one-attempt usage limit remain; no retroactive chat rewrites or retries. - Verification: 279 focused server tests pass, covering naming, runtime updates, provider bindings and queue behavior. Regression fails on the previous implementation. Desktop typecheck, production compilation and 573 focused client transport/service tests pass (with the tests' expected 022 umask; inherited 077 made an unrelated clipboard fixture private). - Exercised the main-based desktop build (source version 0.2.0/build 85) in an isolated offscreen native Electron profile against the branch's authenticated server. Real Codex replies, a follow-up and queued promotion completed while a test-only timing gate held the independent title request. Releasing the gate produced a summarized title, displayed in the header/list and retained after reload, persisted as generated metadata. Provider requests, HTTP/IPC, authorization and persistence were real, not mocked. Cursor lifecycle coverage is synthetic; its provider adapter is unchanged. - Source only: no existing service restart, deployment or release. ## 2026-10-03 — Prefill the public bug-report form - `/feedback` opens the existing AgentsDock GitHub bug-report form with the running app version/build, client platform/OS version/architecture, current server version, agent and selected model. Environment details use the existing additional-information field so no template rollout is required. Description and screenshot uploads remain user-entered. Remove the separate reproduction-steps and expected-behavior questions to shorten the form; this template simplification takes effect when merged into the default branch. - Read app metadata through trusted native IPC. Capture the originating chat and server context before awaiting it. Mark cached offline server versions and server-default models explicitly; do not infer an unknown model. Do not include session IDs, titles, chat content, paths, server addresses or credentials. Opening the form does not submit an issue or send a chat turn. - Five focused component/URL tests, TypeScript and the desktop production build pass. Tests mock the native browser bridge; actual UI-to-browser acceptance remains pending because computer use is disabled in this run. Built and launched an ad-hoc local macOS package (source version 0.2.0, build 85, based on main 3c5e6e59); bundle audit and signature validation pass. Startup reaches the renderer and connected server; user acceptance is pending. Desktop only; no server update, mobile change or public release. ## 2026-10-02 — Align update test fixtures with current idle-update policy - Final CLI merge validation exposed two pre-existing main-branch CI failures after the idle-update scheduling change: isolated update fixtures omitted queued/run-now state, and a legacy assertion still expected automatic jobs to pause for a pending idle update. Supply empty queue state and assert the current policy allowing both automatic and manual work while waiting for idle. - No production update behavior or safety checks changed. All 298 focused update, preparation, scheduling and hardening tests pass with isolated state and a clean environment. The CLI's 147 focused tests and installed-terminal status colors, token selection and cancellation checks also passed. - No release, registry publication or existing-service restart. ## 2026-10-02 — Navigate the token chooser with arrow keys - Use Up/Down to highlight an existing server and Enter to confirm. Esc, Ctrl+C or Ctrl+D cancels without exposing a token. Keep explicit `token NAME` and the numbered/name fallback for basic terminals without cursor control. - Bound the rendered menu to terminal width/height, scroll longer lists and restore terminal input mode and cursor visibility before token/clipboard output, on cancellation and on read failure. No new package dependency. - Validation: 36 JavaScript, 95 instance-manager, five CLI packaging and 11 runtime packaging tests passed. Actual packaged pseudo-terminal tests cover normal/application arrow sequences, returning to the first item, Esc/Ctrl+D cancellation, only-selected-token output and restored terminal modes. Unit checks also cover wraparound, scrolling and interruption/read failure. Reinstalled the matching local packages on macOS; actual terminal Down/Enter, Down/Up/Enter and Esc flows passed with incumbent listeners/configuration unchanged. No provider, server lifecycle or registry publication changes. ## 2026-10-02 — Choose a server before showing its token - Make bare `agentsdock token` list the current OS user's known local instances with numbered names, native service states and ports. Accept a number or exact name, including stopped instances, and display only that instance's token via the existing private-token/optional-clipboard path. Explicit `token NAME` and `token --instance NAME` remain available for direct or scripted use. - Do not default automatically, even with one server. Empty input, EOF and Ctrl+C cancel without displaying a token; invalid choices retry. Refuse an implicit selection when input/output is redirected. Recheck a selected name before reading its token and never fall back to another instance on failure. - Validation: 36 JavaScript tests, 92 instance-manager tests, five CLI packaging tests and 11 runtime packaging tests passed. Actual packaged CLI terminal tests in disposable homes exercise invalid-then-valid selection, name selection, cancellation and only-selected-token output with synthetic tokens. Reinstalled the committed local npm packages on macOS and exercised the actual chooser by number/name, cancellation, redirected-input refusal and explicit selection; incumbent listeners/configuration stayed unchanged and clipboard copying was declined. No provider, service lifecycle or release changes. ## 2026-10-02 — Make CLI status and setup guidance easier to read - Print one readable `agentsdock status` block per instance with native service status, labeled connection addresses, installed runtime version and port. Accept an optional instance name; keep `info` JSON and single-instance token selection unchanged. Status inspection never starts or changes a service. - When setup/install safely refuses existing default state, explain `agentsdock new` and the explicit name/port form. Preserve the error code, nonzero exit and no-installer behavior; do not change other error paths. - Validation: 36 JavaScript CLI tests, 86 instance-manager tests, five CLI packaging tests and 11 runtime packaging tests passed. Repacked committed source and installed the matching local npm tarballs; exercised actual all-instance and named status, terminal header color, setup/install refusal, unknown-instance errors and unchanged JSON info. Existing service listeners and configurations stayed unchanged. No service restart or npm publication. ## 2026-10-02 — Validate the short npm CLI against current main - Merge current main into the prepared CLI branch. The only conflict was two sets of development-log additions; retain both histories without changing the CLI behavior or the existing runtime lifecycle implementation. - Install the actual matching CLI/runtime tarballs globally on Apple silicon macOS. Initial and repeated npm installation detect existing state and leave incumbent services unchanged. Exercise help/version, listing and aliases, default/named token output, info/status, guarded setup/install, no-journal recovery, and missing-target/update-argument errors. - Exercise two disposable native launchd instances: explicit name/port and automatic name/free port, start/stop/restart, authenticated health and wrong token rejection, interactive uninstall cancellation, successful removal and name release. Confirm that a synthetic saved-state marker survives in the private backup and that incumbent listeners/configuration remain unchanged. Terminal output retains blue names, green success, red removal warnings, separate token lines and the optional clipboard prompt (declined). - Validation: 35 JavaScript, five CLI packaging, 11 runtime packaging and 82 instance-manager tests pass. Package the clean merged source; the source version remains unchanged. Both disposable services are removed afterward. - Remaining acceptance: fresh-default macOS auto setup, real signed update activation, migration repair, bulk destructive operations and clipboard copying were not exercised. The existing split-service binding check still refuses the default installation; no bypass or lifecycle fix was added. No desktop/provider workflow, registry publication or incumbent deployment. ## 2026-09-28 — Automatically set up the first server during global npm install - Change the earlier CLI-only install behavior: a direct global `npm install -g agentsdock` now invokes the bundled fresh installer, waits for its health-checked success receipt, and starts the first default server. Existing default/named installation or state is left unchanged; repeated npm installation never selects an upgrade or restarts an existing server. - Keep local/dependency/link/CI installations inert. Respect `--ignore-scripts` and `AGENTSDOCK_SKIP_SETUP=1`; retain `agentsdock setup` for explicit/custom installation. Native prerequisites remain required. A failed automatic setup returns failure, with a recovery command, instead of reporting a ready server. - Filter installer output before it reaches npm logs so access tokens and raw diagnostics are not captured there. Accept the installer's validated local, LAN or Tailscale address; use `agentsdock token` for private token retrieval. Preserve Linux per-user service environment for native systemd operations. - Validation: 35 JavaScript tests, five facade packaging tests and 11 existing runtime packaging tests passed. Real npm lifecycle tests cover fresh setup, reinstall, skips and failure with a labelled service fixture. A separate disposable Linux account exercised actual global npm installation, starting both managed service processes and passing authenticated health; forced npm reinstallation preserved both process IDs, identity, token and synthetic saved data. The real token was absent from npm logs. The native split-aware uninstaller preserved saved state, and the disposable account was cleaned up. - Native testing also found an existing instance-manager gap: its legacy binding validation rejects split-service removal. Split lifecycle controls remain a separate acceptance gap; no validation bypass was added. macOS npm entry/hook tests pass, but macOS first-service creation, provider chats and app workflows were not exercised. No existing user service was changed. This is local source work, not registry publication or integration into the product release pipeline. ## 2026-09-28 — Flatten public CLI instance commands - Make `agentsdock list`, `info`, `new`, `start`, `stop`, `restart` and `remove` the primary instance interface. Accept `new NAME`, `token NAME`, `status NAME`, `uninstall NAME` and `version`; preserve the earlier `servers`/`instances` commands and explicit name flags. Conflicting positional/flag selectors are rejected rather than silently choosing a different instance. - Document the full old/new public command mapping, bulk selectors and actual differences: fresh-only setup, signed updates, guarded recovery, and explicit uninstall targets. Source-only reconfiguration, bulk manifests, Team Hub and developer/transaction controls are not presented as implemented short commands. - Validation: 27 JavaScript CLI tests and four Python packaging tests passed. Actual npm global/local installations exercise the flat command, version, per-instance token reads and rejection of restart/remove without a target. Synthetic configuration is retained and no service is created or controlled. No existing server restart, provider execution, deployment or npm publication. ## 2026-09-28 — Prepare the short agentsdock npm command - Add the `agentsdock` CLI package, pinning the exact same-version `@agentsdock/server` runtime. Global npm installation exposes `agentsdock`; local installation uses `npx agentsdock`. Neither runs installation hooks, changes shell configuration nor starts or updates a server automatically. - Expose `setup`/`install`, `servers` management, `status`, and per-instance `token` commands. Signed updates and guarded recovery retain the existing implementation. Internal helper commands and the source-installer update bypass are not public CLI commands; existing server signing identities, manifests and scoped-package behavior remain unchanged. - Verify 25 JavaScript CLI tests and 15 Python packaging tests. Real offline npm global/local installs of the actual matching packages exercise the executable from another directory, version/help, empty-instance discovery, recovery with no journal, distinct synthetic default/named tokens and safe reinstall refusal. Package inventory, permissions, exact dependency pins, reproducibility and checksum receipts are checked. Add these checks to source CI; hosted execution has not yet run. JavaScript/Python syntax and workflow YAML checks pass. - Availability: source and local packaging checks only. No service creation, restart, provider chat, desktop UI change, npm publication or deployment is included. Fresh service creation/control through the new wrapper still needs disposable native acceptance. The new public package requires owned registry publication and release-pipeline integration after that acceptance; current public registry commands and frozen release candidates are unchanged. ## 2026-10-02 — Mobile fallback for missing selectable text native view - Diagnose a test iOS build rendering Claude assistant Markdown as red `Unimplemented component: ` boxes. The data and server response were intact; the running native app did not expose the `@bsky.app/react-native-uitextview` Fabric views expected by the JS bundle. - Route Markdown and text-file previews through a local `SelectableText` wrapper. When both `RNUITextView` native view managers are registered, the app keeps the iOS range-selection behavior; otherwise it falls back to React Native `Text` so chat content remains readable in stale or mismatched test shells. - Verified against the local 7850 server on the iPhone simulator by opening the affected Claude test chat and confirming the assistant Markdown renders as normal text with no `RNUITextView`/unimplemented component in the hierarchy. Focused text-selection/Markdown tests and TypeScript passed. ## 2026-10-02 — Mobile photo picker accepts videos - Extend the iOS composer library picker from image-only selection to mixed photos and videos. Preserve Files attachments, add video upload naming and MIME inference, request read access before opening the iOS library, and keep iCloud-backed media download enabled for picker results. - Use the existing upload API with video-aware attachment UI and copy. Videos now render with a video icon in pending, failed and uploaded states. - Verified focused upload/composer regressions, TypeScript, iOS export, native prebuild, Info.plist permission text and iOS native dependency parity. Also ran a disposable local AgentsServer and confirmed a video selected through the mobile upload mapping reaches `/api/sessions/:id/files` as `video/mp4`. Built and launched the Debug app on the iPhone simulator, opened a real chat, verified the attachment menu shows `Photos and Videos`, and opened the multi-select iOS media picker from that option. CoreSimulator media import via `simctl addmedia` was not reliable in this run, so selecting a newly imported video inside the real picker remains unverified. Source only; no release or deployment. ## 2026-10-01 — Simplify the repository overview - Group Claude Code, Codex, Cursor and OpenCode in the product introduction. Keep CLI prerequisites in general setup guidance and provider availability scoped to the client and server version. - Remove the separate OpenCode installation callout, backend-name links, standalone server repository links and website development walkthrough. Keep stable 1.0.9 downloads and the maintained server setup documentation. - Documentation only; no provider, mobile, installer or release changes. ## 2026-10-01 — Refresh the README for stable 1.0.9 - Point the repository's desktop badge, downloads and release notes to the published 1.0.9 stable release. Describe OpenCode's desktop availability and Codex Side chat controls, keeping mobile availability separate. - Add the stable npm command for fresh server installations, My Agents setup guidance and the managed update path for existing servers, including the signed legacy bridge. Link backend documentation to the maintained source. - Update the linked desktop OpenCode guide and preserve its previous anchor. Verify the four desktop download URLs and relative documentation links and anchors. Documentation only; no application, service or release changes. ## 2026-09-30 — Desktop workflow usage events - Add 33 aggregate desktop events for side chats, custom API settings, goals, workspace saves and Git actions, uploads, queue/steering, built-in slash selection, and explicit app/server update actions. Include keyboard entry points in the existing Open file metric. The complete catalog and outcome definitions are in `docs/ANALYTICS_EVENTS.md`. - Keep the existing anonymous envelope and `success`-only property allowlist. Do not send content, names, paths, model/provider details, credentials or resource IDs. Team Network instrumentation is unchanged; shared-browser and iOS analytics remain disabled. Update requests are not installation success, and chat-reference submission is not proof of peer delivery. - Verified with an isolated native macOS Electron app, production IPC/HTTP and a disposable AgentsServer: send and queue two real Codex turns; open, send a follow-up and stop a side chat; save/clear a Codex goal; open a file by button and shortcut, edit/save it, stage it and commit the reviewed disposable index; open My Agents and fail a connection check against an unavailable test endpoint; check desktop/server updates and change the desktop channel. Captured analytics requests locally and blocked delivery to Mixpanel; only the documented property keys were present. No production app/server restart. - Validation: 951 focused tests, TypeScript and production compilation passed. Live acceptance is partial for upload, steering, Claude goals, goal pause/resume, credential persistence, Git conflict/abort and actual update install/cancel/retry paths; these were not exercised in this run. No claim of release readiness or completed installation is made. Availability: source and isolated local build only; no release or deployment. ## 2026-09-29 — Preserve Claude history activity access - Keep the activity disclosure when a compact Claude history page uses the final commentary as its only trace anchor. Deduplicate the visible reply while retaining the run and sequence needed to load its activity on demand. - Claude text blocks are projected as assistant output in the server source; the desktop also surfaces legacy report-shaped commentary when those events are present, after ownership filtering and final-answer deduplication. These source changes do not imply a server deployment. - Verified in the actual local macOS package against the connected server: reopened an affected historical chat, expanded the restored activity control, loaded its report and tool details, then collapsed it. Focused timeline/UI regressions (324), 88 Claude SDK/print-runner tests, TypeScript and production compilation passed. - Acceptance remains partial: the running server's compact history can omit legacy report text until expansion. Default-visible recovery through that paging boundary and a new native Claude turn remain unverified. No claim of full native-provider parity or release readiness is made. ## 2026-09-28 — Combine editor line endings with desktop/server fixes - Bring contributor PR #24 into the tested desktop/server follow-up branch, retaining Haokai Ding's authored commit and original commit reference. The runtime/editor test patch is unchanged; resolve only the development log by retaining both histories. The provider/API feature baseline is already in main through PR #47; this integration does not reintroduce it as an unmerged change. - On the combined source, 339 editor/workspace/transport tests and 292 isolated server workspace/history/runtime tests passed, plus TypeScript and production compilation. Actual offscreen Electron exercised CRLF, LF and CR edit/save, undo/save and redo/save against a disposable authenticated AgentsServer: nine exact on-disk byte assertions passed. No clipboard access was used in this integration check. Native Windows UI was not exercised. - Installed test apps and existing server instances are unchanged by this PR preparation. The contributor's fix is newly included in source, not in the previously installed local test package. No release or deployment is implied. ## 2026-09-28 — Manual server connection saves on macOS - Direct macOS builds now use the bounded Keychain credential path for manual server creation and token replacement, not only automatic local discovery. Saving no longer enters synchronous Electron OSCrypt. Failed writes keep the old connection intact, report a recoverable error and never store plaintext. - Rollback does not decrypt legacy ciphertext unnecessarily. Printable tokens containing command-parser punctuation use Keychain's hexadecimal stdin input; secrets remain absent from process arguments. MAS and other-platform secure storage paths remain unchanged. Existing custom server labels are retained. - Validation: 473 focused desktop tests and a full suite of 5,005 tests passed (5 skipped). After extending printable-token transport, all 85 focused settings/Keychain/Team Hub tests passed again. TypeScript, production build and eight package/license checks passed. - Real offscreen Electron UI exercised manual Test connection, Add & switch, token replacement, custom rename and process reopening against two isolated authenticated AgentsServer instances, using the actual macOS Keychain, not a credential mock. Reopened connections and both custom names survived. Failure rollback and secure-store denial are regression-tested; the user's remote machine/network and a physically locked Keychain were not exercised. - Availability: local test build; no server restart, publication or deployment. ## 2026-09-27 — Mailbox input ownership and scoped macOS file discovery - Integrate the existing native-continuation ownership fix: retain the initial Codex input turn ID separately from its later continuation. Older history requires exact, unambiguous assistant-item and completed-run evidence before classifying a replay as internal. Genuine user messages with identical words remain visible; original provider transcripts and audit events are not edited. - Recursive workspace search from Home or an ancestor no longer enters the macOS user's private/system-managed home trees. Skip these paths before opening or inspecting them, including empty queries; bypass Git's recursive untracked-file enumeration for these broad scopes. Explicit project scopes, direct opens, normal same-named project folders and OS permission errors retain their existing behavior. This is not a new sandbox or a permission grant. - Validation: server batches of 290 and 131 focused tests passed (overlapping); desktop batches of 140 and 60 passed (overlapping), plus TypeScript checking and production compilation. New privacy regressions cover broad and explicit scopes, Git fallback, empty/missing searches, aliases, symlinks and denied access. - Isolated offscreen production Electron reproduced both discovery of synthetic app-container files and a duplicate internal input with the old history repair. Through real IPC, authenticated HTTP, WebSocket catch-up and local cache, verify corrected history, older-page loading, refresh and app reopening. One genuine same-text user input and the original replies remain. Opening the file palette, searching, switching workspace and explicitly opening a synthetic private path exercise the new search boundary. Fixture transcript/ledger hashes are unchanged. - Limits: no real private app data or macOS TCC settings were accessed; the reported permission dialog and remote customer's exact ledger were not tested. Provider records are synthetic; continuation execution is covered by controlled native-router regressions, not a new live-model request. No mobile, packaged release, existing service restart or deployment is claimed. A new server candidate must include these changes; frozen release artifacts are unchanged. ## 2026-09-27 — Readable default server names - Unnamed desktop server connections use their URL host and port rather than the opaque server identity. Remember whether a label is automatic or custom; automatic labels follow URL edits, while explicit names remain unchanged. - Normalize exact legacy identity/hostname defaults on settings load without changing profile IDs, server identities, credentials or cached chat ownership. - Validation: 64 focused settings/selector/management tests, TypeScript and production compilation passed. Two full-suite attempts each passed 4,997 tests (5 skipped) but encountered different unrelated UI timing failures; the WorkspaceEditor and WorkingDirectoryPopover files passed all 136 tests on isolated retries. Full-suite green is not claimed. Eight package/license checks also passed. - Actual isolated offscreen Electron exercised legacy-name migration, adding an unnamed server through authenticated native HTTP, manual rename, app restart and switching servers. Dark/light and narrow layout were inspected. No live app replacement, existing service restart, mobile change or release. ## 2026-09-27 — My Agents and clearer chat deletion copy - Rename the desktop settings section to **My Agents** / **我的 Agent**, including navigation, onboarding guidance, accessible card labels and server recovery instructions. Fill Chinese endpoint/model/key wording without translating product names, CLI commands or API protocol identifiers. - Label chat deletion **Delete from AgentsDock** and explain that history in the original agent is retained. Sidebar, header and confirmation use the same wording; deletion behavior is unchanged. Correct the obsolete Codex Forget success message to describe blocked sends until reconnection. - Validation: 4,991 desktop tests passed, 5 skipped, plus 8 package checks; 41 isolated server tests, TypeScript and production compilation passed. The focused bilingual regression set contains 167 passing tests (included in the desktop total). Audited 144 connection/account/model locale entries. - Actual offscreen production Electron with disposable profile/server state exercised English and Chinese settings navigation, Claude/Cursor/OpenCode API forms, authentication-error copy, model discovery against a synthetic loopback endpoint, Forget confirmation/cancellation, and chat deletion cancellation. Production IPC/native HTTP/storage ran intact; no real provider inference or native-history deletion acceptance was repeated for this copy-only change. Light/dark and narrow layout were inspected. Chinese startup was seeded in the disposable language preference; native language-popup interaction was not certified by this offscreen run. - Availability: local source/build only. No live app replacement, existing server restart, mobile change, publication or deployment. ## 2026-09-27 — Desktop line-ending integration accepted locally - Integrated PR #24 from contributor commit `3f29f4d9` as `a4f6b281`, retaining the original author. Only the development-log conflict required resolution. - Reproduced CRLF-to-LF rewriting on the preceding main build. The candidate preserves CRLF, LF and CR through native edit/save, undo and redo; mixed files normalize to the first separator. Native clipboard paste, external reloads, restored YAML folds/cursor positions, UTF-8/CRLF byte-limit boundaries and stale-save rejection/reload recovery passed with byte-level file checks. - Acceptance used isolated offscreen Electron 43.1.1, the production app build and AgentsServer `1.0.7-beta.11`, through real IPC, authenticated HTTP and revision-checked filesystem writes. Only disposable credential storage and fixture setup were synthetic; no provider inference was needed or claimed. - Validation: 339 focused desktop tests and 53 server workspace-file tests, TypeScript checking and production compilation passed. Availability: local source/build only; no package installation, release, deployment or merge. ## 2026-09-17 — Preserve desktop editor line endings - Preserve a file's first line-ending style when emitting edits, including undo and redo. External reloads adopt the newly loaded style; mixed line endings normalize to the first separator on edit. - Count UTF-8 limits against serialized output, including both bytes of CRLF. Keep cursor, fold and replacement ranges in CodeMirror's logical coordinates. - Add regressions for line-ending preservation, external reloads, LF inserts in CRLF documents, byte limits, cursor clamping and fold restoration. ## 2026-09-27 — Custom API live acceptance and API-only chat routing - Cursor and OpenCode advertise cross-chat transport when their CLI is installed and this server has a verified custom API connection, even without native account login. This does not mark native login ready or bypass per-chat admission, credential revocation, CLI compatibility or route authorization. Capability reads use local metadata, not additional provider requests. - Live acceptance used disposable chats and workspaces with real provider requests. Cursor used its official API-key service; Claude Code, Codex and OpenCode used a verified OpenRouter connection. All four produced a response. OpenCode ↔ Cursor and Claude Code ↔ Codex delivered one request and one reply each through the run-bound tool and mailbox, verified against server events. The native desktop rendered replies and recovered a successful Claude tool attempt after an initial model attempt used an unavailable shell command. - A nonexistent-route attempt was rejected with no extra delivery. This is not a substitute for the isolated cross-chat ownership/authorization regressions. Settings rejected unverified credentials without saving them or falling back to native accounts. Direct Anthropic/OpenAI key acceptance remains unverified; gateway success does not certify every endpoint/model combination. - Validation: 373 provider/server tests, 1,126 desktop tests, then 175 targeted routing/binding tests passed (server batches overlap). Native UI checks used a separate desktop profile and the designated test server. No release is published by this acceptance record. - Full CI exposed summary-size and OpenAPI operation-ID regressions, an outdated lifecycle-lock field assertion, and two Chinese terminology inconsistencies. Keep legacy native connection defaults sparse while retaining explicit cache tombstones; keep model inventories out of per-session summaries; register method-specific API routes. The focused follow-up passed 108 server tests and 56 desktop/localization tests without relaxing the original guards. ## 2026-09-27 — Reconnect existing API chats and check on send - Existing custom API chats can resume after explicitly verifying the same original URL and API key again. Forget still blocks new turns immediately; a different endpoint or key cannot silently take over an old chat. Applies to Codex, Claude Code, Cursor and OpenCode, including legacy bindings. - The composer no longer blocks Send based on cached login, model or connection readiness. Server capability and ordinary submission safeguards remain. Actual rejected sends preserve the draft and show Check failed with the Configure API action. Retrying or reconnecting clears the matching old error; equivalent catalog refreshes do not erase a real failure notice. - Verification: 108 isolated server tests and 502 focused desktop tests passed; TypeScript checking and production compilation passed. Isolated native Electron exercised Forget, a rejected send, draft retention, Configure API, reconnect and successful resubmission in the same chat through IPC, native HTTP authorization, production connection storage and send admission. The provider adapter was synthetic: live paid-provider inference remains a separate acceptance check. No real user's chat received a test message. - Availability: local feature branch; authorized local app/server installation tracked separately. No release publication or merge is implied. - Local acceptance: source `a4042410`, desktop `1.0.7-providers.15.local`, server `1.0.7-beta.15+providers.a4042410`. Bundle audit, ad-hoc signing and packaged-source comparison passed; the installed app loaded with its active profile preserved. The designated server retained its identity and provider settings. Read-only verification confirmed the reconnected existing chat's API binding is available, without sending a test message to that chat. ## 2026-09-27 — Forget revokes existing custom API chats - Forget now durably revokes this server's existing custom API bindings for Codex, Claude Code, Cursor and OpenCode. New turns and model checks fail closed without native-login fallback; history and other instances remain untouched. Saving again creates a new connection, not a silent reassignment of old chats. Already-dispatched requests are not forcibly terminated. - Migrate earlier empty Forget records on read. Atomic revision/epoch markers survive restart and later saves. Keep historical key material private for in-flight redaction; revocation is not secure deletion of credential files. - Desktop immediately invalidates associated cached chat readiness and explains the effect in confirmation. CLI Login has no logout/disconnect action. - Verification: 107 isolated server regressions and 443 focused desktop tests passed. Typecheck and production compilation passed. Actual isolated native Electron exercised cancellation, confirmed Forget, recovery navigation and reconnect through IPC, native HTTP authorization and production storage; the production admission function rejected the revoked chat with HTTP 409. Its Cursor success adapter was synthetic, not live inference. A separate installed-CLI loopback probe verified Cursor's dedicated key-exchange protocol and rejection of standalone local-agent flags; arbitrary model URLs remain unsupported. No real user's chat was sent a test message. - Availability: local feature branch; local app/server installation tracked separately. No release publication or merge is implied. - Local acceptance: source `c3784f6d`, desktop `1.0.7-providers.14.local`, server `1.0.7-beta.15+providers.c3784f6d`. Packaged-source comparison, bundle audit and ad-hoc signature checks passed. A read-only installed-server check confirmed a previously forgotten legacy chat remains readable but its API connection is unavailable. Only the authorized test instance was restarted; connection settings were preserved. ## 2026-09-27 — Cursor API-key connection - Add a key-only Cursor API card in AI Providers, separate from CLI Login. Connect validates through the native Cursor model-list command before storing the key; new chats offer `Cursor API key` only while configured and verified. Models belong to that key, with `auto` as the initial default. This is not an arbitrary OpenAI-compatible endpoint. - Store credentials privately per AgentsServer with revision fencing and immutable per-chat bindings. Forget removes the current connection for new chats, not keys already bound to existing chats; no machine logout occurs. Native transport remains administrator-only, keys are never returned, and failed replacement preserves the previously verified key. - Cursor can persist API-exchanged tokens by default. Key-bound runs, checks, and title requests therefore use its in-memory credential store, clear higher-priority auth tokens, force the official service, and disable direnv credential overrides. Require CLI 2026.09.26 or newer (the tested isolation baseline). API-key turns do not change the native login diagnostic. - Verification: 139 focused server tests passed, including a real subprocess runner fixture asserting the pinned key/environment and no key in argv or events. Native Electron → IPC → administrator HTTP → production router/store → rendered status exercised Connect, invalid key, and confirmed Forget with a synthetic Cursor adapter. The actual installed Cursor CLI rejected an invalid key with no native-login fallback. A valid user key and successful Cursor inference are still required for full live acceptance. - Full desktop run: 4,971 passed, 3 failed, 5 skipped. The three failures were in unrelated sharing/editor tests; rerunning those suites together with the changed provider suites passed all 212 tests. Typecheck and native build are checked separately. No claim of a completely green full-suite run. - Final focused reruns passed: 83 server tests and 53 desktop provider tests. Local ad-hoc desktop build `1.0.7-providers.13.local` passed the bundle audit, signature verification, and byte-for-byte comparison of 88 compiled files. Its actual settings page exposes the enabled key-only form. The selected development server was restarted and its authenticated health and Cursor credential route verified; other server instances were not restarted. ### Cursor API-key usage Settings → AI Providers → Cursor → Configure API → paste a Cursor-issued key → Connect. Create a new chat with **Cursor API key**. CLI Login remains a separate choice and identity. The key comes from Cursor Dashboard → API Keys; see the [Cursor authentication guide](https://cursor.com/docs/cli/reference/authentication). Forget is in the API card's three-dot menu and requires confirmation. ## 2026-09-27 — Provider onboarding and safe empty-chat cleanup - Default new Claude API configurations to the official Anthropic endpoint and authentication header; preserve previously configured third-party endpoints. OpenCode no longer defaults to OpenRouter. Link its official Zen documentation and explain protocol selection. Distinguish Cursor-issued API keys from arbitrary third-party model endpoints; settings-based Cursor key storage is not implemented by this change. - Reconcile endpoint Save/Forget immediately and fence older catalog responses. Offer only connected Agent choices, including the direct-new-chat shortcut. Add Agent opens AI Providers; custom-API error recovery opens the matching provider section without rechecking native CLI authentication. - Mark new blank placeholders from both creation paths. On startup or explicit list refresh, preserve drafts, attachments, references, jobs and terminals; discard only proven unused placeholders through a separate conditional server endpoint. Unsupported servers preserve the chat instead of falling back to ordinary Delete. The server checks again under lifecycle/deletion fences and preserves malformed or nonempty history. Native provider history is untouched. - Desktop regression: 4,969 passed, 5 skipped. Server focused regression: 97 passed. Native Electron against a disposable authenticated endpoint fixture verifies official Claude defaults, disconnected choices hidden, Add Agent navigation, successful API setup, and confirmed Forget immediately removing the custom choice. No real account keys were changed during fixture testing. Verify the API recovery button in the native composer opens the matching settings group. Live development-server cleanup removes only disposable empty test chats and retains an unsent draft. This check uncovered the editor's deferred draft persistence; flush live editor refs and temporarily block editing/admission during discard. Re-run 843 affected desktop tests after that fix; all pass. Verify immediate type → refresh preserves the draft and clear → refresh discards the same disposable placeholder, without a debounce wait. The signed local desktop package passes bundle audit; all 88 compiled files match the tested build. Verify AI Providers in the installed app, preserve the active server profile, and leave normal launch/debugging disabled. Official Cursor API-key and OpenCode Zen/Go model calls are not live-tested; their documentation does not imply arbitrary endpoint/model compatibility. ## 2026-09-27 — Endpoint confirmation spacing - Use a compact, centered confirmation with consistent 24px insets, 16px content spacing and an 8px button gap. Wrap the actions on narrow screens; preserve the existing confirmation, cancellation and endpoint-only deletion semantics. - Reproduce zero body padding in the native app, then verify real menu selection, centered geometry, dark/light and narrow layouts, keyboard selection, Escape, Cancel and repeated opening. Recheck the signed packaged code with an isolated profile, then leave the installed app's active server selection unchanged. No live endpoint is forgotten during this appearance-only check; localized visual acceptance remains pending. - All 60 focused tests, type checking, compilation and 8 build/license checks pass. The full desktop suite reports 4,963 passes, 5 skips and 2 failures in unchanged Chinese terminology/OpenCode catalog assertions; retain these for separate follow-up. No server update, publication or merge is included. ## 2026-09-27 — Visible endpoint menu inside Settings - Fix the custom endpoint menu painting behind the Settings dialog. Give this nested menu and its confirmation overlay/dialog explicit, narrowly scoped layers above Settings; CLI Login and endpoint removal semantics are unchanged. - Reproduce the hidden menu in the real desktop before fixing it. Verify visible menu/confirmation, repeated opening, Cancel/Escape and keyboard selection in an isolated native app connected to the development server. Check dark and light/narrow presentation, then repeat menu/confirmation/cancel in the signed installed local package. No real credentials are removed; confirmed deletion transport is unchanged and covered by the preceding disposable-fixture check. - The layering regression fails before the CSS fix; all 59 focused checks, TypeScript and production compilation pass. Local package only; no server update or public release. Localized visual acceptance is not repeated. ## 2026-09-27 — OpenCode chat tools and concise provider settings - Add a run-bound native MCP bridge for OpenCode, preserving server authorization, exact process ownership, permission boundaries and private endpoint settings. Enable OpenCode source/target choices only on supporting servers. - Custom API cards offer a confirmed **⋯ → Forget endpoint** action. CLI Login remains read-only with no Disconnect/logout action. Keep default-model selection visible and save a selection automatically; remove routine reload/manual-toggle, save-default and saved-key-check controls from connected cards. - Verification: 244 isolated server checks and 606 desktop checks pass, with TypeScript, production compilation and signed local-bundle validation. Two real OpenCode A → B → A mailbox round trips complete with nonduplicate receipts and preserved delegation; the second starts through desktop mention selection and Send, including IPC/native HTTP and advertised client capabilities. - OpenCode 1.18.29 executes the run-bound tool with default permissions. Its Plan agent also executes the bridge using a tool-capable custom API model; the free model's Plan request returned a provider-side 403. This does not certify every model or endpoint. Live stopped/busy-turn acceptance remains pending; ownership, cancellation and transport fences have isolated regression coverage. - Actual desktop checks cover connected-card presentation/catalog loading and, against a disposable native-admin API fixture, Forget cancellation and confirmed removal without changing any real account. Default-model autosave has component coverage; native select-menu interaction and light/localized visual acceptance remain unverified. Update only the designated local development server/app; no public release or production deployment is included. ## 2026-09-27 — Independent custom API model catalogs - Keep native CLI model catalogs separate from custom API inventories. Discover endpoint models with bounded pagination, readable labels and conservative non-chat/tool filtering; use OpenRouter's user-filtered inventory. - Add a default-model dropdown and manual-ID fallback to connected API cards. Save only the model/revision, without resubmitting keys or claiming a new authentication check. Never automatically select the first model. - Preserve existing chat bindings and legacy inferred choices for started Codex threads. Fence discovery and saves to the selected server/revision, and expose explicit catalog refresh for each supported custom provider in chat selectors. - Verification: 76 isolated server checks and 96 focused desktop checks pass, with TypeScript, compilation and signed local-bundle validation. The real desktop reads all three endpoint inventories through IPC/native administration; manual default save, reopen and clear succeed without a key prompt. Browser administration is rejected. Dark/narrow layout is checked. Native select-menu keyboard selection, light/localized visual checks and paid model/tool requests remain unverified. The temporary empty UI-test chat is removed; no real provider history is touched. No public release is part of this validation. ## 2026-09-27 — CLI login account cards - Match CLI Login and Custom API card styling in all provider groups. Show available CLI email/plan separately from custom API billing; omit unknown fields. Claude metadata is explicitly a saved profile, not a login probe. - Add a read-only, native-administration account route and profile-fenced desktop transport. Opening a card never renews credentials or runs model inference. - Model discovery/default selection is unchanged; catalog/compatibility UX improvements remain a separate research recommendation. - Verification: 55 focused desktop checks and 17 isolated server checks pass, along with TypeScript, production compilation and local bundle validation. The installed development app exercises expansion/reopening through IPC and native HTTP into the real administration guard: Claude and Cursor return email/plan only in their CLI cards. Dark/narrow layout is checked. Light-theme and full localization visual acceptance remain pending. No inference, login renewal, credential changes or public release is part of this check. ## 2026-09-27 — Folded provider settings and managed local discovery - Present four expandable provider groups. Separate native credential presence from explicit API connection checks, with brief CLI guidance only when absent. Chat choices show connected native and custom methods separately, native first. - Wire Claude and OpenCode custom connections into explicit per-chat runtime bindings. Pin private credentials, preserve native login/settings, reject cross-login resume, inherit bindings on forks and fail closed without a model. Cursor generic third-party endpoints remain unsupported, not falsely verified. - Discover owned managed local installations once at desktop startup; check listener ownership before authentication and verify durable identity. Preserve saved profiles and current selection; never scan LAN addresses or start services. - Keep macOS automatic credential writes in the existing bounded Keychain helper, avoiding concurrent startup OSCrypt/Keychain blocking. Fail closed if secure storage is unavailable; normal explicit credential editing remains unchanged. - Focused synthetic checks cover credential isolation, transport, UI and discovery. TypeScript and production build are checked. Real third-party model/tool compatibility remains user acceptance work; no broad regression or release. ## 2026-09-27 — Visible provider connection state and login guidance - Put API configuration near the top of AI Providers and use the same Configure API action for Codex, Claude Code and OpenCode. Unconnected, unknown, draft and failed states have neutral icons; only verified saved API connections get a green check. A failed settings read is not signed-out or connected evidence. Native account identity remains separate. - Show native sign-in guidance with fixed, copyable commands for the selected server, rather than executing login on a possibly unrelated desktop. Keep installed-but-unconfirmed runtimes neutral. Claude still confirms auth on actual sends; opening settings does not renew its credentials. Cursor's unsupported generic API configuration remains explicitly labeled. - Exercise real Electron mouse/keyboard input through production IPC/native HTTP, native-admin authorization and endpoint storage against an isolated loopback API. Verify all three connections, bad-key retry, key clearing, navigation/reopen, login instructions and light/dark narrow presentation. Fixture credentials do not certify third-party accounts or native login. - TypeScript and production compilation pass. Related settings/CSS tests pass; the full suite found one outdated translated-label assertion and one unrelated workspace timeout. Correct the label, and pass focused settings tests and all 134 workspace tests on rerun. - Update the explicitly designated development server with this feature branch, preserving identity and saved credentials; verify new settings routes and credential-check support. No public release or main-branch merge. ## 2026-09-27 — Clear custom API identity and connection state - Separate the native Codex account from custom API credentials and billing. Hide unconfigured custom choices in new-chat/composer menus without changing existing custom chats. Never query native account quota for a custom chat. - Reduce forms to URL, key and Connect; fold protocol/model controls into Advanced. Default checks use protected read-only APIs, not model inference. Public catalogs cannot verify keys; known OpenRouter connections use its private key-status API. Claude/OpenCode model probes remain explicit opt-ins. - Codex Connect saves only after server-side credential verification. Bind the saved proof to the immutable credential revision; legacy saves remain compatible but unverified. Preserve prior credentials on a failed replacement and clear entered keys after completion. Keep native login unchanged. - Verify real Electron mouse/keyboard → production IPC/native HTTP → actual native-admin guard, router and storage → loopback HTTP checks. All three supported settings cards verify with synthetic keys; bad-key retry, reopen, saved recheck, separate account identity and dark/light narrow layouts pass. No third-party credentials, paid requests or native agent sends are exercised. - Pass 89 related server tests. The desktop suite passes 4,944 tests (five skipped) with one missing CSS token; fix the token and pass its three-test suite. TypeScript, production build, bundle audit and ad-hoc signature pass; all 88 compiled files match the packaged application. - Replace and reopen the local development App as `1.0.7-providers.2.local`, retaining its profile and previous bundle. Confirm the existing server's saved Codex endpoint is unverified, not tied to the native account. That server lacks the new verification/settings routes: deployment remains pending operator approval. Claude/OpenCode are settings-only; Cursor has no generic API form. Availability: local App/source only; no public release. ## 2026-09-27 — Settings-only Claude Code and OpenCode API connections - Extend desktop AI Providers with separate Claude Code and OpenCode endpoint profiles. Explicitly verify a small model request before saving, show a timestamped API-check result, recheck saved credentials, and confirm removal. Keep native login and chat routing unchanged. Add OpenRouter setup guidance and a Cursor notice explaining why a generic model gateway is not a Cursor CLI service endpoint. - Add native-only, revision-fenced server routes with private atomic storage, bounded request/response sizes, timeouts, no redirects or implicit proxy credentials, and credential-free responses. Preserve old credentials on a failed replacement; reject browser-origin and stale-server requests. - Exercise the production desktop build in an isolated native profile through renderer mouse/keyboard input, IPC, native HTTP transport, the actual server authorization functions, new router/storage and real outbound HTTP. Verify both forms, a failed-key retry, saved recheck, removal, cancel, navigation and reopen. Inspect light/dark presentation and narrow layout. The model API, health/session data and initial server profile are synthetic fixtures; no user credentials, native provider login or live chat was changed. - Pass 4,940 desktop tests (five skipped), then targeted profile-isolation coverage; pass 73 related server tests and nine packaging-manifest checks. TypeScript and production compilation pass. Chinese copy, stale responses and disconnected/old-server behavior are covered by component/transport tests. Actual third-party API and agent-tool compatibility remain unverified until a user supplies a test endpoint/key. Availability: local source and development build only; compatible client and server updates are required. ## 2026-09-27 — Dedicated desktop AI Providers settings - Add an AI Providers section with the selected server's existing runtime diagnostics, Codex account information and Custom endpoint controls. Move those controls out of Server settings and retarget endpoint configuration links. Keep authentication, credential storage, transport and server behavior unchanged; do not introduce another provider's endpoint implementation. - Validate the local `1.0.7-providers.local` desktop package, based on main `f50595b4`, through native renderer mouse/keyboard input and the production IPC/HTTP connection. Open and reopen the page, read the actual account and endpoint configuration, enter an unsaved synthetic key, navigate away, and verify it is cleared. Follow the composer configuration link into the new section. Remove the empty disposable chat used for that navigation check. - Inspect dark and light presentation, a narrow viewport, scrolling and keyboard focus. Chinese copy, disconnected controls and stale server responses are covered by component tests, not live multi-server testing. Do not save credentials, probe an endpoint or send provider messages during these UI-only checks; those unchanged workflows are not re-certified here. - Pass 4,932 desktop tests (five skipped), TypeScript, production compilation, package audit and local signature verification. Preserve the existing app profile. Availability: local branch and ad-hoc package only; no server restart, release or publication. ## 2026-09-27 — Skip native questions in no-prompt modes - Skip Codex user-input questions under `never` and Claude SDK questions under `dontAsk` or `bypassPermissions` through the existing empty-answer/skip responses. Do not create waiting interaction cards, invent answers or grant additional tool permissions. - Keep ownership, Stop and deletion checks. Use the current turn's captured policy, including independent Codex side conversations; preserve questions and answer handling in normal prompting modes. - Pass 290 scoped server tests. A broader 299-test run has one existing idle-thread runtime fixture failure, reproduced against unchanged main. New skip regressions fail against unchanged main and pass with this fix. - Integrate with the AI Providers settings work on one development branch. Pass 360 combined isolated server regressions covering question handling, runner policy capture, side conversations, API settings and packaging. - Availability: source-only change; no deployment or release. Native provider and real App acceptance remain pending. See `server/docs/NO_PROMPT_USER_INPUT.md` for the local acceptance checklist. ## 2026-09-27 — Run-bound Cursor chat tools under native permissions - Replace Shell-based helper instructions for Cursor with a private per-run MCP plugin. Preserve native CLI session IDs, login, working directories, stream output, and history. Keep explicit denies and other tool permissions. - Reuse server capability checks and idempotent tool execution. Stop and exit revoke the endpoint; temporary permission/configuration files are removed. Clarify that accepted delivery does not imply reading or replying and must not trigger another send. - Verify real Cursor Default-mode MCP calls and same-ID continuation, native Shell rejection, explicit MCP denial, and unchanged global configuration. These native probes use a synthetic inbox. Server integration tests cover the real IPC and live-run fence with a fixture provider process. - Pass 256 focused Cursor, authorization and mailbox tests, plus targeted packaging/configuration checks and Python/shell compilation. Additional lifecycle coverage rejects non-regular configuration files without hanging. - Activate the patched local beta candidate through the normal authenticated restart path. Preserve server identity and existing chat/native session associations; leave other running server instances untouched. The user subsequently accepted local App testing and requested source integration. Individual manual checklist results were not separately recorded; this does not replace independent full App round-trip or release acceptance evidence. - Availability: locally tested server fix submitted for main integration; no public package or release publication. See `server/docs/CURSOR_PROVIDER_MCP.md` for the contract and limitations. ## 2026-09-27 — Keep Claude model discovery passive - Remove disposable authenticated model-discovery processes and their forced teardown. Reuse bounded model metadata from real SDK initialization, with configuration-scoped expiry and invalidation after native auth failures. - Preserve passive readiness, native alias labels, explicit empty pickers and fallback discovery. Do not promote project-specific model settings into the global catalog or retain private initialization/account fields. - Pass 272 focused server tests covering catalog behavior, readiness, SDK lifecycle and runner integration; compile changed runtime modules and check the diff. Regression tests use synthetic credentials and isolated state. - Exercise the beta candidate's compiled app in an isolated native offscreen Electron window through real IPC/HTTP and the native Claude SDK. Four real requests pass, covering server restart/resume, selecting Haiku, repeated CLI rechecks and app close/reopen with retained native session/context. The picker displays twelve sanitized native options; credentials stay unchanged. - A before/after regression observes one disposable metadata-process attempt in the old path and none in the corrected path. Actual app testing also caught and fixed cache invalidation caused by Claude's startup counters. - Availability: locally tested source submitted for main integration and an isolated patched beta server; no public release. Natural OAuth renewal and coexistence with older servers sharing native login remain unverified; immediate successful requests do not establish that repeated-login incidents are resolved. ## 2026-09-26 — Reauthorize fresh file selections across chats - Let a fresh native drop, paste, or file-picker selection grant an idle file to another chat. Keep selection batches atomic, reject replay of old gestures, and prevent delayed upload requests from consuming a newer selection. Active uploads and managed attachment grants remain protected. - Preserve the shared browser's existing upload bridge when the desktop uses native batch staging. Its real bridge regression caught an incompatible optional-method probe during integration; the corrected tests pass. - Validate source `b3a3fad1` with TypeScript, 4,927 desktop tests (five skipped), eight packaging/license tests, production compilation, bundle audit and local ad-hoc signature verification. Use a native temporary filesystem for tests requiring POSIX permissions and for signing the local package. - Launch an isolated native desktop build `0.2.0` / `85` against an existing test server `1.0.7-beta.11`; application startup and authenticated health, session and job refreshes succeed. The local candidate was subsequently manually tested and accepted. Computer-use permissions prevented independent agent-operated drag/drop and paste verification; individual manual test-case results were not recorded. - Availability: source and manually accepted local test package. No server update is required for this fix; no server restart or release publication. ## 2026-09-26 — Restore ordinary steering during active Codex goals - Send goal follow-ups to the existing native turn even when the model/effort picker has changed for future turns. Preserve the running goal, provider settings, original references, command and authority without Stop/restart. - Remove the obsolete client-capability requirement for plain queued input. Keep the goal steering lane available for turns started from a skill. Report unsupported new actions separately from a turn that is not ready. - Reproduce the reported rejection through signed desktop `1.0.7-beta.14`, build `1216`: start a real Codex goal at Low effort, change the picker to Medium, queue a plain message, and click Send now. The original server responds with the reported 409; corrected server source `810d89f` accepts it. - Exercise two follow-ups on the original running goal, then explicitly Stop, Resume goal, change effort again, and steer the completion marker. All three requests receive 200, native acknowledgements appear once, the first two retain the original run, and the third retains the resumed run. The goal completes with no queued messages or renderer exceptions. This uses native app input, production IPC/HTTP and the actual provider against isolated state. - Pass 211 targeted server regressions covering admission, delivery, original command/reference ownership, legacy queue records, transport races, queue recovery and goal resume; eight focused app goal tests also pass. Regression provider fakes are supplemented by the separate live acceptance above. - Server-source fix only. Existing desktop builds can use it after server deployment; no production server deployment or public release in this pass. ## 2026-09-25 — Keep passive Claude login status out of the composer - Do not show an upfront Claude authentication warning or Recheck button just because login has not been checked, or another chat has a cached login failure. Remove the composer's stale-auth send gate so retrying after external login reaches the native Claude request instead of failing solely on cached status. - Keep actual errors from the selected chat's latest run visible, clear the notice after a successful retry, and preserve missing/broken CLI guidance. Full runtime diagnostics and manual rechecks remain available in Settings. - Regression coverage reproduces the original unknown-state banner, verifies send admission without an auth probe, and covers failures, retry recovery, stale readiness and Settings. Pass 4,912 desktop tests (five skipped), eight packaging/license checks and 30 isolated server authentication/probe tests. Type checking and production compilation pass. The new send regression first exposed the stale-auth gate and passes after the correction. - Availability: desktop source and compiled output only; no installation or server restart. Live desktop interaction remains unverified because computer accessibility permissions are unavailable. This UI correction requires an updated desktop app; restarting the server alone does not change it. ## 2026-09-25 — Hand off Codex work after a native re-login - Detect conservative native file-login revisions on demand and let Recheck CLIs explicitly request a normal-Codex process handoff. Token renewal and file timestamp changes alone do not count as a new login. - Preserve active native work, custom endpoints and stored thread identities. Release idle ownership before the next turn; keep goals, approvals, side chats and background-terminal controls usable while migration waits. - Check pending requests and caller leases again after waiting for the process start lock. Inconclusive release stays retryable; accepted turns are not replayed. - Pass 621 focused server tests and 43 desktop runtime/health tests. Verify isolated lifecycle regressions and a real HTTP/native CLI metadata fixture without real credentials or a server restart. Live OAuth renewal, model/history continuity and graphical client acceptance remain unverified. See [the handoff contract](../server/docs/CODEX_LOGIN_HANDOFF.md). ## 2026-09-25 — Remove standalone Claude auth-status checks - Remove standalone Claude authentication-status subprocesses from startup, catalog refreshes, manual CLI rechecks and turn admission. Installation and capability checks do not claim that the user is authenticated. - Preserve the last actual Claude authentication result and its observation time through installation refreshes. Allow a native retry after external login, while still rejecting missing or broken executables. - Keep the existing desktop contract: unknown Claude authentication allows a request, whose native result updates readiness. This change requires a server update; it does not change the client runtime or publish a release. - Validate 218 focused server tests, 31 desktop contract tests and TypeScript. An isolated HTTP server with a recording CLI fixture exercises startup and repeated catalog refreshes with zero auth-status invocations. Successful renewal using a real signed-in account is not established by these checks. - Preserve native model discovery merged in parallel. Its disposable SDK initialization processes still use the native authentication environment; their renewal/termination behavior needs separate validation. Removing auth-status probes does not establish that all catalog work is passive. ## 2026-09-25 — Keep the Side chat button beside the composer - Align the Side chat launcher with the composer's right edge, including its maximum width and narrow-window gutters. Preserve its existing vertical position and popover behavior as the draft grows or panes resize. - Verify signed local desktop `1.0.7-beta.14`, build `1216`, from committed source `acdad85`. Reproduce the previous 282.5-pixel gap in the prior build; measure zero gap in the corrected native app at wide and narrow widths, horizontal and stacked split views, and with a multiline draft. Open and resize the popover against an isolated real server using native app input. - Pass 4,898 desktop tests, type checking, production compilation, package audit and signature checks. No renderer exceptions during acceptance. This is an app-only local build; no server deployment or public release. ## 2026-09-25 — Repair shared-chat files and live recovery - Restore the combined link-and-token copy action alongside individual copy buttons. Shared-browser attachments support downloads, image previews, and ordinary uploads through the existing session-owned file pipeline. - Remove browser-only attachment count, upload-size and lifetime quotas, and transcript/snapshot size rejection. Keep complete messages in paginated snapshots; migrate existing size constraints atomically without changing stored links, tokens or the database version. - Drag-and-drop and paste use the same upload path as the chooser, without the leftover four-file/eight-MiB restriction. Match browser submissions to their existing request receipts so signed attachment IDs do not leave a duplicate Submitted bubble after acceptance. - A temporary live-stream error now permits the browser's existing reconnect behavior instead of falsely declaring the share unavailable. Actual revocation still ends access. - Pass 216 affected server tests, including migration rollback, attachment ownership, downloads and transient live-stream recovery. Verify real HTTP sharing of a complete 3.64 MB message while excluding a 2.66 MB private tool record. Preserve an existing share byte-for-byte through database migration. - In an isolated Chromium browser against the actual server and Codex provider, drop six files including 9 MiB text, an empty file and an image; preview, send, read them with the provider, and download unchanged bytes. Publish and download a new output through its Markdown link. Repeat a single-file drop onto the conversation, complete a tool-backed follow-up, and observe one accepted message without a lingering Submitted bubble. Running, completion and server reconnect states work. No provider or server mock is used for these checks. - Verify the combined copy action through signed local desktop `1.0.7-beta.13`, build `1215`, from source `4caf335`, including native IPC/HTTP and exact clipboard contents. Its 4,894 desktop tests, type checking, production compilation, package audit and signature checks pass. Follow-up drop/submission fixes pass 28 shared-chat and 23 desktop send tests plus type checking. The server-hosted browser bundle is rebuilt. No deployment or public release. - Integrate browser follow-up `4f95a99` with current main at `8430ba6`; pass 273 affected server, catalog and packaging tests. Stop isolated services, revoke test shares, remove temporary credentials, and verify the original provider credentials remain unchanged. ## 2026-09-25 — Carry forward the provider Delete/release checklist - Port the documentation-only acceptance criterion from AgentsServer PR #117 into the canonical server checklist. Require confirmed runtime release, preserved native history, native resume, failure handling and chat isolation. - Verification: documentation diff reviewed; no runtime behavior or deployment changes. This is an acceptance checklist, not a claim that every provider has already passed those checks. ## 2026-09-25 — Preserve Claude input and repair imported wrappers - Send ordinary slash-prefixed Claude messages byte-for-byte as written, using the provider's native per-message transport flag. Remove the injected literal message instruction. Deliberately selected provider commands keep working. - Reconcile older injected copies and native command XML with their original messages using provider message IDs and transcript ancestry. Preserve original user text, quoted examples and raw transcripts. Imported-only records without matching original-message evidence are left unchanged. - Carry existing same-ID history corrections in semantic pages even when the duplicate's old timeline entry has disappeared. Refresh provider history can replace a previously cached duplicate instead of simply omitting it. - Verify the signed local desktop `1.0.7-beta.12`, build `1214`, with the real Claude provider: an absolute path with spaces reaches the provider unchanged, a native Read retrieves a new file, ordinary slash text reaches the model, and selecting `/context` executes the native command. Refresh provider history removes a cached XML duplicate while retaining the original command, result and file-read response. Pass 83 transport, 98 history/paging and 109 runner, command and goal tests. - Server-source correction only; no production deployment or public release. ## 2026-09-25 — Restore running forks after hidden history repair - Keep the last completed native turn as the fork point when later history repair adds an imported, metadata-only terminal event. That bookkeeping event no longer causes a running Codex or Claude fork to lose its boundary. - Pass 69 fork tests, including a completed turn followed by mailbox work, hidden replay metadata, stopped turns and a currently running turn. - Verify through the signed local desktop `1.0.7-beta.12`, build `1214`, against the corrected isolated server: click Fork while Codex executes a command, receive a native child before the parent completes, and let the parent finish normally. The child excludes the active turn and recalls inherited file-tool output without reading the file again. No renderer exceptions occur. - This is a server-source correction. It is not deployed or publicly released; no desktop rebuild is required for this correction. ## 2026-09-25 — Integrate recovered features with current main - Preserve synchronized side conversations, goal controls, upgraded Codex CLI detection and Python installation fixes alongside the restored provider, installer, history-import and named-instance features. - Include both sets of runtime modules in installation, deployment and archive manifests. Keep API details in the server reference and the README concise. - Move newly added main-branch tests into the recovered test package and align isolated fixtures with bounded provider discovery, import filtering and recipient-bound user delegation. Unattested legacy provenance stays private; oversized attested migration remains recoverable. - Pass 96 focused installation tests, 40 previously failing CI-related checks, 500 affected desktop tests and desktop type checking. Complete source CI is recorded on the recovery pull request. This integration does not deploy or publish a server, desktop application or npm package. ## 2026-09-24 — Accept local beta.12 goal shortcut correction - Accept local arm64 desktop `1.0.7-beta.12`, build `1214`, from committed source `90beac268bda148cbf205a5cf38500feae41f1f7`. Pass all 4,849 active desktop tests, TypeScript, production compilation, compiled-package audit, Developer ID signature and entitlement checks. - In the signed package, click Codex's composer goal shortcut, submit a disposable goal through native IPC/HTTP, observe native completion, and clear it. Open the same dialog again from the Add menu. The unsent main draft survives all actions and never becomes a submitted message. - Verify Claude's goal shortcut still opens its dialog, both composers omit account usage, and the corrected controls fit a narrow window. The isolated packaged app reports no renderer exceptions. Production credentials remain unchanged and temporary credential copies are removed after acceptance. - This local app is signed, not notarized or publicly published, with automatic updates disabled. The correction requires no additional server change. ## 2026-09-24 — Remove account usage preview and align goal shortcuts - Remove the account usage indicator from the composer while provider reporting receives further testing. The context usage meter remains available. - Give Codex the same goal shortcut beside the context meter and in the Add menu as Claude. Both entry points open the existing provider goal dialog directly, without sending a chat message or changing its draft. - Keep public desktop publication on hold. Record local package acceptance separately after testing the corrected build through native controls. ## 2026-09-24 — Publish the signed beta.11 server - Publish server `1.0.7-beta.11` from canonical source `9b04f852c5c6e698edbf47f9e64d9eb9b65a81be` and standalone export `278f88d1d4110b23b1cab3d30013116b7021214d` after all eight release test shards pass: 5,163 cases, including six skipped. - Verify the Ed25519 signature and all 109 packaged runtime files against committed source. Anonymous downloads match all three signed assets. Archive SHA256: `f4e84a2cb93ba4ebcf30f9db2ab77ada4eac496bd06ce85fb4001d2a0cb17a08`. - Both managed deployment targets accept the exact version and finish preparing it. Activation is scheduled for idle; both still report beta.10 at handoff. No running agents are interrupted. This publication changes neither desktop releases nor npm tags. ## 2026-09-24 — Accept local beta.11 desktop - Accept local arm64 desktop `1.0.7-beta.11`, build `1213`, from committed source `9b04f852c5c6e698edbf47f9e64d9eb9b65a81be`, paired with the same server source. Pass all 4,843 active desktop tests, TypeScript, production compilation, compiled-package audit, Developer ID signature and entitlement checks. - Personally exercise the isolated native app and signed package through actual IPC/HTTP with Codex and Claude: inherited tool-result context, follow-ups, shared history in two clients, cross-client cancellation and Clear, app closure while an answer continues, and native side-context retention after a server restart. Verify both providers' goal completion, Codex Clear, Claude Clear & stop, and the new Claude header panel. No provider mock is used in these checks. - Read real account observations from both providers through the packaged UI. The test account reports Codex credit availability and Claude reset times; neither supplies a percentage. Percentage rendering and rejected-window precedence are covered by component/native-event regressions. - The local package is signed but not notarized or publicly published, and its automatic updater is disabled. Public server-candidate validation and deployment are recorded separately; a built candidate is not an installed server. ## 2026-09-24 — Shared provider controls, side-chat sync and account usage - Use the same goal summary, progress and editing layout for Codex and Claude. Add Claude's clickable header status panel, with its native context, pending interactions and goal entry point. Preserve each provider's supported actions. - Persist side conversations on the connected server and reconcile them across native clients using socket notifications. Accepted answers survive app closure; Stop and Clear apply across clients. Codex resumes the saved native side thread; Claude restores its native side history. Side content remains outside the main transcript, and private Codex forks stay out of main-chat import discovery. - Show provider-reported account allowance, reset times and credits when supplied. Missing percentages remain unknown; API and custom endpoints do not inherit ChatGPT allowance. Account changes invalidate observations, and usage updates do not become transcript events or trigger model requests. - Respect Claude's configured data directory when locating native session history and goals. Preserve newly typed drafts during Clear and reject late results from a previous connection. Refresh side history on reconnect transitions and changes, without treating repeated timeline liveness notices as polling triggers. - Source verification includes the desktop suite, provider transport/authentication, persistence and cancellation regressions, installation/package checks, TypeScript and production compilation. Native acceptance uses two isolated desktop clients and real providers through production IPC/HTTP; release-package acceptance is recorded separately after packaging. Both app and server updates are required for synchronized side conversations and the account usage indicator. ## 2026-09-24 — Publish the signed beta.10 server correction - Publish server `1.0.7-beta.10` from canonical source `240e29414a8cc843d593d699d7252e0b7df0c401` and standalone export `2d84e17f8dc7d23f6ef2da8ecc8f0faeb8be6ba0` after all eight release test shards and the paired local app acceptance pass. - Independently verify the Ed25519 signature and all 108 packaged runtime files against committed source. Anonymous downloads match all three signed candidate assets byte-for-byte. Archive SHA256: `ef5f0418ab6e0890c87b346653fcb570256e1f9bb17d0da0afd292650dfec608`. - Managed updates prepare while agents keep working and activate when idle. An accepted update request does not establish completed installation. This publication changes neither desktop releases nor npm tags. ## 2026-09-24 — Remove side-chat answer deadlines - Remove the 150-second answer cutoff from the shared side-chat runtime and Claude's native control path, and the desktop's 210-second HTTP deadline. Long answers retain their native conversation and follow-up context. Codex side chats inherit ordinary Codex transport settings. - Preserve Stop, Clear, request-owner cancellation, disconnection and shutdown cleanup. Remove timeout copy that promised a retry would work on an older server after its native side conversation had already closed. - Regressions fail before the correction and pass after advancing beyond the former deadlines. Exercise a real local HTTP connection, native conversation retention, the full Claude manager/control path and cancellation without stopping the parent. Pass 282 affected desktop and 238 server checks and desktop TypeScript. - Personally reproduce the old cutoff through an isolated native offscreen app and real Codex. With the correction, a 225-second tool completes and the app receives its answer after 234 seconds. The main chat answers concurrently; a side follow-up retains inherited tool-result context. Stop acknowledges in 100 ms and its owned tool exits. - Close a private Codex process promptly when Stop arrives during stalled startup or fork creation. Retain ownership of delayed spawns and avoid restarting a closed transport after a late fork reply. All 172 affected adapter/transport checks pass, including unchanged durable-fork cleanup. - Accept local desktop `1.0.7-beta.10` build `1212`, arm64, from app source `183083a5`, paired with server source `240e2941`. The actual signed package completes native Codex and Claude first questions, contextual follow-ups, Stop and Clear followed by another answer through production IPC/HTTP. The isolated packaged window reports no renderer exceptions. Claude's beyond-deadline control behavior is covered deterministically; the actual 225-second tool check uses Codex. No mocked provider is used for these app acceptance checks. - Pass all 4,804 active desktop tests, TypeScript, production compilation and compiled-package audit, plus all eight server release test shards. Verify the local app's Developer ID signature. This local app is not notarized or publicly published, and its automatic updater is disabled. Both app and server corrections are required to remove both answer deadlines. ## 2026-09-24 — Publish the signed beta.9 server correction - Publish server `1.0.7-beta.9` from canonical source `40902a58873b6a9e298a4c9a24f55b80ebe21b4a` and exact standalone export `ae9d4373eaeeaf6555eefeb6c3b0243329b4f44d`. All eight release test shards and the canonical server, Electron and mobile-source CI checks pass. - Correct incomplete provider-manager test fixtures exposed by the first release validation attempt, then rerun the full suite before signing and publication. No failing candidate is published. - Verify the Ed25519 signature, all 108 packaged runtime files and executable modes against committed source. Anonymous downloads match all three signed candidate assets byte-for-byte. Archive SHA256: `cb442218c9e524bad126190ee1ceb89c28356f41c3d1e91a200c857d75970548`. - Deployment uses the existing managed updater with fresh preparation and when-idle activation. A queued request is not completed installation; running agents retain their current worker until its work finishes. This release changes neither desktop builds nor npm tags. ## 2026-09-24 — Refresh an upgraded Codex CLI without stopping running chats - Recheck CLI and subsequent provider operations detect a replaced CLI. New chats use a new process while existing turns, goals, approvals and background work retain their original owner. Idle chats resume their native thread history on the current process. Rechecking the same version does not restart it, and read-only inspection cannot retain an old process forever. - Keep late notifications and approval requests tied to their emitting process. Include manager identity in goal reconciliation and close every retained process during provider shutdown. - Pass 376 affected checks, then 67 targeted checks after the final inspection correction. The new regression cases cover concurrent routing, idle resume, pending work, delayed callbacks, shutdown and inspection-task lifetime. - Exercise Settings > Server > Recheck CLIs in an isolated native offscreen app through real IPC/HTTP. A fresh GPT-6 Sol chat using ChatGPT authentication completes while the older process continues its existing turn. That turn finishes normally; its old process exits and a contextual follow-up returns the remembered phrase using the same native thread ID on the new process. A repeated same-version recheck creates no additional process. No renderer exceptions or changes to the production authentication file are observed. - The live test changes a wrapper's reported version while both processes use the installed native CLI. It validates handoff and continuity, not historical compatibility between two different CLI executables. The final read-only inspection correction is covered by its focused lifetime regression. - Prepare server `1.0.7-beta.9`; public signing, publication and installed activation are separate checks. No desktop or npm release is included. ## 2026-09-24 — Preserve existing Python permissions during server updates - Accept same-user external Python interpreters and bounded uv runtime trees with group-write permissions during preparation and activation. Record the interpreter's bytes and mode without altering a shared installation. - Reproduce the preparation and durability failures before the correction. Verify candidate and retained releases sharing a `0775` uv prefix with `0664` library files and internal links. Changed interpreter bytes or modes still invalidate the preparation receipt. - Pass 81 focused macOS checks (one Linux-only check skipped) and 61 Linux checks, including real isolated worker/gateway startup and native systemd unit parsing. The corrected scanner also accepts an existing uv runtime and retained release without changing their permissions or service process. Complete installed-service activation remains a separate deployment check. ## 2026-09-24 — Discover current and older Claude model choices - Resolve versioned Claude labels from native SDK initialization while retaining alias values and existing chat selections. Restore selectable older versions through a disposable native picker, not an unfiltered union of static IDs. - Preserve native restrictions, custom gateways and curated picker settings. Metadata probes send no user prompt, disable tools/hooks/MCP, use bounded process lifetimes and output, and never return account metadata. - Include the new runtime module in npm packages and legacy signed-package inputs, installer validation and direct-deploy validation. - Verification: 189 focused server regressions pass. A real Claude Code 2.1.281 metadata probe through the isolated server catalog returns current and older choices without duplicate IDs in 1.78 seconds; user settings are unchanged. A full local npm package contains the exact module bytes and all 78 packaged Python sources compile. - Boundaries: metadata-only native verification, not billed inference or Electron UI acceptance of this monorepo build. This is a source PR; no npm publication, release or service restart is part of this change. - Integration recheck (2026-09-25): retain restored runtime modules and the shared catalog deadline when merging current main; move the added tests into `server/tests/`. All 214 focused tests pass, including three new deadline regressions. A fresh metadata-only native probe returns 15 unique choices in 0.94 seconds. The local npm archive preserves the exact catalog module and all 83 packaged Python sources compile. No archive was published. ## 2026-09-23 — Publish the signed beta.8 server update - Publish the legacy signed server beta `1.0.7-beta.8` from canonical source `8ee941e0d5469a939ffd453acbb42d8e8cfe6132` and exact standalone export `2bdc10afcf3e203e8d56ecf2f8d5dbc2b1a9d4d9`. Preserve the separate OpenCode release branch and existing stable release. - All eight server release test shards pass. Verify the Ed25519 signature, all 108 packaged runtime files and the packaging policy's executable bits. Anonymous downloads of the three published assets match the accepted candidate byte-for-byte. Archive SHA256: `e951e8782ec948fae8562be9980d777b1918c152094464e4fbaa776413f80333`. - The merged canonical source also passes Electron type checking, 4,803 active tests, production compilation, mobile-source checks and all eight server CI shards. This publication changes neither npm tags nor desktop releases. - Server installation uses the existing managed updater's durable when-idle request. Publication and an accepted reservation do not establish completed activation; observe each installation's status and authenticated health. ## 2026-09-23 — Integrate accepted desktop and server work into main - Merge the accepted release-line changes, including native Codex Side chat inspection and latest-message navigation, while preserving main's removal of the Usage analytics screen and clearer saved-server update settings. - Resolve update-control conflicts by retaining cancellation behavior, channel selection state, saved-server inventory and setup only when unconfigured. - Pass 268 focused Settings, restart, coordinated-update and timeline tests, TypeScript, production compilation and the compile-output audit. Personally open General and Updates with native input in an isolated offscreen app; confirm analytics removal, saved server/version display and channel state. This merge check does not exercise an actual update installation. ## 2026-09-23 — Clarify update settings on the main desktop line - List saved servers and their known versions beneath the server update controls. Retain the server heading when a saved server is offline. - Offer setup only before any server is configured, including when an inactive saved server is offline. Selected update channels expose their pressed state, and app update actions wait for a pending channel change. - Adapt the Settings change to this line's existing manual server updater; no server update or backend contract change is required. - Focused Settings component checks, TypeScript and production compilation pass. These checks use component fixtures; native click-through and real update transport acceptance remain pending. Availability: source/local compilation only, with no release or server deployment. ## 2026-09-23 — Validate local beta.8, build 1211 - Build local macOS arm64 `1.0.7-beta.8` from committed source `8a9605aaa4c036061a4912cf5b9ef883d45e5b3a`. Package checks, bundle audit and deep Developer ID signature verification pass. This local build has automatic updates disabled and is not a notarized public release. - Personally exercise the exact packaged app with native input, isolated profiles and real IPC/HTTP. Codex Side chat reads a newly created workspace file, recalls its unpredictable value on follow-up, and clears successfully. The parent gains no conversation turns; normal provider-load metadata is permitted. The provider runs against the corrected isolated server. - On a 600-turn synthetic conversation, scroll upward and use the floating bottom button; navigate into older history and return to the latest message; fork through the real HTTP memory-fork route and verify the child opens at the bottom. Each bottom check measures zero remaining scroll distance. No renderer exceptions occur. Stop all owned test apps and servers. - Packaged archive SHA256: `8b94bf826ed2e3d848f83650f1b8e79c836419b4218ee51af79a6463640ca7fe`. Production server activation and public distribution are separate from this local acceptance; Side chat tool access requires the corrected server. ## 2026-09-23 — Restore side-chat inspection and reliable latest navigation - Match native Codex Side chat: retain the parent workspace and permission settings, allow file inspection and ordinary tools, and keep side questions separate from inherited tasks. Route side approvals through the existing controls without borrowing the main run's helper authority. Stopping the main turn leaves side approvals intact; closing Side chat cleans up its own work. - Verify real Codex reads a file created after the parent turn, remembers the result on follow-up, and performs a separately requested local write. Parent provider history, settings, goals and queues remain unchanged. Personally exercise the app's Side chat with native input through production IPC and HTTP: read another new file, verify its unpredictable value, follow up, and Clear. The native test window is isolated and offscreen. - Keep the floating Jump to latest action visible in older-history windows. Reproduce the missing control in the actual app with a 600-turn fixture, then verify the corrected button reaches the latest message. Verify a new fork opens at the end and ordinary saved reading positions remain intact. These timeline checks use synthetic history and the real HTTP memory-fork path. - Reapply the existing bounded initial bottom alignment when virtualized row heights settle. A focused regression covers a delayed height change and user scrolling cancellation; ordinary fork landing already worked in the baseline native fixture, so it does not establish the intermittent failure's frequency. - Validation: 4,804 active Electron tests, eight stock Node checks, TypeScript, production compilation, and focused server adapter/provider/approval checks pass. Package acceptance is recorded separately. Side-chat tool access needs the server change; the scrolling corrections are app-only. ## 2026-09-23 — Publish the fresh-install npm beta - Publish `@agentsdock/server@1.0.7-beta.5` publicly from committed source `5486cbcb096026e798f6b0bc20743b71d7b4a9c1`, available on `release/npm-1.0.7-beta.5`. This is an opt-in fresh-install server beta; no desktop release, existing-installation migration or server deployment. - Download the public tarball anonymously and verify it matches the tested candidate byte-for-byte: SHA256 `9d9c6b3e69bdb56cae24c38fceb86ed5074d8b576ffbe4dcff384abd6c4a9f88`. With Node 22.18.0 and npm 10.9.3, empty cache and no registry credentials, execute the public beta CLI and confirm `1.0.7-beta.5`; independently pack the public beta and verify its archive hash and bundled version. - Prior isolated runtime acceptance covers authenticated health and session endpoints. Full managed-service installation and migration acceptance remain unperformed; public CLI verification does not establish either boundary. - Publish explicitly with `--tag beta`. The registry also assigns `latest` to this version; two authenticated removal attempts return HTTP 400. Both tags still point to the beta. Use explicit `@beta` testing instructions and do not describe this publication as a stable release or claim tag cleanup succeeded. ## 2026-09-23 — Unify goal editors and validate local beta.7, build 1210 - Give Codex and Claude the same dedicated goal dialog, completion-condition field, progress styling and footer. Open Codex goals directly from the slash command and Edit action; keep other thread controls separate. Preserve native provider behavior and Codex status, token budget and time limit. - Commit app source `c0b233346e513c436ce98967789abf09b2b13320` before building local 1.0.7-beta.7, build 1210. TypeScript, all 4,801 active app tests, production compilation, bundle audit and Developer ID signing pass; five existing tests remain skipped. - Exercise the source UI with native input, production IPC, authenticated HTTP and actual providers. Codex retains paused status and both budgets through edit/save/reopen, then clears. Claude achieves a short goal and clears/stops a second goal. The shared layout fits a narrow light viewport. An initial offscreen renderer loss is not reproduced by the successful sequential retry. - Reject the first local candidate, build 1209, after the actual packaged app exposes a keyboard-focus error when opening Goal from thread controls. Add a regression that fails before the correction, then personally verify build 1210: immediate typing targets the goal field, Escape restores the trigger, both providers open the shared dialog, and switching chats does not preserve an abandoned Codex draft. No renderer exceptions occur in the final packaged check. Close the test app and remove its temporary credential. - Availability is a local Apple silicon app, signed but not notarized, with automatic updates disabled. No upload, installed-app replacement or server deployment. This UI correction requires no server update. Keep public build reservation 1208 separate from these local candidates; builds 1209 and 1210 are consumed locally. ## 2026-09-22 — Keep interrupted mail checks out of user history - Correct Claude mailbox-input ownership proof for stopped and failed runs. A recorded interruption does not change a generated instruction into user input. Keep the exact input hash, provider identity, source checkpoint, unique occurrence and time bounds; assistant replay checks are unchanged. - Cover existing sanitized imports and first imports, stopped and failed terminals, and genuine human quotations. Focused server checks pass. - Reproduce the leak through authenticated HTTP on an isolated server with a persisted provider-transcript fixture. Personally open it in the signed desktop package, update only the isolated server to a new advertised version, and reopen the already-cached chat. The generated input disappears while the identical human quotation and both assistant replies remain. Repeated chat switching stays correct; persisted source and event files are unchanged. - Test transport and history repair are real; the disposable provider transcript is synthetic and no provider inference runs. Missing-terminal or unowned history is outside this correction. A server update with a new version is required to refresh existing desktop caches. No production deployment. ## 2026-09-22 — Validate local desktop beta.5, build 1207 - Build committed app source `92ab2320c1b6b0e41d9cb59fc3887cb3ff90e3ba` as 1.0.7-beta.5, build 1207. TypeScript, all 4,798 active app tests, production compilation, bundle audit and Developer ID signing pass; five tests are skipped by the existing suite. Retain the compact Claude thinking correction. - Personally exercise the exact signed app with native mouse input, production IPC and authenticated HTTP/WebSocket transport. With a real history response held for eight seconds, a cached chat reaches live 30 milliseconds after the click. A new streamed reply arrives before that stale response, remains exactly once afterward, and remains after switching away and back. - No renderer exceptions occur. Close the isolated app and remove its temporary credential. Synthetic persisted messages exercise transport and reconciliation, not provider inference. The separate server history-repair acceptance above uses this unchanged app package. - Availability is a local Apple silicon app, signed but not notarized, with automatic updates disabled. No publication, installed-app replacement or production server deployment. The syncing and thinking corrections are app-only; the mailbox-input correction requires the server change above. ## 2026-09-22 — Reconnect cached chats without waiting for history - Open the live connection immediately when switching to a cached chat, while the existing history refresh checks imports, metadata, queues and repairs in the background. First opens still load their authoritative history page. - Preserve newer live messages and queue changes when a delayed history reply arrives. Reset the stream cursor when server history is replaced, and prevent buffered events from the previous log from returning afterward. - Reproduce the delay in an isolated native desktop app through authenticated HTTP and WebSocket transport: an eight-second history response kept the old app syncing for eight seconds. With the correction, the cached switch reaches live in 28 milliseconds while that response is still pending. A new message arrives before the response and remains visible exactly once after refresh and switching away and back. No renderer exceptions occur. - Focused service, transport and store checks, TypeScript and production compilation pass. The controlled test uses persisted synthetic messages, not provider inference. Package acceptance is recorded separately. - This correction is app-only and requires no server update. ## 2026-09-22 — Validate local desktop beta.4, build 1206 - Build committed source `230f946912c0cdf4f05da0c87e0f106973c0e5b5` as 1.0.7-beta.4, build 1206. TypeScript, all 4,794 active app tests, production compilation, bundle audit and Developer ID signing pass. The first packaging attempt omitted the beta-track environment; correcting that local build configuration passes the unchanged suite. Five tests are skipped by the suite. - Personally exercise a real Claude turn through the isolated native app, production IPC and authenticated server: compact thinking, live setting on/off, completion collapse with the setting enabled, and retained manual history. A 7,206-character received thinking event remains available. Verify long-text fixtures and bounded dark/light rendering, including a narrow viewport. - Open the exact signed package and verify the compact disclosure, full-text expansion, native wheel scrolling and settings behavior on completed history. The panel is capped at 320 pixels, or 40 percent of the viewport height. No renderer exceptions occur; close the isolated test app and remove its temporary credential afterward. - Availability is a local Apple silicon `.app` only, signed but not notarized, with automatic updates disabled by the local-build workflow. No publication, installed-app replacement or server deployment. This correction needs no server update. ## 2026-09-22 — Make Claude thinking compact and optional - Keep Claude thinking in a single-line disclosure by default. Apply the existing thinking visibility setting to active Claude turns as well as Codex; completed and stopped turns collapse while retaining manually readable text. - Bound expanded Claude thinking to a scrollable panel, preserve the user's chat font size, and replace the large colored card with subdued styling. Update the setting's English and Chinese descriptions. - Targeted timeline checks, TypeScript and production compilation pass. Reproduce the oversized panel in an isolated native app and verify the corrected long-text display through production IPC and authenticated HTTP. Local package and live-provider acceptance are recorded separately. - This is an app-only correction. No server update is required. ## 2026-09-22 — Validate unpublished desktop beta.3, build 1205 - Build committed source `4c87d87f28735adedd20295fbb4db4f328c794b1` as 1.0.7-beta.3, build 1205. All 4,792 app tests, TypeScript and production compilation pass. The universal macOS app and installer pass signing, notarization, Gatekeeper, package parity, updater checksum and clean-launch verification. Both Linux architectures and Windows pass their release jobs. Windows first encounters a timeout in an unchanged history-cache test; the single retry passes with the same source and unchanged limits. All 14 assets and updater checksums are verified. Windows remains unsigned. - Personally exercise the actual signed app through native mouse and wheel input, production IPC and authenticated HTTP against isolated synthetic histories. First visits open at latest; returning to an older message in a 600-turn chat preserves its offset within one pixel. Repeated rapid chat switches and scrolling immediately before switching preserve the same row and offset. No renderer exceptions occur; the isolated app is closed and its temporary credential is removed afterward. - The server tree is unchanged from the accepted 1.0.7-beta.2 server candidate. Scrolling needs no server update. Claude Goals and the history/lifecycle changes still require that server candidate; this entry does not record a production deployment. - Keep the desktop candidate local and unpublished, with app updates independent of server updates and no automatic npm migration. ## 2026-09-22 — Preserve chat reading positions - Restore the saved message and pixel offset when returning to a chat. First visits and readers already at the bottom still open at the latest message. - Save the message sequence alongside the existing position. If that message has left the in-memory cache, use the existing history-window request to reload it before displaying the conversation. Do not save the interim tail. - Keep user scrolling and explicit navigation in control of delayed restores. Empty or failed history requests leave the current history usable; saved positions from older apps remain compatible. - Reproduce the old jump in an isolated native desktop app. Verify rapid chat switching and restore the exact message and offset in a 600-turn history beyond the cache limit through production IPC and authenticated HTTP. Delayed history replies do not override a newer chat selection or wheel input. All 4,792 app tests, TypeScript and production compilation pass. - This is an app-only correction for desktop 1.0.7-beta.3; the prepared server remains 1.0.7-beta.2. Package acceptance and availability are recorded separately. ## 2026-09-22 — Validate unpublished desktop beta.2, build 1204 - Build committed source `5fb88e8a09e013eff13f37a706de020aa74ca0d9` as 1.0.7-beta.2, build 1204. All 4,789 app tests pass. Universal macOS signing, notarization, clean launch and the stock artifact verifier pass. Windows and both Linux architectures pass their release jobs; all 14 assets and updater checksums are verified. Windows remains unsigned. - In the actual signed macOS app, complete a native Claude goal, observe Start goal automatically become available without reopening the dialog, and send a normal follow-up that appears exactly once. The isolated server runs 1.0.7-beta.2; no runtime refresh or renderer reload is used. - Prepare the signed 1.0.7-beta.2 server package from source `9ae743b`; its server tree exactly matches the desktop source. All eight test shards, signatures and archive checks pass, with identical runtime files in the npm and legacy packages. - Keep these candidates unpublished while correcting chat-switch reading positions in the next desktop beta. No npm publication or production server deployment is part of this acceptance. ## 2026-09-22 — Correct goal completion refresh and compaction history - Keep Claude runtime subscriptions stable when timeline updates replace the selected chat's session snapshot. A queued completion refresh now survives, so Start goal becomes available when the turn returns to idle. Reproduced the failure in packaged build 1202 and verified the correction through the real desktop IPC, isolated server and native Claude provider. Coalesce immediate shared-chat refreshes with queued event refreshes. - Recognize Codex compaction output using its native response receipt and typed replacement history. Omit the proven summary during the existing parsing pass and repair affected imported rows on read. Ordinary assistant imports do not gain an additional source-prefix scan, and genuine replies with the same text remain visible. - Verify the affected history through the production HTTP and semantic APIs and an isolated native desktop app. The compaction handoff disappears and the surrounding genuine replies retain their text and order. - Preserve beta.1 artifacts as an unpublished candidate. The corrected candidate is 1.0.7-beta.2; package acceptance and availability are recorded separately. ## 2026-09-22 — Integrate native Claude Goals and repair turn transitions - Add desktop Claude Goal controls using the installed provider's native `/goal` command. Read native goal-status records for active, achieved and cleared state; retain completed details without inventing iteration counts. Older servers continue normal chat without the new controls. - Clear a running goal through Claude's native priority command, which also stops that turn. Keep the command receipt separate from the interrupted result, and retire the exact connection if confirmation times out so a later message can start normally. - Preserve interruption provenance across parallel tool-result branches, so native interruption markers do not become ordinary user messages. - Normalize source-proven native goal commands when reopening history and omit native synthetic placeholders and duplicate imported command rows. - Do not promote a queued message to Starting when Stop is still pending. Preserve the queued message and use bounded, exact-run Stop for Claude's Send now path. - Focused provider, runner, queue, transcript and desktop checks pass. The real desktop app, production IPC and isolated server complete a native goal, clear one during a long-running tool, and complete a normal follow-up. Live Send now also completes the replacement turn with an empty queue. Release artifacts are recorded separately; this entry does not claim a deployment. ## 2026-09-22 — Publish desktop 1.0.6, build 1201 - Publish the accepted desktop package as stable 1.0.6 in the public source repository and desktop release mirror, with the same 14 verified assets. - Verify public download links, checksum manifests, Stable updater metadata and Beta discovery of the stable release. Authenticated release metadata checks confirm the source pin after anonymous API requests hit GitHub's rate limit. Withdrawn 1.0.4 and 1.0.5 releases remain absent. - Keep server/npm publication unchanged; stable AgentsServer remains 1.0.3. Stop the owned acceptance VMs and forwards after preserving their evidence. ## 2026-09-22 — Validate desktop 1.0.6, build 1201 - Build committed source `564f38a64a9e748f810de64668e860a4c7badcca` as desktop 1.0.6, build 1201. macOS signing, notarization, clean launch and the stock release verifier pass. Windows and both Linux architectures pass their complete release jobs; all 14 release assets and updater checksums are verified. Windows remains unsigned. - In an isolated macOS VM, the unchanged published 1.0.3 and 1.0.6-beta.1 apps each update through the native updater and automatically relaunch the exact accepted package. Saved connections and the respective Stable/Beta preferences survive. The original authenticated 0.1.25 server keeps its process, identity, credentials, chats and runtime files, with no mutation requests during either app replacement. - The signed package passes native Cancel during checking, download, pre-install refresh and the restart delay; Discard, explicit retry and channel switching also pass. The app process stays alive and late results do not restore the canceled update. These cancellation checks use isolated future-version metadata pointing at the accepted ZIP, without handing that substituted version to the native installer. - Native connection recovery retains the authenticated chat socket during an injected health failure and restores Online before delayed metadata, even during continuous native typing and scrolling. A separate real Claude round trip succeeds. The original intermittent socket trigger remains unconfirmed. Native send-failure checks also preserve drafts and accepted turns without resending, including the Send now queue action. - Finally, the accepted signed app's restored Install button upgrades the isolated original 0.1.25 server to the public, production-signed 1.0.3 archive through one authenticated update request. All 82 runtime files match; the original identity, token, chat and event bytes survive. The app reconnects and shows the retained chat and installed/healthy status. - These are desktop-only artifacts. No npm/server publication or production server restart is part of this release acceptance. ## 2026-09-22 — Cancel desktop updates and recover live chat promptly - Add Cancel during app update preparation/download and Discard after download. Discarded updates no longer lock channel selection. Cancellation stops the download when supported, invalidates late callbacks and prevents a pending restart; native installation handoff remains the final boundary. Normal quit does not install a downloaded update. Focused updater and Settings checks pass; signed-package cancellation acceptance is recorded separately. - Keep existing authenticated chat sockets through transient health failures, while retaining fresh validation for reconnecting sockets and privileged requests. Process health before slow session/job metadata, and allow the existing recovery poll during typing or scrolling without adding a poll loop. - In an isolated native Electron app using real authenticated HTTP/WebSockets, an injected health rejection leaves chat sockets open. Successful health restores Online before an eight-second delayed chat-list response. Focused service/client checks, TypeScript and production compilation pass. - Prevent failed sends from duplicating text already retyped in the composer, preserving current references and attachments. An authoritative live event can confirm acceptance when the HTTP reply is lost; the existing success and steering behavior then runs without resending the prompt. Focused store and Composer checks include lost replies and profile changes. Native UI checks use an authenticated protocol fixture with controlled HTTP/WebSocket faults, not provider inference. - These changes address app recovery and draft handling. The original cause of the intermittent socket failures remains unconfirmed; bounded native error diagnostics preserve evidence for a recurrence. No server deployment or desktop publication is recorded by this entry. ## 2026-09-22 — Preserve native connection failure diagnostics - Record request duration and bounded native socket error codes when desktop server requests fail. Exclude credentials, request bodies, query strings and exception messages; retain the original error and do not retry mutations. - Validate focused client tests, TypeScript and production compilation. In an isolated native Electron app, real server switching and a Claude send/reply succeed; an unavailable local endpoint records `ECONNREFUSED` and the app reconnects after switching back to the healthy server. - This adds diagnostics for intermittent failures. It does not establish the cause of a past disconnect or claim that diagnostics alone fix recovery. ## 2026-09-22 — Bound Windows release test concurrency - Run the Windows release tests with one worker after concurrent disk-heavy suites exceeded their existing deadlines on hosted Windows. Settings suites that previously completed in about one second took about one minute during the affected run. - Retain every assertion and timeout. Product behavior and other platform jobs are unchanged; acceptance requires a fresh complete release build. ## 2026-09-22 — Accept native legacy server update recovery - Exercise the corrected desktop Settings through real UI interaction, production IPC and authenticated HTTP against original 0.1.25 installations with 0755 and 0750 installation roots. Both install the unchanged, publicly downloaded, production-signed 1.0.3 archive through the original server updater. - “Install when idle” waits while a disposable chat runs, then submits exactly one request using the old server's supported fields. Server identity, access token and existing chats survive; all 82 shipped runtime files match the verified archive. A protocol fixture supplies the busy chat lifecycle; this verifies update behavior, not live model inference. - The 0750 installation is an actual managed Team Hub host. Its original owner, team, membership, message, device session, refresh credential and managed host binding survive. The existing access token can read the old message and post a new one after migration, and the maintenance fence is cleared. - An installer failure before takeover preserves the old server and Hub data, clears its exact maintenance fence, and leaves a usable retry path. The retry completes through the native app. A release-check rate limit also leaves the incumbent server healthy. - These are isolated native source-app acceptance runs. Signed desktop package replacement and release verification remain separate, pending checks; this entry does not record publication or a production server deployment. ## 2026-09-22 — Restore updates for existing servers - Restore manual server updates in Settings → Updates when the desktop release has no bundled server operation, including the app-only 1.0.6 beta. - Allow older servers with an authenticated update API to update over remote connections. Their lack of newer request identity fields no longer removes the install action or causes the desktop backend to reject it. - Preserve the app's existing idle-waiting flow for servers without native update scheduling. Send the request format supported by those servers and retain connection ownership, authentication and signed package verification. - Focused settings tests pass, along with update service/client/coordinator checks and TypeScript. Native existing-installation acceptance and release package verification are recorded separately before publication. ## 2026-09-22 — Give the composer model picker available space - Remove fixed model/effort chip width caps so the full selection can use the available toolbar width. Keep the dropdown arrow and send controls visible, and expose the full selection on hover when a narrow pane still truncates it. - Shorten the custom-provider toolbar label to “Codex · Custom”; retain the complete provider name in the menu and accessible button label. English and Chinese are covered. No provider selection or server behavior changes. - Reproduced the clipped model and effort in an isolated native Electron app, then verified the complete label at the same width after the change. Exercised the actual picker twice through production IPC and authenticated HTTP into an isolated 1.0.6-beta.1 server; both effort changes persisted. Checked dark/light, narrow layouts, keyboard focus return, and send control visibility. - Validation: 4,736 desktop tests passed (five existing skips), followed by 221 focused tests after the compact-label refinement; eight build/license checks, TypeScript, and production compilation pass. Synthetic endpoint metadata was used; model inference and a release package were not exercised. All isolated test processes exited. Availability: source only for a subsequent desktop beta. ## 2026-09-22 — Resolve guided server setup from signed release metadata - Remove obsolete Stable/Beta installer pins. Local and SSH guided setup now discover the selected published channel and verify its immutable signed manifest before downloading the exact verified archive. - Keep release discovery cancellable and report its failure before starting an installer. Preserve the existing server channel semantics and check an installed server's version before selecting an older Beta. - Validate focused setup/resolver checks, TypeScript and the production desktop compilation. Live public metadata checks select the published releases; no server version is inferred from an app-only release. - Native Electron guided SSH setup completes against a disposable macOS server: the production renderer selects published Stable 1.0.3, verifies and installs the real archive, reconnects, and displays both saved chats. Server identity, access token, saved histories, protected files and release trust key remain unchanged. This checks an existing 1.0.3 server with Team Network hosting disabled; it does not certify Team Network reactivation or the packaged application updater. This source change does not itself publish a release. ## 2026-09-22 — Publish app-only 1.0.6-beta.1 (1196) - Publish the accepted artifacts unchanged to the [public beta release](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.6-beta.1) and [legacy desktop mirror](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/v1.0.6-beta.1). Both are prereleases with exactly 14 desktop assets and no server descriptor. - Independent anonymous readback verifies all eight beta platform feeds select 1.0.6-beta.1, both direct macOS DMG links respond successfully, and both public checksum manifests match the accepted seal. All stable feeds and latest stable release APIs remain on 1.0.3. - Server and npm publication remain held. Neither a 1.0.6-beta.1 standalone server release nor an npm version is published. Existing server installations are not updated by this desktop beta. - Users on withdrawn app versions 1.0.4/1.0.5 should install the direct desktop download once to replace the old updater. The native 1.0.3 update journey and 1.0.5 direct-install recovery are recorded in the acceptance entry below. ## 2026-09-22 — Accept app-only 1.0.6-beta.1 (1196) - Accepted product source: `d36e1637e6fa6a7cec1b11cf7ffbaf70cff8e17e`. This desktop beta includes no server descriptor or enrollment and cannot resume an older app's saved server-update plan. Server publication remains held; this release does not install the server-side repairs below. - Desktop validation passes 4,736 tests with five existing skips, TypeScript, production compilation and native platform package verification. The universal macOS ZIP and DMG pass Developer ID signing, notarization, Gatekeeper, updater metadata, checksum, package parity and clean-launch checks. Windows remains explicitly unsigned. The 14-file app-only checksum seal is `f7bdb3ebc6cb503839b69c5855b41812d1ce6b36b895e3115b1508db7a11e547`. - The unchanged published 1.0.3 app opts into Beta, downloads the exact signed package, installs through its native updater and automatically relaunches as build 1196. Its existing server remains on 1.0.3 with the same process; a legacy 0755 root and retained rollback journal are unchanged. - The unchanged 1.0.5 app reproduces its old pre-install failure because this app-only release has no server descriptor. Direct replacement with the same signed app then succeeds, preserving its profile and the old update-plan file byte for byte. Users on withdrawn 1.0.4/1.0.5 should use the direct installer. This is not a claim that their old in-app updater was retroactively repaired. - Both native macOS journeys preserve server/runtime files, chats, provider and authority files, Hub records, bootstrap claims and existing mTLS access. No update, restart or stop request reaches either server. A real server-owned terminal worker keeps its process identity and advancing heartbeat through app replacement. This continuity test does not exercise model inference. - Tests use disposable native machines and private feed routing of the exact production-signed bytes. Stable does not offer the beta without opt-in. This entry accepts the unchanged artifacts before upload; publication and public feed readback follow separately. ## 2026-09-22 — Prepare app-only 1.0.6-beta.1; hold the server release - Prepare an opt-in desktop beta with independent app updates and clearer Settings. The server release is held; no npm or standalone server release is included. Previously prepared paired desktop artifacts are superseded. - An app without a bundled server target does not resume a saved server-update plan, contact the update endpoint, or change that saved plan. App updates can proceed independently of existing servers; server repairs below remain unreleased and are not claimed as installed by this app-only beta. - Preserve stable 1.0.3 availability. Verify unchanged 1.0.3 and withdrawn 1.0.5 clients against the signed app-only beta, retaining server processes, versions, chats and Hub data. Verify direct desktop replacement for any old updater whose installed gate prevents self-update. - The app-only behavior passes 316 focused desktop tests and TypeScript checks, including pending, failed, newer and unreadable saved update plans. This entry records preparation only; native package acceptance is pending. ## 2026-09-22 — Repair update blocking after withdrawing 1.0.4 and 1.0.5 - Post-withdrawal checks found both public desktop repositories and the standalone server's stable feed back on 1.0.3. The 1.0.4/1.0.5 release pages and checked assets were unavailable, as were the public npm package metadata and tarballs. Withdrawal does not repair already installed apps or servers. Publication entries below describe the earlier state, not a current upgrade recommendation. - Remove the desktop installation gate entirely: a saved server's release channel, API version, connectivity or failed update cannot prevent an app update. After relaunch, the installed app's signed bundle selects the server target. Ignore stale pre-install plans and an older target's failed operation receipt; keep already newer servers unchanged. Old apps that still contain the gate may need the corrected direct desktop installer once. Publishing npm alone cannot change their updater code. - Remove the artificial Stable/Beta server-channel veto. For older servers that return that exact rejection, use their existing authenticated update route with the bundled target. Other authentication and update failures retain their own handling. Avoid a redundant latest-release lookup for a known target. - Fix the Team Hub operation collision: a new update's maintenance fence could be confused with an older retained rollback journal. Recover the old terminal operation, then continue the same new request automatically. Clean up only the unstarted request when recovery cannot proceed. An already-absent restore receipt no longer requires stopping the live Hub to acquire its runtime lease. - Stop the release-check request burst. Share simultaneous checks, cache recent results, respect GitHub's retry interval and remove the multi-page HTML fallback after HTTP 429. Show the cause and retry delay. Opening Settings no longer triggers app release discovery or unrelated server release checks. - Simplify paired Updates to one app update control, per-server progress and a Retry action. Show concrete causes and keep protocol diagnostics expandable. Remove duplicate server-channel/recovery controls from paired releases. A downloaded withdrawn app is no longer offered when a fresh feed response confirms a different release. - Validation: 196 focused server tests and 103 Hub/activation tests pass, along with installer succession and rollback regressions. The collision regression fails against the withdrawn source. A real native macOS 1.0.3 update API run starts with a 0755 installation root, retained terminal rollback journal and newly admitted Hub operation; it completes both services without a remaining journal or fence. All 107 candidate runtime files match; existing Codex/Claude histories, authority files, synthetic credentials, Hub records and bootstrap claims survive. An existing mTLS peer reads old data and writes/reads new data. - The native migration uses a private QA feed/signing key and a captured source candidate retaining its 1.0.5 test label. The later retry-delay wording change passes nine targeted request tests separately. The 266 focused desktop tests, TypeScript and production compilation cover the installation gate removal. Native desktop UI checks exercise production IPC and authenticated HTTP, scoped Retry and readable dark/narrow layouts with no implicit release checks; that fixture deliberately rejects its test signing key and does not claim an app binary replacement or successful server migration. A subsequent change preserves the server's concrete failure text in two coordinator branches; all 55 coordinator tests, TypeScript and production compilation pass afterward. - This records source repair, not a new release. No live service was restarted or redeployed. A corrected signed desktop package still needs its complete update/relaunch and publication validation before shipping. ## 2026-09-22 — Publish coordinated AgentsDock 1.0.5 (1194) - Publish the accepted build unchanged to the [public desktop release](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.5) and legacy desktop mirror. Both stable update feeds now offer 1.0.5. - Publish `@agentsdock/server@1.0.5` through trusted npm publishing and verify that `latest` resolves to 1.0.5 and the public tarball matches the signed descriptor. Publish the [standalone server bridge](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.5) and verify its production signature, archive and 107-file runtime parity. - Independently read both public desktop feeds without credentials. All four platform feeds, checksums and signed server descriptors match the accepted seal; all 16 uploaded asset digests match. The public release page and macOS ZIP/DMG download links respond successfully. - Native acceptance covers the unchanged stable 1.0.3 app's single-update migration and recovery from a genuine failed 1.0.4 migration. Already stranded servers require `npx @agentsdock/server@1.0.5 recover` on the server computer, followed by Settings → Updates → Retry server update. - The 1.0.4 desktop release remains withdrawn. npm 1.0.4 still exists as an immutable version, but is no longer `latest`; no npm deprecation is claimed. ## 2026-09-22 — Accept direct AgentsDock 1.0.5 (1194) - Accepted product source: `321448f7ac5f7ae393a61660ec6168eedc9791b0`. The paired standalone server is `144d2eaf1690185d6fae1386f793ced32ce18e3f`; its runtime matches the signed npm package byte for byte across 107 files. - Desktop validation passes 4,699 tests with five existing skips. All eight server test shards pass (5,031 cases), together with focused CLI, packaging and publication checks. Windows and Linux x64/ARM64 builds pass their native verification jobs. The universal macOS app and DMG pass Developer ID, notarization, Gatekeeper, exact updater metadata, checksums and launch checks. Windows installers remain unsigned. - The unchanged published 1.0.3 app updates to signed 1.0.5/build 1194 through the real updater, automatically relaunches, and migrates an existing 1.0.3 server with a `0755` installation root. One app-update click completes both; no separate server-update action is used. - The unchanged published 1.0.4 app updates and automatically relaunches into the same accepted package with a genuinely stranded 1.0.4 server migration. Settings displays the pinned recovery command and copies it correctly. The exact npm package recovers the missing-stage transaction without restarting the incumbent. One native Retry click then completes the paired update, and reopening Settings shows both components current with no recovery prompt. - Both paths retain server identity, Codex and Claude histories, authority and credential files, Hub data and bootstrap claims. An existing mutual-TLS peer can read retained data and exchange a new message. Both use original production signatures and the unchanged trust key. Private feed routing supplies the exact signed bytes; public registry/feed verification follows publication. An interrupted VM/VNC harness attempt is retained separately and is not used as upgrade evidence. - Sixteen verified release assets are sealed with matching platform feeds and the same signed server descriptor. This entry accepts build 1194 before upload; publication is performed separately using these exact bytes. ## 2026-09-22 — Prepare coordinated 1.0.5 correction - Remove the withdrawn 1.0.4 legacy desktop mirror after preserving and checksum-verifying its artifacts. Both public stable feeds remain on 1.0.3 until replacement validation finishes. npm 1.0.4 remains published. - Include the dismissible error notification, isolated concurrent downloads, consumed follow-up recovery and merged sidebar version-label removal. - Reproduce the exact published installer failure against an existing 1.0.3 server with a `0755` installation root. The old API stays in `installing` and rejects a new update before downloading its installer. Add an explicit `npx @agentsdock/server@1.0.5 recover` command for this stranded state; conditional Settings guidance names the server computer and copies the pinned command. Retry remains scoped to that server profile. - Verify the packaged recovery command on the untouched failed installation: retire only its exact unfinished transaction, preserving the running process, server identity, trust key and Hub database inode. Recovery cannot start a native recovery owner or restart a service, including a phase-change race. Then complete ordinary authenticated API migration to paired 1.0.5 services and verify all 107 installed runtime files, existing histories, credentials, authority and an existing mutual-TLS peer's new write/read. Candidate delivery uses an isolated QA signing key; final production artifacts remain a gate. - Real Codex steering, test-server restart and a fresh desktop service/window retain an empty consumed queue, authenticated helper access and the visible final answer. Native download overlap and failure-isolation checks pass. - Focused Settings/coordinator tests (88), native history tests (30), and TypeScript checks pass. Correct test-only stale mailbox-text assumptions, macOS temporary-path canonicalization and fixture garbage collection before timed websocket assertions, without changing production deadlines. - This entry records source validation, not release acceptance. Final signed package and desktop update/relaunch checks are required before publication. ## 2026-09-21 — Withdraw 1.0.4 from stable feeds and repair migration and queue recovery - Withdraw the canonical desktop and standalone server 1.0.4 releases to drafts. Return both stable desktop feeds to 1.0.3. The immutable legacy mirror is marked withdrawn and prerelease; its direct downloads remain available pending removal. The npm package remains published while registry authentication for the withdrawal warning is pending. These actions do not change already installed applications or servers. - Correct the acceptance scope recorded below: the previous macOS 1.0.3 migration fixture created its installation under umask `077`, so its root was already private. It missed the normal existing-installation case with a `0755` root. Release validation now explicitly requires legacy `0755` and `0750` permissions, early failure with a live incumbent, and recovery after an older failed installer has already deleted its candidate stage. - Tighten a safely owned legacy installation root to `0700` under its exact installation lock. When activation fails before taking over services, preserve the running incumbent and Hub database, retire only the owned maintenance fence, and verify authenticated health before retiring recovery state. Keep the strict ownership, original-link, configuration and native recovery-service checks intact. Preserve the staged runtime while an activation journal remains unfinished, including the crash window before the installer receives its transaction ID; clean it only after settlement. - Consume durably acknowledged native-goal follow-ups during server queue recovery and in the desktop's persisted queue cache. Retain uncertain deliveries as paused and preserve ordinary queued work. A consumed follow-up must not become a new request after a server restart or chat reopening. - Reproduce the queue failure through an isolated native offscreen Electron app and real server transport, then verify no replay after the fix. A separate genuine Codex goal accepts a typed follow-up, survives server restart without replay, and resumes the same provider thread with a successful authenticated helper read and visible final answer. Recreate the desktop service and reopen the window to check persisted state. Controlled-provider and genuine-provider evidence remain separate. - Verify the frozen installer source on disposable native macOS installations: reproduce the published failure, recover its stranded transaction without a retained candidate, inject a fresh failure before service takeover, and retry the same archive successfully. Preserve the incumbent PID and Hub database inode during rollback, then verify paired services, all 107 installed runtime files, existing identity/history/authority, and an existing mutual-TLS peer's new write and read. This lane uses authenticated server APIs and native services with a guest-only QA signing key; it does not establish final production-signature or app update UI acceptance. A subsequent stage-retention guard passes its focused cleanup regressions; its additional native failure window remains under validation. - Availability: source corrections under validation. This entry does not accept a replacement release; final signed-package migration and desktop relaunch acceptance remain required before publication. ## 2026-09-21 — Keep download errors dismissible and isolate simultaneous saves - Wrap long error paths within the window and reserve a fixed-size close button with a translated label and tooltip. Confirm native mouse, Tab/Enter and Tab/Space dismissal, including repeated errors, in dark and light themes at wide and narrow window sizes. The isolated offscreen app reproduces the previous offscreen close button and passes all six corrected layout cases. - Give each download an exclusively created UUID temporary file. Concurrent saves to the same destination no longer share a partial file, and an interrupted save cannot remove another save's in-progress file. Real AppService/filesystem regressions fail with the previous implementation and pass with the correction, including failure isolation and complete output. - Validation: 384 service/file-action tests, 49 app/design-system tests, and TypeScript checks pass. Production compilation and the toast's native offscreen interaction pass. Error text in the toast check is a fixture; separate native download/HTTP and final packaged acceptance remain pending. - Availability: source only. The sidebar version-label removal from PR #34 is included in the coordinated correction; no replacement is published yet. ## 2026-09-21 — Publish stable 1.0.4 build 1193 - Publish and verify `@agentsdock/server@1.0.4` on npm `latest`, retaining `1.0.4-beta.12` on `beta`. Verify the exact signed tarball's size, SHA-256 and SHA-512 integrity from the public registry. Publish and verify the matching signed standalone bridge, then the unchanged desktop build 1193 on both existing stable feeds. Keep accepted product source `b3bf411c8feea751285a7c9b4e397ec526a61e30`, standalone export `8664a9399f2c282aea7a113771506e2528e6d559`, descriptor and artifact seal unchanged. - Preserve publication run `35681737578`, which stopped at private candidate draft lookup before npm preflight, OIDC authentication or publication. Correct only the publish job's draft-access permission and explicit workflow/source pins. The publishing revision is `52ff3e3a0ee33b1c7106d5924fcf31bfdb01debf`; the accepted product source remains b3bf411. No product rebuild or repacking accompanies this workflow correction. - In corrected run `35682355490`, attempt 1 successfully publishes through npm OIDC, then fails immediate readback while npm processes the package. Preserve that result. Once independent exact-byte registry verification passes, attempt 2 completes with preflight `publish=false`, npm publication skipped, and the unchanged public verifier passing. The package is published once. - Add a follow-up visibility wait for future runs: retry only the unchanged read-only registry verifier, at most 61 attempts with 10-second gaps and an 11-minute step limit. Exhaustion still fails; signature, source, archive and channel requirements remain intact. This polling correction changes no accepted product artifact and was not used by the successful publication run. The follow-up workflow commits are `5f1b503` and `2281c02`. - npm provenance identifies the publishing workflow revision. The original signed descriptor and unchanged tarball identify the accepted product source. Check the public attestation's subject digest and workflow/run identity; independent full Sigstore trust-chain verification is outside this receipt. - Existing stable 1.0.3 users retain one app-update action, followed by automatic server migration. Initial migration and execution replacement wait for idle. macOS is signed and notarized; Windows installers remain unsigned. ## 2026-09-21 — Accept stable 1.0.4 build 1193 - Accept desktop build 1193 from source `b3bf411c8feea751285a7c9b4e397ec526a61e30`, paired with standalone export `8664a9399f2c282aea7a113771506e2528e6d559` and signed npm descriptor SHA-256 `851055682343f7cd97cca1f0341f0b18ffb8ce841807bd0a18409c9300855618`. Verify both original signatures and all 107 runtime files and modes against the committed source and both server archives. - Canonical and standalone server suites each pass 4,995 tests with six existing skips across all eight shards. Native workflow `35679650749` passes Linux x64 and arm64 with 4,656 tests and five skips each, and Windows x64 with 4,620 tests and 11 skips; all stock package verifiers pass. Windows remains unsigned. The universal Mac release passes 4,656 tests with five skips on its first attempt using four workers and unchanged timeouts, then type checking, compilation, Developer ID signing, notarization and stapling, Gatekeeper, mounted-DMG/ZIP parity, updater metadata checks and isolated startup. - Verify the exact signed candidate on native macOS and Linux, including a positively observed candidate worker, induced activation failure, automatic rollback and explicit retry of the same archive. Begin from a preserved dead candidate receipt, retain it through rollback, and verify retry without manual cleanup. Preserve identity, credentials, chats/events, synthetic provider and terminal credential/configuration files, Hub authority/messages and an existing mutual-TLS peer; authenticate preserved content and new peer reads/writes. Installed permissions follow the signed installer, including its Linux `agent_server.py` normalization from mode `0644` to `0755`. - The macOS native server gate starts with the original signed `0.1.26-beta.29` runtime and explicitly selects Stable for this test. It accepts that native migration route, including rollback and retry; it does not promote beta users automatically or establish every historical desktop/feed path. The Linux gate starts with the original stable 1.0.3 managed server. - Verify one actual Update action in the unchanged stable 1.0.3 desktop with its genuine 1.0.3 server. The signed app replaces and relaunches itself, then migrates the server automatically. Both components reach 1.0.4, admission is released, and populated data and the existing peer remain usable. - Verify a separate archive-only HTTP 503 failure before the old server stops. Its PID and boot identity remain unchanged. Normal health callbacks observe the failed operation despite absent legacy health progress; opening and reopening recovery preserves the failure. One explicit coordinated Retry succeeds with Advanced recovery continuously open, updating both status rows without Check or reopening Settings. - Deliver these exact signed packages through an isolated HTTPS discovery fixture. After native app relaunch drops process-only TLS overrides, a feed check exposes the fixture certificate boundary. These results do not establish public feed propagation, which remains a separate publication check. - Retain earlier real Codex/Claude gateway-loss evidence separately: the eight tested execution/provider modules are byte-identical, but these migration checks make no new model calls. Earlier process-loss/reboot and fresh-install proofs retain their original source scope. Execution replacement waits for idle; simultaneous execution generations and live-turn survival through execution-process death or reboot are not claimed. - Seal the 16 distribution assets with SHA256SUMS SHA-256 `e945a6bf07d256e517291774188f282260ee4ce83c0c7c82371b86b118443b89`. Availability: published and verified, with the exact accepted artifacts unchanged. Verify npm `latest`, then the signed standalone bridge, before exposing both existing stable desktop feeds, as recorded above. ## 2026-09-21 — Correct retry after a failed split-runtime migration - Classify the installed runtime before authorizing shutdown or seeding recovery. After a verified rollback to the original server, a candidate's leftover process receipt is accepted as stale only under its private worker lock, with a conclusively absent process and matching authenticated legacy health. Preserve the receipt and all existing native identity, idle and update-owner checks; never send credentials to the stale callback endpoint. - Acquire the installer's authenticated legacy proof when no split execution layout is installed, including after rollback leaves a dead worker receipt. The native failure test exposed this separate shell-path omission before shutdown; retain the Python classifier and its ownership checks. - Pass 62 focused activation, recovery-intent, transaction and management tests, including a real process lease followed by abrupt process death, rollback, and a new admitted retry. Exercise the actual installer shell function feeding both Python admission checks. Cover active or malformed receipts, held or unsafe locks, changed ownership and published-layout races. - Correct an asynchronous Team Network test to await the recovered host-address control independently of bulletin loading. The 119 related renderer tests and type checking pass; product behavior is unchanged by this test correction. - Availability: source corrections awaiting official signing and fresh native macOS/Linux rollback-retry acceptance. Build 1192 remains unpublished and is retained only as preparatory test evidence. ## 2026-09-21 — Correct stable migration recovery found by native testing - Keep expanded server recovery in sync with the active server's coordinated update, boot and version changes. Refresh only authoritative status; preserve the original failure while a refresh is pending or unavailable, without implicitly checking for a new release. - Accept the additional command-display quoting used by newer tmux versions when proving an older macOS updater's ownership. Decode at most one extra serialization layer, then retain exact kernel argument, executable, ancestry, operation and authenticated idle checks before stopping the old service. - Five renderer regressions fail before the fix; 285 related tests and type checking pass afterward. The real private tmux launch regression reproduces the old proof failure on tmux 3.7 and passes with the correction. All 23 helper tests pass on the native framework-Python host; the release interpreter and separate native guest pass with one framework-specific skip. - Availability: committed corrections awaiting a new signed candidate and its packaged migration, failure/retry and rollback acceptance. No stable 1.0.4 release is published from the superseded candidate. ## 2026-09-21 — Prepare the stable coordinated-update bridge - Target the existing stable 1.0.3 app and managed server. The user keeps one app-update action; the updated app requests the matching server automatically, with npm handled by the managed updater. Retain both old stable download channels and the actual app version in About. - Reproduce an older macOS migration rejection through the unchanged published app and an original signed beta.29 server. Verify that its original process, runtime, identity, chats, credentials, Hub data and peer connection survive. Add a read-only native ownership proof for old updaters without a recorded process ID, including the exact Homebrew Python framework launcher mapping. Retain all existing authenticated identity, idle and service ownership checks. - Fix the reproduced stale Updating row for legacy servers that omit update progress from health. Observe an owned active operation through existing health callbacks, and stop when it pauses. Preserve a failed status when opening recovery; checking again remains an explicit action. - Validate the focused coordinator and renderer regressions, TypeScript and production compilation. Native process/tmux proof tests cover ordinary and framework Python. Isolate inherited Hub configuration in four installer test fixtures after reproducing their failures on unchanged published source. - Availability: source candidate. Exact signed stable package, native stable 1.0.3 upgrade, failure/rollback, and public distribution acceptance remain required before release. Beta.12 publication is not stable rollout acceptance. ## 2026-09-21 — Clarify the bridge to npm updates - Document the first coordinated release as a bridge delivered through the existing desktop feeds and signed standalone server updater. Existing managed users update the app; they do not run the fresh npm installer over their data. - Separate the accepted beta.8-app/beta.9-server journey from older unsupported installations, pre-1.0 feed migrations and custom-path macOS prerequisites. Preserve stable/beta channels and retain legacy downloads during transition. - Correct the migration guide and beta.12 release notes to reflect publication. Check the instructions against the shipped coordinator, updater, original one-click acceptance and public-distribution verification. No runtime or released artifact changes accompany this documentation update. - Replay the released beta.29 updater contract in an isolated fixture. Its macOS runner lacks the new installer's admitted ownership proof, and its latest-only selection prevents using the old API to pin an intermediate release. The Linux managed-update environment passes this admission check; neither result establishes a complete native beta.29 migration. Keep that older starting point outside the accepted automatic-migration claim. ## 2026-09-21 — Publish coordinated desktop/server beta.12 - Publish direct desktop `1.0.4-beta.12` build `1189` on the public AgentsDock repository and its compatibility release mirror, with the accepted artifacts from source `2741c05a0772849f6da82f944789060b77ecb91d` unchanged. - Publish `@agentsdock/server@1.0.4-beta.12` to the npm beta channel and the signed standalone migration bridge to AgentsServer. Verify the public npm archive against its signed size, SHA-256 and SHA-512 integrity, then verify all 106 runtime files and modes against the public legacy archive before exposing either desktop release. Both desktop mirrors retain the accepted checksum manifest and signed paired-server descriptor. - Configure npm trusted publishing for the public repository's protected `server-npm-publish.yml` workflow and `npm-release` environment. Existing installations retain the signed legacy migration path; the packaged app requests its matching server automatically after updating. - macOS is signed and notarized; Windows remains unsigned. Initial migration and execution-runtime replacement wait for idle. Running native goals on old servers can retain execution ownership between replies, so a final reply alone does not guarantee an idle migration window. ## 2026-09-21 — Accepted desktop beta.12 build 1189 candidate - Accept the direct desktop 1.0.4-beta.12 build 1189 from committed source `2741c05a0772849f6da82f944789060b77ecb91d`. Native workflow `35661716190` passes Linux x64, Linux arm64 and Windows x64 build and package verification. Windows remains unsigned. The local universal Mac release passes 4,646 tests, type checking, compilation, Developer ID signing, Apple notarization and stapling, Gatekeeper, mounted-DMG/ZIP parity, updater metadata and blockmap checks, and an isolated startup. Five existing tests are skipped. - Retain all release checks. Fix two asynchronous UI tests to wait for their rendered result or effect callback, compare signing keys independently of checkout line endings, and cap release test concurrency at four workers. - Bind the app to signed npm descriptor SHA-256 `18a4bc7c54dc749b93235bda4e0c3247123e03d85d7b2e2dc525087b7014b5b5`. The npm archive is byte-identical to the server candidate that passed all eight CI shards (4,971 cases), real Codex and Claude foreground/subagent gateway-loss checks, and native Linux/macOS migration and recovery checks. Both signed distributions contain the same 106 runtime files and modes. - Verify the production-signed legacy archive with the original trust key on native Linux and macOS. Preserve identity, credentials and saved state through forward migration, updater/installer loss and automatic rollback. An abrupt macOS VM power loss at the durable activation boundary recovers automatically after reboot without an HTTP recovery trigger or manual repair. - Verify one real update click in the unchanged published beta.8 app: Squirrel replaces it with the exact build 1189 ZIP and relaunches it automatically. The new packaged coordinator migrates the genuine beta.9 server to beta.12 without a separate server-update click. Check the installed signature, executable, application archive and descriptor against the accepted package; retain the same saved profile, server identity and credential, with both components current and execution admission released. Discovery uses a private HTTPS fixture; public release propagation remains a separate publication check. - Close the populated-data acceptance gap before publication. A genuine Linux beta.9 Hub host survives an interrupted signed upgrade, automatic rollback and retry with its saved chat/events, Hub records, approved peer, keys, provider paths and credential files preserved. A separate joined-server migration keeps its active mutual-TLS connection and content access without pairing again. The Mac host migration preserves both Codex and Claude histories, Hub messages, board content and peer authority; existing peer credentials authenticate saved reads and new writes. Provider credential contents are synthetic preservation fixtures, while Hub/peer authentication runs against the actual native services. Verify retired bootstrap authority remains retired; a revoked proof file is not required to survive snapshot recovery. These checks need no product edits. - This records accepted build artifacts before desktop upload. Public publication and the live Studio upgrade are still pending. Execution-runtime replacement waits for idle; simultaneous execution generations are not claimed. ## 2026-09-21 — Prepare updates during work and recover interrupted activation - Stage and verify server dependencies while agents continue working. Preserve the existing pending-update protocol for older clients, then acquire the exact idle execution hold before activating the prepared candidate. - Extend the existing installer transaction to both native services, retaining configuration, previous runtime and state recovery. Bind automatic recovery to the admitted candidate and journal. Register an independent native recovery job before stopping the main services, so recovery also runs while the app cannot connect. Join that owner through the existing update action. Keep incomplete recovery fenced and distinguish verified rollback from successful installation. - Require both component versions and released execution admission before reporting completion. Exercise the coordinator and existing updater endpoints, exact transaction recovery, failed launch, stale ownership and rollback results. - Verify one real Codex turn and one real Claude turn while actual dependency preparation and receipt validation run. Each original foreground command and provider stays alive and completes exactly once, with ordered event delivery; execution admission remains open. These isolated tests do not activate an installed release or establish signed delivery acceptance. - Real fresh installations and migration from the released legacy updater have passed on disposable Linux systemd and macOS launchd hosts. A Linux fault test kills the updater and installer after service shutdown, then verifies that the independent native owner restores the previous installation and reports a failed, retryable update without an HTTP trigger. Native testing exposed and fixed directory permissions, generated-cache validation and recovery ownership. - Repeat native acceptance with credentials pinned to the connected process before transmission and runtime durability checked before service shutdown. An abrupt macOS VM shutdown during activation restores the previous release automatically after reboot, retaining identity, credentials and saved state. Correct enablement parsing for both launchd output formats found during testing. - Exercise the production renderer, preload, native transport and coordinator in an isolated offscreen Electron app against an installed Linux server. Add and authenticate it through the UI, request the signed npm candidate, and verify automatic reconnection, both updated components, released admission and the Up to date result. Registry, signing key, version enrollment and credential storage are controlled QA boundaries; no app binary replacement is asserted. - Repeat a genuine beta.9 migration after a prior rollback, then abruptly stop the Linux VM after the old service is disabled. After reboot, the independent systemd owner automatically restores beta.9, retains identity, credentials and saved state, and retires its recovery job without an HTTP recovery trigger. Cover carried-over legacy intent and pre-arm retry failures with regressions. - Fix the Linux lock-inode reuse and closed-transport races exposed by the full CI suite; rerun their regressions on both Linux and macOS. Final signed-package acceptance remains pending. These results do not establish simultaneous execution generations. No release has been published or deployed by this entry. ## 2026-09-21 — Verify both server components before update completion - Keep a coordinated update incomplete until both the gateway and execution runtime report the paired release. Reject inconsistent component health, retain failed-operation recovery, and observe gateway changes independently of the execution process's boot identity. Existing single-process servers retain their compatibility path. - Pass focused coordinator, update and restart settings regressions, TypeScript checking and production compilation. Exercise the current production renderer, preload, service, native transport and coordinator in an isolated offscreen Electron window against real worker and gateway processes from the committed execution foundation, with controlled release version files and a QA signing key. Add and authenticate the isolated server through the UI, open Updates, verify the incomplete result, replace only the gateway, and explicitly refresh. Verify the same worker and server boot, both current component versions and the resulting Up to date row. All owned processes exit and no model turn runs. - The QA harness substitutes credential storage and release enrollment. This establishes component-status handling, not installed migration, app binary replacement, automatic reconnect-only behavior or production signing. The installer and runtime migration integration remains in development; this change is source only and no release has been accepted or deployed. ## 2026-09-21 — Persistent execution foundation (source only) - Separate the public gateway from the process owning chats, provider transports, pending approvals and tool execution. Preserve authenticated request semantics, event ordering and private provider callbacks during gateway replacement. - Give gateway and execution independent process/release identities. Keep the actual execution version visible while an older runtime remains active. Retirement requires an idle worker and a durable admission hold; an API restart does not close provider managers or cancel accepted commands. - Complete one disposable real turn each with Codex and Claude. Their foreground tools survive both graceful and forced gateway termination, then the original turns complete without duplicate execution. Reconnected WebSockets receive the complete ordered event sequence. Check both CLI logins remain valid. - Verify separate native subagent runs for both providers: one child continues through both gateway replacement modes and completes its tool once. Claude's unanswered tool approval retains its request identity through another restart and resolves once after reconnect. - Lock chat state before loading it or sweeping provider children, across both maintained entry points. Test both startup orders with actual processes: a competing server is refused while the incumbent and its registered controlled child remain intact. Retain ownership through shutdown stragglers. - Start a copied production runtime through a real pending activation journal. Verify recovered queued turns and due jobs reach their admission checks and remain deferred, with no provider launch. Reject an incorrect release of the admission hold; permit a normal zero-turn mutation after the exact release. - Add transport, admission, recovery, controlled subprocess and production application regressions. Verify existing managed-service proof and pending update admission behavior in isolated state. Package the seven execution modules in both server distributions and retain explicit single-service installer protection for experimental split installations. - Exercise real launchd and user systemd replacement with controlled application and child-process fixtures: gateway upgrade, failed gateway rollback, busy worker refusal and exactly one approved side effect. Retain the worker and chat identities. Fix systemd working-directory rendering and exact file-mode restoration exposed by these native tests; restore the disposable baseline. - Exercise the existing built desktop through an isolated offscreen Electron window with production renderer, preload, IPC and server transport. Connect a server, create a Claude chat without a turn, navigate owned history and use explicit reconnect after both gateway replacements. Keep selected chat and final text. Credential storage is substituted in this QA harness; this does not establish automatic reconnect-only or packaged-feature-build acceptance. - These changes do not yet enable rolling execution generations or the normal npm/app migration path. The application routes still live in the retained execution process. Native service fixtures and desktop checks do not establish production state/Team Hub migration. No release build is accepted, published or deployed by this entry. See [the implementation boundary](PERSISTENT_EXECUTION.md). ## 2026-09-21 — Integrate completed desktop work into main - Merge the completed desktop branch through `8d7745d` with main `33f9355`, including the Side chat popup, copy and scroll behavior, provider settings, reasoning controls, Claude subagent visibility and live synchronization fix. - Retain newer main changes for analytics privacy, optimistic send feedback, shared-chat recovery, artifact-open errors, media layout and release workflows. Combine live reasoning overlays with pending-send presentation. - Pass all 4,634 desktop tests (five intentional skips) with four workers, TypeScript validation, eight compile/license checks, production compilation and the compiled-entry guard. - Re-exercise the merged production renderer/preload in native offscreen Electron with production IPC/service/native HTTP: Side chat opening and sizing, native copy/paste, scroll restoration, independent chat positions, pending close/reopen, cancellation and a subsequent question. Check light/dark and narrow layout; restore the clipboard. These requests use a controlled loopback response server; live model execution is covered by earlier feature acceptance. - Exercise the merged desktop and isolated real server together with synthetic histories and controlled health failures: live updates recover in 727 ms during continuous native typing, including an injected 700 ms server delay. Preserve all 48 draft characters and retain deferred metadata application. - Availability: source integration. This does not create a new app package or update a running server; paired server changes are integrated separately. ## 2026-09-21 — Resume live chat updates during input — source acceptance - Resume requested chat subscriptions immediately after the server is healthy, before waiting for the foreground input pause used by background session/job metadata. Preserve scope and subscription ownership; hidden chats stay closed. - Pass six focused service checks, TypeScript validation, production compilation and compiled-entry verification. The regression fails on the old ordering. - Exercise actual offscreen Electron with production service, packaged renderer and preload, native sidebar clicks and typing, and an isolated real server. Synthetic histories contain 5,000 and 300 events; normal cached switches are already fast and are not reported as a reproduced stall. - Inject one HTTP health failure and a 700 ms healthy response delay. On the old service, selected-chat recovery waits 8.93 seconds, including three seconds after typing stops. The correction recovers in 732 ms during continuous input; an independent repeat records 729 ms, preserves all 49 typed characters, and leaves background metadata deferred. No provider turn is part of this check. - Availability: accepted source correction for the coordinated desktop release. A separate compatible server optimization avoids redundant full-history fork scans on timeline refreshes. Production apps and servers remain unchanged. ## 2026-09-20 — Revised beta.12 package and public release validation - Build the revised server package from clean committed source `751c1e0`. Its SHA-256 is `bc69cb8817d3f085330306a463f61b0353186193b2357600f1613f56288995ea`. The standalone export preserves upstream history and all 88 runtime files; packaging that export through npm produces the identical archive. - Pass native macOS fresh-install retry over the exact failed candidate's empty folders without cleanup. Refuse another install without changing the running process, identity, token or synthetic provider files. Pass candidate activation and forced incompatible-API rollback with the exact prior runtime and plist. - Pass the app-driven Linux update and rollback through the production renderer, preload, IPC, coordinator, systemd service and detached installer. Verify all 88 installed runtime files, identity, token and six synthetic state/history files. The failed update remains paused after repeated health refreshes, and the isolated HTTPS registry records only the two intended package downloads. Inspect dark and light layouts, including the minimum supported window width. - These results cover disposable native services and controlled app-replacement, signing-key and distribution endpoints. Production-signed app replacement, public registry transport and live provider work are not established by them. - Exercise the complete legacy bridge route from an old managed server without npm update support. The production app checks and starts a signed legacy update; the old detached updater installs the paired runtime and reconnects with npm update capability. The app marks the equal-version bridge current without a redundant npm download. Discovery and signing endpoints are controlled within the disposable guest; no public legacy release was published. - Correct source CI to use runner paths in step environment variables and run for maintained release branches. All eight public server test shards pass. Update legacy release assertions for protected public workflows and reviewed source-branch ancestry; retain release identity and mirroring checks. - Correct settings test fixtures to provide the required typed app-update status, settle initial loading and distinguish app controls from server controls. Pass 4,630 Electron tests with five existing skips, eight script tests and type checking. Retain all server operation and recovery assertions. - Reproduce delayed app-update status dismissing an already open server restart confirmation or clearing a restart error. Reset these controls when Settings opens, preserving user actions while status finishes loading. Keep server polling and profile/boot checks unchanged; add regressions for both cases. - No released app build or public npm version is accepted by this entry. Native signing credentials still need to be supplied to the public release environment, and signed publication checks remain pending. ## 2026-09-20 — Current server integration and transition signing — candidate follow-up - Merge the six newer commits from the maintained standalone release branch, preserving automatic Codex/Cursor chat titles and the shared-chat Cursor fix. Preserve the updated shared-browser bundle. The earlier `8a52b14` archives below are historical QA artifacts and are superseded for release preparation. - Include automatic title requests in update and restart blockers. Prevent new optional title requests after update admission closes, and allow unstarted queued requests to retry on a later turn. Verify an actual disposable provider subprocess delays update advancement and that shutdown reaps its process group. - Pass 291 targeted title, provider-background, update and restart tests. Pass another 189 focused provider-isolation, storage, shared-chat, terminal and packaging checks, plus 33 Node publication/staging/CLI checks. Reproduce the terminal cancellation regression with a single-worker executor. - Make npm packaging work in both the combined repository and standalone compatibility export. Select legal notices from the checkout boundary, reject unrelated parent files and retain exact canonical license copies in the export. Nine packaging tests include real offline npm archives in both layouts. - Include the license and notice in both published server distributions and installed runtimes. Pass 16 focused packaging/manifest tests, including actual npm and legacy archive comparison of the legal files and their permissions. - Reproduce macOS device-number changes across reboot breaking interrupted activation recovery. New journals bind their filesystem coordinates to a persistent volume UUID while retaining inode, ownership, content and live race checks. Negotiate the new guard-path option with older recovery helpers. Pass 75 activation/UUID tests and 17 installer recovery tests. In a disposable macOS VM, interrupt the real installer, reboot across an actual device-number change and verify unchanged installer retry restores the previous runtime, exact service plist, identity, token and six synthetic state/history files. A second orderly reboot retains that rollback and starts the restored service. Legacy journals without saved volume proof still require manual recovery if their device numbers changed; specialized interrupted Hub reactivation after remount remains unsupported. Do not describe those boundaries as accepted. - Prepare transitional signing through the standalone repository's existing release secret. Its prepare-only workflow can produce both signed server distributions after all server test shards pass, without publishing them or moving the private key. Publication and native acceptance remain separate. - Move desktop signing and publication automation into public AgentsDock. Keep signing credentials in the branch-restricted `direct-production` environment and npm OIDC in `npm-release`; ordinary CI and fork pull requests receive neither. Rebase the public workflow's native build counter above 1185 and retain exact source, signature, immutable-asset and server-runtime checks. Preserve the private repository's history and retire its release workflows when the public pipeline becomes the active publisher. Secret values must be supplied again from their original source; they have not been copied or logged. - Pass 60 release-orchestration tests, parse both public native workflows and check all 50 shell steps. Verify manual/canonical/trusted-branch guards on all 13 jobs and the release environment on all seven jobs that use secrets. - Reject the packaged `d9c1f50` candidate after a pristine macOS install exposes a missing LaunchAgents parent during volume binding. Bind a safe existing ancestor until publication creates and verifies the destination directory. Pass 77 activation tests, including missing-parent recovery checks. - Allow retry after that failure without deleting the empty configuration and state/admin directories it leaves behind. Both launcher and locked installer reject existing data, credentials, links, locks and registered services; fresh installation creates no legacy migration alias. Pass 12 CLI tests, 14 installer admission tests and two actual installer regressions from a clean source snapshot. Exact-package macOS retry acceptance remains pending. - Availability: committed source candidate. Updated packaged migration and recovery verification are in progress; no public release or production service has changed. ## 2026-09-20 — npm publication and native migration validation — beta.12 candidate - Reserve `1.0.4-beta.12` for the coordinated candidate. Do not publish the earlier beta.9 QA package under an already-used server release version. - Add manual unsigned preparation and protected OIDC publication of an exact signed npm candidate. Verify reviewed source, accepted descriptor hash, signature, package identity, immutable version, channel and registry bytes. Keep private signing separate; inspecting or signing a candidate does not establish native acceptance or publish a desktop release. - Reproduce a failed fresh installation leaving only empty runtime folders. Allow the npm launcher to retry only safely owned empty scaffolding, while retaining rejection of state, configuration, files, links, releases, locks and registered services. Delete no existing data and retain the installer's repeated admission check under its lock. - Exercise Update through the actual production renderer, preload, IPC, service, coordinator and native HTTP in isolated offscreen Electron against a disposable Linux systemd service. Verify the detached updater downloads a signed HTTPS archive, validates it, activates the candidate and reconnects with the same identity and token. Preserve six synthetic state/history files. - Send an intentionally incompatible signed API contract through the same desktop path. Observe candidate activation, rejection and real rollback to the prior runtime. Verify the UI pauses with an explicit retry action and repeated health refreshes do not download or install it again. Preserve an independent offline profile and inspect light/dark minimum-width layouts. - Separately exercise real macOS launchd in disposable virtual machines: legacy installer to candidate, wrong-API health rejection and restoration of the previous runtime and exact service plist. Verify identity, token and six synthetic provider/configuration/history files remain unchanged. Fresh installation through actual offline npx succeeds; a second installation is refused without changing the running service. - Test boundaries: ephemeral signing key and guest-only HTTPS registry for Linux; controlled app download/replacement; synthetic provider data rather than live model work. macOS dependency caches are preloaded after guest outbound network failure. These tests do not establish public npm transport, production-signed packaged-app acceptance or actual app replacement. - Pass 33 focused Node tests covering publication, packaged metadata and the npm CLI, plus five actual npm packaging tests and nine installer admission tests. Preserve explicit unsupported-boundary notes instead of treating a dry run or simulated app replacement as an accepted public release. - Prepare the exact beta.12 npm archive from committed source `8a52b14` in a clean detached checkout. Its SHA-256 is `ce392cf842774eb55fcd889a36e5c875d18e7a70551a77163744413fd9252241`. On a third pristine macOS VM, reproduce the actual failed first install and retry with this unchanged archive through npx, without removing the leftover folders. Verify beta.12/API 28 activation, then refuse repeated installation while preserving the running process, identity, token and provider sentinels. - Keep native build/draft staging possible before registry publication. Final release verification requires both the signed npm package and matching legacy bridge to be public, with identical runtime files and executable bits. - Verify the compatibility export preserves the standalone repository's ancestry and exact server tree. Build the beta.12 legacy archive from that export and compare it with the committed npm archive: all 85 runtime files and their executable permissions match. Both manifests remain unsigned until production signing; these local archives have not been made available to installed users. - Availability: committed source candidate after focused validation. Registry publication, trusted-publisher execution and the complete signed native release remain pending. Production services and CLI credentials are untouched. ## 2026-09-20 — Coordinated npm updates — source candidate - Import the maintained server under `server/` with its complete history. Retire the frozen snapshot and its legacy Swift server-text assertions. Verify that the initial subtree export reproduces the original standalone commit; require subsequent compatibility exports to preserve ancestry and exact contents. Keep legacy signed downloads available during migration. - Prepare `@agentsdock/server` from the exact runtime allowlist, with no npm installation hooks. Stage a separate signed descriptor tying the app's public version to an immutable npm archive, integrity hashes and API contract. Keep source package metadata private and publication disabled in preparation. - Add authenticated, identity-bound reconciliation through the existing managed updater. Persist signed bytes across queued work and restart, queue while busy, and validate candidate identity, version and API before activation commits. Fresh npm installation refuses existing state and services, including a repeated check under the installer lock. - Add desktop coordination with durable per-server receipts, independent offline recovery, explicit enrollment, exact downloaded app version pinning and a compatibility gate before restart. Ordinary unenrolled builds keep the existing update behavior. The app and server retain their native packaging formats. Failed or canceled owned attempts stay paused until an explicit scoped retry; enrolled releases keep legacy controls under Advanced server recovery. - Pass 233 focused server checks, including real HTTP authentication and identity guards, signed metadata, queued-work recovery, installer protection and candidate health rejection. Verify actual offline npm packing, CLI native transport, exact payload bytes and executable permissions, paired artifact staging and Git export rejection on divergent history. - Pass 722 affected desktop tests, TypeScript validation and production compilation. Exercise the production renderer, preload, IPC, profile service, updater, coordinator and native HTTP against an isolated production FastAPI server in native offscreen Electron. Click Update, reopen a second process, preserve the queued receipt, display an independent offline profile, pause on failure/cancellation and retry explicitly. Verify dark/light minimum-width layouts and no automatic legacy release lookup. Feed/download, signing key, provider work, server activation and app quit are controlled test boundaries; simulated completion is not recorded as a real managed update. - Separately exercise the real installer in a disposable Ubuntu systemd VM: legacy beta.9 to guest-stamped beta.12, then a deliberately incompatible beta.13 candidate rolls back to beta.12. Authenticated health verifies exact version and API, stable identity and preserved token. Six synthetic provider, configuration and history/state files remain byte-identical. This validates Linux service activation and rollback, not real provider sessions or the full registry-to-app update journey. - Install the committed local npm tarball through actual offline `npx` in a second disposable Linux user account. Verify its independent real systemd service, identity and token; a second fresh-install attempt is refused and both services remain unchanged. Package retrieval from the public registry and same-user multiple-server installation are not claimed by this test. - Compile the legacy Swift guardrail executable successfully. Its unchanged React mobile source-text assertion still fails before later checks; this is not recorded as a passing full Swift guardrail run. - Availability: source candidate only. macOS launchd migration and rollback, active real-provider work and retained live chat data, registry publication, the complete coordinated upgrade and packaged native acceptance remain required before a coordinated release. No production service or published release is changed by this source work. ## 2026-09-20 — Side chat scroll memory — 1.0.4-beta.16 local acceptance - Accept signed local Apple silicon macOS app **1.0.4-beta.16 / 1190** from `9451baae691a2f9ad93bbef8e05b60c3e1a38a12`. Verify bundle audit, Developer ID signature, runtime entitlements, exact version/build and isolated startup. All 88 compiled files in the package match the tested production output. - Exercise the exact packaged renderer/preload with native offscreen mouse input. Verify exact scroll restoration after reopening, stable reading position when replies arrive, bottom following after a reply arrives while closed, Jump to latest and independent positions across two chats. The controlled server and unchanged provider boundary are described below. - Availability: signed local `.app` with automatic updates disabled, not a notarized public release. This scroll correction needs no server update. ## 2026-09-20 — Remember Side chat reading position — source acceptance - Restore each Side chat's reading position across closing/reopening and chat switches. Keep positions scoped to their server and parent chat, and reset them when the side conversation is cleared or its server is removed. - Open new conversations at the bottom. Follow replies while already at the bottom; preserve the reading position while scrolled up and offer a compact Jump to latest control. Restore after the popup measures its available space. - Pass 38 focused component/controller/layout checks, TypeScript validation and production compilation. In native offscreen Electron, exercise actual scrolling, close/reopen, delayed replies while reading older text, replies arriving while closed, Jump to latest and independent positions in two chats. The controlled server exercises production preload/IPC/native HTTP. Provider execution and server-picker transitions are outside this acceptance; a focused ownership check covers server identity changes and revisits. - Availability: source correction. No server change is needed for scroll state. ## 2026-09-20 — Side chat button spacing — 1.0.4-beta.15 local acceptance - Accept signed local Apple silicon macOS app **1.0.4-beta.15 / 1189** from `d3e4877a6b83eae994e4ee910d0e998808465850`. Verify bundle audit, Developer ID signature, runtime entitlements, version/build and isolated startup. All 88 compiled files in the package match the tested production output. - Reproduce overlapping controls in beta.14, then exercise native offscreen scrolling and clicks in the corrected source and exact packaged renderer. At normal and narrow widths, Side chat sits 30 pixels lower with a clear gap beneath Jump to latest. Both controls work, and the popup opens above the unobscured composer. The controlled timeline uses production preload/IPC and native HTTP; provider execution is outside this layout-only acceptance. - Availability: signed local `.app` with automatic updates disabled, not a notarized public release. No server update is required. ## 2026-09-20 — Side chat button spacing — source correction - Lower the Side chat button into the folder row above the message composer, separating it from the timeline's jump-to-latest arrow. Reserve room beside the folder control and keep the button clickable above the composer layer. - Pass 30 existing component/layout checks, TypeScript validation and production compilation. No server change is required. ## 2026-09-20 — Side chat copying — 1.0.4-beta.14 local acceptance - Accept signed local Apple silicon macOS app **1.0.4-beta.14 / 1188** from `3a6905319b229aaa08f1e3012af09c2d08756ef3`. Verify bundle audit, Developer ID signature, runtime entitlements, version/build and isolated startup. All 88 packaged compiled files match the tested production output. - Reproduce disabled selection in beta.13. In the corrected source and exact beta.14 packaged renderer/preload, use native offscreen mouse dragging to select user messages, assistant prose and inline code. Native copy commands produce the exact selected text; paste inserts it into the composer while the popup remains open. Restore the original clipboard after verification. - Keyboard verification uses Meta+C/Meta+V with Chromium native edit commands; the hidden window does not exercise macOS global menu accelerators. The production menu retains its standard copy/paste roles. The controlled server exercises production native transport; no model-provider execution changes. - Availability: signed local `.app` with automatic updates disabled, not a notarized public release. No server update is required. ## 2026-09-20 — Side chat text selection — source correction - Restore normal text selection in Side chat history so user messages, assistant replies and inline code can be copied with the native shortcut. The popup no longer inherits the app chrome's selection-disabled style. - Pass 27 existing component/theme checks, TypeScript validation and production compilation. Native clipboard and packaged-app acceptance are recorded above. No server change is required. ## 2026-09-20 — Compact popup — 1.0.4-beta.13 local acceptance - Accept the signed local Apple silicon macOS app **1.0.4-beta.13 / 1187** from `c3aab200f66c5b36526b899b8e087658c543711a`. Verify bundle audit, Developer ID signature, runtime entitlements, version/build and clean isolated startup. All 88 packaged compiled files match the tested production output. - Exercise the exact packaged renderer, preload and CSS in native offscreen Electron with keyboard focus enabled. Verify compact empty presentation, growing drafts, long-answer scrolling, Clear, help, close/reopen and Escape in dark/light themes and a narrow window. The controlled server and service harness preserve the boundary described below; no model provider runs. - Availability: signed local `.app` with automatic updates disabled, not a notarized public release. No server update is required for this layout change. ## 2026-09-20 — Compact Side chat popup — source acceptance - Size the popup to its content instead of reserving a full-height empty panel. Start with a single-line composer and grow it with the draft. Keep long conversations scrollable within the existing maximum popup height. - Remove the duplicate input focus outline, manual resize grip and repeated explanatory text. Keep one subtle composer focus treatment, a circular Send button and direct Clear/Close icons. Expand context help inline when requested. - Pass 27 existing component/theme checks and TypeScript validation. Exercise native offscreen Electron with production CSS ordering and keyboard focus: empty and long drafts, long-answer wheel scrolling, Clear shrinking the popup, context help, close/reopen and Escape, dark/light themes and a narrow window. Requests use production preload, IPC and native HTTP into a controlled local server; provider execution is unchanged and outside this visual acceptance. - Availability: source correction; no server change is required. ## 2026-09-20 — Side chat popup and Claude agents — 1.0.4-beta.12 local acceptance - Accept the local Apple silicon macOS app **1.0.4-beta.12 / 1186** from `a83a18e85e0b6207f5583e317dc00f902076912b`, including the popup and Claude subagent corrections described below. Pass 381 service/projector checks, 100 focused popup checks, TypeScript validation and production compilation. - Verify Developer ID signing, bundle audit, hardened-runtime entitlements, exact version/build and clean startup with isolated user data. All 88 packaged compiled files are identical to the accepted production output. - Exercise the exact packaged renderer and preload in native offscreen Electron through production bootstrap, store, IPC and read-only HTTP: opening/reopening Claude agents, opening the popup, retained drafts, direct Clear and Escape. The harness compiles the service from the same committed source and suppresses read receipts. The signed main binary is checked separately at startup. Real Claude side-question acceptance precedes packaging. - Availability: signed local `.app` with automatic updates disabled. This is not a notarized public release, cross-platform acceptance or server deployment. Claude subagent visibility works with the existing server; the separate side question configuration correction still requires a server update. ## 2026-09-20 — Claude subagent refresh — source acceptance - Fetch authoritative subagent state when opening Claude chats, as already done for Codex. Seed live tracking from that state and retain progress and completion in the local cache, including native `task_updated` messages. - Exclude explicitly identified background shell and workflow tasks from the agent list. Reject older snapshots and replayed events after newer activity so reopening cannot roll an agent's status backward. - Reproduce the missing snapshot request in isolated native offscreen Electron. Verify the corrected full app through production preload, IPC and native read-only HTTP: cold-open a real Claude chat, inspect its active and historical agents, open details, navigate away and reopen. The authoritative state and agent activity persist; background shell tasks do not flood the list. - This read-only check does not launch a new provider agent. Focused service checks cover progress, both completion formats and stale-event races. - Availability: source correction using the existing server API. No server update is required for this subagent visibility fix. ## 2026-09-20 — Side chat popup — source candidate - Move Side chat to a single button beside the composer. Open the conversation in a floating popup without changing the main chat width, including split chat panes. Keep Clear and Close directly accessible in its header. - Preserve side conversations, pending answers and drafts when the popup is dismissed. Retain the existing per-chat and per-server ownership rules. - Pass 100 focused checks, TypeScript validation and production compilation. Exercise the full app in isolated native offscreen Electron: popup placement, input focus, Escape and outside-click dismissal, reopening drafts and answers, direct Clear, pending request retention and cancellation, split-pane isolation, dark/light themes, narrow layout and Chinese text. Requests cross production preload, IPC and native HTTP. Also exercise real native Claude through an isolated production server: ask about a fact present only in a completed tool result, ask a contextual follow-up, close/reopen and clear the popup. The running parent stays active, and side requests do not alter its transcript. The test bootstrap uses a seeded native profile; full server-picker setup is outside this check. - The paired server correction preserves the connected Claude parent when saved effort settings change for a future turn. Existing servers require that correction to avoid the related side-question configuration conflict. - Availability: source candidate. No published build or live production server is changed by this acceptance. ## 2026-09-20 — Compact running command blocks — 1.0.4-beta.11 local acceptance - Keep the active Codex tool inside its compact command group. Update the group's single row to the latest running call and retain previous calls behind its disclosure. Preserve commentary boundaries and visible reasoning chronology. Only the current activity pulses. - Hide extra trace-history controls in compact live Codex turns. Keep manual pagination in expanded live traces and explicitly opened completed history. - Retain live-only reasoning display and identical collapsed completed/stopped history. This supersedes the undelivered beta.10 candidate, whose running call could still appear beside a separate completed-command group. - Pass 4,598 source tests (10 skipped), 164 focused timeline checks, TypeScript validation and production compilation. Verify real native Codex commands separated by reasoning within one commentary block, including a delayed second call: one compact pulsing row retains both calls. Confirm a later commentary creates its own chronological block. Exercise the actual Settings entry and switch, completion, stop, unchanged completed history, reduced motion, both themes and no automatic trace requests in isolated native Electron through the production server and authenticated transport. Provider Responses are controlled fixtures; this is transport and display acceptance against the supplied visual reference, not external-model output or native GUI pixel parity. - Accept signed, notarized universal macOS **1.0.4-beta.11 / 1185** from `494c91fd7236a430d498cb90481fc06697b5a500`. All 88 compiled files match the frozen source, and the final ZIP's ASAR matches the exact packaged full-app completion and interruption replays. Verify actual Sidebar → Settings interaction, compact command blocks, no extra default live controls, identical terminal history, reload, reduced motion and supported minimum width. The packaged replay uses captured native events and offline IPC; production transport was exercised separately in the source acceptance. - Pass Developer ID signature, Gatekeeper, stapling, universal architecture, version/build, ZIP updater hashes and feed checks. The actual signed app launches cleanly for ten seconds with isolated user data and is then closed. Matching ZIP SHA-256: `60792d4c880914d9c05c7769289532b6d9ffe93d972fce94071d2325e923f3e9`. - Availability: accepted local `.app` and matching ZIP. Public desktop publication remains blocked by the private Actions budget. Local app acceptance does not imply cross-platform or DMG installer acceptance. ## 2026-09-19 — Live-only reasoning display — 1.0.4-beta.10 candidate - Apply the reasoning display preference only during an active Codex turn. Collapse finished and stopped turns under both settings; changing Settings leaves completed history unchanged. Explicit history expansion retains all available text in chronological order, independently of the preference. - Group adjacent commands across hidden reasoning entries. Retain visible commentary and reasoning boundaries instead of moving or dropping content. Update the English and Chinese setting descriptions. - Exercise completion and interruption through actual sandboxed Codex, an isolated production server, authenticated WebSocket, desktop service, preload and timeline in native offscreen Electron. Click the real Settings switch while running and after completion; verify both settings converge to collapsed history, manual expansion retains text, and hidden reasoning produces one command group. Check dark/light narrow layouts and no automatic history requests. Responses are controlled fixtures; this validates native transport and presentation, not external-model output or native GUI parity. - Pass 4,594 tests (10 skipped), eight build/license guard checks, TypeScript validation and production compilation. Keep release verification temporary extraction under the configured temporary directory, with a writable-path check. This presentation change requires no server contract update. - Availability: source candidate; signed local package acceptance is pending. ## 2026-09-19 — Codex activity and per-chat limits — 1.0.4-beta.9 - Present Codex commentary and command rows inline, with one muted pulsing current activity. Stop animation on completion or interruption and honor reduced motion. Preserve earlier summaries under a compact disclosure. - Add a persistent **Show reasoning traces** switch in Settings → General. Keep compact presentation by default; optionally expand summaries and separately labeled plaintext supplied by Codex. Do not decode encrypted content or imply that unavailable reasoning can be recovered. - Carry the distinct plaintext event through live transport, timeline projection, completion, interruption and history. Keep transient updates outside durable cache/cursors and distinguish summaries sharing an item ID. Verify the previous desktop renderer ignores the new plaintext event even when its trace is expanded, while continuing to display ordinary summaries. - Add optional Codex and Claude sub-agent limits in the chat Inspector. Fence saves to the original server identity and generation, reject old servers before mutation, and retain drafts after failure. Saving during active work is allowed; explain each provider's application boundary. - Verify the actual Settings entry and toggle, native Chromium animation, reduced motion, complete text expansion, persistence, and light/narrow layouts in isolated offscreen Electron. Exercise limit saves and clearing through production preload, IPC, HTTP authorization and server persistence; preserve a sibling chat and active-work status and reject unauthenticated writes. Provider execution is verified separately in the paired server. - Pass 4,585 source tests (10 skipped), eight build/license guard checks and TypeScript validation, including scoped and unscoped preload compatibility. - Accept local universal macOS **1.0.4-beta.9 / 1182** from `1641cb97dfba17b6c3b3807c79c0f63e4a36cca3`. Developer ID signatures, notarization, Gatekeeper, DMG/ZIP parity and updater checks pass. The signed executable passes its clean CI launch; all 88 packaged compiled files match the frozen source. Replay the native capture through the exact packaged renderer and preload, including Settings, persistence, reduced motion and completion. This packaged replay uses offline fixture transport; native provider and authenticated transport checks are recorded separately. - Verify the copied app on a second Mac: matching archive SHA-256, version, build, deep signature and Gatekeeper acceptance. Preserve its existing app. - Correct a Linux arm64 test that checked an unread callback before its React effect committed. Pass all 111 affected module tests and TypeScript checks. Follow-up source `5ff80a385957722c99026084036750934fcc564f` changes only test synchronization and type declarations; its 88 compiled files are byte-identical to accepted build 1182. - Desktop publication remains blocked: the replacement prepare for build 1183 could not start because of the GitHub Actions budget. No desktop beta.9 release was published. The local Mac acceptance does not certify the incomplete cross-platform release. - The paired [AgentsServer beta.9](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.4-beta.9) is published and independently verified. Managed updates are queued to apply when active work finishes; scheduling is not deployment acceptance. ## 2026-09-19 — Live thinking summaries — 1.0.4-beta.8 - Display thinking summaries directly and retain expansion through completion. Opening historical traces loads one bounded activity page automatically; additional pages remain available without loading all history in the background. - Carry live summary snapshots through the native WebSocket, profile-scoped service and renderer state. Keep them outside SQLite and durable read cursors, replace them with authoritative completed items, and restore current snapshots on reconnect. Retain interrupted summary text with a partial marker. - Show custom-model summary support separately from basic tool compatibility. - Pass 822 transport, service, state, locale and custom-provider checks, plus 173 timeline and projection checks and TypeScript validation. - Native acceptance follows controlled Responses through the real Codex app-server, production server, authenticated WebSocket, desktop service, preload, state and timeline. Compare rendered text and completed sections exactly with native public-summary notifications. Verify live visibility, summary/tool/summary order, reconnect, authoritative replacement and retention after completion without persisting transient rows. - Stop an actual native turn that omits item completion; retain its received text as a partial summary. Reopen SQLite in a fresh desktop process and expand its historical trace to verify the partial marker and bounded load. Inspect light/dark layouts at narrow width. Retain finite Chromium ResizeObserver notifications in the evidence; geometry and warning counts settle, with no application errors. This is public-summary validation, not native GUI pixel parity or a claim about unavailable internal reasoning. - Accept desktop **1.0.4-beta.8 / 1181** from `3bb296e1ea076f76f235905d3d2deb965b28e08b`. All 88 packaged compiled files match the source fingerprints frozen before artifact download. Replay the accepted native capture through the exact packaged full-app renderer, preload and state; live/final text and chronological tool placement match. This packaged replay uses offline fixture transport. - Universal Developer ID signatures, notarization, Gatekeeper, mounted DMG/ZIP parity and updater checks pass locally; the signed executable passes its clean CI launch. Preserve the previous accepted app separately. - Pass all four platform builds and package checks, then all four publication replay checks. macOS and both Linux suites pass 4,564 tests (10 skipped); Windows passes 4,528 (16 skipped). Publish the Windows installer under the documented unsigned beta policy. - Publish after the matching signed AgentsServer **1.0.4-beta.8** is publicly accepted. Verify 14 exact assets and authored notes on both desktop feeds, every public asset digest and size against independently hashed held files, the exact source tag, anonymous asset availability, and downloaded checksum manifests/updater metadata. Release/tag metadata uses authenticated API reads. - Releases: [desktop beta.8](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.4-beta.8), [legacy Beta feed](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/v1.0.4-beta.8). ## 2026-09-19 — Custom endpoint model compatibility — 1.0.4-beta.7 - Separate optional saved-model compatibility checks from endpoint saving. Display unverified, unsupported and basic-check-passed states without treating model discovery as proof of native Codex compatibility. - Respect explicit per-model effort capabilities, including empty effort lists, and retain manual entry for unfamiliar model IDs. Fence saved-model checks to the selected server and credential revision. - Pair this client candidate with AgentsServer 1.0.4-beta.7. Its 72 focused provider and side-chat checks pass, including stale effort cleanup and retained credential ownership. Native loopback capture confirms the production override helper clears inherited effort while preserving thread instructions and unrelated thread settings. - Exercise Settings, New chat and Composer in native offscreen Electron through production preload, service, native HTTP, server middleware and provider routes. Save without a model or test, and while a connection or compatibility check is pending; verify late results cannot relabel saved settings. Complete repeated checks through isolated native Codex against a controlled streaming endpoint. Filter an embedding model, clamp advertised efforts, and clear effort for unfamiliar/manual models. Inspect light/dark narrow layouts, with no overflow or typing/idle requests. Ordinary account status and full profile bootstrap are fixtures; chat-turn execution is covered separately by server regression and native request capture. - Pass focused app, service and transport regressions, TypeScript, production compilation and output verification. Also complete a basic check against a configured external provider while preserving its saved credentials and ordinary account configuration. This does not certify every model or tool. - Accept desktop **1.0.4-beta.7 / 1180** from `26d6e4586e077997b7a5de203fa3b0d41e876699`. All 88 packaged compiled files match the reviewed source and the fingerprints recorded before download. Inspect the exact packaged renderer/preload in isolated offscreen Electron. Universal Developer ID signatures, notarization, Gatekeeper, mounted DMG/ZIP parity and updater checks pass locally; the signed executable passes a clean CI launch. Recheck the native endpoint workflow against the final paired server code, using isolated native Codex and a controlled provider. - Pass all four platform build and package checks, followed by all four publication replay checks. macOS and both Linux release suites pass 4,550 tests (10 skipped); Windows passes 4,514 (16 skipped). Publish the Windows installer under the documented unsigned distribution policy. - Publish matching sets of 14 assets and authored notes to both Beta feeds. Match every public asset digest and size against independently hashed held files, verify the exact source tag, and anonymously check all public asset URLs. Download both checksum manifests and all updater metadata anonymously and verify byte parity. Release/tag metadata uses authenticated public API reads after the shared anonymous API rate limit is reached. - Releases: [desktop beta.7](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.4-beta.7), [legacy Beta feed](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/v1.0.4-beta.7). ## 2026-09-19 — Center the Team Network mail reader — 1.0.4-beta.6 - Center mail threads in a wider reading column instead of pushing sent messages against the far-right edge. Align incoming and sent messages, increase message spacing, and soften the sent-message background. - Reproduce the previous layout at 2,000 pixels and 70% zoom in native offscreen Electron using the production mail and Markdown components. Inspect the corrected reader at 2,000, 1,200 and 600 pixels in light and dark themes, including long text, code blocks and attachments. Verify no page overflow, native navigation and scrolling, code copy and attachment preview. The fixture uses synthetic read-only mail and an isolated clipboard; delivery and production server data are outside this check. - Pass all 114 existing mail/style tests, TypeScript, production compilation and the compiled-output verifier. This layout change needs no server update. - Accept desktop **1.0.4-beta.6 / 1178** from `7e9ec89e007bc32e1f7889c4f558160385773ba9`. All 88 packaged compiled files match the reviewed source. Inspect the packaged renderer/preload in isolated offscreen Electron. Universal signing, notarization, Gatekeeper, mounted DMG/ZIP parity and updater checks pass locally; the signed executable passes a clean CI launch. All four native platform builds and package checks pass, and the release test suite passes 4,547 tests (10 skipped). - Publish matching sets of 14 reviewed assets and authored notes to both Beta feeds after all four platform replay checks pass. Verify public asset digests, updater metadata, exact source tag and canonical/legacy parity. - Release: [desktop beta.6](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.4-beta.6). ## 2026-09-19 — Preserve side chats across servers — 1.0.4-beta.5 - Keep side-chat state and its native conversation owned by the saved server identity and chat, across connection generations. Preserve drafts, replies, pending work and follow-up context when switching away and back. Retain dispatch checks and explicit cancellation, removal and shutdown cleanup. - Reproduce lost history/drafts and a switch-triggered close request in native offscreen Electron. Exercise the production panel/controller/lifecycle, preload, AppService switching, isolated settings/cache and native HTTP to two controlled servers with matching chat IDs. Verify background completion, follow-up continuity, server separation and cancellation after returning. Inspect dark and narrow/light views. The picker and store hydration are outside this focused fixture; provider responses are explicitly controlled. - Also exercise a real Codex side chat against an existing authenticated server, using one disposable main chat with a random verification fact. Switch servers while its side question runs, restore its answer and draft, and complete a contextual follow-up in the same native side conversation. Verify both answers, then remove the disposable chat. Existing conversations and login settings remain untouched. - Remove Electron's internal fork-error prefix. Exercise the production chat menu through HTTP and native Codex: reproduce a valid symlink-workspace fork rejection, then verify repeated forks and a child continuation with the corrected standalone server while the parent continues running. - Pass 111 focused side-chat checks, four fork-error checks, TypeScript and production compilation with output verification. The side-chat navigation fix needs no server contract change; the fork workspace correction is in AgentsServer 1.0.4-beta.5. Include authored notes for both releases. - Accept desktop **1.0.4-beta.5 / 1177** from `56f268292b889f173f7d1e0a10bdc43796545575`. All 88 packaged compiled files match the reviewed source. Inspect the packaged renderer/preload in isolated offscreen Electron. Local universal signing, notarization, Gatekeeper, mounted DMG/ZIP payload parity, checksums and updater metadata pass; the signed executable passes a clean CI launch. The macOS release suite passes 4,547 tests (10 skipped). - Publish matching sets of 14 reviewed assets and authored notes to both Beta feeds after all four native platform checks pass. Verify public asset digests, updater metadata, source tag and canonical/legacy parity. - Publish the signed standalone server 1.0.4-beta.5 and submit its managed update for idle installation. The running service remains on beta.4 while active work continues; its pending update has no error. - Releases: [desktop beta.5](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.4-beta.5) and [server beta.5](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.4-beta.5). ## 2026-09-19 — Inter-chat chronology — 1.0.4-beta.4 - Accept desktop **1.0.4-beta.4 / 1176** from `d2b40e5b3e8f208d9b36f20be1efcb460ba821f4`. - Keep inter-chat cards among the work that happened around them, ahead of a later final answer even when the turn has early-created files or media. Give the trailing media group a presentation anchor consistent with its displayed position, and refresh cached rows when that anchor advances. Preserve original message timestamps and attachment metadata. - Reproduce the incorrect order with the production Timeline, virtualizer and row components in native offscreen Electron using synthetic event snapshots. Exercise native controls for live/completed work, late read receipts, cold reopen, a genuinely later send and card expansion. Inspect light/dark output. Provider execution and production chat data are outside this renderer check. - Pass 229 focused timeline tests, TypeScript and production compilation with output verification. No server contract change or deployment is required. - Pass 4,538 desktop tests (10 skipped) in release CI. Verify all 88 packaged compiled files against the committed source and inspect the packaged renderer/preload. Universal macOS signing, notarization, Gatekeeper, DMG/ZIP parity and clean executable launch pass. Linux x64/arm64 and Windows x64 package and launch checks also pass; Windows remains an unsigned preview. - Publish identical sets of 14 assets, checksum manifests and authored version-specific notes to both desktop Beta feeds. Independent public download checks confirm the exact source tag, asset digests and all four updater metadata files after publication. - Release: [desktop beta.4](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.4-beta.4). ## 2026-09-18 — Custom endpoint model controls — 1.0.4-beta.3 - Accept desktop **1.0.4-beta.3 / 1175** from `c1dc59c56a666fe881317ff36781f78fcde6decd`, with matching standalone AgentsServer **1.0.4-beta.4** from `b4b116d022ba9d73949e56476cbfc46fdba27160`. - Configure a Codex endpoint with its URL and separate key. Saving no longer requires a successful test or a model ID, and a pending test does not block it. - Discover the endpoint's models and choose a model and reasoning effort in the normal chat controls. Keep an explicit model entry for endpoints without discovery, and keep custom catalogs separate from ordinary Codex. - Require the matching server capability before using the new controls. Existing custom chats retain their endpoint when the default is edited. - Validate production compilation and focused renderer, service and transport regressions. Exercise native Electron input through the production service and HTTP boundary with disposable state and controlled provider endpoints. - The matching server scopes messaging instructions to its helper contract; the harness enforces messaging access instead of broad prompt restrictions. Remove the blanket identifier prohibition without adding a special permission paragraph for local log diagnosis or changing messaging grants. - Exercise simultaneous normal/custom native Codex threads against controlled endpoints. Save and reset during active turns and a pending test; retain the original endpoint through model/effort changes, follow-ups and native forks. Stop one custom turn without interrupting the normal turn. Verify zero account-login calls and unchanged normal runtime identity. Live external gateway credentials and production background startup remain outside these disposable acceptance fixtures. - Pass 4,536 desktop tests (10 skipped), TypeScript and production compilation. Verify all packaged compiled files against the committed source and inspect the packaged renderer/preload. Universal macOS signing, notarization, Gatekeeper, DMG/ZIP parity and clean executable launch pass. Linux x64/arm64 and Windows x64 package and launch checks also pass; Windows remains an explicitly approved unsigned preview. - Publish identical sets of 14 assets and checksum manifests to both desktop Beta feeds. Independent publication checks revalidate every native platform and confirm public updater metadata and Beta discovery. The matching signed server release is published and its managed update is accepted for idle installation; it remains pending while active work continues. - Releases: [desktop beta.3](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.4-beta.3) and [server beta.4](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.4-beta.4). ## 2026-09-18 — Codex credential isolation — 1.0.4-beta.2 - Accept published desktop **1.0.4-beta.2 / 1173** from `f87790cf69e66261a2271c2056bd5bcd06ff666a`, with matching standalone AgentsServer **1.0.4-beta.2** from `dfc05e997b7c97b366c87400f35e4c640f8f2f85`. - Remove shared API-key sign-in from Settings and every desktop transport layer. Normal Codex account status is read-only. The matching server rejects the legacy login route before accessing the account manager, including requests from older clients. This prevents Settings from overwriting the credentials used by ordinary Codex chats and the CLI. - Keep one explicit Custom endpoint flow: enter URL/model/key, Test, then Save. Explain that a new chat must select **Codex · Custom endpoint**. Normal account-status failures no longer block endpoint configuration. - Reproduce the original shared-login call in a disposable native runtime. Exercise the corrected Settings with native mouse/keyboard input through the full production server module and actual Codex process against a controlled Responses service. Verify failed tests and retry, exact URL/model persistence, invalidated tests after edits, busy-save rejection, Remove, account-status failure recovery and legacy-route rejection. Confirm zero native login calls and unchanged ordinary account state. - Verify simultaneous normal/custom native threads and follow-ups retain separate credentials and models. Inspect the packaged renderer/preload and all 88 compiled files; the removed shared-login paths are absent. UI fixtures use disposable profiles and controlled endpoints; production background lifecycle and a live external provider account are not claimed by these checks. - Pass focused regressions, TypeScript and production compilation. The macOS release run passes 4,528 tests (10 skipped), universal signature/notarization, DMG/ZIP parity and clean executable launch. Linux x64/arm64 and Windows x64 packaging and native launch checks also pass. Windows remains an explicitly approved unsigned preview. - Publish the verified desktop packages to the canonical and legacy Beta feeds with identical assets and checksum manifests. Independent publication checks verify every platform again and confirm public Beta discovery. Matching signed server artifacts are published; the managed server update is queued for idle installation without interrupting active work. - Releases: [desktop beta.2](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.4-beta.2) and [server beta.2](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.4-beta.2). ## 2026-09-17 — Per-chat native Codex endpoint selection (beta candidate) - Accept local desktop **1.0.4-beta.1 / 1175** from `59451a6` and matching standalone server **1.0.4-beta.1** from `46d72a4a`. Stamp only the desktop package metadata in the clean build snapshot; verify all 88 compiled files byte-for-byte, ARM64 Developer ID signature and disabled local updater. The server archive matches all 85 allowlisted source files and its checksum. - Add **Codex · Custom endpoint** beside ordinary Codex in the composer and New chat. Configure its base URL, exact model and separate masked key in Settings; test explicitly before saving. Ordinary Codex sign-in remains available and existing chats retain their original provider. - Persist the choice per chat and reject unsupported older servers before they can silently ignore it or replace a global provider. Started chats cannot switch providers. Custom readiness does not require ordinary OpenAI sign-in; removing the endpoint leaves custom chats unavailable, not rerouted. - Run a real native Codex manager with simultaneous normal/custom threads and repeated follow-ups against two controlled Responses endpoints. Verify separate credentials/models, unchanged process defaults and no tool calls, external requests or production account/history changes. The earlier real gateway probe verifies the configured native protocol separately. - Add no polling, automatic model requests or per-keystroke network work. Focused transport, renderer, persistence and provider-isolation checks pass. Scheduled jobs use the same provider-specific readiness and label. - Exercise native offscreen mouse/keyboard input through the production picker, New chat and Settings, real preload/main HTTP transport and extracted production session routes/store. Verify Save, default/custom switching, persisted selection, locked-thread rejection, older-server refusal and no typing/idle requests. Inspect light/dark narrow layouts. Full application bootstrap/cache/profile lifecycle remain fixture boundaries; packaged production-profile startup is not claimed. - The clean-source package pass has 4,539 desktop tests passing (10 skipped), TypeScript and production compilation. The clean standalone snapshot passes 175 focused tests and the real same-manager native provider check. These are local test candidates, not published/notarized releases; no installed app, production server or active chat was replaced or restarted. ## 2026-09-17 — Native Codex custom endpoint controls (source only) - Add endpoint base URL, exact model ID and a masked provider key to Codex account settings, with explicit Test connection, Save and reset actions. Test uses native Codex Responses behavior rather than a replacement agent. - Keep provider credentials separate from normal Codex sign-in. Require a successful test of the current form before saving; invalidate it on edits and fence late responses by server/profile generation. Never reuse a saved key for a newly entered endpoint or show raw provider errors. - Add no polling, per-keystroke requests or automatic retries. Test does not save configuration; Save/reset reconcile runtime readiness once. - Exercise native offscreen Electron Settings navigation and the real renderer/preload/service/HTTP/router/native Codex path against a controlled Responses endpoint. Verify failed tests and retries, unchanged unsaved configuration, stale-result invalidation, busy/authorization failures, save/reset, missing-credential recovery and narrow localized themes. - Full profile bootstrap, account status and runtime refresh callbacks remain fixture boundaries in UI acceptance; isolated tests cover admission and readiness reconciliation. A separate authorized probe also completes a native Codex response against a real external Responses gateway with its exact model ID. That verifies a small model request, not every tool or billing capability. Requires matching standalone server endpoints. No release, production deployment or existing account change is included. - Full desktop tests, TypeScript and production compilation pass after the recovery changes. The standalone server's selective source snapshot passes its focused authentication, provider, side-chat and manifest regressions. ## 2026-09-17 — Native Codex API-key authentication (source only) - Add Settings → Codex account with masked API-key sign-in, account status and an explicit Recheck action. Explain server-wide account scope, native Codex credential storage and separate API billing. Support English and Chinese. - Use the installed Codex app-server's native account API through the selected server's operator-only HTTP endpoint. Do not replace Codex with a model API client or put credentials into settings, histories, logs or command arguments. - Clear credentials on submit, cancel, close and server switch. Fence requests by profile/generation, reject redirects and use fixed secret-free errors. Add no polling, automatic login retries or per-keystroke network activity. - Preserve active/queued Codex work during authentication changes. Refresh runtime readiness once after a successful save, including same-timestamp health records and a pre-login probe that was already in flight. - Exercise the real UI with native offscreen Electron keyboard/mouse input, production preload, scoped service methods, native HTTP, server authorization and the actual Codex process using synthetic credentials in an ephemeral store. Verify save/recheck/repeat, busy and permission failures, stale replies, clearing secrets, idle traffic and narrow light/dark localized layouts. - Full Settings/profile bootstrap and runtime-catalog behavior remain isolated fixture boundaries in UI acceptance; focused service tests cover readiness reconciliation. TypeScript, production compilation and desktop regressions pass. Native credential acceptance alone is not a live model/billing test. - Requires matching standalone server authentication endpoints. Availability is source only: no public release, production deployment or account switch. ## 2026-09-17 — Native provider side conversations (local build) - Build **1.0.3-local.1174 / 1174** from `36ba482`. Verify all 88 compiled payload files, ARM64 Developer ID signature and local-only updater marker. - Place Side chat below Media & files, after Subagents, within one inspector. Verify expanded/collapsed media and narrow/light/dark native Electron layouts. - Replace visible-text snapshots with native context: a persistent ephemeral Codex fork, or Claude's native side-question control used by `/btw`. Include provider tool results without injecting a new message into the main chat. - Keep follow-up identity and history on the server. Clear closes only the selected side conversation; late requests cannot recreate it. Profile/chat ownership, cancellation and provider-generation fences protect the main task. No polling or per-keystroke network work is added. - Require the matching native-context server capability. Do not silently fall back to a copied transcript on an older server or incompatible provider. - Real disposable-provider checks cover hidden tool-result recall, follow-ups, cancellation while the main request runs, unchanged parent history/goals and Claude cold resume without a main query. Focused transport and lifecycle regressions cover cancellation, duplicate requests, expiry and cleanup races. - Exercise native offscreen Electron mouse/keyboard input through production preload, native HTTP authorization, router, provider binding and Claude SDK manager into a real authenticated Claude provider. Verify first answer, follow-up, cancellation, Clear/new conversation and continued main work. Both provider adapters also pass real disposable-provider checks; synthetic full-boundary UI fixtures cover both providers. Session store/SDK option construction and full app-profile bootstrap remain fixture boundaries. - TypeScript, production compilation and the clean-source desktop suite pass (4,412 tests, 10 skipped). Matching standalone server source is `39e59aad`. Packaged startup is not exercised; payload/signature validation does not claim an installed production-profile test. No server deployment, publication or replacement of the running app is included. ## 2026-09-17 — Side chat in the shared inspector (local only) - Build **1.0.3-local.1173 / 1173** from `85471ed`. Verify the ARM64 Developer ID signature, local-only updater marker and exact compiled payload against the accepted source. Do not replace the installed app or publish. - Place Side chat directly below Subagents in the existing inspector scroll; remove its separate tab and nested inspector landmark. Keep Review available. - Preserve per-chat drafts, replies and pending requests across panel/review navigation. Only the explicit Side chat shortcut focuses the composer; merely opening the inspector cannot steal main-chat input focus. - Validate focused component/App regressions and TypeScript. Exercise the production Inspector and Side chat with native offscreen Electron input in light/dark and narrow layouts, using synthetic sessions and provider replies. This is layout/interaction acceptance; transport and providers are unchanged. - Catch and correct a narrow-window clipping case by revealing the complete composer rather than only its textarea. Verify long-history wheel scrolling, localization and no extra transport/global-store writes while typing. - The final package pass has 4,387 tests passing, 10 skipped. One earlier run encountered an intermittent pre-existing Team Network address-label timing assertion; its focused rerun and final full run pass without changing that feature. Package startup and live-provider behavior are not retested here. - Prepare a local desktop package only. No publication, server update or replacement of the running application is part of this change. ## 2026-09-17 — Desktop 1.0.3 stable accepted - Publish [AgentsDock 1.0.3](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.3), build **1172**, from `fa5d815118c28a1c18fa6d9afc3cc77f7d089fff`. Include Workspace Changes, native Side chat transport correction and the small 9px sidebar app-version label. Matching AgentsServer 1.0.3 is published for Git controls and the Codex Side chat startup correction. - Preparation `35277907016` and publication `35279799800` pass all four platform gates. macOS is Developer ID signed and notarized; Windows retains the approved unsigned policy. Verify the exact downloaded universal Mac package's 1.0.3/build 1172 metadata, signature, notarization, public Stable update feed, bundled feature code and version typography. - Verify identical 14-asset public and legacy releases, the pinned source tag, sealed checksums and fresh public updater metadata for every platform. Checksum-manifest SHA-256: `6ede115f1bceb8acbfcb37cb951b3bc10753582051db378f57913a945b44f39a`. - Native offscreen interaction covers Sidebar themes/widths, real-repository Changes workflows through the production client/server boundary, and real provider Side chat answers, follow-ups and cancellation. A disposable Codex overlap check confirms cancelling the side turn leaves its separate test main turn active through normal completion. Official package clean-start checks run on disposable CI machines; local native fixtures do not claim a full packaged production-profile Changes journey. - Preserve concurrent uncommitted work. Publication does not replace the user's running app, install on another machine, or restart either server. ## 2026-09-17 — Quiet sidebar version label (unreleased) - Keep the installed app version beside the brand at a small, muted 9px size; preserve the title and control layout and keep the full version in its tooltip. - Check production Sidebar rendering in isolated offscreen Electron at narrow and normal sidebar widths, light and dark themes. Verify no title/control overlap, working keyboard navigation, one local metadata read and no network requests. This is renderer acceptance, not packaged-release acceptance. ## 2026-09-17 — Workspace Changes first cut (unreleased) - Add a lazy Changes workspace tab for repository-wide staged, unstaged, untracked and conflicted files, on-demand diffs, whole-file staging, staged-set commit review, and text conflict resolution. Continue reports further conflicts honestly; abort requires explicit confirmation. - Use native operator-only, profile-scoped requests and repository revisions. Preserve conflict drafts on stale writes, reject late responses from another workspace, and add no polling or per-keystroke Git requests. - Exercise production workspace entry/renderer, preload, HTTP client, server Git router and native authorization with mouse/keyboard in isolated offscreen Electron against disposable real repositories. Verify actual commits/index, stale stage and resolution rejection, merge completion, confirmed abort, tab switching/closing, light/dark and narrow views, and zero idle requests. Session lookup and app-shell context are fixtures; full installed-app startup and production-profile acceptance are not claimed. - Validate focused desktop regressions, TypeScript and production compilation. This feature needs the matching standalone server Git endpoints. PR/MR, push and branch creation remain outside this first cut. Not published, installed or deployed; existing applications and research jobs remain intact. ## 2026-09-17 — Desktop 1.0.3 acceptance checkpoint (not published) - Validate clean source `f387a1ac874ab3e153a648bd53e42bf4347ae4e4` with TypeScript, production compilation, 4,369 passing desktop tests (10 skipped) and eight package/license guards. Keep unrelated working-tree changes out of the pin. - Exercise Side chat with native Electron input, production preload IPC, request ownership and HTTP client, the unchanged server authorization/router, and a real Claude provider. First answer, contextual follow-up and cancellation pass; the old generic transport reproduces 403. Session context and the main task are synthetic: this does not validate a concurrent live main agent or full production profile bootstrap. - The equivalent real Codex check uncovers a separate server adapter startup failure before a thread starts: fresh temporary state indexes existing provider history synchronously and initialization times out. Do not classify that as an app authentication failure or extend deadlines to hide the work. - Hold publication pending the known Codex limitation and release decision. No installed app or production server was replaced or restarted. ## 2026-09-17 — Desktop 1.0.3 side-chat transport and visible versions - Correct side-question POST and cancellation to use the existing native HTTP transport. Generic fetch added a browser-style header that the server's native-only authorization rejects. Preserve the authentication boundary, response ownership, follow-up history, cancellation and timeout behavior. - Replace mock-only HTTP checks with loopback wire checks reproducing the 403 and exercising the corrected requests, redirects and cancellation. - Show the app version and selected server version in the sidebar using local app status and scoped server metadata, without polling. Validate light/dark, narrow layout and server switching in an isolated native Electron renderer. - Add the app development operational manual to the repository rules, requiring hands-on workflows and explicit accounting for real versus mocked boundaries. - These desktop changes do not require a matching server update. Source preparation and focused checks are not publication or live-provider acceptance; record those separately when completed. ## 2026-09-17 — Desktop 1.0.2 stable accepted - Publish [AgentsDock 1.0.2](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.2), build **1170**, from `94659a201581a2adf1f0938f157e519a859e1459`. The public and legacy repositories carry identical sets of 14 release assets and Stable updater metadata. Preserve unrelated uncommitted work. - Native release preparation `35195524801` and publication `35197194846` pass all platform gates. macOS universal is Developer ID signed and notarized; Linux x64/arm64 and the explicitly unsigned Windows x64 installer are verified. - Validate the exact committed desktop source with TypeScript, production compilation, 4,360 desktop tests (10 skipped) and eight package/license guards. Deterministic regressions cover both first-click races found by the initial native release attempt; the failed candidate was never published. - Inspect the actual isolated Electron recovery dialog and full Team Network surface in light/dark and wide/narrow layouts. Verify cancel/focus, wrong-host errors, offline-to-workspace recovery and no typing-triggered requests or global store writes. Synthetic endpoints never mutate live networks. - Pair with the published standalone AgentsServer 1.0.2 recovery contract. Members must update their own server and explicitly change a moved host's saved address; publication does not migrate addresses or restart services. ## 2026-09-17 — Preserve the first recovery and attachment click - Native release checks expose a commit/passive-effect ordering race: a late identity reset can close the freshly opened host-address dialog or invalidate the first explicit attachment preview request. - Reset only these identity-bound local states in layout effects, before the controls can be used. Preserve the existing stale-request fences and exact dependencies; add no timers, polling or per-keystroke work. - Add deterministic commit-phase click regressions. Both reproduce the old failure and pass after the correction; all 300 affected renderer tests and TypeScript checks pass. Rebuild the held desktop 1.0.2 source rather than publishing the failed candidate or weakening its tests. ## 2026-09-17 — Desktop 1.0.2 endpoint recovery prepared - Add a single Change host address action for the current Team Network and saved approved connections, including offline members. Reuse one localized dialog with the previous address prefilled; preserve approval and routes. - Require the additive member-side AgentsServer 1.0.2 capability. Verify the exact local server instance, saved connection and remote trust before a write; reject stale replies and leave inactive connections inactive. - Retire old authenticated state only when a validated write begins, then revalidate the current connection once. A lost write response permits one status check, not a repeated mutation. Add no polling or per-keystroke work. - Inspect the actual isolated Electron actions and dialog in light/dark, wide/narrow, error, pending and unsupported-server states. Exercise explicit save/cancel and confirm typing makes no global store writes. These UI checks use synthetic endpoints and cannot mutate real Team Networks. - Prepare stable notes against desktop 1.0.1. The release also includes the committed independent Side chat, indexed Mail/Bulletin search, visible changed-file summaries and quiet syncing status. Unrelated unfinished changes remain excluded. Publication acceptance is recorded separately. ## 2026-09-16 — Local desktop 1.0.1 build 1170 accepted - Package committed source `7f116b2` as an Apple Silicon local desktop build with the Details / Side chat inspector layout, independent follow-ups and retained drafts. Exclude unrelated unfinished worktree changes. - Validate full desktop tests, type checks, production compilation and package guards. Confirm all 86 compiled files match the packaged archive and the Developer ID signature verifies; keep local auto-updates disabled. - Exercise the actual isolated Electron panel in light/dark and wide/narrow layouts, including follow-ups, hide/reopen, cancellation and stale replies. Provider responses are mocked in these UI checks, not live model runs. - Prepare the matching standalone server `1.0.1-beta.2` package. Side chat needs that server capability installed; neither server deployment, publication nor replacement of the running desktop application is part of this local build. ## 2026-09-16 — Side chat in the inspector dock - Replace the temporary question dialog with a full-height Side chat tab next to Details. A labeled chat-header action opens and focuses it directly, leaving the main conversation visible and usable. - Keep per-chat drafts and side answers in memory when the dock is hidden or Details is selected. Follow-ups carry only the side conversation's bounded completed question/answer pairs; Clear and Cancel affect only Side chat. - Keep typing local and introduce no polling, main timeline subscription or normal chat turn. Fence in-flight work by server profile, generation, chat and request identity; unsupported servers explain the missing capability. - Pair with the standalone server's additive side-history support. Packaging does not install, publish or restart either running application or server. ## 2026-09-15 — Independent side questions (unreleased) - Add a localized Side question entry for Codex and Claude chats. Questions and answers stay in a temporary panel, with independent cancellation and explicit context limits; they never become a normal prompt, queued turn or goal steer. - Use one request per question, with no polling or event subscriptions. Fence answers and cancellation by server profile, generation, chat and request ID; discard stale replies when a panel closes or the user changes chats/servers. - Require the additive standalone server capability. Older servers explain the missing support instead of silently forwarding a question to the main agent. - Validate request/cancellation races and the actual renderer in light/dark, wide/narrow, keyboard focus, pending, answer, error and localized states. No installed application replacement, server restart or publication. - Guard cross-chat Markdown whitespace and preserve the exact sent body. The standalone sender guidance discourages joining words and technical values; historical text is not rewritten by speculative spacing corrections. ## 2026-09-15 — Keep file changes visible outside collapsed progress (unreleased) - Keep known changed-file summaries and Review accessible when a turn's progress is collapsed. Show a compact filename list with aggregate line counts while leaving full diffs and tool activity behind their existing disclosures. - Reuse recorded diff metadata without fetching activity or parsing arbitrary tool output on chat open. Keep scheduled-job summaries single-rendered. - Validate completed and live-to-completed turns, collapse/expand, exact Review targets and scheduled-run identity. Check actual renderer layouts in light and dark themes at wide and narrow sizes with network access disabled. Type checks, desktop regressions and production compilation pass. - Desktop-only presentation change; no polling, provider runs, server update, publication or running-app replacement. ## 2026-09-15 — Local desktop 1.0.1 build 1169 accepted - Package committed source `9ca56af` as an Apple Silicon local desktop build, including explicit Mail/Bulletin search and the quiet synchronization status. Exclude unrelated unfinished worktree changes. - Verify TypeScript, desktop regressions, compile/license guards and production compilation. Confirm all 86 compiled files match the packaged archive, the hardened Electron bundle passes its audit, and the Developer ID signature verifies. Keep local auto-updates disabled. - No publication, running-app replacement or server deployment. Indexed search still requires the matching standalone server update; the synchronization status change works without it. ## 2026-09-15 — Quiet chat synchronization status (unreleased) - Replace the uncertain incoming-delivery warning during chat synchronization with a muted “Syncing…” status. Keep known incoming-delivery notices and explicit Stop/Send now confirmations unchanged. This is a desktop-only presentation change with no new requests, subscriptions or server changes. ## 2026-09-15 — Indexed Mail and Bulletin search (unreleased) - Add explicit Search/Enter and Clear controls to Inbox, Sent and Bulletin. Search current subjects, message contents and sender names across accessible history, with indexed server queries and explicit result pagination. Typing stays local; no polling, timers or per-keystroke requests are introduced. - Keep filtered results separate from ordinary snapshots, unread counts and notification acknowledgements. Fence old-query and old-connection results; preserve route, read, edit and delete behavior for individual results. - Capability-gate the feature on the matching standalone server contract and migration. Older hosts retain ordinary Mail with an update explanation. See [Mail search](TEAM_MAIL_SEARCH.md) for compatibility and migration notes. - Validate renderer, IPC, direct and secure-peer paths, pagination, stale responses, notification isolation and current-content indexing. The actual renderer passes eight isolated light/dark and wide/narrow journeys: typing and idle issue no requests; Enter searches, open/back retains the query, and Clear restores the normal list. Type checks and production compilation pass. No published artifact, installed app or running server is changed. ## 2026-09-15 — Desktop 1.0.1 accepted - Published stable [AgentsDock 1.0.1](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.1), build `1168`, from reviewed source `9a13649eb5a8f01d83c4c33b8bd3af3a70163cda`. Includes member self-rename, durable join observation and force-update confirmation recovery. See [release notes](RELEASE_1.0.1.md). - All four native build and package gates passed, followed by independent package replay before publication. macOS is universal, Developer ID signed, notarized and Gatekeeper-accepted. Windows remains unsigned under the release owner's standing distribution policy; its status is explicit in both public release listings. - The canonical and legacy repositories publish the same fourteen artifacts. Both stable update feeds resolve to `1.0.1` for macOS, Linux x64, Linux arm64 and Windows. The sealed checksum manifest's SHA-256 is `b1fcdc86950f190891ee6c0ea4e9bd1fd8a4495886b821362ed270d6e5250b58`; the verified Mac update ZIP is `c0649157f5f31a7b9cea37f447f7a3e465666d3f5a4f5a057b764353f01dbfc9`. - Confirmed the downloaded Mac package contains the member-rename renderer, preload and main-process implementation. No installed app, live server, mobile distribution or unrelated worktree change was included. ## 2026-09-15 — Member self-rename, desktop 1.0.1 candidate - Committed member self-rename in `ee277b162b34a341d2d771e92bea2697342bcc3e`. The owned member's directory menu now offers Rename. Saving changes its Team Network display and recipient names in place, without changing its host role, peer identity, connection or local profile label. Other members remain protected; stale identities and mismatched receipts are rejected. - Verified the real paired-service principal, whose identity is distinct from its directory node, along with legacy node-shaped sessions. The existing published server API passes isolated self-rename, read-only, reprovision, current-mention and new-mail-label checks; no server changes were required. - A clean archive of the committed source passes TypeScript, 4,218 desktop tests (10 platform/intentional skips), eight compile/license guards and production compilation. The actual renderer passes isolated offscreen dark/light and narrow-layout checks for Save, Cancel and editable failures; saving preserves row identity and adds no polling or role-switch operation. - Release notes are prepared in [RELEASE_1.0.1.md](RELEASE_1.0.1.md). Native release packaging and publication remain pending per-release Windows signing approval. This is not yet an accepted or published application package. No installed app or live server was changed. ## 2026-09-15 — Durable team join waiting (unreleased) - Keep one automatic-join observer attached across long HTTP observation windows when the server confirms a durable pending approval or activation. Renew only the same held read, with unchanged request, transcript, server and cancellation fences. Reject early responses instead of creating a hot retry loop; do not add inbox polling, repeated Join requests or UI refresh timers. - Make legacy expired incoming requests discoverable in a collapsed section, separate from pending approvals and without approval controls. A replacement request hides the stale attempt. New non-expiring joins require the matching standalone server change on both host and joining server. - Accepted locally: 600 focused desktop checks, type checks, production compilation and compile-output verification. Inspect the real host panel offscreen at narrow width in light and dark themes, including expired and replacement requests; no overflow, new network calls or approval mutations. No package, installation, publication or live server restart performed. ## 2026-09-14 — Force-update status recovery (unreleased) - Recover from a force-update confirmation refused because the queued update changed while the confirmation was open. Read status once; never retry a restart or update automatically. Follow an already-started update only when its schedule, target and track match the approved reservation. - Show actual installer/preflight failures instead of a stale confirmation error. If status cannot be verified, re-enable Check server. Clear only the handled recovery notice after a successful check or server-scope change; preserve unrelated failures and ignore responses from an old server/boot. - Support the existing beta.8 response, including bridges that preserve only its error prose. No server contract change or background polling is added. - Validate focused confirmation-race and recovery regressions, the full desktop suite, type checks, production compilation and compile/license guards. Inspect the actual Settings dialog in isolated offscreen Electron in both themes, covering install progress, preflight failure and manual-check recovery. - Source fix only: no new package, publication, installation or live restart. ## 2026-09-17 — Surface native artifact-open failures - Propagate operating-system errors when opening an artifact so the existing desktop action handlers can report the failure. - Add focused service coverage for successful opens, native error responses, and download failures using synthetic data and mocked native boundaries. ## 2026-09-14 — Desktop 1.0.0 replacement accepted - Published desktop `1.0.0`, build `1167`, from committed source `45eb06c9db69cf0afad0fe8a1a38cf0823f1c035`. The canonical release is [AgentsDock v1.0.0](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.0). This includes the native subagent setting and corrected untitled-child headings. The separate server history-proof correction requires the rebuilt standalone server; installing the desktop alone does not repair server imports. - The explicitly approved same-version replacement preserves a verified backup of the withdrawn package. Existing `1.0.0` installations need a manual download/reinstall; same-version automatic discovery is not claimed. - All four native package gates pass. macOS is universal, Developer ID signed, notarized and Gatekeeper-verified; Linux x64/arm64 and Windows x64 artifacts pass their package and launch checks. Windows remains an explicitly approved unsigned preview. The platform source suites pass: 4,130 checks on macOS and each Linux architecture, and 4,094 on Windows, with platform-specific skips. - Both public repositories contain the same fourteen sealed artifacts. All eight public update-metadata files match the accepted replacement bytes. GitHub permanently retired the deleted immutable legacy `v1.0.0` tag, so the [legacy compatibility mirror](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/1.0.0) uses the exact tag `1.0.0`. App versions, package filenames, signatures and canonical `v1.0.0` are unchanged; the frozen legacy tag was not modified. - Actual native updates from Stable `0.2.12` and Beta `1.0.0-beta.2` both install the new `1.0.0` ZIP, whose SHA-256 is `f6914dfc7cba0ab9d762e0185c41e948a98709afb87bf113b091c3abcab919c2`. Real UI download/install and native replacement/relaunch pass; each journey retains its profile, language and original channel, then checks the canonical feed. Final UI screenshots were reviewed. Validation used disposable CI profiles, not live user apps or servers. ## 2026-09-14 — Partial-page cross-chat replay regression - Reproduce two old, owned async deliveries reimported as provider user text. Full-history desktop projection can correlate the original receipts; a recent-only page cannot establish that proof locally. - Add regression coverage for the server's existing source-proven replay marker on recent-only pages, stale cache merges, older-page overlap and a genuinely user-authored quotation. Original agent messages and answers keep their positions. This adds no desktop runtime change or background polling; the correction requires the rebuilt standalone server package. ## 2026-09-14 — Untitled subagent headings - Correct the inspector fallback for native children without an explicit title: show a readable task/path heading and retain the provider nickname underneath. Explicit titles, child identities, ordering and selected output remain unchanged; no polling or provider-state writes are added. - Regression cases reproduce the former nickname-first behavior and cover null/omitted titles, title clearing, legacy identity fields, locale changes, separator-only tasks and sixteen simultaneously active children. - Inspect the actual inspector in isolated offscreen Electron at narrow width in light and dark themes, including completed children and open output. - Prepare an explicitly approved replacement desktop 1.0.0, not a version bump. Existing 1.0.0 installations need a manual reinstall to receive it. The desktop naming correction is separate from the rebuilt server's history-proof correction. Package and publication acceptance is recorded above. ## 2026-09-14 — Local subagent-settings build accepted - Accepted Apple Silicon local `1.0.0-beta.2`, build `1165`, from committed source `a1ef7b0f6314b74c1244d8504f9643099fbf3eb7`. This supersedes local build `1164` with caller-bound server selection for settings requests. - Type checks, focused native HTTP/service scope tests, production compilation, and the full desktop suite pass. One existing secure-peer cancellation test failed in the initial concurrent full run, then passed independently and in the full rerun; its source was not changed in this work. - Isolated offscreen Electron validation covers light/dark narrow layouts, local typing, save/disabled states, compatibility hints and profile-switch races. It does not launch the production app or access a live server. - All 86 packaged compiled files match validated output. Developer ID signature, hardened runtime, entitlements, fuses and packaged version checks pass. Existing app bundles are preserved; this package is not installed, notarized or publicly released, and local automatic updates are disabled. - The control requires standalone server `1.0.0-beta.8`. Publishing that server is separate from updating a live installation; no live runs are interrupted. ## 2026-09-14 — Native Codex subagent setting - Add Codex subagent limit to desktop Settings > Server. A positive integer sets the server's native provider override; clearing it uses Codex's default, not an unlimited sentinel. The main agent is excluded from this count. - Read on opening the settings page and write only on Save. No polling, keystroke requests, chat-state writes or provider restarts are introduced. - Show when an older server, non-admin connection or legacy exec transport cannot change the setting. Preserve drafts on failed saves and ignore stale replies after switching servers. Explain new/reloaded-thread scope and chat-specific override precedence; provide English and Chinese labels. - Bind settings reads and writes to the server selection displayed by the renderer, checking it before dispatch as well as after the response. A stale screen cannot send its old draft to a newly selected server. - Use exact native-admin GET/PUT transport with request validation. Real loopback transport and service-scope checks cover token framing, reset, errors and profile races; component checks cover local editing and no polls. - Requires AgentsServer 1.0.0-beta.8 for the setting. That server release also fixes plain goal steering with automatically attached saved chat routes; the active goal's owner and existing permissions remain unchanged. ## 2026-09-14 — Project licensing - Add the Apache License 2.0 and project attribution notice. Preserve existing Expo, SwiftTerm, native-module, and other third-party licensing terms. - Document licensing in the README and contribution guide, and declare it in first-party desktop, mobile, website, and Team Hub package metadata. - Include the project LICENSE and NOTICE in future Electron packages. Existing published artifacts and tags are unchanged; no release is cut by this change. - Validate the canonical license text, package metadata, preserved component licenses, and Electron's actual resource-copy behavior with focused checks. ## 2026-09-13 — Local live-activity build - Accepted local-only Apple Silicon `1.0.0-beta.2`, build `1163`, from `f76c1f4087d855924a4498f831d8c30e05b8ac0b`, including live activity freshness, quiet-run/compaction indicators and native Codex subagent display names. - Type checks, the full desktop suite, production compilation and isolated light/dark renderer journeys pass. All 86 packaged compiled files match the validated output; strict Developer ID signature, entitlements and fuse checks pass. Existing app bundles were preserved. - This local app has automatic updates disabled and is not notarized or published. It was not launched or installed over a running app. - Source-proven duplicate answers and leaked native wake imports still require the matching standalone server update; this app alone cannot fix the older server's history reconciliation. ## 2026-09-13 — Live activity and mailbox replay - Preserve newer streamed run ownership when an older health request finishes late. Connection-only notifications do not replace live activity with cached health. Fresh idle health can still settle a run whose terminal was missed. - Show a compact Working indicator when an owned run has no visible trace yet, including a quiet mailbox wake. Show Compacting context during live compaction and animate the Running header; completed and historical views stay settled. - Keep activity reconciliation local and scoped to the connected server. No polling, provider requests, synthetic chat messages or minimap rows are added. - The separate large-history server correction also suppresses source-proven duplicate answers imported after mailbox wakes, retaining the original answer, peer deliveries and real human messages. It requires a server update. - Type checks, the full desktop suite and production compilation pass. Isolated full-renderer checks cover quiet wake, live compaction, resumed progress and actual completion in both themes, with stable timeline geometry and no added user messages. The duplicate correction was also checked against source-proven native/import pairs without changing stored transcripts or message delivery. - No installed app, running server or published release has been changed in this pass. ## 2026-09-13 — Codex subagent display names - Show the explicit Codex child-thread title before its nickname, task or path fallback. Names stay literal in every locale. Nicknames and paths remain available in details; no names are generated from prompts. - Accept optional `subagent_title` snapshots from the standalone server's existing thread metadata and name-update stream. Omitted or malformed fields retain the known title; explicit clearing restores the existing fallback. - Keep one stable child row and update an already-open output panel's title in place. Naming does not introduce polling, new provider requests or UI timers. - Desktop type checks, production compilation, the full suite and focused rename/clear/lifecycle regressions pass. Isolated full-renderer checks cover dark/light and narrow layouts, including renaming and clearing a title while the output panel stays open. The change requires the matching standalone server update to supply explicit titles. No installed application, running server or published release has been changed. ## 2026-09-13 — Large-history cron and mailbox replay correction - Correct the standalone server's native-history proof path for large chats. Tool-heavy logs no longer bypass duplicate checks at a total-file-size limit. Incomplete or cancelled proof defers import without advancing its cursor. - Existing desktop projection keeps the original scheduled-job group and purple mailbox delivery while suppressing only source-proven imported copies. Genuine human messages remain intact, including identical quoted text. - Add a combined desktop regression fixture covering interior same-ID repairs, stale event replay, overlapping older pages and SQLite reopen. Type checks and focused tests pass. Isolated full-renderer checks verify the read receipt, chronology, chat reopen and renderer reload in light and dark themes. - This requires a server update; no desktop runtime change, package, deployment or release was made. Delivery, wake behavior, jobs and provider transcripts are unchanged. No polling or per-event filesystem reads were added. ## 2026-09-13 — Video playback in shared chats - Interactive shares now reuse the existing chat video thumbnails and player for videos attached to sent messages or explicitly published by the agent. View only snapshots include native video players for those captured videos. - The matching standalone server provides token-checked playback and seeking for the exact shared chat. Unused uploads, unrelated files and workspace paths remain inaccessible. Revocation checks are demand-driven; no polling or extra stored video copies were added. - Preserve native event identities and chronology through current, historical and trace views. Unsupported native file actions stay hidden in shared mode; the normal desktop file controls are unchanged. - Desktop type checks, focused media/bridge tests, the full desktop suite and production compilation pass. Isolated server checks cover ownership, token entry, byte ranges, revocation, file mutation and cancellation cleanup. - Isolated browser acceptance passed for an uploaded WebM and a distinct agent-published H.264 MP4 in both viewers: decoded frames, playback, seeking, separate token entry and blocked media requests after revocation. Testing used synthetic chats and hidden browser windows, not live user sessions. - This requires a server update, including the generated shared-web bundle. Existing Interactive links can be refreshed afterward; older text-only View only snapshots must be recreated to include videos. Browser codec support still applies. No desktop package, server deployment or release was made. ## 2026-09-13 — Choose a LAN address for chat sharing - Add one localized Share address field to the existing View only / Interactive dialog, defaulting to the selected server connection. Operators may choose another reachable HTTP or HTTPS address of the same server when creating a share. Previously created links and tokens are unchanged. - Keep share management and native credentials on the authenticated connection. The chosen browser address is validated body data, never a probe or a new credential destination. Creation remains one-shot; malformed or mismatched responses cannot silently produce a link at the wrong address. - Editing stays local to the dialog, with no polling, address discovery or background refresh. The existing standalone server contract supports this client change without an update or restart. - Accepted local-only Apple Silicon `1.0.0-beta.2`, build `1162`, from `b8a100ce75782d22edb9529d90811b5360d67a6d`. Desktop type checks, the full desktop suite, focused address/response regressions, and production compilation passed. All 151 packaged compile-output files match; bundle/fuse audit and strict Developer ID signature verification passed. - Isolated offscreen renderer checks passed for both modes, light/dark themes, narrow layouts, localized labels, keyboard activation, invalid addresses, busy state, draft reset, and unchanged existing links. These used synthetic services; no real chat share or recipient connection was created for testing. - This build is not notarized or published and has automatic updates disabled. The running app and servers were left untouched. ## 2026-09-13 — Local main integration build - Accepted local-only Apple Silicon `1.0.0-beta.2`, build `1161`, from `bc73ab326c5a6fcc4a8425be41384faf9ec3ff1a`, which merges public main `935a76b64af2b2a7c263a15d91af87342dd88381` into the release branch. - Include the local-chat import label correction while retaining the released Team Network changes. Scheduled-job controls remain in each row's context menu; this fetched main does not contain visible inline action buttons. - Type checks, focused import/menu checks and production compilation passed. All 151 packaged compile-output files match the source build; bundle/fuse audit and strict Developer ID signature verification passed. - This local build is not notarized or published and has automatic updates disabled. No running app, installed server or published artifact was replaced. ## 2026-09-13 — Published server goal-steering correction - Published [AgentsServer `1.0.0-beta.4`](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.0-beta.4) from immutable source `b5fa0728ede6e99f228685f625198b5bdcde20a0`. The full release gate, downloaded signature, public asset digests and all 76 packaged source files passed verification. - Preserve active native Codex goals when steering during the initial turn or a continuation, including uploaded attachments and between-turn delivery. Keep one chronological follow-up across acknowledgement rollover and prevent stale-prompt recovery after accepted or uncertain steering. No polling added. - Existing desktop `1.0.0-beta.2`, build `1160`, supports this server correction; there is no new desktop binary for the regression-only app changes below. - An accepted idle deployment schedule is not a completed installation. Live-provider acceptance remains separate from the isolated protocol tests. Stable readiness also requires resolving the idle-goal mailbox-authority limitation and completing the stable-upgrade and platform acceptance gates. ## 2026-09-13 — Active-goal steering regression coverage - Verify Send now and Cmd/Ctrl+Enter while a goal is active across a pending server response, acceptance and rejection. The goal and other queued work remain unchanged; retries cannot duplicate prompt submission or promotion. - Confirm native goal follow-ups retain attachment rendering when reopened. These are renderer regression checks accompanying a standalone server fix; no desktop runtime change or new app build is required. - Full Composer tests, focused goal timeline tests and desktop type checks pass. No app or server installation, deployment or publication is part of this change. ## 2026-09-13 — Published Team Network and cross-chat beta - Accepted direct desktop `1.0.0-beta.2`, build `1160`, from the immutable source `516213ddb27b4fc641c418609a7667d14e391918`. Native build and package verification passed for universal macOS, Linux x64/ARM64 and Windows x64. macOS is signed and notarized; Windows remains an unsigned beta preview. - Published the identical 14-asset set to the [public release](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.0-beta.2) and [legacy mirror](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/v1.0.0-beta.2). Both match checksum-manifest SHA-256 `6fa6cb2877d411ef423ef0c808a10286cd919f0acc5727c364e1de0620b5c445`. The protected publisher independently replayed all platform verifiers; publication recovery retained the original artifacts and source identity. - Include quiet, separately tracked Mail and Bulletin indications through one metadata-only connection. Refresh stays explicit; arrivals do not fetch content, move the current view, interrupt an agent or change a draft. Bulletin edits retain version history, and incomplete refreshes cannot clear attention to unseen changes. - The matching [AgentsServer `1.0.0-beta.3`](https://github.com/ZhengyiLuo/AgentsServer/releases/tag/v1.0.0-beta.3) fixes Chats provider-tool stdin replies and returns an original canceled receipt on exact retries without resending or waking the recipient. The complete server release gate, downloaded signature, exact source contents and older-updater compatibility passed. An app-only update does not apply the server corrections. - The isolated full-desktop lifecycle journey passed 15 checkpoints using actual renderer, IPC and SQLite cache with synthetic transport: sent/read/ replied messages, disconnect/backfill, reopening, single canceled-message identity, explicit Bulletin refresh, author edit/history/delete and scope retirement. A 1,000-hint burst preserved drafts, focus and scroll position without content requests. Real helper/TLS server journeys additionally covered access revocation, concurrent reads, late arrivals and idle wakes. - No user's active provider run or live server was used for these checks. Publication did not install or restart either component. Stable and mobile channels are unchanged. ## 2026-09-13 — Quiet Team activity and cross-chat delivery regression - Local, unreleased candidate: extend the existing single Mail notification connection with an independently tracked Bulletin cursor. Posts, revisions and deletions produce metadata-only hints; older servers retain Mail v1. - Show quiet navigation indicators and explicit refresh affordances. Arrival never fetches content, navigates, interrupts an agent, or changes a draft. Main-process publications coalesce bursts, with no recurring idle timer; only small indicator components subscribe to pending state. - A Bulletin refresh acknowledges the head captured before its complete fresh traversal. Partial, failed, cached or stale-scope loads cannot clear it, and an update arriving during refresh remains pending. Author-only, versioned Bulletin revisions remain supported. - Add a cross-chat projection regression: an assistant's claim is not a send receipt. Actual registered/received events appear on both sides immediately, before recipient wake or read, without duplicate rows or repositioning. The matching standalone server corrects Chats provider-tool stdin handling; that fix and Bulletin v2 require a server update. - Validation passed: TypeScript, the full desktop suite (3,972 tests passed, 10 skipped), four compile-output guard tests, and production compilation. An isolated Electron UI journey checked dark/light and narrow Chinese layouts. A 200-hint burst caused no content/receipt requests or acknowledgments and preserved the draft node, text, focus and scroll container. Explicit refresh performed only the expected content requests and local acknowledgments. - This source-change entry preceded release acceptance; see the published beta entry above. No live-server deployment is included. ## 2026-09-12 — Published 1.0 desktop migration bridge - Published `1.0.0-beta.1`, build `1159`, from the exact reviewed source `885bfa9a382c734f7b66e9a2b6bb025333ba9d0a`. This is a beta migration bridge; stable `1.0.0` is not published. - Native release builds and platform verification passed for universal macOS, Linux x64, Linux ARM64, and Windows x64. The macOS artifacts are Developer ID signed and notarized. Windows remains an explicitly unsigned beta preview, not a signed stable distribution. - The release embeds the public AgentsDock desktop feed and preserves application identity, saved update-track preference, and explicit installation. Its Beta track can select a newer stable release without adding background polling or changing Team Network behavior. - The exact 14-asset set is published in the [canonical public release](https://github.com/ZhengyiLuo/AgentsDock/releases/tag/v1.0.0-beta.1) and [legacy mirror](https://github.com/ZhengyiLuo/AgentsDock-Releases/releases/tag/v1.0.0-beta.1). Release API checks verified both releases' matching asset digests and sealed checksum-manifest identity. The source tag resolves to the reviewed commit above; checksum-manifest SHA-256 is `237bfd4ef1ed42fbe8e16b5549510eaea88a935296fa6b96903b5335486d169f`. - The protected publisher replayed all four native platform verifiers and published the original sealed artifacts without rebuilding or replacing them. Canonical beta download links now point to the public source repository. - Installed legacy-feed → public-feed upgrade acceptance passed on an isolated macOS runner using the signed production packages and actual updater UI. `0.2.13-beta.33` downloaded the bridge, completed native replacement and relaunch, and then checked the public feed successfully. The profile, Beta preference, and language were preserved. Installed application payload matched the published bridge; screenshots and a machine-readable receipt were retained. No operator app or live server was used for this acceptance journey. - Current stable `0.2.12` downloads and the Android release feed remain on the legacy repository. No standalone server deployment or mobile release is part of this desktop migration. ## 2026-09-12 — Public desktop release migration - Prepare the `1.0.0-beta.1` migration bridge with the public AgentsDock repository as its canonical desktop download and update destination. Keep the legacy release feed available for older installations and Android; publish the same verified desktop artifacts to both repositories. - Beta subscribers can receive a newer stable release without losing their Beta preference. Selecting stable metadata during a normal Beta check does not enable downgrades. Existing startup and four-hour checks are unchanged; downloads still require an explicit install action to restart the app. - Preserve application identity, signing requirements, saved settings and connections. Merge the current public UI and pinned-message navigation changes while retaining shared-chat restrictions and attribution. - Validate release ordering against both feeds. Publication must pin the public source tag to the reviewed commit and resume mirrors only when the sealed source identity and artifact checksums match; published conflicts fail closed. - Stable `1.0.0` remains a separate release gate. Existing stable downloads continue to use the verified `0.2.12` artifacts until that promotion. ## 2026-09-12 — Reusable-token chat sharing - Keep share addresses separate from access tokens. Copy invitation includes both on separate lines; opening the interactive address requires manual token entry. Existing browser sessions can resume without storing raw tokens in browser storage. View-only sharing also offers an explicit token-in-link copy. - Use one reusable token per share for multiple collaborators. Remove the one-person claim wording, and move revoked entries into collapsed history. - Preserve initial-latest timeline positioning without new scroll timers or background refreshes. The matching server supplies paginated full-log snapshots and starts their viewer at the latest page. - Focused component, bridge, URL-validation and timeline-position checks, TypeScript, the desktop suite and production compilation passed. Actual compiled desktop clicks verified both share actions, separate URL/token, copy/open dispatch, snapshot-only token links, and revoked-history folding against isolated fixtures without creating real shares. - Accepted local arm64 candidate: `0.2.13-beta.37`, build `203`, source `699911b`. Developer ID signing, strict signature verification, bundle audit, and all 151 packaged compiled-file byte comparisons passed. This local build is not notarized or publicly released; automatic updates are disabled. The running desktop was not replaced. - These changes require AgentsServer `0.1.26-beta.66` or later. Server release and installation are tracked separately; the local app does not update a server automatically. ## 2026-09-12 — Two-action HTTP chat sharing - Replace the preview/checkbox workflow with View only and Interactive actions. Each explicitly creates, copies and opens a link; existing-link revocation is available on demand without background refreshes. - Derive share addresses from the selected native server connection. Support direct HTTP as well as HTTPS without requiring a separately configured domain. This requires the matching standalone server change; an app update alone is insufficient. HTTP is intended for trusted networks and is not encrypted. - Serve a styled static snapshot with user bubbles, assistant Markdown, dates and responsive light/dark layouts. Interactive sharing reuses the chat UI and keeps explicit one-time Join, chat-scoped access, same-origin and CSRF checks. - Verified actual compiled desktop right-click, both creation buttons, exact copy/open dispatch and lazy existing-link management with isolated fixtures. Verified served pages on a genuinely non-secure HTTP browser origin: Join, reload, prompt, queue Send now, Stop, wrong-origin/CSRF denial and used-invite denial. Static pages were checked at desktop and mobile widths in light/dark themes. These synthetic journeys did not execute providers or expose real chats. - Accepted local arm64 candidate: `0.2.13-beta.36`, build `202`, source `ce1e0d9`. Production compilation, desktop checks, strict Developer ID signature, bundle audit and exact packaged main/preload/renderer byte comparisons passed. Automatic updates are disabled; this local candidate is not notarized or published. The running desktop was not replaced. Server changes are committed separately and were not deployed by this validation. ## 2026-09-12 — Quiet idle-mail wake presentation - Keep the server-generated mailbox availability input out of the user-message timeline while retaining the receiving agent's progress and final answer. Suppression requires exact native wake metadata, not matching message text. - Preserve genuine user inputs and unproved imported history. Verified cold and incremental projection for Claude and Codex with focused checks and TypeScript. - Idle execution requires the corresponding standalone server change; this desktop change alone does not wake an idle recipient or update its server. - Validated the compiled Electron renderer with synthetic Claude and Codex live-progress, completion and sidebar-reopen journeys. Each retained one genuine user bubble and both original and wake answers, without duplicate output or an internal wake notice. No live provider execution was exercised. - Accepted local arm64 candidate: `0.2.13-beta.35`, build `201`, source `ffe6168`. Production compilation, desktop tests, strict Developer ID signature, bundle audit and matching packaged main/preload/renderer bytes passed. The app has automatic updates disabled; it is not notarized or published. The running desktop was not replaced. ## 2026-09-12 — Local desktop beta.34 candidate - Accepted local universal macOS installer: `0.2.13-beta.34`, build `157`, exact committed source `85654170df60df954d0efea1f68d0fd34e37a2a2`. - Verified Developer ID signing, notarization ticket, Gatekeeper acceptance, embedded version/build and updater ZIP checksum. This candidate includes the large-share preview and browser-sharing changes described below. - Desktop publication was canceled at the user's request. No public beta.34 release was created; the installer was retained locally and the running app was not replaced. The signed candidate has normal direct-update support. - Full cross-platform release acceptance is not claimed: macOS and Linux x64 build jobs passed, an ARM renderer fixture exposed a passive-effect assertion race, and the remaining workflow was canceled. The fixture correction is retained separately and is not part of this installer. ## 2026-09-12 — Large chat sharing - Preview and confirm text snapshots independently of raw tool-log size and message count, retaining bounded UTF-8 snapshot and individual message sizes. The matching server release preserves existing shared links and revocations when upgrading snapshot storage. - Page the review dialog in groups of 20 messages, resetting scroll position on each page. Confirmation still covers the complete reviewed snapshot, not just the visible page. Strip native IPC boilerplate from share errors. - Give snapshot requests transport headroom beyond the server's bounded scan; do not change unrelated request deadlines or add retries or background work. - Verified the actual desktop right-click, preview, paging, confirmation and old-server error journey with a synthetic 2,384-message snapshot and a source boundary larger than 64 MiB. Only 20 message elements were mounted at once; confirmation from the second page retained the original full-snapshot digest. - Interactive sharing requires the matching standalone server implementation; publishing the desktop alone does not add missing server endpoints. ## 2026-09-12 — Shared-browser recovery and control parity - Resume an already joined browser session on reload without consuming another invitation. Preserve session identity when loading older timeline pages. - Disable shared provider controls and close permission popovers when access is lost; retain a usable Close action and explain the disconnected state. - Stop further writes after an uncertain acknowledgment instead of offering an automatic duplicate send. Do not misreport an accepted action as failed when only its subsequent refresh fails. Release failed local upload staging slots. - Expand exact cross-chat message bodies on demand within the shared chat, preserving recipient edits without sending messages or marking mail read. - Wire the web file-drop surface into the existing one-way upload path. Reject oversized selections before staging a partial selection. Native desktop file selection behavior is unchanged. - Rechecked the actual compiled shared renderer in an isolated Chromium browser: join/reload, older history, expanded message scrolling, queue controls, Stop, goal controls, settings/models/permissions, approvals, schedules, uploads, pushed owner results, one-use invitations and read-only/interactive revocation. The idle stream made no additional chat-history requests across its heartbeat. Provider callbacks and owner state were synthetic, not production research jobs. - The final web pass also verified Claude approval/Stop and revoked-popover behavior, actual rejected/accepted file drops, a narrow light-mode layout and lost-response duplicate prevention. Confirmed reload after a retained draft created a fresh document and restored one accepted queue receipt without a resend. Claude permissions saved and re-rendered through browser form events; native OS popup interaction remains unverified because the isolated window could not take keyboard focus. - Accepted local arm64 candidate: `0.2.13-beta.33` build `200`, committed desktop source `d44e1ad`. Developer ID signature and all 86 compiled archive files verified. The running app was preserved; this candidate is not notarized or published and has automatic updates disabled. The subsequent web-only drop change is included in the separately packaged shared renderer at `6839cf2`. - The matching standalone server changes and web bundle are committed locally, not deployed. Public HTTPS ingress and real provider execution remain separate deployment acceptance steps. No direct file, terminal or other-chat API was added to guest access; sharing remains trusted agent collaboration, not a sandbox. ## 2026-09-12 — Receipt-based outgoing chat status - Replace the unconditional outgoing “Sent to” heading with “To” and an explicit status. Distinguish unconfirmed delivery, stored unread mail, agent-read receipts, cancellation and failure; preserve legacy queue status. - Correlate receipts to the exact message and participants. A default unread state or registration event alone is not proof of mailbox storage. Reading is not represented as processing or replying. - Preserve target navigation, on-demand message expansion and existing controls. Add no polling, automatic resend, agent invocation or server change. - Verified synthetic receipt states in the actual desktop renderer, including on-demand expansion and target navigation, plus focused component checks. These checks do not establish the delivery of any particular live message. - Accepted local arm64 candidate: `0.2.13-beta.33` build `198`, source `2734708`. Developer ID signature and all 86 compiled archive files verified. The running app was preserved. This candidate is not notarized or publicly released and has automatic updates disabled. ## 2026-09-12 — Full control within a shared chat - Expand the interactive invitation disclosure and required confirmation to full control of the one shared chat: prompt/upload, stop/steer, send queued messages now, manage the queue, change model/chat settings and permissions, manage goals, and create/edit/delete scheduled jobs. - Retain the boundary against direct terminal access, file browsing/download, other chats and server administration. The collaborator can still ask the existing agent to use its tools and context; this is not a tool sandbox. - Explicitly explain that revoking access does not undo accepted work or scheduled jobs. Snapshot sharing remains read-only and unchanged. - Reuse the native timeline, composer, queue, approvals, goal and schedule components in the token-scoped browser view. Keep file browsing/downloads, working-directory controls, terminals, other chats and administration out. - Receive shared-chat changes through one demand-open stream, not polling. Keep model choices across updates and refresh cached Codex/Claude status after a snapshot commits. Use the native positioned settings modal. - Validated the compiled browser UI against the isolated token router and native adapters with synthetic provider/store mutations: prompt queueing, edit/reorder/send-now/stop, goal pause/resume, settings and permissions, model choices, schedule create/edit/delete and approval responses. Verified that a second browser cannot reuse an invitation and revocation disables further guest actions and ends the live stream. No production provider run or live user job was started by these checks. - The expanded guest controls require the matching standalone server update; this source change does not deploy or publish that server. - Accepted local arm64 candidate: `0.2.13-beta.33` build `197`, desktop source `ed1b20e`. Developer ID signature and all 86 compiled archive files verified. The running app was preserved. This candidate is not notarized or publicly released and has automatic updates disabled. The subsequent settings-modal correction affects only the separately packaged web renderer. ## 2026-09-12 — Explicit chat sharing - Add right-click **Share chat** with two separate choices: a reviewed, read-only text snapshot and a one-time invitation to a live chat. - Interactive sharing explicitly grants trusted use of the existing agent's tools/context. The guest web surface has no native terminal, file browser, downloads, other chats, or administration; it is not a provider sandbox. - Use exact native-authenticated, server-scoped management requests with no redirects or automatic retries. Show missing HTTPS hosting honestly, retain new link secrets only in the open dialog, and provide exact revocation. - Mark shared-chat prompts as Collaborator in the timeline and queue. Do not add an inbox poller, background navigation, or automatic sharing. - Validated the actual isolated Electron right-click/preview/confirmation, copy/revoke and collaborator-label journey against synthetic share responses, plus actual loopback native-header transport and affected regression checks. New interactive sharing requires the separately updated standalone server; no server deployment or public app release is included in this source change. ## 2026-09-12 — Preserve proven assistant replay corrections - Keep exact server-proven Codex and Claude assistant replay corrections in the local SQLite cache when an older page or buffered stream repeats the same event. Reuse the existing identity/provenance merge checks; do not infer duplicates from similar text or remove original scheduled reports. - The standalone server also corrects Codex history comparisons when native delivery removed leading decorations from the same provider message. That server correction is separate and requires deployment before historical duplicates can be repaired. No polling or background refresh was added. - Verified the actual isolated Electron service, SQLite cache and renderer: repair, stale replay, switching chats and reload preserve corrections while all original scheduled reports, genuine messages and the running job remain intact. Provider transport was synthetic. Claude's equivalent cache path was verified with SQLite close/reopen and shared provenance checks, not a second graphical journey. - Accepted local arm64 candidate: `0.2.13-beta.33` build `193`, source `c9c5210`. Developer ID signature and all 86 compiled archive files verified. The running app was preserved; this candidate is not notarized or publicly released and has automatic updates disabled. The standalone server patch is committed separately and has not been deployed. ## 2026-09-11 — Cross-chat heading navigation - Make the sender or recipient name on purple cross-chat messages open the exact referenced chat. Grouped inbox messages keep a separate unread-count disclosure, so clicking the sender does not expand or consume mail. - Reuse normal chat navigation without additional requests, polling or route grants. Guard the originating server scope; do not resolve duplicate display names or turn imported label-only messages into guessed links. - Verified actual isolated Electron navigation for Codex and Claude, both directions, keyboard activation, historical exchange legs and grouped inbox messages. Unread mail remains unread and collapsed after returning. The synthetic transport recorded no mail or execution actions. - Accepted local arm64 candidate: `0.2.13-beta.33` build `192`, source `9fae5cff`. Developer ID signature and all 86 compiled archive files verified. The running app was preserved; this local candidate is not notarized or publicly released and has automatic updates disabled. No server change is required. ## 2026-09-11 — Lazy history repair after server upgrades - On an observed server-version change, invalidate only that server's cached history verification. Opening a chat then uses the existing bounded authoritative history check so corrections to older messages are received. - Do not clear cached content on version change or refresh every chat. Repeated health responses do not trigger another invalidation. Version and verification changes share one transaction, preserving retryability if storage is exhausted. Failed or offline history requests keep cached content. - Verified the actual isolated Electron health-upgrade journey: one lazy history audit corrects older runtime records, genuine inputs and answers remain visible, and the unsent draft survives. Repeated health responses, switching chats and reloading do not repeat the audit or fetch unopened chats. Transport was synthetic; no live provider run was started. - Accepted local arm64 candidate: `0.2.13-beta.33` build `190`, source `fb986357`. Developer ID signature and all 86 compiled archive files verified. Automatic updates are disabled; this candidate is not notarized or publicly released. The running app was left untouched. Historical corrections require standalone server `0.1.26-beta.61` and reopening the affected chat after the upgrade. - Published standalone server `0.1.26-beta.61` from `34b66875` after its full release validation passed. Verified the downloaded manifest signature, archive digest and all 67 packaged files against the committed source. Managed activation is when-idle; publication does not imply installation. No public desktop or mobile release was made in this pass. ## 2026-09-11 — Typed provider notice coverage - Extend the same source-proven runtime metadata contract to known provider compaction summaries and infrastructure notices. A provider user-role record is not automatically a message authored by a person. - Keep real user text, quotations, assistant output and live activity intact; preserve exact repairs through stale event replay and disk-cache reloads. No background refresh or polling is added. Historical repair requires the matching standalone server update. - Verified mixed Codex and Claude messages in the actual isolated Electron service, SQLite, IPC and renderer: user input, quotations, public progress, expanded tool details, final answers, goal activity, scheduled output and passive purple mail remain visible. Runtime notices stay out of user bubbles across stale replay and reopening. Transport was synthetic; unknown future provider formats and live provider execution are not certified by this check. - Accepted local arm64 candidate: `0.2.13-beta.33` build `189`, source `864e15de`. Developer ID signature and all 86 compiled archive files verified. Automatic updates are disabled; the app is not notarized or publicly released. Historical repair requires standalone server `0.1.26-beta.61`; the running app was left untouched. ## 2026-09-11 — Codex interruption notice provenance - Apply the source-proven runtime notification contract to typed interruption notices as well as subagent completions. Neither becomes a message from the user, and an old imported notice does not stop current work. - Preserve genuine user quotations and retain exact corrections across stale event responses and disk-cache reloads. Historical corrections require the matching standalone server update; this is a local desktop change only. - Verified both runtime notice types together in the actual isolated Electron service, SQLite, IPC and renderer: source-proven corrections, stale replay, switching chats and reload preserve genuine quotations and assistant output. Provider transport was synthetic; no live provider turn was started. - Accepted local arm64 candidate: `0.2.13-beta.33` build `188`, source `9cbed1e3`. Developer ID signature and all 86 compiled archive files verified. Automatic updates are disabled; this candidate is not notarized or publicly released. The running app was left untouched. Historical repair requires standalone server `0.1.26-beta.61`. ## 2026-09-11 — Codex subagent notification provenance - Treat source-proven imported subagent completion notifications as runtime metadata, not messages authored by the user. Preserve genuine quotations, assistant answers, native subagent activity and original timestamps. - Keep the exact correction across stale same-ID events and disk-cache reloads; require source identity, a full-text digest and no positive human provenance. - Verified the actual isolated Electron service, cache and renderer: the legacy bubble is corrected, a genuine identical quotation remains, and stale replay, switching chats and reloading do not restore the bogus input. Transport and provider history were synthetic; no real provider turn was started. - Historical repair also requires the matching standalone server correction. No mobile or public desktop release is included in this local change. ## 2026-09-11 — Storage recovery and native history replay - Keep the desktop open after local storage exhaustion. Preserve the original database and saved drafts; use temporary storage only for rebuildable cache data. Failed draft saves block window close, with an explicit Retry saving action after space is freed. An incomplete legacy credential migration keeps its original settings protected and requires reopening after recovery. - Prevent failed event-cache batches from advancing the durable history cursor. Handle browser storage quota errors in layout controls without crashing React. Recovery is user-triggered, with no new storage or inbox polling. - Recognize the server's exact source-proven native replay marker. Keep original human messages and scheduled reports, suppress only verified imported copies, and preserve the correction when an older cached response arrives later. - Verified disk-full startup, preserved drafts, failed-close handling and a successful explicit retry using the actual desktop service, SQLite, IPC, preload and renderer with isolated fault injection. The system disk was not filled, and no provider turn was started by this check. - Verified the native replay correction in the actual isolated Electron UI: duplicate history is removed while the original human message, scheduled report and genuine later question remain, including after switching chats and reloading. Preserve unmatched output across silent imported boundaries. - Accepted local arm64 candidate: `0.2.13-beta.33` build `186`, source `a4fc5ba1`. Developer ID signature and all 86 compiled archive files verified. Automatic updates are disabled; the app is not notarized or publicly released, and the installed/running app was not replaced. Historical corrections and passive mailbox reads require standalone server `0.1.26-beta.61`. ## 2026-09-11 — Passive agent mailbox - Show passive incoming agent messages as compact purple sender groups in chronological position. Only adjacent messages are grouped; human follow-ups and progress remain boundaries. Preserve individual identities and replies. - Expand and delete exact messages on demand. Opening a message does not mark it read by the agent. Long expanded bodies have a bounded scroll surface. - Use the existing event stream, without inbox polling or duplicate execution queue rows. Grouping is linear and preserves immutable cached inputs. - Verified the actual desktop main process, cache, IPC and renderer with synthetic transport: grouping, edited full-body scrolling, exact deletion, read-state updates and reopening. This is not a live-provider execution test. - Requires the matching standalone server mailbox contract. No server deployment or public release is included in this local implementation. - Accepted local arm64 candidate: `0.2.13-beta.33` build `185`, source `bac1876e`. Developer ID signature and all 86 compiled archive files verified. Automatic updates are disabled; this candidate is not notarized or published, and the installed/running app was not replaced. ## 2026-09-11 — Scheduled history ownership - Keep explicitly job-owned output inside Scheduled Job cards even when a paged response omits the start event or scheduled-purpose field. Match cold history, incremental updates, and late ownership metadata. - Preserve ordinary imported/user messages and independently rendered emergency, Mail, and cross-chat receipts. - Older Claude imports also require the standalone server's source-proven historical-page correction; this desktop change does not replace that repair. - Local arm64 candidate: `0.2.13-beta.33` build `184`, source `3088ead6`. Production build, signature and all compiled archive files verified; not notarized or published. The installed/running app was left untouched. ## 2026-09-11 — Expanded queued messages - Keep expanded message rows at their natural height so the queue scrolls instead of clipping the body. The queue retains its existing height limit. - Verified long Codex and Claude messages with wheel scrolling to the final paragraph, reachable queue actions, and collapse in light/dark narrow views using the actual desktop renderer and synthetic transport. - Accepted local arm64 candidate: `0.2.13-beta.33` build `183`, source `a300d1d9`. Signature and compiled archive verified; not notarized or published. No server update is required. ## 2026-09-11 — Chronological agent messages - Place agent messages between the progress before and after their send or delivery-start event. Keep the sender's live or stopped state below the message, and retain that position when delivery receipts arrive later. - Preserve earlier answers during incremental receipt updates. Keep tool calls and results together when activity is split around a message, including explicitly loaded trace pages. - Preserve full queued message bodies, recipient edits, and compare-and-swap revisions during stream updates and stale receipt replay. - Do not display scheduled-job input as a public result. Preserve genuine user-authored quotations of provider control text. - Validated live, completed, queued-reply, and reopened views in the actual desktop renderer with synthetic Codex and Claude transport. Checked both themes, narrow layout, and typing/scrolling in long cached conversations. This does not certify live provider execution, server history pagination, or minimap navigation. No server deployment or public release is included. - Accepted local desktop candidate: `0.2.13-beta.33` build `182`, arm64, source `22a19ce1`. Developer ID signature and compiled archive contents verified. Final desktop UI pass includes stop-after-send, late delivery receipts, and reopening for both providers. This candidate is not notarized or published, and its automatic updater is disabled. ## 2026-09-11 — Current desktop development - Bring current desktop changes into this repository: source-proven history deduplication, readable asynchronous agent-message queues, revision-safe recipient edits, explicit Send now priority, and actual chat names. - Include on-demand Mail threads, agent-only replies, searchable chat routing, guarded Host rename, negotiated unlimited chat routes, and Team Network translations. Quiet arrival hints update a badge without polling Inbox contents or navigating away from the current chat. - Include native provider skills and commands in the slash palette, retaining capability negotiation and unsupported-server fallbacks. - Preserve synthetic test data, the source-only CI boundary, and the existing public binary update feed. No release, native mobile build, or server deployment is part of this source migration. - Validation: desktop type checks, the complete desktop test run, and the production bundle build passed. Local validation used Node 26; source CI remains pinned to the documented Node 24 environment. ## Source verification - Mobile CI now runs the cross-chat protocol, projection, route/queue race, rendering, recipient-picker, and native-workspace resolution regressions. - The additional checks use synthetic data and mocked native boundaries; source CI still does not build signed applications or publish releases. ## Documentation - Reorganized the README into a product overview, installation steps, and separate desktop, mobile, and website development workflows. - Clarified the client/server boundary, current versus legacy client sources, and the distinction between local builds and release publishing. - Checked development commands against package scripts and source CI, and checked installation guidance against the standalone server documentation. - Refreshed the public README hero with a centered product introduction, website, community, and release badges, and an approved desktop and mobile product image. Placed the introduction below a more compact product image to keep the opening layout focused. - Updated the introduction to name the currently supported agent backends, speak directly to AI researchers, and provide direct current download links for every available platform with a matching desktop release badge. - Verified the README with GitHub's Markdown renderer and checked every new destination and badge URL before review. Reviewed the supplied image and its metadata before inclusion. ## Mobile cross-chat parity — 0.1.1 (171) - Apple validation and processing completed successfully; build 171 is active for internal TestFlight testing with automatic notifications enabled. External beta review was not submitted. Binary source: `aa1153be`. - Align mobile with the current desktop async agent-message protocol: one Markdown card per message, pending incoming messages in the queue, and delivery-time chronology without duplicating internal provider prompts. - Show granted chat access, pending grants, route limits, loading/errors, and revision-safe Revoke controls. Reconnects fence stale requests and callbacks. - Add exact queued-message removal with truthful confirmation, duplicate-tap protection, and desktop purple pending-message styling. - Include offline server inboxes in `@@` discovery and distinguish capability-gated `@@bulletin` posts from `@@all` inbox broadcasts. - Validate real component handlers against synthetic native hosts, projection and store/API race regressions, broad library tests, and native build checks. Synthetic rendering does not substitute for physical-device touch/pixel QA. - Release preparation uses generated, ignored native projects and resolves the configured workspace name rather than assuming the legacy project name. - Signed arm64 iPhone/iPad archive and exported IPA passed deep signature, framework ABI, version, production-entitlement, and matching-symbol checks. The production JavaScript bundle contains the new features and excludes the visual test fixture. ## Source snapshot - Includes the Electron desktop and React Native mobile clients, legacy Swift targets, compatibility fixtures, and project documentation. - Private development history, operational incident notes, and unreviewed screenshots and recordings are not included. - This source snapshot does not itself publish or change any installed release. ## Electron workflow controls - Added native provider skills and commands to the composer slash palette. - Refined scheduled-job status, direct actions, working-directory navigation, and compact unavailable-agent guidance. - Added a grouped keyboard-shortcuts page to Settings with localized labels. - Documented privacy-preserving usage events for these workflows. - Validated the affected Electron behavior with focused tests, type checking, a production build, and a local desktop UI pass. Future entries should describe public-facing changes and validation without including credentials, user data, private infrastructure, or internal history. ## Recovered Codex turns do not duplicate user messages History reconciliation now recognizes exact public Codex message items even when Stop or server-restart recovery omitted the normal completion receipt. It repairs already imported copies and prevents duplicate inputs on subsequent imports without rewriting the transcript. Genuine repeated messages in distinct turns remain visible. The new recovered-stop regression failed before the change and passed afterward; all 40 native history-repair tests passed. The signed desktop build 1245 was exercised against an isolated patched server over its real HTTP interface using a synthetic recovery ledger/provider transcript: the original input and answer remained visible, the replay was hidden, and a genuine later repeat survived reopening. No live provider turn, user installation, release, or service restart was part of this replay acceptance. ### Pending idle updates leave scheduled work running An update reserved for when the server is idle no longer blocks automatic scheduled jobs or automatic titles. These use the same update admission policy as ordinary turns. Previously parked, unchanged schedule occurrences are rearmed on startup or update-status reconciliation without cancelling the pending update or rewriting edited schedules. Idle updates also wait for queued messages, including already persisted ones, instead of treating their ability to survive a restart as permission to jump ahead of them. Existing active-work checks remain in place. Explicit force updates keep their separate, confirmed behavior. The focused scheduler, update, admission, status-reconciliation, and generated title suites passed (248 tests). The unchanged signed desktop build 1245 was exercised against an isolated patched server with real Codex: an automatic scheduled turn completed and appeared in the chat while another chat's native tool continued and the update stayed pending. The pending reservation was a test fixture with updater polling disabled; no installation or live deployment was exercised. Provider credentials were copied only into disposable test state and the original remained unchanged. This is a source fix, not a published beta. ## 2026-10-02 — Codex browsing and unchanged sign-in rechecks - Selecting, reopening, or reconnecting a chat now reads Codex status without automatically resuming its native thread. Explicit provider controls still load a persisted thread; normal Send uses the existing server resume path. Browsing therefore does not hold the provider lifecycle lock needed by Send. - Recheck CLIs no longer retires a native Codex process when its known sign-in revision is unchanged. Subagent inspection can read the owning process while a genuine login handoff is pending. Native request/callback ownership scopes idle handoff to the affected chat, preserving other chats' ongoing work. - Validation: 54 desktop controls tests, TypeScript checking, production compile and compiled-entry verification; 130 targeted server tests. A local compiled Electron app exercised real authenticated HTTP and native Codex in isolated state: first message and follow-up both produced the expected answers, browsing an unloaded persisted chat issued zero native-load requests, and an unchanged sign-in recheck allowed the follow-up while another chat's shell work completed without interruption. Original credentials were unchanged; test copies removed. - Acceptance boundary: this validates unchanged-sign-in rechecks, not migration between different real accounts. An earlier forced process-retirement probe encountered Codex's retained native writer after unsubscribe; that probe is failed, and genuine cross-process credential handoff remains unaccepted. No public release, installed-app replacement, or live-service deployment. ## 2026-10-22 — Draft coordinated busy-server update verification - One product Update action now prepares and verifies saved servers before the native app installer. Disposable managed-service upgrades confirmed one internal continuation and a final reply for each of Codex, Claude, and Cursor. A non-chat job and a persistent goal deferred forced interruption; a persistent terminal survived, and simultaneous client requests caused one server replacement. - An isolated packaged-app replay confirmed that an unreachable server leaves the old app installed, but found its error message could be overwritten by a background release check. The source fix retains an explicit incomplete update message across checks and identifies the failed connection. Focused desktop tests, type checking, and the production compile passed; this revised UI message has not yet passed packaged-app acceptance. - Native replacement remains unaccepted: a repacked QA app downloaded the matching update, but did not exit automatically. Its installer completed and relaunched only after the disposable old app was explicitly terminated. The QA shell was modified for isolated feeds and is not an exact-source signed release candidate. A clean packaged replay, multi-server partial failure, rollback, and busy-chat UI journey are still required. No public release or production installation changed. Follow-up verification covered a partial multi-server update in the desktop coordinator: after one server reaches the target and another rejects the update, the plan retains both results; repeating the main Update action retries only the failed server. The affected desktop suites passed 130 tests and type checking passed. Server activation, handoff and rollback fixtures passed 112 tests. A fresh locally signed QA package was built with the isolated feed, but it did not reach the application UI on this machine, so the revised offline message and automatic app exit are still unverified in a packaged App. This QA package is not a release candidate. The disposable server was restored.