# Changelog All notable changes to Kite are documented here. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [Unreleased] ## [1.0.10] - 2026-11-01 ### Added - Full benchmark baselines/comparison and runtime tracing: see [`kite bench`](kite_commands.md#harness-timing-kite-bench) and [`KITE_TRACE_JSONL`](kite_commands.md#runtime-event-tracing). - Isolated worktrees, fast gates, offline end-to-end QA, and CLI profiling: see [contributor workflows](CONTRIBUTING.md#fast-isolated-agent-loops). - `kite sessions --fork ` creates a separate resumable conversation with inherited runtime settings. - Unknown subagent profiles suggest close matches; profile details show model and tool settings. ### Changed Representative before/after measurements from controlled optimization runs (not portable timing guarantees): - Streaming answer processing ~13 → 4.6 µs/chunk; reasoning accumulation 119 → 0.42 µs/chunk at 80k events. - `import kite.ui` 30 → 0.2 ms; configured model resolution 546 → 72 ms cold; `LitellmModel` construction 1.15 s → approximately zero; headless time-to-first-query ~1.9 → 0.6 s. - Path completion on 10k files 34 → 4 ms; picker on 10k options 166 → 35 ms; directory sketch on 20k files 68 → 4 ms; grep benchmark ~26 → 7 ms. - Redaction of 5 MB plain text 300 → 30 ms; `ProcessRunner` peak memory for 200 MiB of output 525 → 1.3 MB. - Agent-loop accounting at 200 turns 0.39 → 0.27 ms/turn. - Session tail resume with a single 50 MB row 7.4 s → 75 ms; recent-event reads 41 → 0.05 ms; semantic-memory rendering 16 → 2 ms. - Warm development install 0.77 → 0.12 s; see [locked development installs](CONTRIBUTING.md#how-to-set-up). - Startup/import boundaries, model capabilities, context/history processing, and session internals: see [architecture.md](architecture.md). - Named behavioral tests replace count-driven batching; see the [test policy and dev loop](AGENTS.md#tests--ci). - The system prompt uses bounded Kite reads/search on every platform and documents scoped-worker write limits. - Bundled progress interval matches the 2-second runtime default. ### Fixed - Prompt caching sends Anthropic `cache_control` blocks only to Anthropic-compatible routes. - Cache-hit ratios no longer double-count normalized cache-creation tokens; failed LiteLLM prewarm imports can retry. - Ordinary `git push` now reaches the serious-action approval flow; force-push remains blocked. - CLI/REPL rendering, completion, jobs, session recovery, and verification behavior: see [kite_commands.md](kite_commands.md). - Prompt-history correctness and episodic-store lock handling: see [architecture.md](architecture.md#startup-and-request-work) and [context & memory](architecture.md#context--memory). ### Security - Bounded output capture, redaction-before-truncation, private trace files, and durable atomic writes: see [SECURITY.md](SECURITY.md). ## [1.0.9] - 2026-10-23 ### Changed - Durable session events append to the transcript only; the sidecar's atomic write + fsync per tool event was unnecessary because events change neither runtime identity nor message count, and transcript mtime already carries recency. Illustrative local run: session event appends ~7.3 → ~0.95 ms/event. - `record_event` resolves the persistence policy once per event and reuses it for redaction. - `KITE_HOME` overrides cache their resolved path, and `UserConfig.load` uses one stat instead of two, avoiding repeated filesystem round-trips on config/session reads. - REPL and headless startup load plugin, skill, and markdown-command machinery only when a slash command needs it. Plain REPL input no longer builds the full slash index, and plain `kite run` tasks skip the slash module entirely. Cold `import kite.ui.repl` fell from ~655 ms to ~590-630 ms locally; isolated `import kite.cli.slash` fell from ~333 ms to ~110 ms. - Windows session appends skip `chmod` (it only toggles the read-only bit there; inherited profile ACLs already restrict the file). Combined with the sidecar change, the local append path is ~7.3 → ~0.95 ms/event. - Test suite consolidated duplicate composer/persistence coverage and merged the question-tool, stream-stall, and diff-render cases into cohesive behavioral tests. Collected cases: 465 → 460. ### Removed - Unwired application modules (`application/context.py`, `application/persistence.py`, `application/ui.py`, `application/event_payloads.py`), token-report/routing/tool-tier experiments, and dead UI wrappers. - Unused `HarnessDependencies` slots and miscellaneous dead helpers. ### Fixed - LiteLLM 1.106 compatibility: plain function-tool requests skip LiteLLM's MCP bridge, which otherwise imports the optional `mcp` package. - Test names that claimed "token efficiency" now describe the live behavior they exercise (reasoning/cache, registry/read, spill/errors). ## [1.0.8] - 2026-10-12 ### Added - Filesystem write scopes for subagents; out-of-scope mutations are denied before execution, and scoped workers cannot use mutating bash commands. - Escaped untrusted prompt boundaries for user memory and automatically gathered repository data. ### Changed - Antigravity transient service, network, rate-limit, and 5xx failures use the existing bounded retry; auth, quota, and malformed-response failures remain permanent. - Provider and model errors now show sanitized failure categories and next steps; Antigravity sessions are linked only after verification succeeds. - Restricted-mode `git push` goes through the network approval policy instead of being unconditionally blocked. Force-push remains blocked. ### Fixed - Safe HTTP fetches no longer use environment-configured proxies, keeping DNS and connected-peer validation aligned with the requested destination. ## [1.0.7] - 2026-10-09 ### Added - Offline end-to-end smoke coverage exercising the real CLI with a loopback recorded provider; isolated worktrees; faster CI tooling; and CLI profiling. - `KITE_TRACE_JSONL` opt-in redacted JSONL runtime tracing. - `kite bench --suite full` with 43 measurements and MAD-aware `--save` / `--compare`. ### Changed - Runtime hot paths improve model construction, headless startup, streaming, reasoning accumulation, completion/picker performance, and large transcript/session operations. - Test suite refactored into named, isolated tests with deterministic fixtures and no external network calls. ### Fixed - Prompt-cache stale history, missing repository symbol maps, and narrow-terminal stream boundary/width handling. - Session recovery, transient SQLite lock handling, resume exit codes, background-job watchdogs, shell output limits, pre-truncation secret redaction, interpreter selection, and invalid task argument prewarm. - Test environment leakage that could trigger real external provider requests. - Added developer workflow documentation for performance/behavior changes, runtime traces, and test policy. ## [1.0.6] - 2026-10-05 ### Added - `/last` — re-print the previous tool call in full: the record is captured at tool-end (bounded at 200k chars per payload) and rendered through the existing formatters, uncollapsed and uncapped, so a truncated diff or output body is retrievable after the fact. - `/tasks` now prints the full itemized task checklist alongside the running turn and queued follow-ups (`commands.py` `BUILTINS` is the single source feeding both `/help` and completion). - `SessionUiState.active_flash` — flash expiry checked on read, not only inside `touch()`, so a flash set at an idle prompt still ages out. ### Fixed - Scrollback density: the live task list is one line (`render_plan_tasks(compact=True)` by default) instead of one row per item. - `todo_write` / `todo_read` no longer echo their payload under the checklist that already shows it — no start card, no JSON body; status, timing and exit code still print. - Scroll-printed tool output, write/edit previews, bash blocks and diffs cap at `PREVIEW_LINES` (5) with an overflow marker naming a command that actually exists. Approval previews keep the 40-line window. - Overflow markers named a `/diff` command that does not exist; a regression test now cross-checks every marker against the registered command list. - `render_diff` `IndexError`: the leftover-pair index math walked past the line-number table whenever a cap cut a `-`/`+` run. Latent at the old 40-line cap, near-universal at 5. - Deleted sessions leaked their `.stats.json` sidecar forever; `delete_all_sessions` also sweeps sidecars orphaned by a crash between unlinks. - `list_sessions` parsed metadata for every session before applying `limit`; it now pre-ranks by transcript/sidecar mtime and parses a bounded head (sidecar matters — `note_runtime` bumps `updated_at` without touching the transcript). - Session crash tolerance: readers raised `AttributeError` on valid-JSON-non-object rows and let `UnicodeDecodeError` escape (not an `OSError`); `list_session_events` discarded *every* event on one torn line. - One saved checkpoint counted twice (`context_checkpoint` row + `checkpoint` event for the same id); deduplicated by id. - A turn ending within 125 ms of its last paint could freeze the footer on the final `working` line until the user typed — `clear_running()` now forces the repaint. - A flash set at an idle prompt outlived its 8 s TTL; the composer toolbar now reads `active_flash`. - `context_pct` divided by a negative context window, rendering `ctx -800%`. - `_trajectory_path` was duplicated byte-for-byte across two modules. - `scripts/bump_release.sh` passed an MSYS path to native Windows Python when splicing CHANGELOG.md, aborting the release mid-run. ### Changed - Test suite 119 → 141 collected, still under the 150 CI budget. New: scrollback density, tool-output retrieval, session hygiene, analytics hygiene, footer throttle. ## [1.0.5] - 2026-10-03 ### Fixed - Infinite-thinking hang: bounded cancel-responsive provider establishment (TimeoutError → retryable ProviderFault), loop-level watchdog, `agent_end` always fires, REPL busy/running reset in `finally`; `stream_thinking_long` live label, coalescer flush on `turn_end`, running cleared on interrupt/error. - Answer streaming: no split-style flicker, 4k streamed code-block cap with explicit marker, central 12k tool-output cap with spill pointers, `…` table truncation. - Long tasks: headless progress heartbeats/checkpoints/retries/approval gates with elapsed times, continuity brief + resume hint on every non-success exit, phantom running jobs eliminated. - Submit gate: Done/Changed (+ Verification with ✓ when checks exist) required for edited sessions; unrelated test passes no longer count; `py_compile` / `node --check` recognized. ### Changed - Prompts require evidence citations and `submit`-tool finishes; casual chat stays literal. ## [1.0.4] - 2026-10-03 ### Security - Project-local `.kite/extensions` gated behind explicit project trust (F-01): untrusted and headless runs load only global `~/.kite/extensions`; `no_extensions` disables both. - `GH_TOKEN` / `GITHUB_TOKEN` injected only into a lone `gh …` subprocess, never into chained shells (F-02). ### Fixed - Session rewrites preserve durable event rows across `set_exit` / compaction (F-03). - Session snapshots write atomically via temp-file + replace; interruption keeps the prior transcript (F-04). - Non-object JSONL rows skipped during session load/resume (F-05). - Cache hit ratio and token totals no longer double-count OpenAI-style cached input (F-06). - Project-context cache fingerprinted on instruction files + commit; `git status` refreshed on hits (F-07). - Plain `git push` / `clone` route to approval (SERIOUS); only `--force` / `-f` stays hard-blocked (F-08). - Agy prompt enforces the 24k-char ceiling after composing system + turns; agy child env filtered (F-09). ### Changed - Test suite consolidated 305 → 114 collected with identical assertions (under the 150 CI budget). ## [1.0.3] - 2026-09-29 ### Added - `kite theme` / `kite font` CLI subcommands — same source as `/theme` and `/font` (list, typed pick, or set by name). - `kite variants` and `/variants` — Pi-style thinking-variant menu strictly gated to levels the current model supports (unsupported names rejected, no clamping). Saved as the fresh-session default and shown as `provider/model#variant` in the status line. - Opencode-style `question` tool — the model asks clarifying questions mid-turn (number, free text, or skip). The REPL answers live; headless runs get an empty set and proceed on assumptions. Serial-only, read-only, in plan and build modes. - Antigravity subscription turns via the signed-in `agy` CLI (`agy --mode plan`, prompt on stdin, text-only). Quota exhaustion maps to `ProviderFault` with the reset clock; `GEMINI_API_KEY` upgrades to direct LiteLLM calls with tool support. - Orca-relay compact UI — narrow rendering (pick lists, status, banner, console width) activates only inside Orca-managed sessions (`ORCA_*` env, `KITE_COMPACT_UI` override); desktop defaults untouched. - Picker rows render a cursor, multi-select checkmarks, per-option descriptions and an explicit key hint. - Diffs show a dim old/new line-number gutter derived from the hunk headers. - Clarifying questions (`question` tool) use the arrow-driven picker with mouse support and Space-to-toggle multi-select on an interactive TTY. ### Changed - The system prompt tells the model to submit with the `submit` **tool** and warns that writing `COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT` as prose is the usual reason a finished task reports "Stalled — no progress". The parallel tools section now states that the runtime preserves call order and why (dependent calls stay sequential), rather than implying everything batches. - Auto-approve now matches what it always claimed to be. In `auto`/`trust`/ `yolo` the agent runs on its own and only risky effect classes prompt — destructive shell, network, package/skill install, durable memory, and nested agents on untrusted projects. An in-workspace `write`/`edit` and a slow command no longer stop the turn for a decision the human cannot make meaningfully. `supervised` still prompts on every mutation, and sandbox denies (outside workspace, protected paths, sudo) are unchanged. - Slash-command completion is cached per index: the visible menu and every rendered row are built once instead of on each keystroke. Typing `/` went from ~10ms to ~0.8ms per keystroke. - Theme/font selection is back on the full picker (arrows, wheel, click) instead of the typed-only fallback introduced earlier in this release — the relay reader that made it unsafe is fixed (see Fixed). `KITE_TYPED_PICK=1` still forces typed input. - Antigravity default model refreshed to a live `agy` id (`gemini-3.7-flash-medium`); `GEMINI_API_KEY` falls back through the antigravity provider like Claude's `ANTHROPIC_API_KEY` fallback. - `tools.progress_interval_seconds` defaults to 2.0 (was 5.0) so a running tool never sits silent for long. ### Fixed - **The REPL never streamed the answer.** With the composer pinned, `_on_stream_delta` buffered tokens and painted nothing until teardown, so a turn could sit silent for minutes and read as a hang. Tokens now stream live through `patch_stdout` (held for the whole turn) and teardown appends a submit report only when it is not already on screen — never a duplicate copy. - `agent_end` never flushed the stream coalescer, so a short trailing chunk under the flush threshold was silently discarded. - `X stopped` printed once per interrupt event, stacking up to four identical rows when several cancel paths fired. One line per turn now. - Pydantic serializer `UserWarning` noise from provider usage objects no longer dumps a traceback into the middle of the transcript. - Windows picker click-release now selects the row (was highlight-only, unlike the POSIX path). - Pickers no longer pick their input reader by `sys.platform`. A Windows relay (Orca, mintty, conpty hosts) is a PTY, and driving it with `msvcrt` blocked forever — arrows did nothing and the terminal had to be killed. Terminal type now decides: native console vs VT reader. - Mouse reporting is no longer enabled on relay sessions (`ORCA_*`), which echoed the enable bytes back as escape junk. - A picker panel taller than the screen repaints with a clear+home instead of relative cursor moves that landed in the scrollback. - Mouse row hit-testing works off Windows: the picker asks the terminal for its cursor row instead of relying on the Win32 console API. - A turn no longer goes silent while it works. The spinner was hard-disabled whenever stderr was not a TTY, so long runs looked frozen in pipes, relays and CI; it now emits newline heartbeats there. The answer tail mirrors into the footer running line during generation, and a finished tool hands the spinner straight to the next model call. - `_normalize_font` accepts canonical `unicode`/`ascii` — font choice applies and persists again. - Long agy turns passed the prompt as argv and died with WinError 206; the prompt travels on stdin now. - Bare `_typed_pick` calls crashed with `TypeError` (missing `show`/`refreshable`) — same defaults as `numbered_pick`. ## [1.0.2] - 2026-09-25 ### Added - Eric Zakariasson token efficiency pass: telemetry, prompt split, spill management, tiering, error handling, history compaction, and routing. - Antigravity auth verification via official `agy` flow; cached OAuth status probes. - Crew coalescing, whole-tree cost aggregation, and richer worker context (`feat(agents)`). ### Changed - Background prewarming of LiteLLM imports during startup and inlined CLI subagents parser structure. - Detached background handoff for Windows self-update, shim-aware helpers, install lock retries, and clean uninstall flows. - Documented NVIDIA compatibility coverage and provider fallback/test coverage. ### Fixed - Gated parallel tools on explicit support for NVIDIA. - Normalized CRLF on read across platforms so output is byte-identical. - Hardened provider history, event flow, and streaming/model discovery responsiveness. - Removed `` eager binding that swallowed composer keystrokes. - Preserved instructions and bound memory; fell back stalled streams and loaded gateway keys. ## [1.0.1] - 2026-09-22 ### Added - Antigravity (Google subscription) OAuth via the official `agy` CLI: `kite login antigravity` (browser / SSH URL loop), link status/logout, catalog provider, fallback Gemini model list. Kite model calls need `GEMINI_API_KEY` (`kite keys --set gemini`) — subscription stays CLI-owned. - Composer fold for long pastes: first lines + `+N lines` placeholder (F9 toggles, auto-folds sudden pastes, any key expands, submit always sends the full text). - `/plan ` and `/build ` act immediately; bare `/plan` reports a resumed checklist like `/build` does. - Cost fallback: usage-based pricing from LiteLLM's cost map when a provider omits `response_cost` (footer no longer stuck at $0.000). - Terminal-responsive truncation: running line and toolbar queue preview follow the live terminal width (windowed vs fullscreen). - Thinking levels re-detect on provider/model switch (model-keyed cache). ### Changed - Steering a busy turn interrupts the agent but keeps the pinned composer alive; a `steering — continuing turn` event clears the stopped flag. - Steer/queue messages fall back to the inbox when harness inject fails, so they are never silently dropped. - Busy toolbar and `/tasks` hints follow `KITE_BUSY_ENTER` (steer default). - `uninstall --purge` retries read-only files and reports leftover counts. - Test suite consolidated 425 → ~217 collected with identical assertions; suite-size references updated across docs. ### Removed - `examples/clamp` untracked (`.gitignore`) and purged from git history. ### Fixed - Websearch tests force the free path (paid chain mocked) so a live `TAVILY_API_KEY` in the environment cannot hijack hermetic asserts. ## [1.0.0] - 2026-09-20 ### Added - MiniMax-style project bootstrap: `kite init`, REPL `/init`, init skill, `examples/clamp`, `/context` and `/status` digests, memory-layer prompt, `resolve_verification_command()` for discovery/status/AGENTS scaffold (#96). - SoL-Pi optional harness efficiency layer (`src/kite/sol_pi/`): Action Fusion (`then_run` on edit/write), ObservationPack (`obs_recall` + projection), Evidence-Preserving Reducer receipts, Online Context Compact at plan boundaries — all **disabled by default** via JSON config (#97). - `docs/RELEASE-1.0.0.md`, `src/kite/data/sol-pi.example.json`. ### Changed - Version stamps and docs aligned to **1.0.0** across pyproject, README, AGENTS.md, and install scripts. - README badges and design-doc links no longer reference PyPI (git/uv install only); release notes pointers aim at `docs/RELEASE-1.0.0.md`. - Bootstrap scripts (`download.sh` / `download.ps1`, aligned `install.*`): clearer private-repo 404 fallbacks, `--` / switch pass-through, Windows ExecutionPolicy one-liner, post-install version line on Windows bootstrap. - Setup auto-prompt runs only on a true fresh install (TTY); stops after `~/.kite/onboarding.complete` or any linked provider key/BYOS session. ## [0.9.9] - 2026-09-15 ### Added - Subagent orchestration upgrade (Codex/Pi/Claude-inspired): role model tiers (`fast`/`coder`/`smart`), per-profile tool scopes, depth cap (1) + spawn budget (64), context packets, bounded summaries, quality gate with one-shot revise, thread-tree registry, `abort_on_failure`, per-worker timeouts, input-ordered results. - Resume hint on interactive exit (`kite resume `, persisted sessions only) — closes #84. - `kite resume ` renders the full chronological transcript and restores context — closes #83. - `docs/RELEASE-0.9.9.md`. - Split CI (lint/test/bench/release-check), `fail-fast: false`, `sync_version --check` fast path + `--release` for tags; `ci_check.sh/ps1` `--release` flag. ### Changed - Single tool pipeline (executor only), single policy truth (sandbox), single slash registry (`ui/commands.py`), single approval/pick/theme sources. - Informational Q&A turns keep the answer; the turn report is an adjunct or suppressed when nothing changed — closes #81. - Pre-tool budget guard, schema-repair hint, summed multi-turn cost, verification-wired `submit`, stable-prefix prompt cache, single memory render budget. - Composer layout test hermetic via `DummyOutput` (headless Windows CI). ### Removed - Optional Textual TUI (`kite/ui/textual/`, `KITE_TUI=1`, `kite[tui]` extra) — the Rich + prompt_toolkit REPL is the only interface. - Retired `/fullscreen` slash command and Textual-only shortcuts (`Ctrl+K`, `Ctrl+J`, `Ctrl+\`). ## [0.9.8.5] - 2026-09-13 ### Changed - Lean CLI is the default surface (Pi-style header, `!` / `!!` shell, `-c` / `-r`). Fullscreen Textual TUI is opt-in (`kite[tui]` + `KITE_TUI=1`). - `kite --help` lists every shipped subcommand so each one is discoverable and callable. - Unknown first tokens are treated as an opening prompt (Codex/Pi). ### Added - `docs/RELEASE-0.9.8.5.md`. - `scripts/ci_check.sh` / `scripts/ci_check.ps1` — same gates as GitHub Actions (`sync_version`, ruff, pytest, `kite bench --check`). ### Fixed - CLI pick/select lists (`kite models --select`, `kite select`, sessions, setup) no longer use a prompt_toolkit overlay that cancelled immediately on Windows. Typed + arrow picker; CI sets `KITE_TYPED_PICK=1`. ## [0.9.8] - 2026-09-13 ### Added - Subagent model override on the `subagent` tool (`model` / `provider` and crew arrays). The system prompt dispatches workers when users ask for parallel agents or name a model. - Project trust via `~/.kite/trust.json`, `/trust`, or `.kite/project.toml` with `[project] trust = true`. Trusted workspaces skip nested-agent approval. - Release notifications from GitHub on interactive startup (24h cache; `KITE_OFFLINE=1` disables). - Textual: `Ctrl+K` fast-path approval, `Ctrl+J` agents panel, `/reload` for skills and slash index. - `docs/RELEASE-0.9.8.md`. ### Changed - Harness performance: cached LiteLLM capability probes, static `prepare()` cache, compaction tool-token reuse, prompt-cache memoization, lazy loop prompt loading, shared parallel tool executor. - Textual poll intervals for status and sidebar (less idle CPU). - Tighter `cli_import` and `repl_chat_init` budgets in `kite bench --check`. ### Fixed - Textual 8: theme CSS applies via stable `read_from` key + `refresh_css` (removed `Stylesheet.clear()` crash on mount). ## [0.9.7] - 2026-09-11 ### Added - **Paid web backends** — optional `TAVILY_API_KEY` / `EXA_API_KEY` / `FIRECRAWL_API_KEY` in `~/.kite/.env` via `kite web-keys set tavily|exa|firecrawl` (or `kite keys --set …`). `websearch` auto-tries Tavily → Exa → Firecrawl → DuckDuckGo; Firecrawl also upgrades `webfetch` / `webcrawl`. - **`docs/RELEASE-0.9.7.md`** — release notes for this version. ### Changed - **Layout** — nested one-file application packages, `eval`, `tasks`, and plugin extensions are now modules (`kite.application.cli`, `kite.tasks`, `kite.plugins.extensions`, …). Public names stay the same. - **Test suite** — collapsed ~555 one-assert tests into ~174 domain modules (`tests/README.md`). Same contracts, less slop. - **Scripts** — only install/download (unix + Windows), `sync_version.py`, and `bump_release.sh`. Dropped `pkg.*`, `lint.sh`, `download-macos.sh`, and `verify_release_pr.sh`; CI and `uv tool` cover those jobs. - **Docs** — keep architecture/glossary/commands plus current `docs/RELEASE-X.Y.Z.md`. Dropped duplicate atlas, old release notes, `plan.md`, and one-shot merge/bench notes. Version history stays in `CHANGELOG.md`. - **`/help` / `kite help`** — list the remaining docs (`kite_commands.md`, `CONTEXT.md`, `architecture.md`, `SECURITY.md`, current release notes) instead of deleted atlas/memory files. - **Headless success** — `kite run --headless`, `kite exec`, and `kite tasks` succeed only when the run submitted **and** leftover jobs were torn down. - **Startup load path** — CLI parser no longer imports setup/dashboard/bench suites; catalog/runtime TOML parse once; credentials skip BYOS until needed; LiteLLM and approval policy load on first use. ### Fixed - **REPL** — blank Enter no longer starts a turn; “Setup needed” no longer shows when a usable key exists but `default_model` is unset (catalog groq has an empty default). - **`kite tasks run`** — requires a file (or `--stdin` / `-`); `--steps` / `--cost` / `--time` pass through. - **`kite exec`** — always headless and quiet unless `-v`. - **Claude subscription login** — CLI link still needs `ANTHROPIC_API_KEY` for Kite model calls; login success text says so. - **ChatGPT BYOS hang** — LiteLLM expects flat `auth.json`; Codex stores nested `tokens`. Kite now materializes a LiteLLM-compatible copy under `~/.kite/oauth/chatgpt/` from the official Codex store so the harness no longer blocks on device-code login. ## [0.9.6] - 2026-09-09 ### Added - **Themes** — palettes (`kite`, `dark`, `light`, `dim`, `mono`, `catppuccin`, `ember`, `forest`, `hues`, `transparent`) wired through TUI chrome; `/theme` and `/font`. - **Subagent orchestration** — auto sync/async dispatch, `wait_for`, crew TUI, `/agents` board, `/live agents`. - **Personas** — bundled scout/coder/reviewer/context/shell profiles; `kite subagents` CLI; `~/.kite/subagents/.md` overrides. - **Global user context** — `USER.md` / `PROFILE.md` (`/user`, `/profile`) plus fluid working rhythm (`WORKING.md`, `/working`). - **Goal + recovery** — `/goal` persistent objective; resume/retry after provider errors (`kite resume --retry`). - **Headless tasks** — `kite tasks` JSONL batches and `kite run --headless`. - **Web tools** — webfetch readable extraction, websearch ranking/dedupe, DuckDuckGo unwrap, crawl budgets. - **`docs/RELEASE-0.9.6.md`** — release notes for this version. - **`scripts/download-macos.sh`** and **`scripts/lint.sh`** — macOS bootstrap and contributor lint. ### Changed - **BYOS authentication** — provider-delegated auth for ChatGPT/Codex (`openai-codex` SDK), Claude Code (`claude auth` CLI), and Grok (`grok` CLI). Removed LiteLLM private OAuth internals. - **`kite logout`** — unlink BYOS subscriptions (`kite logout codex|claude|grok|xai`). Claude subscription stays separate from Anthropic BYOK; tokens are never copied into `~/.kite/.env`. - **Slash startup** — first `/` completion prewarms so the REPL menu opens faster. - **CI** — pytest runs on all pull requests, not only those targeting `main`. ### Security - Auth errors sanitize token-shaped strings; OAuth secrets are not logged or written to project `.env` files. - **Recursive secret redaction** — nested audit/event/session/tool payloads sanitized before persistence or display. - **Session persistence policy** — `full` | `redacted` (default) | `disabled`; owner-only session files when enabled. - **Process-tree teardown** — timeout/cancel kills child process groups (Unix `killpg`, Windows `taskkill /T`). - **Environment isolation** — sensitive env keys stripped from child processes; `extra` overrides cannot re-inject credentials. - **Skill trust model** — bundled trusted; npm/git/project/user untrusted with provenance metadata. - **SSRF hardening** — resolve → validate all IPs → connect-time peer check; redirect and DNS TOCTOU re-checks. - **Harness OS protection** — `/proc` `/sys` `/dev` blocked for read/write/bash; restricted mode blocks all network tools (bash, webfetch/websearch/webcrawl, Context7). - REPL `/privacy` and CLI `kite privacy` / `kite config --session-persistence`. - System prompt documents harness limits, attachment handling, and user interrupt behavior. - REPL **F8** / **Esc v** clipboard attach, **Ctrl+L** clear screen; shortcuts in `/help`. ### Fixed - Windows permission / process-group tests skip Unix-only assertions; bash cancel tests use `sleep`. - Ruff unused-import failures in BYOS tests; duplicate dict keys in `build_design_pdf.py`. ## [0.9.5] - 2026-09-07 ### Added - **Pi-style runtime steering** — mid-turn corrections inject into the active run and continue the same turn; follow-ups queue at turn boundaries (`RunMessageQueue`, `harness.inject_user_message()`). - **REPL message inbox** — Enter queues a follow-up while busy; Ctrl+G steers; Ctrl+U dequeue-to-composer; `/steer` and `/tasks`; steer vs follow-up labels on the toolbar. - **Compaction lifecycle events** — `compaction_start` / `compaction_end` on the agent event bus (durable in session JSONL). - **Harness timing budgets** — `kite bench --check` and `tests/test_bench.py` enforce median-ms ceilings; budgets in `src/kite/bench/budgets.py`. - Expanded `kite bench` suite (18 cases): `user_config_load`, `catalog_load`, `slash_index`, `repo_map`, `runtime_prepare`, plus existing startup/context/tool benchmarks. - **Session browser** — `kite sessions` table (date, time, title, model, status); `kite sessions -q` filter; `kite resume ` with prefix match and suggestions. - **`~/.agents/skills`** — global Agent Skills library loaded on any machine (sandbox read allowed). - **`docs/RELEASE-0.9.5.md`** — release notes for this version. ### Changed - **REPL cold start** — startup banner uses `assess_setup_status_fast()` (no `resolve_model` on boot). - **Busy composer** — unified slash dispatch while a turn is running; live activity preview on the toolbar during bash/jobs. - **Resource use** — session rewrite on compact (not append-only snapshots); bounded `read` (256 KB / 400 lines); TTL cache LRU caps; discovery/slash/skills cache limits; job registry prune (24 h finished); checkpoint cap (5); repo-map walk capped at 600 files. - **Application run state** — `ApplicationRunService` tracks run transitions from loop events; `after_prepare` hook fired from runtime. - **Test suite** — ~320 focused tests (down from ~540); merged `test_util`, `test_ui_basics`, `test_cli_misc`, `test_agent_loop`, `test_agent_runtime`, `test_repl_misc`; removed redundant/trivial modules. - **Docs** — removed stale `docs/adr/`, `docs/superpowers/`, `cli-ux.md`, `ideal-cli-spec.md`, `kite-0.9-architecture-program.md`; updated references in AGENTS, README, architecture, CONTEXT, kite_commands. - **CI** — `pytest` and `kite bench --check` on every push/PR; release verify script checks bench gate. ### Fixed - **Streaming TUI** — busy composer no longer restarts on every agent event; hints use toolbar flash instead of scrollback spam. - Provider retry countdown in the running line; compaction start/end render in the transcript. - Status tail rendering aligned with `format_status_tail` expectations in tests. - `compaction_end` always emitted after `compaction_start`. ## [0.9.4] - 2026-09-07 ### Added - **`/live`** — toggle real-time bash and background-job output in the REPL (`SessionUiState.live_terminal`). - **Auto venv** — `discover_venv()` + `prepare_child_env()` prepend project `.venv`/`venv` to bash `PATH` when `pyvenv.cfg` is present (`[environment] auto_venv` in config). - Execution context prompt line documents detected `python_venv`. ### Changed - Bash output routes through `tool_output` / `job_output` events when a subscriber is attached (Rich TUI); stderr fallback for headless runs. ### Fixed - (Includes all fixes shipped on main since v0.9.3: guardrails, SSRF, REPL responsiveness, approval composer keys, token defaults, agent/persistence reliability, CI gates.) ## [0.9.3] - 2026-09-06 ### Fixed - **False "work complete" on task requests** — assistant replies like `Hey! 👋` no longer auto-submit; only the **user's** casual turn (hi/thanks/short Q&A) may end in text-only submit. - Footer **running line** clears on `agent_end` so idle chrome does not show `working · 1 task` after completion. ### Changed - System prompt + `CONTEXT.md` + `kite_commands.md` document the completion decision table (user intent vs model reply). - `_is_casual_user_turn` recognizes `hi kite` and task keywords (`test`, `lower`, `can you`, …). ## [0.9.2] - 2026-09-06 ### Added - **Opt-in durable memory** — MEMORY.md and episodic notes inject only when the user asks (`/remember`, `/memory`, or `[memory] inject = "always"`). - **Working-state continuity** — compact/budget-continue briefs inject separately from durable memory; no auto-pin to MEMORY.md. - **Daily-driver loop** — unverified edits get verify nudges instead of idle stall; blocked submit includes suggested verification command; fuzzy `edit` fallback; compaction preserves edited paths. - **Web tools overhaul** — `webfetch` returns extracted readable text (title, metadata, JSON pretty-print); DuckDuckGo redirect unwrapping; search dedupe + HTML parser fallback; private-network URL blocking; `tests/test_web.py`. - **Composer `@` completion** — `@path` attach tokens complete like `/attach` (word-boundary `@`, skips email addresses). - **Persistent verification badge** — footer shows live `verification_status` (e.g. `unverified edits`) during runs. ### Changed - System prompt clarifies memory is not instructions unless loaded. - **`webfetch`** consolidated into `web.py` (shared fetch/extract with `webcrawl`); no longer returns raw HTML soup by default. - **`websearch`** tries GET lite fallback when HTML POST fails; unwraps tracking URLs from DuckDuckGo/Google. - Approval toolbar shows **`[a]/[n]/[q]`** keys during prompts instead of queue/steer hints; composer placeholder switches to approval keys. - **80% cost warnings** flash on the toolbar while the busy composer is pinned (no scrollback jump). - Failed tool rows show **collapsed multi-line errors** (same collapse as bash output). - Flash notes on the footer **expire after 8s** so stale messages do not linger. - **`Tab`** explicitly cycles slash and `@` completions; **Enter** always sends the line. ### Fixed - Approval wait no longer counts as idle no-tool turns. - Relative cache deletes on Windows stay out of mandatory approval in auto/yolo. ## [0.9.1] - 2026-09-06 ### Added - **Interactive budget floors** — chat uses ~80 steps / $10 when on package defaults; honor explicit lower user caps. - **Budget auto-continue** — up to 2 resumes on `LimitsExceeded` with unfinished work + continuity brief; inbox queue wins. - **Continuity memory** — Codex/Pi-style briefs (mission / done / next / todos) after compact and before budget continue; episodic store + prompt inject. - **Structured `submit` tool** — preferred completion path with `message` (Done / Changed / Verification); legacy bash marker still supported. - **Repo map** — `context/repomap.py` injects Aider-style symbol sketch; git-changed files ranked first. - **ToolExecutor production cutover** — agent loop routes tools through `PolicyEngine` + `ToolExecutor` by default. - **Live verification UX** — `verification_status` events update footer during runs; `submit_blocked` when claims outrun evidence. - **Replay acceptance** — `ReplayBundle.events` + `acceptance` criteria for transcript-level eval without live providers. - **EvidenceVerifier wiring** — bash check commands feed evidence ledger; summary included in verification payload. ### Changed - Default `compaction_ratio` **0.80 → 0.75** (auto-checkpoint remains ~72%). - `LimitsExceeded` / `TimeExceeded` render as soft pauses with continue hints (not hard error spam). - Documentation refreshed for adaptive budget, continuity, and busy composer chrome (`kite_commands.md`). ### Fixed - `/compact` (and other zero-arg slash handlers) no longer raise `TypeError` when dispatch passes an empty arg. - WaitSpinner no longer wipes the pinned busy composer; activity updates the toolbar instead. - Status toolbar no longer crashes on pending approval (`bits` → `parts`). ## [0.9.0] - 2026-09-04 ### Added - **Application layer** — `RunSpec`, `EventEnvelope`, `ApplicationRunService`, `HarnessDependencies` (`src/kite/application/`). Existing `Harness` is a compatibility adapter; production CLI/REPL still use it. - **Context assembler** — `ContextItem` / `ContextBudget` / `ContextSnapshot` with provenance, source budgets, and untrusted-content delimiters. - **Policy and execution seams** — `PolicyEngine`, `ToolExecutor`, `ProcessRunner`, `ChangeJournal` (restore without `git reset --hard` on the journal itself). - **SQLite event store** — WAL append/load/redact/resume reconstruction (`SQLiteEventStore`). JSONL sessions remain the production store. - **ModelGateway** — typed provider error categories, retries, `BudgetLedger` (subagent cost counted once). - **EvidenceVerifier** — verification from tool results, not model claims. - **CLI/REPL contracts** — `CliResult` exit codes, `ReplEventReducer` (adapters; live REPL still projects legacy events). - **Recorded replay** — `kite.eval.ReplayBundle` runs without live providers. - **CI** — Linux and Windows × Python 3.11 and 3.12; ruff on the application layer. - **Busy composer** — pinned prompt while a turn runs; Enter queues; `/tasks`; footer tok/s and cache hit. - **Skill library links** — `/skills add ./path` (or an absolute folder) **symlinks** into `~/.kite/skills` (Windows directory junction if a symlink is refused); project `.kite/skills/` points at the global copy. Restricted mode may **read** that library (and symlink targets); writes stay sandboxed. Reinstall unlinks; it does not delete the real tree. ### Changed - Milestone docs describe adapters as **landed**, not full production cutover. - Packaged execution mode stays **host**; `PolicyEngine` defaults to restricted when used as the new seam. - Sibling-prefix path checks are separator-aware; host mode does not clamp to the workspace (protected paths still blocked). ### Fixed - ChangeJournal reports a conflict when the user deletes a file the agent wrote, instead of rewriting it. - `BudgetLedger.total_cost()` no longer double-counts subagent usage. ## [0.8.2] - 2026-09-02 ### Added - **JobRegistry** - unified background bash (ackground=true) and live subagents; /jobs, /kill [id|all], kill-on-quit; footer jobs N. - **Busy composer** - Esc/Ctrl+C//stop, Ctrl+G//steer, Enter queues; busy chrome on footer/composer; main-thread SIGINT. - **Left-bar pickers** - numbered TTY menus for approve/theme/skills/sessions/models and related CLI flows. - **Bundled skills** - /orchestrate, /research, /pr. - **SYSTEM.md / APPEND_SYSTEM.md** - project or user system prompt replace/append. - **Package scripts** - scripts/pkg.sh / pkg.ps1 update|reinstall|uninstall. ### Changed - Plan/build handoff - plan writes checklist + risks; build continues the list. - Thinking traces expanded by default; mouse capture off by default; clipboard shortcuts in composer. - Submit gate blocks narrated done without a recorded passing check. ### Fixed - BYOS login opens browser + left-bar device/code flow; interactive select/setup polish. ## [0.8.1] - 2026-09-01 ### Added - **`kite login`** — CLI for BYOK (hidden API key, double-entry for new keys) and BYOS OAuth (`chatgpt`, `claude`, `grok`). - **`kite dashboard`** — per-user session analytics: tools, tokens, cache, cost, attention queue (`--session`, `--json`, `--watch`). - **`--long` task mode** — higher step/cost limits, phase checkpoints, `mode_long.md` for multi-hour work. - **Context7 docs tools** — built-in `context7_resolve` / `context7_docs` (optional `CONTEXT7_API_KEY`); session UTC/local time in system prompt. - **Token-efficient tools** — bash-first inspection; lean `read`; plan-mode read-only bash via `is_inspection_bash()`. - **UI polish** — unified approval panel; terminal-style bash cards; thinking collapsed by default (`/expand-thinking`, Ctrl+T); colourful task progress strip. - **Provider retry/continue** — exponential backoff on transient faults; session preserved (`kite resume `). - **`load_kite_env()`** — project `.env` first; `~/.kite/.env` fills unset/empty keys (fixes `.env.example` placeholders). - **BYOK login UX** — validation, permission warnings, masked key fingerprints in `/keys` and success messages. - **Mandatory approval** — high-risk bash (sudo, rm, installs, git writes, curl, …) always prompts; no yolo/auto/trust bypass. - **Agent completion** — no early submit; idle nudges; stop on tool errors with logs. - **Git approval split** — `git status`/`log`/`diff` auto-allow; `git add`/`commit`/`push` always gated. - **Memory `ForgetResult`** — `/forget` and `memory` tool report removed notes **and** episodes. - **Tool arg repair** — malformed model tool JSON repaired before execution (`models/tool_args.py`). ### Changed - **Setup wizard** — BYOK vs BYOS paths; OAuth during setup; credential-ready banner. - **Provider picker** — BYOK/BYOS labels, auth state (`linked`, `key set`, `login required`). - **`/help`** — canonical slash list + **legacy aliases** section (`/select` → `/model select`, etc.). - **Memory prompt** — slimmer injection; semantic `MEMORY.md` + episodic sqlite. - **LiteLLM retries** — `num_retries=0` when agent loop handles `provider_max_retries`. - **Anti-bloat** — dead code removed; GitHub tools off by default; trimmed prompts. ### Fixed - **REPL `/keys`** — OAuth shows `linked` / `login required`, not `missing oauth`. - **Duplicate provider fault UI** — `agent_end` no longer re-renders after `provider_fault`. - **Credential env loading** — empty project `KEY=` no longer blocks `~/.kite/.env` keys. ## [0.8.0] - 2026-09-01 ### Added - **BYOS (Bring Your Own Subscription)** — OAuth login for ChatGPT/Codex, Claude, and Grok via `kite keys --set ` or `/login` in the REPL; dynamic model lists from subscription APIs (`src/kite/providers/byos.py`). - **Approval modes** — `yolo` (no prompts), `auto` (auto inside workspace, ask outside), `supervised` (reads free; all mutations need approval); aliases on CLI and `/approve`. - **Submit gate** — blocks `COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT` when edits lack passing test/lint artifacts, tests failed, or summary claims success without evidence. - **Verification hardening** — progress-aware loop guard (reset on changed output, hard-stop at 5 repeats), post-edit verification nudges, broader test-command detection. - **Token optimizations** — 80% compaction ratio, fast deterministic compaction below 92% context, summary-aware observation elision (8k default), tool-pair-safe compaction tail. - **Durable session events** — JSONL event stream alongside transcript messages. - **Subagent timeout** — configurable `orchestrator_timeout_seconds` (default 300s). - **`api_styles` config** — per-provider route hint (`chat` | `messages` | `responses`) on `ResolvedModel`. - **BYOK model picker** — radiolist TUI for API-key providers; subscription providers skip live picker after OAuth. ### Changed - Default `max_bash_output_chars` lowered to 32k; `system.md` trimmed with token-aware peek guidance and evidence-first verifiability rules. - `mode_build.md` documents harness submit blocking and required `## Verification` section. ### Fixed - Circular import in `config/readiness.py` when loading OAuth provider modules. ## [0.7.2] - 2026-08-31 ### Added - **`/restricted on|off`** (alias `/sandbox`) — REPL toggle for path sandbox; **off by default** (host mode). - **Slash menu scroll** — mouse wheel and scroll-key bindings on the `/` completion dropdown. - **Compaction ctx meter** — `/compact` and auto-compaction refresh the footer context bar immediately. ### Removed - **MCP integration** — stdio MCP client and `[[mcp]]` runtime config removed; use `Harness.extra_tools` for custom tools. ### Changed - **Default execution mode** — `host` instead of `restricted` in runtime TOML and guardrail defaults; use `/restricted on` for a tighter sandbox. - **Runtime assembly** — merged extra-tool wiring; audit listener registered once; cached `UserConfig` and bundled prompts. ### Fixed - **Bare `kite` launch** — `readiness` used `os.stdin.isatty()`; now routes through `sys.stdin`. - **`set_cwd` outside repo** — sessions can move to Desktop or sibling dirs; sandbox follows execution cwd (protected paths still blocked). ## [0.7.1] - 2026-08-31 ### Added - **Tool cards UI** — structured tool rows (`▸ read path`), parallel batch headers, read line-count summaries, stream coalescing for less flicker. - **Setup readiness** — `config/readiness.py`; first-run prompt on bare `kite`; REPL `/setup` wizard; `kite providers` / `kite keys` show ready/not-ready status. - **Install `--setup`** — `./scripts/install.sh --setup` and `install.ps1 -Setup` run the wizard after install (TTY only). - **`kite help`** — grouped quick reference CLI map; slimmer `/help` builtins with legacy aliases preserved. ### Changed - **CI** — pytest on every push and PR to `main` (Python 3.11 + 3.12); isolated `KITE_HOME` + `KITE_SKIP_SETUP` in CI. - Provider picker sorts configured and recommended providers (groq, openrouter, ollama) first. - Install scripts point new users at `kite setup` instead of editing `.env` manually. ### Fixed - `UserConfig.path` property for setup wizard config display. ## [0.7.0] - 2026-08-31 ### Added - **`kite bench`** — repeatable harness timing (CLI import, config, context, tools, prompt assembly) with `--json`, `--save`, and `--compare` for BEFORE/AFTER deltas. - **`set_cwd` tool** — session execution cwd separate from project root; file tools and bash resolve relative paths from execution cwd. - **`[guardrails] execution_mode`** — `restricted` (default sandbox) or `host` (broader filesystem access; protected paths still blocked). - **`ToolResult` contract** — structured tool outcomes and scheduling metadata (`tools/metadata.py`). - **Parallel read-only tools** — safe fan-out for concurrent `read`/`grep`/`glob`/`ls` in one model turn. - **Bash cancellation** — Ctrl+C / interrupt propagates to long-running bash subprocesses (`CancelToken`). - **Context checkpoints** — auto snapshot at ~72% context; `/checkpoint save|list|restore|show` in the REPL. - **`/handoff`** — export `.kite/handoff-.md` + `.json` for another agent or machine. - **Preserved-fact compaction** — shared `run_compaction()` path; facts block survives summarization. - **Lazy REPL init** — defer `resolve_model` until first task; reuse `Harness` across turns when config unchanged. - **Bundled `/handoff` prompt**; expanded `system.md`, mode/role prompts, and install scripts. - **`scripts/bump_release.sh`** — version bump + CHANGELOG stub + annotated tag helper. ### Changed - Guardrails follow live execution cwd; host mode allows explicit external paths when configured. - System prompt documents execution context, structured finish format (Done/Changed/Verification), and session continuity. - Install scripts bootstrap `~/.kite/` (including `checkpoints/`), support `--verify` / `-Verify` for post-install pytest. - Docs refresh: `CONTEXT.md`, `kite_commands.md`, `AGENTS.md`, `architecture.md`, design specs. ### Fixed - Missing `HarnessSlots` / `HookBus` imports in `agent/runtime.py` after merge. ## [0.6.8] - 2026-08-30 ### Fixed - Long-running `bash` tools stream output live instead of appearing hung until exit. - Slow tools emit `tool_progress` heartbeats; model API calls honor a 180s timeout. - Default `bash_timeout_seconds` raised to 120s. - CI commit-count job handles rebased/force-pushed batches; pytest runs via `uv run`. ## [0.6.7] - 2026-08-30 ### Added - `kite setup` and `kite keys` onboarding with hidden API key input to `~/.kite/.env`. - REPL `/login`, `/logout`, `/keys`, and `/select` for credentials and model selection. - Keyboard shortcuts: Ctrl+O expand tools, Ctrl+P plan, Ctrl+B build, Ctrl+S status. - `/thinking` and `/fast` level menus when the provider advertises both effort modes. - Maintainer-only `kite maintainer dashboard` (requires `KITE_MAINTAINER_KEY`). - `CONTEXT.md` and `AGENTS.md` for agent-readable repo guidance.