# Changelog ## [0.10.5](https://github.com/langchain-ai/managed-deepagents-sdk/compare/v0.10.4...v0.10.5) (2026-11-18) ### Features * **mda:** export a Fleet agent's MCP tools, connections and approval rules ([345f030](https://github.com/langchain-ai/managed-deepagents-sdk/commit/345f030869a646f9f7d9bc65eabe9f86e1f304a5)) * **mda:** export a Fleet agent's sandbox and access profiles ([3c19b2f](https://github.com/langchain-ai/managed-deepagents-sdk/commit/3c19b2f37d5ec3cc8684226aa6aae90b7de08562)) * **mda:** export a Fleet agent's Slack channel ([370a8b6](https://github.com/langchain-ai/managed-deepagents-sdk/commit/370a8b633afe9f7538c3bfc7cad6b2aea4870483)) * **mda:** export a Fleet agent's subagents ([5af9b34](https://github.com/langchain-ai/managed-deepagents-sdk/commit/5af9b34f63a36bb286ae6fbf72c985f0607aa3ae)) * **mda:** export Fleet skills and memory ([a83d050](https://github.com/langchain-ai/managed-deepagents-sdk/commit/a83d050a6dabec054b7bccd1091b57e92ed17c23)) * **mda:** mda init --from exports a Fleet agent ([fd9181a](https://github.com/langchain-ai/managed-deepagents-sdk/commit/fd9181adce116660f2176498dd8c6e884133cc98)) * **mda:** mint agent access tokens from the pod token in TypeScript ([a1b436e](https://github.com/langchain-ai/managed-deepagents-sdk/commit/a1b436e1899987eabd2b4119c2e1d0b75508a929)) * **mda:** print the agent identity at the end of mda deploy ([d6b0b81](https://github.com/langchain-ai/managed-deepagents-sdk/commit/d6b0b8185d9e603190f3648daac37eb90da92c51)) * **mda:** Tool Gateway servers declare auth "langsmith" ([#1353](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1353)) ([33a1251](https://github.com/langchain-ai/managed-deepagents-sdk/commit/33a12516fcd468f8402bfdf6d86e41c44c53e66c)) * **runtime:** add agent access tokens to Gateway and Managed Tools calls ([#1365](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1365)) ([8287ba4](https://github.com/langchain-ai/managed-deepagents-sdk/commit/8287ba41439796fae958d33037477c61f4c14dff)) ### Bug Fixes * **cli:** let build runners compile without credentials in `mda prepare` ([#1313](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1313)) ([b67d8a7](https://github.com/langchain-ai/managed-deepagents-sdk/commit/b67d8a729a16cec5d3246542337ad0a0ed2bfda4)) * **mda:** address review findings across the Fleet export stack ([28a0e69](https://github.com/langchain-ai/managed-deepagents-sdk/commit/28a0e696c5d4a3367f60656b97f2fe2821c6dc29)) * **mda:** delete integration test agents in CI cleanup ([9b35f1b](https://github.com/langchain-ai/managed-deepagents-sdk/commit/9b35f1b390ffb7dae3ac8921bb8be2f7b901ba7b)) * **mda:** keep MCP tool calls on the cached catalog with FastMCP 4.2 ([bf3862d](https://github.com/langchain-ai/managed-deepagents-sdk/commit/bf3862dbd1c1960782ef2a11976b676e6c690da5)) * **mda:** return tool errors to the model instead of ending the run ([649c2db](https://github.com/langchain-ai/managed-deepagents-sdk/commit/649c2db4e4403076af328548547aae5cf3519eca)) * **mda:** stop integration test leaks in MDA CI ([8cbff69](https://github.com/langchain-ai/managed-deepagents-sdk/commit/8cbff693fe75f81e63812f14af96da3e509d123b)) * **mda:** stop mda dev restarting the server on every save when LangSmith settings are exported ([c6d01ca](https://github.com/langchain-ai/managed-deepagents-sdk/commit/c6d01caec2b0b6fe08015e57b2177b4ec54d1fc1)) * **mda:** tag launched async workers as subagents in traces ([4110827](https://github.com/langchain-ai/managed-deepagents-sdk/commit/4110827cddfc1dde35b4bdfbe7ebf5e03f65af61)) * **runtime:** give sync subagents their own credential gate ([f815ec2](https://github.com/langchain-ai/managed-deepagents-sdk/commit/f815ec2583a572a73c6a29c031e7b0bd3d46522a)) ### Chores * **cli:** default scaffolded agents to openai:gpt-6-luna ([#1364](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1364)) ([5952530](https://github.com/langchain-ai/managed-deepagents-sdk/commit/59525309f10ea8b3d2d169ca0c9af05f1fa0269d)) * **deps:** bump @hono/node-server ([#1339](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1339)) ([36a5ce7](https://github.com/langchain-ai/managed-deepagents-sdk/commit/36a5ce7a3f4f65ecc2d598012a72e0ac6d5c43b4)) * **deps:** bump dnspython from 2.9.0 to 2.9.2 in /packages/pypi ([#1351](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1351)) ([0972c37](https://github.com/langchain-ai/managed-deepagents-sdk/commit/0972c37a9d779517fc478bce0763bd9ab217880c)) * **deps:** bump opentelemetry-api ([#1350](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1350)) ([a0598c3](https://github.com/langchain-ai/managed-deepagents-sdk/commit/a0598c3cd7e01532023fe94bc6860dcd092cda67)) * **deps:** bump opentelemetry-api in /packages/pypi ([#1352](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1352)) ([a31eb87](https://github.com/langchain-ai/managed-deepagents-sdk/commit/a31eb878eef86bd6bb72fd2aa860e6d11c8a816d)) * **deps:** bump oxc from 0.157.0 to 0.158.0 in /sdks/managed-deepagents ([4897f27](https://github.com/langchain-ai/managed-deepagents-sdk/commit/4897f27e73a2abc5506f24970edbfec8bab871e4)) * **deps:** bump the minor-and-patch group in /sdks/managed-deepagents with 5 updates ([9f0d635](https://github.com/langchain-ai/managed-deepagents-sdk/commit/9f0d63524d0627492c774481c1febbf4ed876966)) * **deps:** bump the minor-and-patch group in /sdks/managed-deepagents/packages/npm with 11 updates ([32f2be4](https://github.com/langchain-ai/managed-deepagents-sdk/commit/32f2be4826790196b747362bf0011f2e75bc846f)) * **deps:** bump the minor-and-patch group in /sdks/managed-deepagents/packages/pypi with 10 updates ([b1c48a1](https://github.com/langchain-ai/managed-deepagents-sdk/commit/b1c48a1c6966d63d4212a6375f5caf518bf68314)) * **deps:** bump the npm_and_yarn group across 3 directories with 4 updates ([66abee8](https://github.com/langchain-ai/managed-deepagents-sdk/commit/66abee8bea68c693b5632eec7d392c8615c78412)) * **deps:** bump uv-pep440 from 0.0.99 to 0.0.100 in /sdks/managed-deepagents ([0dad531](https://github.com/langchain-ai/managed-deepagents-sdk/commit/0dad53121ab9277cae60f0fe5f3f471694919310)) * hand the changelog to release-please ([#1367](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1367)) ([e60c3ce](https://github.com/langchain-ai/managed-deepagents-sdk/commit/e60c3ceb524c222e76d159ba04c87093a3edcb7e)) * **mda:** move managed-deepagents-sdk into sdks/managed-deepagents ([3b82a54](https://github.com/langchain-ai/managed-deepagents-sdk/commit/3b82a543bb110e8f340521c4501a64bd37948f1c)) * **release:** 0.10.4 ([#1354](https://github.com/langchain-ai/managed-deepagents-sdk/issues/1354)) ([232ebb9](https://github.com/langchain-ai/managed-deepagents-sdk/commit/232ebb9348b99b535e61bcc3740a01ab770da109)) ### Documentation * **mda:** drop the PyPI hosts note for browser recording ([8660dfe](https://github.com/langchain-ai/managed-deepagents-sdk/commit/8660dfef136e1c6c98e0288c359bf464fdad5a0c)) ### Refactors * **mda:** delete dead runtime code in npm and pypi ([e3e26ae](https://github.com/langchain-ai/managed-deepagents-sdk/commit/e3e26ae27cdf1e11d250ca73827f6d81eafafc5e)) ## 0.10.4 - 2026-11-14 ### Added - Limit the managed browser to an allow list of hosts (`browser`) (#1250) - Let a GitHub channel handle its own App's events (`channels`) (#1143) - Add runtime-resolved assistant memory scope (`memory`) (#1297) - Scaffold AGENTS.md in mda init (`cli`) (#1270) ### Changed - Bump hono from 4.13.14 to 4.13.16 in /packages/npm (`deps`) (#1344) - Use Agent Server naming and flags with hidden deployment aliases (#1323) ### Fixed - Pin langgraph-api 0.19.0rc1 instead of 0.19.0.dev8 (`cli`) (#1347) - Fit the memory cache breakpoint to the request's budget (`memory`) (#1341) - Summarize a public App's installations in the channel setup output (`github`) (#1151) - Keep a finished run's result when its status read gets no answer (`channels`) (#1326) - Read pasted Slack message links from the preview Slack adds (`channels`) (#1315) - Call Tool Gateway as the agent in agent-owned schedules (`runtime`) (#1317) - Use Stagehand 4.4's watchdog and disconnect() instead of 4.3.0 workarounds (`browser`) (#1213) - Recover from credential outages during tool execution (`mda`) (#1298) ## 0.10.3 - 2026-11-09 ### Added - Let authenticated users create crons that run as themselves (`runtime`) (#1290) - Let local mda dev callers manage api crons (`runtime`) (#1288) - Let api key callers create and update untagged crons (`runtime`) (#1282) - Mount authored backends with define_deep_agent(routes=...) (`runtime`) (#1272) - Slack status false hides the status and stop button (`channels`) (#1271) ### Changed - Bump the major group with 2 updates (`deps`) (#1293) - Bump the minor-and-patch group with 3 updates (`deps`) (#1292) ### Fixed - Retry transient channel reply delivery failures (`mda`) (#1299) - Keep credential-bearing redirects within their origin (`cli`) (#1300) - Avoid project conflicts with agent tracing (`dev`) (#1285) ## 0.10.2 - 2026-11-06 ### Added - **cli:** `mda init` scaffolds an `AGENTS.md` for coding agents working in the project. It lists the project's commands and asks the agent to end each session by asking whether to keep the agent on `mda dev` or deploy it. It tells the agent not to run `mda deploy` unless the user asks. Build and deploy leave the root `AGENTS.md` out. - **cli, npm, pypi:** A persistent schedule no longer needs a thread `id`. Declare `thread: { mode: "persistent" }` / `thread={"mode": "persistent"}` and MDA derives a stable thread ID from the agent and schedule name, then creates that thread on deploy, so every run shares one history without a Studio step or a second deploy. Deploy also creates a missing thread for an explicit `id`; a thread that another caller owns is still refused. An explicit `id` must now be a UUID, which `mda deploy` checks at compile time. (#1238) - **npm, pypi:** Declarative async subagents accept an optional `mcp` selection that narrows their MCP tools. Omit it to inherit the agent's MCP servers, list `defineMcp` / `define_mcp` definitions to keep only those, or pass `[]` or `null` / `None` for none. An async subagent loads only the servers it selects and requests only their credentials. Other connectors and authored tools are unaffected. (#1200) - **npm, pypi:** Sync declarative subagents also accept `mcp` to select tools from servers the agent loads. Omission keeps normal tool inheritance; `[]` or `null` / `None` removes MCP tools. Authored and non-MCP connector tools are preserved. Sync `mode: "fork"` subagents cannot set `mcp`. (#1201) - Steering / deliver channel messages to a running agent at its next model call (`mda`) (#1206) - Cache compiled graphs across runs with `cached_agent` (`runtime`) (#1248) - Ask for LangSmith tool kit connections on first use (`mcp`) (#1118) - Open a thread's managed sandbox (`sandbox`) (#1148) - Pull remote instructions and skills into local projects (`cli`) (#1203) - Add notification metadata to handoff notifications (`browser`) (#1091) - Add self-hosted langsmith development mode to mda-local (`mda-local`) (#1072) ### Changed - Bump the minor-and-patch group across 1 directory with 12 updates (`deps`) (#1221) - Bump the minor-and-patch group across 1 directory with 10 updates (`deps`) (#1220) - Bump pyasn1 from 0.6.4 to 0.6.5 in /packages/pypi (`deps`) (#1209) - Bump joserfc from 1.7.5 to 1.7.6 in /packages/pypi (`deps`) (#1207) - Bump the minor-and-patch group across 1 directory with 3 updates (`deps`) (#1210) - Bump the uv group across 2 directories with 1 update (`deps`) (#1208) - Bump the major group with 2 updates (`deps`) (#1187) - Bump docker/login-action (`deps`) (#1188) ### Fixed - **npm, pypi:** `auth.langsmithApiKey()` / `auth.langsmith_api_key()` treats a personal access token as a person instead of a service principal. Its caller can use user-owned connections, is eligible for user memory, and resolves to the same Agent Auth principal as that user in deployed Studio. Service keys stay workspace service principals, and thread ownership keys are unchanged. (#1240) - Run async subagents in mda eval (`eval`) (#1172) - Read the mda dev project directory once, off the run's event loop (`sandbox`) (#1242) - Never give a thread a second sandbox on first open (`sandbox`) (#1214) ## 0.10.1 - 2026-11-02 ### Added - Handle GitHub merge queue events (`channels`) (#1133) - Keep a Linear agent session alive while its run works (`channels`) (#1135) ### Changed - **cli:** `mda init --channel`, `mda channels init` and `mda channels setup` warn when they add or set up an experimental GitHub or Linear channel, their `--help` marks both experimental, and `mda channels setup --json` reports `experimental` for each channel. (#1144) - Recover from transient Agent Auth credential lookup failures (#1170) - Copy new attachments concurrently (`slack`) (#1165) - Keep MCP tool listings for five minutes (`python`) (#1164) - Load credential-gate consumers concurrently (`python`) (#1163) - Bump rustls from 0.23.45 to 0.23.46 (`deps`) (#1130) - Bump the uv group across 2 directories with 1 update (`deps`) (#1128) - Bump undici (`deps`) (#1102) - Bump dnspython from 2.8.0 to 2.9.0 in /packages/pypi (`deps`) (#1100) - Bump authlib from 1.8.0 to 1.9.1 in /packages/pypi (`deps`) (#1101) ### Fixed - Reuse credential tools within a verified run (`python`) (#1156) - Name the deployment and channel on run failures, and count closed streams as cancelled (`runtime`) (#1173) - Tell a Linear session at once when its run can't start (`channels`) (#1136) - Say which signed Linear delivery the replay window refuses (`channels`) (#1134) - Stop copying the bot's own uploads into the sandbox (`slack`) (#1158) - Upgrade vulnerable Zod and KaTeX dependencies (#1166) - Save TypeScript screenshots into thread state with deepagents 1.16.1 (`browser`) (#1138) - Connect afresh when Chrome restarted since the last call (`browser`) (#1107) - Tell the model when its browser was replaced or a recording lost (`browser`) (#1097) - Save screenshots without a sandbox (`browser`) (#1088) - Stop Slack channel deploys that use a service key, and say which key to use (`cli`) (#1140) - Let agents delete memory files (`mda`) (#1142) - Warn when deploy drops a reserved env var (`mda`) (#1141) - Say why a channel reply failed (`channels`) (#1131) ### Experimental - **cli:** `mda channels setup [slack|github|linear]` runs the channel setup that `mda deploy` runs, against the deployed agent, without deploying. In a terminal it asks which channels to set up. Without a terminal, or with `--json`, it only reports where each channel stands, and it exits non-zero unless every channel is set up. `mda deploy` now sets up an existing deployment's channels before it uploads the new revision, so a channel that cannot be set up stops the deploy before the upload. (#1132) ## 0.10.0 - 2026-10-29 ### Added - **Async subagents** (`npm`, `pypi`, `cli`). A subagent marked `async: true` runs in the background as its own graph on the same deployment, in TypeScript and Python projects. It shares the lead agent's sandbox and trace, loads the skills its `skills` names, and passes `update_strategy`, `webhook`, `notify_parent` and `mode: "fork"` through to deepagents. Every served agent also gets a background copy of the general-purpose subagent. On a channel thread, a subagent's updates and approvals post to the lead agent's thread, and an answer there resumes it. (#860) (#728) (#921) (#888) (#894) (#945) (#977) (#951) (#980) (#1018) (#1020) (#1022) - **Browser use** (`npm`, `pypi`, `cli`). A root `browser.ts` / `browser.py` exporting `defineBrowser()` / `define_browser()` gives the agent a LangSmith browser and Stagehand's `browser_run`, `browser_snapshot` and `browser_screenshot` tools. A browser belongs to a thread, or with `scope: "user"` to a signed-in person across their threads, logins included. Factories can replace it or turn it off with `browser`. `browser_handoff` hands the browser to the user and wakes the agent when they hand it back. `browser_record` records the browser tab as an mp4 in the agent's sandbox, for example to show a person a UI change. `mda` adds Stagehand only to projects that declare a browser. See [Browser](README.md#browser). (#967) (#1064) (#1010) - **npm, pypi:** `runtime.thread` (`id`, `origin`, `participants`). Managed ingress keeps a thread record in thread metadata (`ls_mda_thread`, `ls_mda_participants`) on every channel message: the channel the latest message came from, who sent it, and the people on the thread (`owner` or `member`). Every run on the thread reads it, and the run posts its own answer and its approvals to that origin when it ends. So a LangSmith Studio resume, a scheduled run and an async subagent's wake-up run all answer the channel conversation. The record never holds message text or provider events. (#987) - **npm, pypi:** Runs can be stopped. Slack threads get a native Stop button that cancels the thread's runs. Studio can cancel runs on Slack threads, which used to answer `404 no matching runs to cancel`. Cancelled runs report a `run.cancelled` lifecycle event instead of an empty `run.interrupted`. (#870) - **npm, pypi:** Slack channels take a `triggerMode`: `"mention"` (the default for new apps), `"thread"` to keep replying in a thread after one mention, or `"all"`. When unset, a deployed agent keeps its current mode. `triggerOnAllMessages` is deprecated in favor of `triggerMode: "all"`. (#744) - **cli:** `mda deploy` asks which Slack app a new Slack channel uses: one LangSmith creates and manages (naming the workspace you authorized it in), a new one you create from a manifest, or one you already have. Self-hosted LangSmith offers your own app only. Your app's credentials come from prompts, or from `MDA_SLACK_BOT_TOKEN` and `MDA_SLACK_SIGNING_SECRET` without a terminal. LangSmith keeps them, so later deploys ask nothing. (#994) (#1053) - **npm, pypi:** `BrowserMiddleware.browser_name(runtime)` / `browserName(runtime)` gives the name of the run's LangSmith browser, so an app can open the same browser before the agent has browsed. (#1108) - **cli:** `--env-file` chooses the env file per environment. (#1075) - **runtime:** Logs how long each phase of the graph build takes. (#1054) ### Experimental These APIs ship in 0.10 and may change in a minor release. Their docs in code say so, and `mda dev`, `mda deploy` and `mda build` show a warning when a project declares the GitHub or Linear channel. - **GitHub channel** (`npm`, `pypi`, `cli`). `channels.github()` starts runs from GitHub App events, one thread per issue or pull request, as the sender. Handlers get the conversation's thread and the handles an event mentions, and `onReceived` hears each delivery first. `mda deploy` creates the App or uses one you bring. (#839) (#861) (#892) (#895) (#931) (#934) (#938) - **Linear channel** (`npm`, `pypi`, `cli`). `channels.linear()` answers Linear agent sessions (mentions and delegated issues) with session activities, cancels the run when someone presses Stop, points a run that stops on interrupts to LangSmith Studio, and starts runs from `Issue` and `Comment` webhooks. `mda deploy` sets up a new Linear app or uses one the workspace has, and keeps it as an Agent Auth app the deployment owns. See [`examples/linear`](examples/linear). (#828) (#830) (#831) (#832) (#834) (#995) (#1023) - **npm, pypi:** Slack channel handlers. `onEvent` decides what each `mention`, `directMessage`, `message` or button `interaction` starts, with the verified sender as `event.actor`. `multitaskStrategy` sets what a message does while its thread runs, `onFinish` hears once each message is done, `status` sets what Slack shows while a run works, and `reply: "none"` leaves replying to the agent. (#947) (#949) (#950) (#988) (#989) (#990) (#997) - **npm, pypi:** HTTP channel hooks. `input` decides what a message starts, `onReceived` and `onFinish` hear each delivery, `multitaskStrategy` and `reply` work as for Slack, and `parse` may return `cancel` to stop a thread's runs. A message may carry a `deliveryKey` and start in the `background`. (#826) (#913) - **npm, pypi:** Handler run options and endpoints. A channel handler may return run `metadata` and a completion `webhook` with its input, or be an endpoint of your service (`EventEndpoint`) that answers what to start. (#899) (#922) (#911) (#929) (#948) - **npm, pypi:** `runtime.channel.post(message, { newThread: true, run })` posts at top level, which starts a new provider thread, and can start a run in it. (#1024) - **npm, pypi:** `runtime.schedules.create(..., run_as=participant)` (`runAs` in TypeScript) runs a schedule as another participant of the thread. - **npm, pypi:** `resolveUser` / `resolve_user` maps a channel's verified sender to one of your users, with a function or an endpoint MDA asks. Set it on the identity declaration for every channel, or on one channel. Without a mapping, a GitHub user runs as their Agent Auth principal for `(github, github.com, )` instead of `github:`, and a Linear user as theirs for `(linear, linear.app, )` instead of the shared `mda:channel`, as a Slack user already did. Bots and apps still run as the channel. User memory a GitHub sender kept under `github:` is not carried over. Unmapped GitHub and Linear senders still get no user connections. (#1061) (#1080) (#1081) ### Changed - **npm, pypi:** `runtime.channel` is deprecated in favor of `runtime.thread.origin.channel` and will be removed in 0.11. It now has the thread's origin in every run on a channel thread, not only in the run a message started. `event` and `raw_event` (`rawEvent`) are set only in the run the message started, so they can be `None` / `null`. - **npm, pypi:** Trigger (Slack) channels post the reply from the run, not from the ingress process. Same-thread runtime schedules report `channel` and `target` as `None` / `null`, because the thread decides where they post. - **npm, pypi:** An HTTP channel message to a thread that waits on an interrupt starts nothing. It used to start a new run. - **cli:** `mda deploy` sets up GitHub, Linear and Slack channels the same way. Each channel's CLI-only env vars skip its questions with or without a terminal, are checked before anything is uploaded, are never uploaded, and can be removed once deploy says the channel is ready. A failed setup step offers to try again or continue without the channel for Slack too, and one channel's failure no longer skips the others. A brought GitHub App's webhook secret now comes from `MDA_GITHUB_WEBHOOK_SECRET` (`GH_WEBHOOK_SECRET` stays the deployment's own override), and `MDA_GITHUB_APP_ID`, `MDA_GITHUB_APP_PRIVATE_KEY_FILE` and `MDA_GITHUB_APP_INSTALLATION_ID` bring an existing App without a terminal. Without a terminal, a new Slack channel stops before uploading unless LangSmith can create its app in a workspace you already authorized. (#998) - **cli:** When Slack still needs you after `mda deploy` has the deployment live (authorizing LangSmith, or installing or approving its new Slack app), deploy prints the link and waits until Slack reports it done, instead of failing without a terminal or asking you to press Enter in one. In a terminal the link opens in your browser and Ctrl-C skips the channel; without one, deploy waits up to an hour. (#999) - **npm, pypi:** deepagents `^1.16.0` (npm) and `>=0.7.30` (PyPI) are the lowest supported versions. - **cli, breaking:** Shell env wins over project `.env`. `LANGSMITH_API_KEY`, `LANGSMITH_ENDPOINT`, `LANGSMITH_WORKSPACE_ID`, `LANGSMITH_PROJECT` (in `mda dev`) and `--secret-from-env` values now come from the shell first and the project `.env` second, so `LANGSMITH_ENDPOINT=… mda deploy` and CI-injected variables are no longer silently shadowed by the file. If you export one of these globally (e.g. for tracing) and keep a different value in a project `.env`, `mda` now uses the shell value: unset it in the shell or align the two. When both set a key to different values, `mda` warns once and names the key, never the value. Values deployed as secrets still come from `.env` only, so shell variables never leak into a deployment. (#769) - **npm, pypi, breaking:** A factory that passes `sandbox` per run keeps the thread's sandbox when the selection's options change. Snapshot, image and sizing apply only when MDA creates the thread's sandbox, and MDA logs a warning when a run asks for other options. Only different `mountConfig` / `mount_config` mounts give a thread a separate sandbox. A thread with a live sandbox from 0.9 reconnects to it. (#1115) - **cli:** A self-hosted deployment's Slack app is set up the same way as any other: `mda deploy` asks for it or reads `MDA_SLACK_BOT_TOKEN` and `MDA_SLACK_SIGNING_SECRET`. 0.9.5 read `SLACK_BOT_TOKEN` and `SLACK_SIGNING_SECRET` from `.env` for this; rename them. - **cli:** `mda deploy` creates a `serverless` deployment by default, which can scale to zero. Pass `--deployment-type dedicated` for the old behavior. `dev` and `prod` still work. Serverless needs a supported GCP SaaS environment, and other environments reject it instead of falling back to dedicated. Existing deployments do not change. (#1099) - **pypi:** The Python package loads its exports on first use, so importing it is faster. (#1039) - Bump the major group with 2 updates (`deps`) (#1035) - Bump the minor-and-patch group (`deps`) (#1037) - Bump crc32fast (`deps`) (#1034) - Bump the minor-and-patch group across 1 directory with 3 updates (`deps`) (#1036) - Bump the npm_and_yarn group across 4 directories with 4 updates (`deps`) (#1031) - Bump fastmcp from 4.1.0 to 4.1.2 in /packages/pypi (`deps-dev`) (#1030) - Bump opentelemetry-api in /packages/pypi (`deps`) (#1029) ### Removed - **Breaking (`cli`, `npm`, `pypi`):** `connectors.mcp(...)` and the `connectors/` MCP file layout, deprecated since 0.7 and announced for removal in 0.8.0. Declare MCP servers in `tools/mcp.ts` / `tools/mcp.py` with `export const mcp = defineMcp(...)` / `mcp = define_mcp(...)`. The `connectors` export is gone from both runtimes, and `mda` fails the build when a module under `connectors/` still exports `connector` or has a default export. (#924) - **Breaking (`cli`, `npm`, `pypi`):** the sandbox `scope` option, announced for removal in 0.7.0. Sandbox reuse is always one box per thread. `mda` fails the build when the sandbox declaration sets `scope`, and `defineSandbox` / `define_sandbox` throw when it is passed at runtime. Remove it from the sandbox declaration. - **`cli`:** the generated Python MCP connection emitter no longer reads the legacy `mcp_servers` config key, which the runtime already rejects. - **Breaking (`cli`):** the ignored `mda evals init [TASK_NAME]` positional, which is now rejected. `mda evals init` also fails on Harbor job configs written by mda 0.4.0 to 0.5.3, and says how to regenerate them. (#926) ### Fixed - Verify the Trigger `identity_context` off the event loop (`pypi`) (#1004) (#1042) - Clear a stale Stagehand claim by reloading the extension, not reinstalling it (`browser`) (#1085) (#1121) - Give mda dev boxes stable names (`sandbox`) (#1110) - Keep LangSmith agent address vars out of deployment secrets (`deploy`) (#1112) - Drop the runtime's uv source when bundling it (`cli`) (#1109) - Recognize managed-tools MCP URLs with query strings (#1092) - Leave reasoning blocks out of channel replies (`channels`) (#1094) - Report a rejected LangSmith API key as an authentication failure (`cli`) (#978) - Check and read a channel icon through one open file (`runtime`) - Verify the Slack actor against the signed principal (`channels`) (#1059) - Report scheduled run errors to the channel (`channels`) (#1089) - Log why a channel run input is refused (`channels`) (#1070) - Continue a scheduled run's thread from its new Slack thread (`mda`) (#1019) - Fall back to LANGCHAIN_API_KEY for the deployment key (`runtime`) (#1057) - Name local agent connection variables so shells keep them (`connections`) (#824) - Read sync subagent spreads in `subagents=[...]` (`mda`) (#941) ## 0.9.5 - 2026-10-24 ### Added - Set up self-hosted Slack channels with customer apps (`mda`) (#937) ### Changed - Decouple explicit connection targets from local agent builds (#683). Store API requests from ordinary users with an empty or wildcard namespace now return 403. - Send runtime logs to Datadog over OTLP (`runtime`) (#907) - Trace `deploy` and `dev` commands by stage (`cli`) (#946) - Bump the uv group across 2 directories with 1 update (`deps`) (#957) - Bump certifi from 2026.7.22 to 2026.10.19 in /packages/pypi (`deps`) (#956) ### Fixed - Fix connections auth on BYOC deployments (#984) - Return tool errors for writes to read-only Context Hub mounts (`runtime`) (#876) - Treat the instructions.md mount as a file in agent file tools (`runtime`) (#875) - Use authoritative LangSmith dashboard URL (`cli`) (#993) - Honor fs_capacity_bytes when baking a docker_image sandbox (`cli`) (#979) - Guide first runs past a missing model key and a wrong directory (`cli`) (#982) - Hash long deployment names in bake builder sandbox names (`mda`) (#983) - Trace mda dev runs to the agent's local environment (`mda`) (#952) ## 0.9.4 - 2026-10-20 ### Added - Report failed child stderr as the error stacktrace (`mda`) (#902) - Send CLI logs to Datadog over OTLP (`mda`) (#886) ### Changed - Bump @modelcontextprotocol/client in /packages/npm (`deps`) (#954) - Bump the npm_and_yarn group across 4 directories with 1 update (`deps-dev`) (#953) - Bump the minor-and-patch group in /packages/pypi with 4 updates (`deps`) (#918) - Bump the minor-and-patch group (`deps`) (#917) - Bump actions/create-github-app-token (`deps`) (#916) - Bump the major group with 2 updates (`deps`) (#915) - Bump the minor-and-patch group with 4 updates (`deps`) (#914) - Bump the minor-and-patch group across 1 directory with 9 updates (`deps`) (#798) - Bump the minor-and-patch group across 1 directory with 11 updates (`deps`) (#799) ### Fixed - Suppress MCP import beta warning (`python`) (#940) - Delete each integration job's own deployments on cancel (`ci`) (#866) - Drop other package managers' lockfiles from the build (`mda`) (#919) - Fail writes to unmounted memory layers instead of dropping them (`memory`) (#745) ## 0.9.3 - 2026-10-17 ### Added - **cli, npm, pypi:** Per-call human-in-the-loop approvals (`HumanInTheLoopMiddleware(interrupt_mode="per_call")`) work end to end. `mda chat` reviews one tool call at a time, offering the decisions in its `response_schema`, and can edit the arguments or answer for the tool. Trigger channel `run.interrupted` events carry each interrupt's `response_schema`. `mda dev` requires `langgraph-api` 0.16 or later. (#710) - Provision `schedules/` on the local server in `mda dev`, with `--no-schedules` to opt out (`cli`, `npm`, `pypi`) (#825) ### Fixed - Read schedule limits from env vars (`schedules`) (#862) - Send the tracing project on runtime events (`channels`) (#748) - Cap runtime schedules per owner (`schedules`) (#840) - Show each schedule's cron expression in the `mda dev` banner instead of `(?)` (`cli`) (#825) - Connections in a deployed agent no longer fail with `LANGSMITH_WORKSPACE_ID is not set` when the deploy never named a workspace: the runtime falls back to the `LANGSMITH_TENANT_ID` LangSmith sets on every hosted revision. `LANGSMITH_WORKSPACE_ID` still wins when set (`npm`, `pypi`) (#905) - Handle self-hosted runtime mounts and credentials (`channels`) (#901) - Interrupt for credentials a partial tool kit lists in _meta (`connectors`) (#889) - Preserve self-hosted Trigger API base paths (`channels`) (#887) - Require langgraph-cli 0.4.36 (`mda`) (#882) - Run Python agents on the checkout's langgraph-api (`local-stack`) (#883) - Bump @langchain/openai to 1.6.2 (`examples`) (#872) - Bump langchainplus pin and regenerate patches (`local-stack`) (#871) - Give mda dev its own build directory (`mda`) (#850) - Bump langchain in TS fixtures so gpt-6 uses Responses (`test`) (#867) - Keep the Python emit site across dev rebuilds (`mda`) (#849) - Route self-hosted channel setup through API prefix (`mda`) (#845) - Cache, retry, and keep 503 for LangSmith key verification (`auth`) (#853) - Use the Responses API for deployed TypeScript fixtures (`test`) (#857) - Sweep leaked dockerbase/sourcebase bake snapshots (`mda`) (#829) - Show the Trigger Server's reason when channel setup fails (`mda`) (#841) - Route self-hosted Agent Auth through API prefix (`mda`) (#842) - Pass Slack message attachments and blocks to the agent (`channels`) (#836) - Route self-hosted sandbox runtime calls (`mda`) (#844) - Recognize LangSmith tool kit URLs (`mcp`) (#847) - Retry deployment status polling errors (`mda`) (#843) - Let the local stack deploy without --push-to (`mda`) (#838) ## 0.9.2 - 2026-10-14 ### Added - **`setup.sh` for dynamic agents** (`cli`, `npm`, `pypi`). A dynamic agent's `sandbox/setup.sh` is now baked at deploy even without a sandbox declaration, on the default image. A sandbox the factory selects clones that recipe when it sets no base or the declared base. A run that selects a different base fails with an error, since `setup.sh` was never baked on it. - Add support for Agent Auth user-scoped secrets (`connectors`) - Build and push images for self-hosted deploys (`mda`) (#790) - Support setup.sh for dynamic agent factories (`sandbox`) (#755) - Include local packages from outside the project (`mda`) (#802) - Surface local stack startup errors in deploy status (`tools`) (#821) - Interrupt for API-key credentials from LangSmith Managed Tools (`connectors`) (#779) ### Changed - Report local stack preparation failures without waiting for builds (#822) - Resolve toolchain programs with one candidate helper (`mda`) (#812) - Reuse copy_dir_recursive for the vendored TS runtime (`mda`) (#813) - Drop unreachable schedule input branch and cover schedule validation (`mda`) (#815) - Share plugin-module discovery between connectors and channels (`mda`) (#808) - Use shared channel types in delivery and Slack transport (`pypi`) (#807) - Deserialize Python agent name kind into an enum (`mda`) (#811) - Reuse LangSmith client helpers (`mda`) (#814) - Use oxc helpers for import attribute and export names (`mda`) (#816) - Reuse json_expression for static array elements (`mda`) (#817) - Use vitest allTests and drop dead helpers (`integration-tests`) (#805) - Use mda_channel for ingress context (`mda`) (#754) ### Fixed - Route mda's /api-host deployment calls to the mock host (`local-stack`) (#823) - Link Context Hub repos to their agents (`mda`) (#819) - Treat failed git status as dirty in eval provenance (`mda`) (#809) - Reject malformed schedule deliverTo destinations (`mda`) (#804) - Report timeouts in logged run() errors (`local-stack`) (#806) - Stop `mda logs -f` reprinting lines that share a timestamp (`mda`) (#810) - Report an invalid schedule timezone the same way as Python (`mda`) (#803) - Route self-hosted APIs independently (`mda`) (#781) - Install vendored Python runtime separately (`mda`) (#795) - Preserve self-hosted auth API prefix (`mda`) (#797) - Link self-hosted deployment dashboards (`mda`) (#793) - Honor N_JOBS_PER_WORKER locally (`dev`) (#796) ### Fixed - Sweep leftover `--dockerbase-` / `--sourcebase-` bake snapshots on deploy, dev and delete (`cli`) ## 0.9.1 - 2026-10-13 ### Added - Enrich CLI and runtime attribution metadata (`telemetry`) (#743) - Accept any JOSE asymmetric algorithm chosen by the JWKS key (`channels`) (#772) - Let local Studio answer a channel's interrupts (`tools`) (#770) - Local stack server provenance, ref fetching and doctor warnings (`tools`) (#760) - Add a local LangSmith stack for fast mda deploy loops (`tools`) (#742) - Expose schedules on the managed runtime (`schedules`) (#719) - Support a user context on crons (`schedules`) (#740) ### Changed - Faster requirement emit — install once, keep across builds, local tsx (`deploy`) (#721) - Bump the minor-and-patch group with 3 updates (`deps`) (#763) - Bump uuid (`deps`) (#764) - Bump the major group with 2 updates (`deps`) (#765) - Use LangChain-ecosystem types instead of local copies (`runtime`) (#762) - Bump sharp (`deps`) (#758) - Bump rustls from 0.23.44 to 0.23.45 (`deps`) (#717) - Bump the npm_and_yarn group across 4 directories with 1 update (`deps`) (#716) - Reuse LangChain message content types (`channels`) (#734) ### Fixed - **Dynamic sandbox identity** (`npm`, `pypi`). Dynamic sandbox selection now derives sandbox identity from image and box-shape options only (`snapshotId`, `snapshotName`, `dockerImage`, `registry.url`, `vCpus`, `memBytes`, `fsCapacityBytes`, `mountConfig.mounts`). Rotating a proxy secret or changing `defaultTimeout` / `idleTtlSeconds` reuses the thread's existing sandbox; lifecycle options such as `idleTtlSeconds` and `defaultTimeout` apply to new sandboxes only. - **0.9.0 sandbox adoption** (`npm`, `pypi`). Live sandboxes created by 0.9.0 under the previous identity formula are adopted on upgrade. Threads that started on 0.9.0 keep the 0.9.0 identity. Changing any sandbox option for such a thread, even after the upgrade, starts a new sandbox. Wait until those sandboxes idle out before changing options. This compatibility lookup is scheduled for removal in 0.10. - **Stale proxy rules** (`npm`, `pypi`). Each run sets its desired proxy configuration before its first sandbox command, including empty rules when `proxyConfig` is absent. A failed update blocks the command; later attempts retry. This requires Host support for `PATCH /v2/sandboxes/boxes/{name}`, including for sandboxes without proxy rules. - Boot docker-base bake builder by snapshot id only (`cli`) (#759) - Omit TypeScript emit node_modules from deploy archive (`cli`) (#768) - Reuse Agent Auth HTTP connections (`mda-py`) (#746) - Keep sandbox identity stable across secret and lifecycle changes (`sandbox`) (#741) - Use LangGraph API 0.16.0rc5 (`cli`) (#738) - Recover deleted sandboxes without replaying operations (`mda`) (#731) - Honor dynamic snapshot selections (`sandbox`) (#724) ## 0.9.0 - 2026-10-07 ### Added - **Runtime schedules SDK** (`npm`, `pypi`). Use `schedules` from tools or middleware to create, list, get, update, and delete user-owned or agent-owned schedules. Use `cron` for recurring runs or `at` for a single run. Choose a new or existing thread; schedules created from a channel send replies there and survive redeploys. Scheduled runs cannot create more schedules. - **Per-run graph factory** (`npm`, `pypi`, `cli`). Export a sync or async `agent(runtime)` factory that returns `defineDeepAgent(...)` / `define_deep_agent(...)`. Select models, tools, middleware, instructions, skills, MCP servers, sandbox, and memory for each run. The factory receives `ManagedServerRuntime` and also supports graph inspection without an active run. Static agent definitions remain supported. - **Slack batch file uploads** (`npm`, `pypi`). Send multiple workspace files as one attachment batch with `upload_files(paths)`, in the supplied order. This requires a managed sandbox. - React to the inbound Slack message with an emoji when a channel run picks it up, so the sender knows it was seen before the agent has anything to say. Defaults to a fixed `eyes`; pass `reactions: false` to turn it off, another short name to use that emoji, or a function to choose one per message (`npm`, `pypi`). Return `null` (TypeScript) or `None` (Python) from the callback to skip a reaction for one message. ### Changed - **npm, pypi:** Memory `allow` callbacks are deprecated. Select memory in the agent factory with `defineMemory(...)` / `define_memory(...)` instead. Root memory declarations remain the fallback; factory-selected memory cannot include `allow` callbacks. - **npm, pypi, breaking:** Channel tools, middleware, and factories read delivery from `runtime.channel`. Replace `runtime.context.channel["event"]` with `runtime.channel.event` in Python and `runtime.context.channel.event` with `runtime.channel.event` in TypeScript. Read provider data from `runtime.channel.raw_event` (Python) or `runtime.channel.rawEvent` (TypeScript). Python channel context is `None`; TypeScript tool and factory context is `undefined`, and native middleware context is empty. The reply target is private; use `runtime.channel.post` to send a reply. Direct API context and validation stay unchanged. See the [Python](packages/pypi/README.md#run-context-and-channels) and [TypeScript](packages/npm/README.md#run-context-and-channels) channel guides. - **npm, breaking:** TypeScript MCP connectors run on `@langchain/mcp-adapters` 2.x: `defineMcp` takes `servers` and `transport` (not `mcpServers` and `type`), tool results use LangChain content blocks, and modern servers can pause a run to ask the user for input. See the [migration notes](packages/npm/README.md#mcp-servers) and the adapter's [2.0.0 changelog](https://github.com/langchain-ai/langchainjs/blob/main/libs/langchain-mcp-adapters/CHANGELOG.md#200) (#645). ## 0.8.6 - 2026-10-06 ### Added - Pass --agent-environment flag to deploy and delete CLI (`CLI`) (#686) - Add deploy --yes to confirm first-deploy overwrites unattended (`mda`) (#688) ### Changed - `mda chat` no longer needs a dev server to be running already. It reuses one that is, and otherwise starts its own for the session, logging to `.mda/dev-server.log`. Run `mda dev` yourself to keep its output in view — chat attaches to it [MDAC-79] (`cli`) - Bump urllib3 from 2.7.0 to 2.8.0 in /packages/pypi (`deps`) (#685) - Bump the minor-and-patch group (`deps`) (#682) - Bump the major group with 2 updates (`deps`) (#680) - Start a dev server when mda chat finds none (#654) ### Fixed - Warn when a watch rebuild needs a LangSmith key that is missing (`dev`) (#695) - Stop owned child processes on SIGTERM and SIGHUP (`mda`) (#689) ## 0.8.5 - 2026-10-01 ### Changed - Bump fast-uri from 3.1.7 to 3.1.8 in /packages/npm (`deps`) (#667) - Bump the npm_and_yarn group across 3 directories with 2 updates (`deps`) (#669) - Bump ip-address from 10.7.0 to 10.7.2 in /packages/npm (`deps`) (#668) - Bump the uv group across 2 directories with 1 update (`deps`) (#659) - Bump pyjwt from 2.13.0 to 2.15.0 in /packages/pypi (`deps`) (#658) ## 0.8.4 - 2026-10-01 ### Changed - Bump the minor-and-patch group in /packages/npm with 4 updates (`deps`) (#641) - Bump the minor-and-patch group in /packages/pypi with 5 updates (`deps`) (#642) - Bump the major group with 2 updates (`deps`) (#640) ### Fixed - Lower pydantic floor so Fireworks deploys resolve (`pypi`) (#655) ## 0.8.3 - 2026-09-26 ### Added - Add mda chat terminal client (`cli`) (#601) ### Changed - Bump LangGraph.js release versions (`deps`) (#635) ### Fixed - Finalize from packed npm metadata (`release`) (#636) - Catch unsupported agent settings during build (`mda`) (#628) ## 0.8.2 - 2026-09-25 ### Fixed - Support older Linux systems with static musl binaries (`cli`) (#627) - Clarify PAT requirement for user memory [MDAC-72] (`dev`) (#629) ## 0.8.1 - 2026-09-24 ### Fixed - Allow requested client credentials scopes (`auth`) (#619) - Stop isolated dev server processes (`tests`) (#621) ## 0.8.0 - 2026-09-24 ### Added - **Agent and user memory** (`npm`, `pypi`, `cli`). Declare shared agent memory and private user memory as separate Context Hub layers. Each layer supports a sync or async `allow(context)` policy. See the migration notes below. - **Sandbox file API** (`npm`, `pypi`). Authored tools and middleware can use `runtime.backend` to list, read, write, edit, search, upload, and download files in the managed sandbox. Binary transfers use `uploadFiles` / `downloadFiles` in TypeScript and `upload_files` / `download_files` in Python. The backend is absent when no sandbox is declared; it does not expose Context Hub memory or skills. - **Python deployment version** (`cli`, `pypi`). Set `[tool.mda].python-version` in `pyproject.toml`, or override it with `MDA_PYTHON_VERSION`, to select Python 3.11–3.14 for the deployment image. Without an explicit setting, MDA selects the newest supported version allowed by `requires-python`. This setting does not change the local or sandbox interpreter. - **HTTP channels** (`npm`, `pypi`, `cli`). Define a webhook adapter with `channels.http`. Verify requests, decode JSON, form data, text, or bytes, and optionally send replies through the adapter. Verification and parsing each receive a separately readable request body. Deploy output lists the webhook URLs. Failed runs can send a default error reply or use an optional error callback. - **Agent-owned prompt responses** (`npm`, `pypi`). Accept `user_prompt_response` channel events for prompts posted by the application. Match a response to an interrupt with `INTERRUPT_CORRELATION_KEY` and `correlation_id`. Without a correlation ID, a response resumes the only pending interrupt or starts a new turn if none is pending. Multiple pending interrupts require an exact match. - **Slack file transfers** (`npm`, `pypi`). Save incoming files under `/workspace/attachments/` and send completed workspace files with `attach_file(path)`. This requires a managed sandbox. The runtime also checks a limited number of thread-history pages for attachments and reuses files already saved. History access requires the bot's history scope for the conversation. Files are saved before an OAuth interrupt so they remain available when the run resumes. - **OAuth client credentials** (`cli`). Create agent-owned connections with `mda connections create --grant-type client_credentials`. MDA acquires the token for the managed deployment without a browser authorization flow. - **Connections in sandbox proxy headers** (`npm`, `pypi`). Use Agent Auth connection references directly, or wrap them with `bearer` or `basic`, in proxy header configuration. MDA resolves credentials before each command and updates changed values. Missing user grants use the existing authorization interrupt flow. Runs in one thread share the proxy configuration; long-running commands do not receive token refreshes during execution. - **Custom OAuth token headers and Stripe Link** (`cli`). Pass `--token-request-header KEY=VALUE` to `mda connections create` for providers that require headers on token and revocation requests. Add the Stripe Link OAuth service to the connection catalog. - **LangSmith Managed Tools** (`npm`, `pypi`). Connect to managed MCP servers, including Parallel Search, through `defineMcp` / `define_mcp`. MDA supplies the trusted caller identity and LangSmith authentication. Managed OAuth tools use the existing authorization interrupt flow. API-key tools require the caller to connect the key in LangSmith Tools. - **Run completion telemetry** (`npm`, `pypi`). Report `run_completed` events from both runtimes through the shared telemetry client and PostHog transport. Set `MDA_TELEMETRY_DISABLED=1` or `DO_NOT_TRACK=1` to disable telemetry. ### Changed **Channel API migration** (`npm`, `pypi`) - Direct API runs use the authored context schema. Managed channel runs use MDA's channel schema, with delivery data and the verified reply target in `runtime.context.channel`. Tools and memory policies that handle both sources must support both context shapes. Supply run context through Agent Server's `context` input. - Use `runtime.channel` for the channel name, provider, normalized event, optional raw provider event, and optional `post` function. Trusted caller information remains in `runtime.serverInfo` (`server_info` in Python). - Give each outgoing message an explicit `type`: use `post({ type: "content", content: "..." })` or `post({ type: "native", native: { ... } })` in TypeScript. Use the same message fields in a Python dictionary. Native JSON requires an adapter that supports it. Built-in Slack `post` sends text through Trigger; use `attach_file` for file delivery. - Channel posts use the verified reply target. Remove uses of channel `update`, capabilities, destination overrides, and final-post options. A post sends an additional message; the final agent reply remains automatic. **Memory migration** (`npm`, `pypi`, `cli`) Declare the enabled layers in the root `memory.ts` or `memory.py` module: ```ts import { defineMemory, memoryLayer } from "managed-deepagents"; export const memory = defineMemory({ agent: memoryLayer(), user: memoryLayer(), }); ``` Python uses `define_memory(agent=MemoryLayer(), user=MemoryLayer())`. - Omitted layers are disabled. Agent memory is shared across the deployment at `/memories/agent/`. User memory requires an identity declaration and a trusted person, and mounts that person's Context Hub repo at `/memories/user/`. - The default user policy permits managed Slack one-to-one DMs. Other sources, including direct API runs, require an explicit `allow(context)` policy. A policy cannot grant user memory to a service principal or select another person's memory. Studio users can inspect memory regardless of run policy. - Policies run once per run, including resumes, before memory mounts. A denied layer contributes neither its mount nor its hot memory. Inspection does not call the policy. - Hot memory loads once per run and is injected into each model call. It is not saved in thread state. Deploy preserves existing memory, and a missing hot file stays empty until the first write. - Legacy object-based `scope` declarations remain supported, including a user `id` assertion. Do not combine them with named layers. String scopes are rejected. The generated `MemoryConfig` uses version `2` and `scope` instead of `slices`. - The graph factory selects context and memory access for each run. Native Node Agent Server must pass the run's context to the factory. Direct users of the managed compiler must build a fresh graph for each execution. - Release 0.8 (#623) - Add OAuth client credentials connections (#598) ### Fixed - Resolve the local caller identity for LangSmith Managed Tools during `mda dev` (`npm`, `pypi`). - Accept slashes in connection names (`npm`, `pypi`). - Keep OAuth token header values out of CLI error output (`cli`). ## 0.7.4 - 2026-09-22 ### Changed - Bump the minor-and-patch group (`deps-dev`) (#588) - Bump the minor-and-patch group (`deps`) (#587) - Bump oxc from 0.149.0 to 0.150.0 in the major group (`deps`) (#586) - Bump the minor-and-patch group with 2 updates (`deps`) (#585) ### Fixed - Return sign-in guidance for missing credential-gate callers (`mda`) (#565) ## 0.7.3 - 2026-09-16 ### Added - Add lmt for mda python + ts [closes TI-48] (`mda`) (#551) - Name deploy target and link it earlier (`cli`) (#544) - Open the prefilled activation page (`cli`) (#463) ### Changed - Surface deployment metadata earlier (#544) - Revert "feat(mda): add lmt for mda python + ts [closes TI-48] (#551)" (#555) - Modernize TypeScript dependencies and fix sharp (`deps`) (#541) ### Fixed - Remove MDA's default recursion limit (`runtime`) (#564) - Explain init language selection [closes MDAC-47] (`cli`) (#554) - Fix GitHub preparation auth and Context Hub routing (`mda`) (#523) - Limit OpenWiki merge checks to wiki updates (#532) ## 0.7.2 - 2026-09-10 ### Changed - Mount the project's skills in eval trials (#527) - Bump the minor-and-patch group across 1 directory with 9 updates (`deps`) (#513) - Bump the minor-and-patch group across 1 directory with 5 updates (`deps`) (#506) - Bump oxc from 0.147.0 to 0.148.0 in the major group (`deps`) (#504) - Bump the minor-and-patch group with 2 updates (`deps`) (#503) ### Fixed **Eval trials mount the project's skills** (`cli`, `npm`, `pypi`) Harbor eval trials ran with no skills library, so an eval whose expected behaviour is defined by a skill failed on the harness rather than on agent quality. Skills were already packaged into the eval artifact, but the generated eval entry never told the runtime to mount them. The eval entry now passes `has_skills` the way the dev and deploy entries do. A trial has no Context Hub: the Harbor adapter copies the compiled project into the trial workspace, so `/skills/` resolves to the project's own `skills/` tree. Memory stays off and the authored sandbox is still not instantiated in eval mode. - Require local provider keys for direct models [closes MDAC-29] (`mda`) (#529) - Use separate tracing projects for local runs (#530) - Merge OpenWiki updates after CI (#524) ## 0.7.1 - 2026-09-10 ### Changed - Bump @vitest/mocker (`deps`) (#519) ### Fixed **Studio callers resolve to an authenticated principal** (`npm`, `pypi`) Every Studio request resolved to no actor at all. The Agent Server injects `langgraph_api.auth.studio_user.StudioUser`, which reads like a mapping but is not a registered `collections.abc.Mapping`, so an `isinstance` gate in the runtime's Studio detection dropped the principal. `build_runtime_identity` returned `None`, and everything downstream — `server_info.principal`, `server_info.user`, `source.provider: "studio"`, actor-scoped memory and store namespacing — behaved as though the run were anonymous. Detection now prefers `langgraph_sdk.auth.is_studio_user` and otherwise reads the signal structurally, matching what the auth handler already did. Both runtimes use the verified LangSmith `ls_user_id` for Studio user identity, thread ownership, and user-owned connections. Deployed Studio uses the `langsmith` identity provider; `mda dev` uses `langsmith-dev` so development credentials stay separate. - Guide Markdown link formatting for Slack replies (`channels`) (#518) - Detect the Studio principal the server injects (`runtime`) (#510) - Wait for npm registry propagation (`release`) (#511) ## 0.7.0 - 2026-09-08 ### Added **User-owned connections in `mda dev`** (`npm`, `pypi`) User-owned connections (`connections.get(slug, { type: "user" })`) resolve through Agent Auth in every environment, including `mda dev`. The runtime maps the verified ingress identity to an Agent Auth principal (`POST /v1/agent-auth/principals`) before grant lookup or OAuth session create. Channel `identity_context` principals skip that resolve. OAuth sessions use canonical `owner_type` / `owner_id`. `LANGSMITH_HOST_PROJECT_ID` is optional for user owners; agent-owned secrets still use `MDA_DEV_` under local development. For Studio, `mda dev` requires a personal LangSmith API key (or prompts for browser sign-in), resolves its verified LangSmith user through `POST /v1/agent-auth/principals`, and stages that principal for the local run. Other service principals cannot own user OAuth grants. **Agent-owned OAuth authorization** (`cli`) `mda connections create --authorize` now creates the workspace connection, starts an OAuth authorization session for the deployed agent, and prints the provider authorization URL. It supports catalog or manual OAuth and explicit or project-inferred MCP OAuth discovery. OAuth connections without `--authorize` keep the per-user runtime authorization flow. **OAuth service catalog** (`cli`) `mda connections create --oauth ` now resolves the service's authorization endpoint, token endpoint, token endpoint auth method, default scopes, and authorization params from a table compiled into the CLI, so an author supplies only their own client credentials. Covers `atlassian`, `bitbucket`, `box`, `click-up`, `discord`, `dropbox`, `facebook`, `figma`, `github`, `gitlab`, `google`, `hubspot`, `huggingface`, `linear`, `linkedin`, `notion-api`, `patreon`, `reddit`, `salesforce`, `slack`, `spotify`, `twitch`, and `x`. An unrecognized service is refused and names the ones that are known. The Notion catalog name includes `-api` to distinguish its BYOT API integration from an MCP server discovered through DCR. Tenant-specific issuers (Auth0, Okta, Entra tenants, GitLab self-managed) are omitted; authors still override any entry with `--authorize-url` / `--token-url`. `mda connections catalog [--json]` lists the table. It reads the compiled-in data, so it takes no workspace id, no API key, and makes no network call. `--auth-method`, `--allowed-scope`, and `--authorization-param` make the remaining provider registration fields author-controlled. Client credentials are named with `--secret-from-env` or `--secret-from-file`. A successful OAuth create prints the redirect URI to register with the provider before prompting for the client secret, and again after the connection is stored. The error for a missing `--client-id` names the provider's app registration page and leaves the URI in the block above. The URI follows the configured host rather than being fixed to production. A successful OAuth create also reports the scopes it registered, since those alone decide whether the provider's consent screen accepts the authorization request. A manual registration (`--authorize-url` / `--token-url`) that named no scopes is called out: it produces an authorization request with no `scope` parameter, which most providers refuse. A catalog service that defines no scopes, such as Notion, stays quiet. `mda connections create --mcp ` registers an OAuth connection via MCP discovery. A scheme-less value such as `mcp.acme.com/mcp` is stored as `https://mcp.acme.com/mcp`. If the server cannot register a client automatically (CIMD or DCR), the CLI says so and points at `--oauth` or a manual client id. Agent Auth reports a failed discovery the same way whether the host does not exist or simply registers no client, so a mistyped URL used to read as a provider that cannot register a client. The CLI now asks the server itself and distinguishes a host that never answered, a host that answered 404, and a reachable server that supports neither CIMD nor DCR. `mda connections create ` infers MCP OAuth discovery when that slug is used by exactly one user-owned remote MCP server in the project. The CLI passes the server URL through so Agent Auth can discover and register a client instead of requiring `--oauth` or a client id. Agent-owned MCP connections and ambiguous slugs are not inferred. Deploy applies the same inference to missing user-owned MCP connections. It leaves existing connections unchanged, automatically creates each unique missing MCP connection through OAuth discovery, and fails before deployment with guidance when discovery or registration is unavailable. `mda connections --help` (and the `connection` alias) lists grouped examples for catalog OAuth, extra scopes, a manual provider, MCP from a server URL, MCP from a project connector slug, opaque secrets, list, get, and delete, each labeled with a `#` comment. `mda connections create --help` presents opaque secrets, general OAuth, and MCP OAuth as separate modes, with grouped flags and examples for each. `Examples` and `Docs` use the same underlined heading style as clap's `Commands` and `Options`. - Authorize agent OAuth during connection create (`mda`) (#493) - Resolve user OAuth grants in mda dev (`connections`) (#500) - Forward --tunnel to the LangGraph dev server (`dev`) (#470) - Expand the compiled OAuth service catalog and improve `--help` docs (`connections`) (#490) - Add --mcp for explicit DCR create (`connections`) (#488) - Infer MCP DCR from user-owned connector slugs (`connections`) (#485) - Add `mda prepare` for platform-run GitHub builds (`cli`) (#489) - Reject user credentials without an OAuth provider (`deploy`) (#447) - Resolve local agent secrets from env (`credentials`) (#445) - Prompt for LangSmith login during `mda dev` (`cli`) (#397) - Show the wordmark intro in dev and delete (`cli`) (#443) - Expose caller on serverInfo.user instead of identity (`runtime`) (#400) - Add the agent-auth credential client (`cli`) (#419) - Schedules go through channel server [closes TI-20] (`mda`) (#413) - Report a bounded error category on failed commands (`cli`) (#406) ### Changed **Authoring — `connectors.mcp` → `defineMcp` / `define_mcp`** (`npm`, `pypi`) MCP is the only connector, so it is a resource factory like `defineSandbox` rather than a `connectors.*` provider. That keeps `connections.get(...)` as the sole `connect*` namespace for secrets and credentials. **Authoring — `mcpServers` / `mcp_servers` → `servers`** (`npm`, `pypi`) The MCP server map is now `servers`. `mcpServers` (TypeScript) and `mcp_servers` (Python) remain as deprecated aliases and will be removed in `0.8.0`. The stored definition always uses `servers`. The LangChain adapter constructor still receives `mcpServers`. **Channel failures name missing connections** (`npm`, `pypi`) When a channel run cannot start because a declared workspace connection is not configured, the runtime sends a safe, actionable failure message naming the connection and the `mda connections create` command. Other runtime failures keep the generic channel error rather than exposing arbitrary exception text. **Channel threads are owned by their source conversation** (`npm`, `pypi`) A managed channel thread is now owned by `mda:source-thread:{provider}:{key}` rather than by the principal that delivered its first message. A Slack thread is addressed by its source conversation and its deliveries carry whichever principal `identity_context` resolved to — none on an interrupt resume, a different one when another participant replies — so per-principal ownership denied every delivery after the first. Channel conversations that were already running do not carry over. Their threads were created with the delivering principal as owner, and nothing in a later delivery can present that principal again, so the run is denied. There is no automatic migration: the thread id is a hash of the source thread key, so existing threads cannot be mapped back to the conversations that would re-own them. Start a new conversation in the channel after deploying this version; in-flight ones stop responding. **Runtime caller — `runtime.identity` → `serverInfo`** (`npm`, `pypi`) Tools and middleware now read the caller from LangGraph's standard `serverInfo` path, reshaped to match Agent Auth so the surface does not move again when `resolve-principal` starts supplying it directly. `runtime.identity` is gone from the tool/middleware runtime; there is no compatibility shim. On PyPI the identity fields are an attribute view (`principal.id`, `subject.authority`, `link.status`) for parity with TypeScript, while `principal.claims` stays a mapping so namespaced IdP claims remain reachable as `claims["https://acme.com/tenant"]` — the two named claims carry the same normalization guarantee as TypeScript's `PrincipalClaims`, documented rather than declared because the bag is open. The envelope's `link.from` is exposed as `link.from_` because `from` is a reserved word in Python. - TypeScript: `tools/mcp.ts` exports `const mcp = defineMcp({ servers: { … } })` - Python: `tools/mcp.py` defines `mcp = define_mcp(servers={…})` - `connectors.mcp(...)` remains as a deprecated alias and will be removed in `0.8.0`. The old project `connectors/` directory and `connector` export remain available with a warning during `0.7.x`. - `runtime.identity?.user.id` becomes `runtime.serverInfo?.principal?.id` (`runtime.server_info.principal.id` on PyPI). - `identity.user.kind` becomes `principal.kind`, and gains a `"channel"` variant alongside `"person"` and `"service"`. - `identity.user.email` and `identity.groups` fold into `principal.claims` as `claims.email` and `claims.groups`. Authorization checks that read groups must move. Both are normalized whether or not the deployment mapped the claim, so `claims.groups` is always a string array even when the IdP sent a delimited string, and a value that cannot be normalized is dropped rather than passed through under a name the type has promised. - `identity.source` becomes `serverInfo.source` (`provider` plus an optional `threadId` / `thread_id`), resolved from the trusted ingress stamps. Reading `serverInfo.user.mda_source_provider` is no longer necessary. - `serverInfo.assistantId` / `graphId` and `subject.agent_id` are absent when the platform supplied no id, rather than reported as an empty string. - New exported types: `Principal`, `PrincipalClaims`, `Subject`, `Link`, `IdentitySource`, `PrincipalEnvelope`, and `ManagedServerInfo`. `principalEnvelopeFromServerInfo` reassembles the wire envelope for Agent Auth mint. - `RuntimeIdentity` is unchanged and still describes what HTTP connector route handlers receive from `resolveRequestIdentity`. - Replace connectors.mcp with defineMcp (`authoring`) (#483) - Update README with correct LangSmith Gateway link (#486) - Drop mda connections check (`connections`) (#499) - Bump qs from 6.15.3 to 6.16.0 in /packages/npm (`deps`) (#479) - Bump fast-uri from 3.1.5 to 3.1.7 in /packages/npm (`deps`) (#480) - Standardize static command output (`cli`) (#476) - Pluralize resource commands (`cli`) (#471) - Remove --client-secret-env and extract secret reading into a module (#465) - Add workspace-scoped connections runtime for Python (#454) - Add OAuth connection create, check, and deploy gates (#453) - Add workspace-scoped connections runtime (TypeScript) (#451) - Add workspace-scoped mda connection CLI (opaque) (#450) - Bump the minor-and-patch group with 2 updates (`deps`) (#437) - Bump oxc from 0.146.0 to 0.147.0 in the major group (`deps`) (#438) - Bump the minor-and-patch group (`deps`) (#439) ### Fixed - Make automatic Slack replies explicit (`channels`) (#508) - Explain missing workspace connections (`channels`) (#494) - Add x oauth catalog entry and show scopes (`connections`) (#484) - Mda connections list displays workspace scoped connections instead of deployment scoped (`connections`) (#491) - Let Studio read and search channel-owned threads (`runtime`) (#482) - Quote Agent Auth problem+json hints in store errors (`connections`) (#481) - Scope channel threads and agent secrets by conversation and deployment (`connections`) (#477) - Stamp Trigger identity_context principal on ingress runs (`channels`) (#474) - Resolve user grants via connection-for-owner (`connections`) (#472) - Drop --app, rename secret flags, print the OAuth redirect URI (`connections`) (#478) - Use LANGCHAIN_ENDPOINT for Agent Auth on cloud (`connections`) (#469) - Warn instead of failing when schedule reconcile fails (`deploy`) (#446) - Speak nested Agent Auth credential shape (`connections`) (#466) - Use event identity for actions (`channels`) (#442) - Keep deploy URLs on one line (`deploy`) (#464) - Stop unscoping Studio principals in managed auth (`runtime`) (#427) - Default-deny unhandled LangGraph auth resources (`runtime`) (#426) - Clear crons before deleting the deployment (`delete`) (#430) - Speak the project's language in CLI guidance (`cli`) (#429) - Cap connector Events bodies at 1 MiB (`runtime`) (#416) - Require matching CORS scheme for credentialed origins (`runtime`) (#415) - Delete stale crons before recreating schedules (`deploy`) (#410) ## 0.6.1 - 2026-08-26 ### Added - Report long-running command completion on Ctrl-C (`cli`) (#403) - Add PostHog telemetry and completion reporting (`cli`) (#378) - Add anonymous installation identity and analytics opt-out gate (`cli`) (#395) ### Fixed - Reconnect 0.5.x thread sandboxes after 0.6 upgrade (`sandbox`) (#404) - Use LC palette and dots for Slack icons (`cli`) (#396) - Accept legacy LangSmith keys during init (`sdk`) (#399) - Start mda deploy with the same wordmark and version as init (`cli`) (#385) - Use public channel terminology (`channels`) (#393) ## 0.6.0 - 2026-08-24 ### Added - Generate Slack icons from app names (`cli`) (#381) - Add Slack app background color [CLOSES AB-000] (`fleet`) (#379) - Log Trigger channel traffic (`channels`) (#369) - Support custom Slack channel icons (`mda`) (#372) - Scaffold Slack from mda init --channel (`cli`) (#371) - Warn when channel/ is used instead of channels/ (`cli`) (#370) - Standardize guidance output sequences (`cli`) (#373) - Guide slack responses with mrkdwn (`sdk`) (#365) - Define analytics event contract (`cli`) (#359) - Add Slack channel init command (`cli`) (#364) - Add Harbor evaluation workflow (`evals`) (#351) - Identify HTTP requests with User-Agent (`cli`) (#357) - Prototype Trigger-backed Slack channels [closes LSD-1857] (`sdk`) (#348) - Bake setup.sh into a recipe snapshot at deploy (`sandbox`) (#339) - Introduce interactive init mode (`cli`) (#349) ### Changed - Bump uuid in the minor-and-patch group (`deps`) (#386) - Bump oxc from 0.144.0 to 0.146.0 in the major group (`deps`) (#387) - Bump oxfmt (`deps-dev`) (#388) - Bump the minor-and-patch group (`deps-dev`) (#389) - Update npm dependencies (`security`) (#368) - Pin LangGraph API image to stable 0.12.6 (#350) - Remove LANGSMITH_GATEWAY_API_KEY preflight requirement (#347) ### Fixed - Fail deploy when sandboxes are disabled for the org (`cli`) (#391) - Make slack app metadata optional (`channels`) (#390) - Send interrupt envelopes on run.interrupted (`channels`) (#383) - Emit run.interrupted and accept interrupt_resume (`channels`) (#380) - Skip recipe snapshots still used by sandboxes (`cli`) (#377) - Normalize authored assistant ids (`sdk`) (#376) - Pause deploy for Slack authorization (`fleet`) (#366) - Avoid redundant Python dev reloads (`cli`) (#358) - Show install-specific docs link (`cli`) (#362) - Require Gateway provider keys (`sdk`) (#352) - Improve missing Slack channel error (`cli`) (#346) - Sync Python lockfile metadata (`release`) (#345) ## 0.5.3 - 2026-08-18 ### Added - **npm:** Ship ambient `*.md` typings via the package `"types"` entry (`types/markdown-md.d.ts`) so `import prompt from "./file.md" with { type: "text" }` type-checks without a project-local `declare module "*.md"` (matches MDA's compile-time text-module inlining for schedule prompts and similar). - Upsell a plan upgrade when the org lacks the entitlement (`deploy`) (#274) - Improve mda init onboarding experience [closes LSD-1863] (`cli`) (#338) - Support LangSmith Gateway provider (`models`) (#336) - Remove credentials authoring and runtime overlay (`identity`) (#322) - Ship ambient *.md typings for text-module imports (`npm`) (#306) - Watch project sources and reload in mda dev (`cli`) (#300) ### Changed - **cli, breaking:** `mda deploy` no longer reads deployment secrets from the process environment. A provider key, gateway key, ingress secret, or connector / channel secret is now taken only from the project `.env` or LangSmith workspace secrets. Previously a value exported in your shell was copied out and uploaded as a persisted deployment secret, so an `OPENAI_API_KEY` set in a shell profile for unrelated local work silently became long-lived credential material on the deployment. Deploys that relied on the shell now fail with guidance; move the key into `.env` or add it under Settings → Workspaces → Secrets. - Remove leftover custom-credentials authoring and runtime overlay: `defineIdentity` / `define_identity` reject `credentials`, the `credentials` namespace (`credentials.github`) is gone, and tools no longer receive `runtime.credentials`. Use env tokens (for example `GITHUB_TOKEN`) in authored tools (`npm`, `pypi`). - **cli:** Inline only `with { type: "text" }` text-module imports. The undocumented `type: "markdown"` alias is no longer accepted. - Remove Connect-with-X leftovers after that surface was deleted (`npm`, `pypi`, `cli`): - Drop `/identity/{provider}/callback` auth bypass and `/identity` CORS allowlisting (no identity HTTP routes remain). - Stop injecting or documenting `MDA_GUEST_SIGNING_KEY` (OAuth `state` signing had no remaining readers). - Keep the `mda_github_credentials` store ACL reserved (no Connect writers remain; leftover tokens must stay unreadable until scrubbed). - Remove empty `repoPrefix` mount branches that existed only for per-user Hub roots. - Stop vendoring a nonexistent `./evals` package export and stop injecting the unused `uuid` TypeScript dependency. - Delete orphaned Python extractors (`name.py`, `ingress.py`) and refresh Connect / per-user memory wording in READMEs, init scaffolds, and examples. - **integration-tests:** Restore `COVERAGE.md`, re-enable `channels-skip-non-exports.standalone.test.ts`, and delete unused `evals/scaffold/{contracts,harbor-custom,identity,overlays}` trees that no test referenced. - Guard Context Hub edits during deploy (#257) - Stop reading deployment secrets from the process environment (`cli`) (#340) - Bump cliclack (`deps`) (#329) - Bump the major group across 1 directory with 2 updates (`deps`) (#333) - Bump the minor-and-patch group across 1 directory with 6 updates (`deps`) (#334) - Bump the minor-and-patch group (`deps`) (#332) - Update nanoid to 3.3.18 (`deps`) (#308) ### Fixed - Honor `default_tool_timeout` on Python HTTP/SSE MCP connectors instead of forwarding it into the adapter session constructor and crashing `mda dev` (`pypi`). Non-positive timeouts are now rejected in both runtimes (`npm`, `pypi`). - Default dev traces to assistant project (`cli`) (#337) - Omit redundant current-directory paths (`cli`) (#328) - Require audience for JWKS tokens (`identity`) (#324) - Use the project venv for Python connector emit (`cli`) (#315) - Keep Slack manifest URL selectable (`sdk`) (#319) - Say Slack manifest template was created (`cli`) (#316) - Stop HTTP MCP from crashing on default_tool_timeout (`pypi`) (#317) - Float TypeScript LangChain dependencies (`compile`) (#271) - Use uv for Python deploy emit [closes LSD-1814] (#299) ## 0.5.2 - 2026-08-11 ### Changed - Bump the minor-and-patch group (`deps-dev`) (#297) - Bump the minor-and-patch group (`deps`) (#296) - Bump the minor-and-patch group with 2 updates (`deps`) (#295) - Bump Swatinem/rust-cache in the minor-and-patch group (`deps`) (#294) - Fix Python sandbox timeout serialization (#292) - Bump npm langsmith to ^0.8.9 (#290) - Bump oxc from 0.142.0 to 0.143.0 in the major group across 1 directory (`deps`) (#250) - Bump deepagents (`deps`) (#285) - Bump the npm_and_yarn group across 3 directories with 2 updates (`deps`) (#284) - Bump the npm_and_yarn group across 3 directories with 1 update (`deps`) (#268) ## 0.5.1 - 2026-08-08 ### Added - Revive MCP connector (#262) ### Changed - Update python deps (#283) ## 0.5.0 - 2026-08-07 ### Added - `auth.langsmithApiKey()` / `auth.langsmith_api_key()` explicitly enables `x-api-key` ingress while retaining MDA's thread and store authorization callbacks. The managed runtime verifies keys against the platform-owned LangSmith auth endpoint, enforces the deployment tenant, and maps accepted keys to scoped service identities (`npm`, `pypi`, `cli`). - Identity declarations now require an explicit `auth` selection. `mda init` scaffolds LangSmith API-key authentication by default in both languages; argument-free `defineIdentity()` / `define_identity()` declarations fail during authoring and compilation (`npm`, `pypi`, `cli`). - `mda init --gateway` scaffolds a project that runs on LangSmith Gateway (`cli`). The agent file constructs an OpenAI-compatible client against `https://gateway.smith.langchain.com/v1` on a model LangSmith hosts, billed to the workspace's Gateway Credits, and authenticates with a LangSmith API key read from `LANGSMITH_GATEWAY_API_KEY` — `LANGSMITH_API_KEY` is reserved and never reaches a deployment, so `.env` is seeded under both names. Because the model is constructed rather than named as a `provider:model` string, deploy pre-flights no provider key. Mutually exclusive with `--model`. - `mda deploy` now pre-flights `LANGSMITH_GATEWAY_API_KEY` for a project whose agent runs on LangSmith Gateway (`cli`), the same way it already pre-flights a model provider key: blocked only when the key is confirmed absent from `.env`, the process environment, and LangSmith workspace secrets. Previously a project scaffolded with `--gateway` but no key set deployed cleanly and failed on its first message with an opaque gateway 401. - `mda delete` (alias `mda destroy`) tears down a deployment's LangSmith resources: the deployment, its tracing project, its Context Hub repo including per-user memory and legacy child repos beneath it, and the managed sandboxes it created. Confirms before deleting unless `--yes` is passed (`cli`) - Managed sandboxes are now named `{deployment}--{digest}` so they can be attributed to the deployment that created them. `mda deploy` passes the deployment name to the runtime as `MDA_DEPLOYMENT_NAME` (`cli`, `npm`, `pypi`) - `mda logs` reads a deployed agent's Agent Server logs. In a terminal it streams until Ctrl-C; piped or redirected it prints the most recent lines (1000 by default, `--lines`) and exits. Also supports `--level`, `-f`/`--no-follow`, and `--tenant-id` (`cli`) - In a terminal, `mda init` prompts for a missing project name. Adding `--interactive` / `-i` also offers a coding-agent handoff after scaffolding. `mda init ` remains non-interactive for scripts and coding agents, and default completion prints language-specific quickstart and deploy steps (`cli`) - `mda init` takes the project's shape as flags, so a coding agent or a CI script can scaffold a specific project without answering prompts: `--instructions`/`--instructions-file` (`-` reads stdin) for `instructions.md`, `--scope ` for an identity declaration, `--model` for the agent's model, and `--no-sandbox`/`--no-evals` to leave those out. The generated README describes whatever was scaffolded (`cli`) - `identity.credentials` accepts a map keyed by target name, so one deployment can integrate several platforms without hand-writing the dispatch. Each entry keeps declaring its own Connect-with-X routes, which a hand-written wrapper would have hidden (`npm`, `pypi`) - Connect-with-X routes are inferred from the rest of the identity declaration: GitHub when a credential chain reads user tokens on a user-scoped deployment, Slack when a Slack channel is declared on one. `connect` remains for custom providers (`npm`, `pypi`) - Slack Events uses a **bring-your-own Slack app**: put `SLACK_SIGNING_SECRET` and `SLACK_BOT_TOKEN` in `.env` (deploy preflights them via `SLACK_CHANNEL_REQUIRED_ENV`), point Slack Event Subscriptions at `/connectors/{name}/events`, and subscribe to the bot events implied by `on`. Fleet/LangSmith no longer provisions the bot app; the optional authoring `app` branding bag and deploy-owned secret strip are gone (`cli`, `npm`, `pypi`) - Default identity to LangSmith API key auth - Add `mda init --gateway` (`cli`) (#255) - Scaffold identity on init (`cli`) - Create a default Slack manifest template - Add Slack channel manifest command - Generate Slack app manifests during deploy - Remove guest ingress entirely (`identity`) - Remove auth.guest from the public authoring surface (`identity`) - State memory instructions at runtime, not in the memory file (`memory`) - Declare durable memory in memory.{ts,py}, off by default (`memory`) - Make eval task workflow explicit (`cli`) - Personalize Slack Events via Connect-with-Slack (`channels`) - Invoke Slack Events as mda:slack agent principal (`channels`) - Put Slack Events on channels.slack (`channels`) - Drop unused threads, partition, and access axes (`identity`) (#220) - Require bring-your-own Slack app secrets (`slack`) (#218) - Require confirmation for resource collisions - Remove organization identity from MDA (#213) - Slice-based user and agent memory (`identity`) (#210) - Warn on first-deploy resource collisions - Opt into langgraph-api version via env (`cli`) (#198) - Thread LANGSMITH_ENDPOINT into device login (`cli`) - Enable login auth flow for deploy and connect (`cli`) (#191) - Add mda connect for OAuth and workspace API keys (`cli`) (#190) - Add Fleet marketplace MCP bridges and LinkedIn Connect (`connectors`) (#189) - Connect Google per actor and provision Slack in mda dev (`identity`) (#181) - Fold channels into connectors (`connectors`) (#177) - Create a channel's Slack app through LangSmith (`channels`) (#176) - Reach every LangSmith tool server integration (`connectors`) (#175) - Add onboarding TUI (`cli`) (#158) - Add mda logs and print the Agent Server URL on deploy (`cli`) (#160) - Add mda delete and clean up the sandboxes it created (`cli`) (#159) ### Changed - `mda init` now scaffolds the required `identity.ts` or `identity.py` declaration by default. The redundant `--identity` flag is removed; both it and the previously retired `--scope` flag are rejected (`cli`). - A first `mda deploy` now asks for confirmation before overwriting a same-name resource it does not own: a Context Hub repo created outside `mda deploy`, or a non-MDA LangSmith deployment. It fails closed in non-interactive terminals. Deploy-owned resources are recognized by the `managed-deepagents` repo tag and the deployment's Managed Deep Agent marker, so redeploys — and retries after a deploy that failed partway — are never prompted (`cli`) - `mda deploy` now prints the Agent Server URL next to the dashboard link on success, so there is a URL to call without opening LangSmith (`cli`) - A restarted or redeployed container now re-adopts its existing sandbox for the same scope instead of creating a replacement and stranding the old one until its idle TTL expires. The sandbox name digests how it was provisioned, so changing `setup.sh` or the snapshot still gets a fresh sandbox; only a sandbox built from the same recipe is adopted, and adopting one skips `setup.sh` (`npm`, `pypi`) - Fix required Harbor task staging files (#278) - Fix repository format and lint checks - fix Rust tests after main merge - fix Python checks after main merge - fix loopback auth and CI after main merge - add changes - Move eval scaffolds under evals workspace - Make sandbox definitions backend neutral - Expose sandbox definition factories - Clarify API key auth resolution - adding api key auth - Enable Slack app direct messages - Remove init identity flag (`cli`) - Source Slack app settings from template - Bump the minor-and-patch group (`deps`) (#251) - Bump the minor-and-patch group (`deps-dev`) (#252) - Bump clap from 4.6.4 to 4.6.5 in the minor-and-patch group (`deps`) (#249) - breaking(identity): expose only Supabase auth - fix Slack test formatting - breaking(identity): remove x-mda-groups header - breaking(identity): remove the auth.github() provider - Preserve custom Harbor tasks during eval compilation - breaking(channels): drop `on` from channels.slack (#230) - chore/remove-scopes-from-interface - Drop the migration for memory files older deploys already have (`memory`) - Remove per-caller memory, and stop the store gate depending on it (`memory`) - Drop comments the code already states (`cli`) - Update standalone eval staging assertions - Bump cryptography from 49.0.0 to 50.0.0 in /packages/pypi (`deps`) - Simplify MDA eval task scaffolding - change evals - git ignore - Bump ip-address from 10.2.0 to 10.4.0 in /packages/npm (`deps`) (#206) - Bump fast-uri from 3.1.4 to 3.1.5 in /packages/npm (`deps`) (#203) - Bump postcss from 8.5.22 to 8.5.25 in /packages/npm (`deps-dev`) (#202) - Bump @hono/node-server in /packages/npm (`deps`) (#205) - Bump hono from 4.12.32 to 4.13.0 in /packages/npm (`deps`) (#207) - Bump @hono/node-server and @modelcontextprotocol/sdk (`deps`) (#201) - Default `defineIdentity` scope to agent (`identity`) (#196) - Rename organizations to organization (`identity`) (#197) - Clean up high-priority Rust duplication and inefficiency (`cli`) (#195) - Clean up runtime duplication and async Connect-hint I/O (`pypi`) (#193) - Dedupe Connect OAuth helpers and drop marketplace wrap (`npm`) (#192) - Replace LANGCHAIN_* env vars with LANGSMITH_* (`cli`) (#194) - Bump deepagents (`deps`) (#186) - Bump pypa/gh-action-pypi-publish (`deps`) (#185) - Bump base64 from 0.22.1 to 0.23.0 in the major group (`deps`) (#187) - Reach every provider as .() (`primitives`) (#173) - Bump oxc from 0.141.0 to 0.142.0 in the major group (`deps`) (#169) - Bump ty in /packages/pypi in the minor-and-patch group (`deps-dev`) (#171) - Move compile's tests out of compile/mod.rs, colocating where a stage owns them (`mda`) (#161) - Scope workspace via --workspace-id (`deploy`) (#147) - Bump the minor-and-patch group in /packages/npm with 3 updates (`deps`) (#156) - Bump the minor-and-patch group in /packages/pypi with 3 updates (`deps-dev`) (#157) - Bump the minor-and-patch group across 1 directory with 4 updates (`deps`) (#153) - Bump actions/setup-python in the major group (`deps`) (#154) - Bump fast-uri from 3.1.3 to 3.1.4 in /packages/npm (`deps`) (#149) - Bump the minor-and-patch group with 2 updates (`deps`) (#152) - Bump oxc from 0.140.0 to 0.141.0 in the major group (`deps`) (#155) - Bump the npm_and_yarn group across 4 directories with 3 updates (`deps`) - Bump the npm_and_yarn group across 2 directories with 2 updates (`deps`) - fix hot memory integration marker - fix npm deepagents dependency - fix npm langsmith peer dependency - Bump the minor-and-patch group across 1 directory with 9 updates (`deps`) #### BREAKING (eval gateway credential) - `mda evals compile` forwards the gateway credential to Harbor trials as `LANGSMITH_GATEWAY_API_KEY`, replacing `LANGSMITH_API_KEY_GATEWAY` (`cli`). One credential now has one name across the CLI: the scaffolded agent reads it, `mda init` seeds it, `mda deploy` pre-flights it, and evals forward it. If you export the old name in a shell or CI job for `harbor run`, rename it. Keys in the project's `.env` are unaffected — those are forwarded whatever they are called, which is how a `--gateway` project's trials already worked. #### BREAKING (channels Slack primitive) - Slack Events ingress, messaging, and deploy requirements move back to a public `channels` primitive under project `channels/` (`export const channel` / module-level `channel`). Author with `channels.slack()` (`npm`, `pypi`, `cli`). - Events URL is `POST /channels/{name}/events` (no `/connectors/.../events` alias for Slack) (`npm`, `pypi`, `cli`). - `connectors.slack` is tools-only (`includeTools` / `excludeTools`). Event options (`on`, `autoReply`, …) are rejected; use `channels.slack` instead (`npm`, `pypi`). - Compiler discovers `channels/`, emits `_mda_channels`, and writes `channel-manifests.json` for deploy secret preflight alongside `connector-manifests.json` (`cli`). - `channels.slack` no longer accepts `on`. Bot Event Subscriptions and OAuth scopes are configured on the BYO Slack app; MDA accepts Phase 1 event kinds when they arrive. Authoring keeps only MDA runtime knobs (`autoReply`, `mentionBehavior`, `filters`, `conversation`). `requiredPermissions` is the fixed recommended scope list for app setup docs (`npm`, `pypi`, `cli`). #### Memory instructions - Memory instructions moved out of the memory file and into the system prompt. The Hub hot-memory file (`memories/agent/AGENTS.md`) is now seeded **empty** and holds only what the agent wrote; a new `mda-memory-instructions` middleware states the policy — the mount paths, which tools reach them, that only `/memories/agent/` is durable, that the tree is shared by every caller, and that memory is data rather than instructions — on every model call, but only when a slice is mounted. The instructions previously committed into the file rendered as the customer's "agent memory" in the Context Hub UI and went stale in place: a file seeded before the slice layout landed still taught `/memories/AGENTS.md` (`npm`, `pypi`, `cli`). - Deploy does not rewrite memory files. A deployment created before this release keeps whatever an older CLI wrote into its hot file, stale paths and all — the injected policy is what the model follows, and it says to treat file contents as notes rather than instructions. Delete the old block by hand if you want it gone (`cli`). - **Fix:** a Context Hub directory response whose shape drifted (a renamed or re-nested `files` field, or files without a `commit_hash`) parsed as "no files and no parent commit", which would have let deploy seed an empty memory file over a real one with the conflict check waived. It now fails the deploy instead (`cli`). - Authoring guidance changed: do not restate memory mechanics in `instructions.md`. The runtime says them on every turn, so a Memory section should cover only what is worth remembering in that project (`cli`, agent-builder skills). #### BREAKING (memory declaration) - Durable memory is now opt-in and owned by one root declaration: export `memory` from `memory.ts` / `memory.py` with `defineMemory({ scope: "agent" })` / `define_memory(scope="agent")`. A project without that module mounts nothing — previously every project with no `identity.{ts,py}` got agent memory implicitly (`npm`, `pypi`, `cli`). - `scope` is typed and slice-shaped rather than a boolean: `"agent"`, `"none"`, or `["agent"]` today. `"user"`, object slices (`{ kind, mount, access }`), `partition`, `access`, `mount`, `retention`, `provider`, and `sleepTimeCompute` are rejected with errors that name them as not-yet-supported, so later slice kinds land as data in the same shape (`npm`, `pypi`). - `disableMemory` / `disable_memory` on the agent definition is removed and now fails the build for **any** value: `false` used to mean "memory on", so honoring it under an opt-in contract would invert what was written. Delete it — memory is already off unless declared (`npm`, `pypi`, `cli`). - `scope.memory` on `defineIdentity` / `define_identity` is removed and fails the build with a migration message. Identity says who is calling; memory says which slices exist. `scope: "user"` keeps its threads/credentials meaning and no longer implies memory, and `IdentityConfig["scope"]` no longer carries a `memory` axis (`npm`, `pypi`, `cli`). - Per-caller memory is **removed**, not just unauthorable: the `/memories/user/` mounts, the per-user Context Hub repo (`{deploy}--u--{user}`), the lazy per-run remount (`LazyPerRunContextBackend`), the per-user hot-memory seeding, and the user-slice helpers are gone from both runtimes. `ResolvedMemorySlice` is agent-only, so a future private slice is rebuilt rather than re-enabled. The resolved config stays a versioned slice list, so adding a kind later does not change what consumers read. `mda delete` still tears down legacy `--u--` child repos, because deployments made before this release still have them (`npm`, `pypi`). - **Security:** the store authorization gate no longer depends on memory. The reserved managed tables (`mda_identity_links`, `mda_github_credentials` — account links and OAuth tokens for every caller) were checked *after* a per-caller memory check that returned early whenever no user slice was declared, so on any deployment without one — every deployment, in this release — an authenticated caller could read or overwrite other callers' entries. The reserved-table check now runs first and unconditionally, and `buildManagedAuth` / `build_managed_auth` no longer take a memory declaration (`npm`, `pypi`, `cli`). Reported by Corridor on #229. - A memory `scope` that is not a static literal fails the build: deploy seeds the hot file from that value, so it cannot be decided at runtime (`cli`). - The shared slice is documented as a trust boundary. Hot memory is injected into every run and ordinary runs can write it, so one caller can persist content — including instructions — that every later caller reads. The launch mount stays read/write; the seeded `AGENTS.md` now tells the model that memory is recorded data rather than orders and cannot widen what it may do, and the authoring helpers, scaffolded templates, and builder skill say who should not enable it. A privileged-writer policy is future work (`npm`, `pypi`, `cli`). - `mda init --memory ` scaffolds the declaration; without the flag no memory module is written. `mda dev` seeds `.mda/__contexthub__/memories/agent/` only when memory is declared, and deploy warns when Context Hub already holds memory files a project no longer declares (they are preserved, not deleted) (`cli`). #### BREAKING (identity scopes) - Organization-scoped identity is removed. `scope: "organization"`, the `organization` authoring option / requirement, organization claim mappings, organization fields on `runtime.identity`, and user-memory `partition: "organization"` are rejected. Trusted ingress no longer accepts `x-mda-organization-id` as an identity field (the header and `mda_organization_id` stay reserved so request bodies cannot spoof them). Private memory is always `{deploy}--u--{user}`; the CLI still deletes legacy `--t--` Hub children during teardown (`npm`, `pypi`, `cli`). - The `conversation` thread scope and `mda init --scope shared-conversations` preset are removed. Threads are always user-owned; the `scope.threads` authoring/resolved axis and the `ThreadScope` type are gone. Shared-bot recipes use user-owned threads and memory, with `credentials: "agent"` when they need the shared vault (`npm`, `pypi`, `cli`). - User-memory `partition` is gone from the resolved slice shape. User memory is always per-caller keyed; omit `partition` on `{ kind: "user" }` (any value, including `"user"`, is rejected). The `MemoryPartition` type and `userMemoryPartition` / `user_memory_partition` helpers are removed (`npm`, `pypi`). - Memory slice `access` (`"rw" | "ro"`) and the `MemoryAccess` type are removed. Mounts are always read/write; omit `access` on slice objects (`npm`, `pypi`). - Authoring `scope.memory` slices are string shorthands only (`"user"` / `"agent"`). Object form (`{ kind }`) is rejected until there are real per-slice options; resolved config still uses `{ kind }` (`npm`, `pypi`). - `defineIdentity()` / `define_identity()` now default to `scope: "agent"` (shared memory + LangSmith vault) instead of `"user"`. Per-caller Connect needs an explicit `scope: "user"`. `mda init --scope agent` scaffolds the argument-free declaration; `--scope user` writes the scope explicitly (`npm`, `pypi`, `cli`). - Override bags with no `default` also inherit `"agent"`. - Fleet tool errors on agent-scoped / no-identity deployments point at `mda connect` / LangSmith Integrations instead of Connect-with-X (`npm`, `pypi`). #### BREAKING (primitive authoring surface) Every primitive that has more than one provider is now selected as `.(…)` from a first-party namespace, and the old authoring names were removed rather than deprecated. Nothing has shipped to users, so there is no migration window. `specification/primitives/` has the design; `docs/` has the contributor guides. **Authoring** (`npm`, `pypi`) | Was | Now | | ------------------------------------------------------- | ------------------------------------------------ | | `defineSlackChannel` / `define_slack_channel` | `connectors.slack` | | `defineGitHubChannel` / `define_github_channel` | removed (use custom tools + `credentials.github`) | | `channels.slack` / `channels.github` | `connectors.slack` / removed | | `defineMcpServers` / `define_mcp_servers` | `connectors.mcp` | | `github.connector` | removed | | `langsmith.connector` / `connectors.langsmith` | removed (see above) | | `sandboxes.langsmith(…)` | `defineSandbox(…)` / `define_sandbox(…)` | | `providers.*` | `auth.*` | | `github.credentials(…)` | `credentials.github(…)` | | `github.connect()` / `slack.connect()` | `connect.github()` / `connect.slack()` | | `channels/` directory + `export const channel` | `connectors/` + `export const connector` | | `POST /channels/{name}/events` | `POST /connectors/{name}/events` | | GitHub channel `handlers` | removed (custom tools + `credentials.github`) | - **Channels → connectors (BREAKING):** the public `channels` namespace and `channels/` project directory are gone. Slack Events and messaging declare under `connectors/` with `export const connector` / `connector =`. Events URLs are `/connectors/{name}/events`. - `defineSandbox(…)` / `define_sandbox(…)` implicitly select LangSmith, so a project does not import `LangSmithSandbox` from `deepagents`. MDA still supports only that backend; a backend selector can be added if more are supported later. - Connector modules under `connectors/` export a `connector` in both languages — `export const connector = connectors.mcp(…)` in TypeScript, a module-level `connector = …` in Python. The earlier named `mcp` export is gone, and a TypeScript `export default` is now a build error naming the fix rather than a module the compiler skips in silence. - Python's internal connector/channel implementation packages are `_connectors` and `_channels` (runtime helpers), so they do not shadow the public `connectors` namespace. **Contribution API** (`npm`, `pypi`) - The primitives are now built on one internal `defineExtension(kind, options)` (`define_extension` in Python), narrowed by kind for `channel`, `connector`, and `sandbox`. Each declares its input as `args`: any [Standard Schema](https://standardschema.dev) (Zod, Valibot, ArkType) or a plain function in TypeScript, and a pydantic model / `TypeAdapter` — anything exposing `model_validate` or `validate_python` — or a callable in Python. Both runtimes validate through those interfaces, so a contributor can bring their own library. - The first-party primitives now validate with schemas — Zod in the npm package, pydantic in the PyPI package — instead of hand-written checks. `zod` is a dependency of `managed-deepagents` (npm) and `pydantic>=2` of `managed-deepagents` (PyPI), which previously had no runtime dependencies. Validation errors now name the failing path (`mcpServers.docs.url`) and read slightly differently for plain type mismatches, which Zod and pydantic word themselves; the messages for MDA's own rules are unchanged in substance. - It is **not exported**, and neither are the extension points it replaces: `defineChannel` and the `managed-deepagents/channels` subpath are gone, and the connector brand and hook context types now live only on `managed-deepagents/runtime`. A project selects primitives from the first-party namespaces and cannot mount an adapter, HTTP route, or sandbox backend of its own. #### BREAKING (identity) The identity surface was reworked and the old spellings were removed outright rather than deprecated, since nothing has shipped to users yet. Everything below is a rename or a removal; no capability was dropped. **Authoring — `identity.ts` / `identity.py`** (`npm`, `pypi`) - `defineIdentity` / `define_identity` no longer accept the `{ ingress, tenancy, scoping }` config. Declare `auth`, `scope`, `credentials`, and `connect` instead; unknown keys now throw. - `defineIdentity.preset(...)` / `define_identity.preset(...)` are removed, along with the `IdentityPresetName` type. `private-assistant` and `internal-tool` become `defineIdentity()` (they were identical in behavior), shared bots use `defineIdentity({ scope: { memory: "user", credentials: "agent" } })`, multi-tenant apps use per-user isolation, and `service` becomes `defineIdentity({ scope: "agent" })`. - `auth: "trusted_backend"` is removed; the value is `"backend"`. - `providers.guest({ actorPrefix })` is now `userPrefix` only. - `github.credentials(...)`: the `"actor"` token source is removed in favor of `"user"`, and `resolveActorToken` / `resolve_actor_token` is now `resolveUserToken` / `resolve_user_token` only. - Claim mappings take `user`, `groups`, and optional `email`; actor, tenant, and organization identity claims are removed. **Resolved config — `IdentityConfig`** (`npm`, `pypi`) The resolved contract now uses the same words as the authoring surface, so a value read while debugging is the value that was written. - `scoping` becomes `scope`; tenancy and organization requirements are removed. - Axis values use `user`, `agent`, `none`, and `custom`; thread ownership is always user-scoped. - `ScopingConfig` becomes `ScopeConfig`. - Context Hub user memory uses `{deploy}--u--{user}`. The CLI still recognizes legacy `--t--` child repos during deletion so old resources can be cleaned up. **Runtime envelope — `runtime.identity`** (`npm`, `pypi`) - `identity.actor` becomes `identity.user`; tenant and organization fields are removed. The aliases are gone, so `runtime.identity?.actor.id` must become `runtime.identity?.user.id`. - `user.type: "user" | "service"` becomes `user.kind: "person" | "service"`. - New `identity.groups` carries every group the caller's token asserts for authorization decisions in tools and middleware; isolation keys on the user. **Transport** (`npm`, `pypi`, `cli`) - The trusted ingress header is `x-mda-user-id`; tenant, organization, and group identity headers are not accepted. Group memberships come only from validated token claims. - `langgraph_auth_user` uses `mda_user_id`, `mda_user_email`, `mda_user_kind`, and `mda_groups`. Tenant and organization fields are removed, while reserved-key filtering still prevents request bodies from spoofing retired fields. - Hosted credential token-exchange assertions carry `user_id` and `user_kind`; tenant and organization identity fields are removed. **Plugin contracts** (`npm`, `pypi`) - `IdentityConnectProvider`: `status`, `startAuthorize`, and `unlink` take `userId`; the `shouldMount` context field `scoping` is now `scope`. - Slack identity links: record field `actorId` → `userId`, `slackActorId` → `slackKey`; `getByActorId` → `getByUserId`, `getBySlackActorId` → `getBySlackKey`, `removeByActorId` → `removeByUserId`. - Slack channel filters: `includeActors` / `excludeActors` → `includeUsers` / `excludeUsers`; filter reasons `actor_filtered` → `user_filtered` and `ambiguous_actor_namespace` → `ambiguous_user_namespace`. **CLI** (`cli`) - `mda init --identity ` is removed. Use `mda init --scope `; the retired preset names are no longer accepted as aliases. - Eval `identity.json` fixtures use `user` and optional `groups`. Unaffected on purpose: `LANGSMITH_TENANT_ID` and the CLI's `X-Tenant-Id` header refer to the LangSmith workspace, not identity, and keep their names. ### Removed #### BREAKING (validated-token provider helpers) - `auth.auth0`, `auth.clerk`, `auth.okta`, `auth.cognito`, `auth.entra`, `auth.google`, `auth.oidc`, and `auth.langsmith` are removed from the public authoring surface. Use `auth.supabase` for validated-token ingress (`npm`, `pypi`). #### BREAKING (guest ingress) - Anonymous/guest ingress is removed in full (`npm`, `pypi`, `cli`). Gone: the `auth.guest(...)` authoring helper, `ValidatedTokenProvider.guest`, the `POST /identity/guest` issuance route and its legacy Chat LangChain `POST /api/auth/guest` counterpart, HS256 guest-token minting and verification, and `validateIdentity`'s `guest` verifier branch. Callers must authenticate with `auth.supabase` or use trusted-backend ingress. - `MDA_GUEST_SIGNING_KEY` is **not** removed. Despite the name it is also the signing key for Connect-with-X OAuth `state` JWTs, so deployments using Connect must keep setting it. Only its guest-token usage is gone. - `issueGuestToken` (npm `managed-deepagents/runtime`) and `issue_guest_token` / `issue_legacy_guest_token` / `verify_legacy_guest_token` (pypi `managed_deepagents.runtime`) are removed. The Python-only `_legacy_chat_langchain` module is deleted — it existed solely for guest. - Internal rename: the auth-middleware bypass principal for public routes is now `mda:public-route` (was `mda:guest-issuer`), exposed as `publicRouteBypassUser` / `public_route_bypass_user`. It is a per-request synthetic identity, never persisted. #### BREAKING (`auth.github`) - The `auth.github()` validated-token provider is removed. It verified opaque GitHub OAuth access tokens by calling `GET https://api.github.com/user` on every request. Declare the same shape inline on `defineIdentity({ auth: ... })` if you need it — the `introspect` mechanism itself is unchanged (`npm`, `pypi`). - `credentials.github(...)`, `connect.github()`, and the Connect-with-GitHub routes are unaffected; they are separate features that only share the name. #### BREAKING (identity `scope`, `credentials`, `connect`) - `defineIdentity` / `define_identity` no longer accept `scope`, `credentials`, or `connect`. Identity is managed: threads and downstream credentials are always per-caller. `auth` is the only remaining option, and `scope` fails with a message pointing at `memory.{ts,py}` for the axis that moved there (`npm`, `pypi`). - This changes the resolved contract for existing projects that declared nothing: `defineIdentity()` used to resolve to `{credentials: "agent"}` and now resolves to `{credentials: "user"}`. Downstream calls carry the caller's own tokens through per-caller Connect rather than the shared LangSmith vault from `mda connect`. A deployment that relied on the shared vault has to move those integrations to per-caller Connect. - Durable memory is unaffected: it is declared in `memory.{ts,py}` and was never part of this change. - `mda init --scope ` is replaced by the `mda init --identity` flag, which scaffolds the argument-free declaration (`cli`). - Connect-with-GitHub no longer mounts from a `credentials.github(...)` chain, since that chain can no longer be declared. It mounts only for a user-scoped deployment that declares a GitHub connector (`npm`, `pypi`). #### BREAKING (`connectors.github`) - The `connectors.github` first-class connector is removed (sandbox repo checkouts, `gh` CLI install, GitHub App webhook Events ingress, and `POST /connectors/github/events`). Use custom tools or MCP for GitHub API access. Fleet/CLI `github-oauth-provider` is unchanged (`npm`, `pypi`). #### BREAKING (`connectors.langsmith`) - The `connectors.langsmith` capability-delegation connector is removed (authoring builders, HTTP routes under `/connectors/langsmith/capabilities/…`, and `CredentialTarget.kind: "langsmith"`). Custom LangSmith capability routes are out of scope while connector boundaries settle (`npm`, `pypi`). Sandbox backend behavior is unchanged, but declarations now use `defineSandbox(…)` / `define_sandbox(…)` instead of `sandboxes.langsmith(…)`. Definitions are data-only: provider classes and provider-generic sandbox types are internal, while TypeScript authors use the backend-neutral `SandboxOptions` type. ### Fixed - A syntax error in an authored file is now reported as `failed to parse ::: ` instead of being read as configuration that was never written. `mda build` used to discard parser diagnostics, so a stray brace in `agent.ts` produced "agent `name` is required" for a file whose name was right there, and a broken `channels/*.ts`, `connectors/*`, or `identity.*` was silently dropped from the deployment or deployed without its ingress secret. Applies to both languages; the Python extractors no longer swallow `SyntaxError`, and a missing `python3` is now distinguishable from "not configured" (`cli`) - `mda build --out ` no longer deletes a directory it does not own. The target must be missing, empty, or a directory a previous build wrote, and it can no longer be the project directory or an ancestor of it — `--out .` used to recursively empty the project, `.env` included (`cli`) - Harbor evals: invoke Python trials asynchronously so async-only tools (MCP connectors) no longer fail the first tool call, matching the Node runner - Harbor evals: forward every shell-exportable project `.env` key to the trial container so connector headers, MCP tokens, and tool API keys reach the agent - Keep handoff credentials on loopback - Generate internal handoff credentials - Avoid remote verification for server handoffs - Allow dynamic identity auth options - Reject null LangSmith auth responses - Recognize deploy-owned Context Hub repos in the collision gate (`deploy`) - Limit channel event request bodies - Drop removed defineIdentity options from fixtures (`integration-tests`) - Allow empty params for slack (`npm`) - Drop `scope` from the defineIdentity error text (`identity`) - Repair scope-removal fallout (`identity`) - Stop fixture and example prompts promising a user slice (`memory`) - Migrate channel fixtures off the identity memory axis (`memory`) - Refuse a scope a spread could decide (`memory`) - Resolve conflict markers committed into .gitignore - Match toolserver slugs case-insensitively in `--list` (`cli`) - Drop void operator that fails oxlint no-void (`npm`) - Repair clippy, oxfmt, and authoring export checks after rebase (`ci`) - Discover OAuth provider slugs and scopes per install (`cli`) - Strip tsx/cts from channel and connector stems (`runtime`) - Discover OAuth provider slugs from toolserver catalog (`cli`) (#217) - Persist LangSmith key and infer local public API URL for mda dev (`cli`) (#211) - Keep execute mounted under user/org memory remounts (`runtime`) (#199) - Mint LANGSMITH_API_KEY after device login for connect and dev (`cli`) - Allow empty Enter for LangSmith browser sign-in (`cli`) - Resolve npm/python on Windows and drop TS file: runtime dep (`cli`) (#178) - Surface parse errors instead of deploying misread config (`cli`) (#167) - Bound every outbound HTTP call in the TypeScript runtime (#166) - Keep server-local requests and the ingress secret on-host (`runtime`) (#164) - Stop mda build --out from deleting directories it does not own (`cli`) (#165) - Clear a prior actor's token on sandbox reuse (`github connector`) (#163) - Stop publishing a corrupted changelog and stale wheels, and gate release automation in CI (`release`) (#162) - Invoke Python trials async and forward project env (`eval`) (#148) ## 0.4.2 - 2026-07-24 ### Fixed - Harbor evals: always template the model provider API key, honor project `.env` - Harbor evals: write file tools to real disk under `/app` instead of StateBackend - Ignore OS junk that breaks UTF-8 Context Hub sync (`cli`) (#146) - Forward Harbor credentials and write trial files to disk (`eval`) (#143) - Skip co-located tests and non-exports at startup (`connectors`) (#144) - Use dotenvy for env file read (`cli`) - Remove allowlist parameter from github connector (`runtime`) (#141) ze `/app/...` paths so virtualMode does not double the root (`runtime`) - Harbor evals: print the jobs directory and note that re-running the same command resumes cached trials (`cli`) - Skip `connectors/` and `channels/` unit-test filenames and modules without the required plugin export so co-located tests and helpers are not imported at startup (`cli`, `npm`, `pypi`) ## 0.4.1 - 2026-07-22 ### Fixed - Forward LANGSMITH_API_KEY_GATEWAY to harbor output (`cli`) ## 0.4.0 - 2026-07-22 ### Added - Connector `sandbox` hook and per-connector plugin directories (`mcp`, - Neutral channel manifest envelope + `ChannelProviderAdapter` registry; Slack - `defineGitHubChannel` / `define_github_channel` — GitHub App webhook channel - Channel **plugin** contract (`defineChannel` / `define_channel`, brand, - Reference docs under `docs/` for building channel/connector plugins; product - Plugin connectors/channels and GitHub App channel (`channels`) (#131) - Add connector credentials, sandbox provisioning, and Connect-with-GitHub (`github`) (#93) - Reuse LangChain and deepagents types in authoring APIs (`pypi`) (#123) - Add Slack Events ingress with identity-scoped auto-reply (`channels`) (#101) - Local Context Hub mock and identity-scoped remounts (`memory`) (#114) - Add support for running evals with MDA (`cli`) (#100) ### Changed - Clean up dead code (#139) - Bump the npm_and_yarn group across 2 directories with 1 update (`deps`) (#138) - Collapse duplicated helpers across CLI and runtimes (#128) - Bump actions/setup-node in the major group (`deps`) (#117) - Bump the minor-and-patch group with 3 updates (`deps`) (#118) - Bump oxc from 0.139.0 to 0.140.0 in the major group (`deps`) (#119) - Bump the minor-and-patch group (`deps-dev`) (#121) - Bump typescript in /packages/npm in the major group (`deps-dev`) (#122) - Export typed ManagedDeepAgentRuntime for tools and middleware (`pypi`) (#113) - Fix release script CodeQL alerts (#107) - Harden Rust CLI toward idiomatic patterns (`cli`) (#108) ### Fixed - Require static name for deploy and assistant id (`cli`) (#137) - Fail build on missing connector/channel exports (`compile`) (#134) - Make Slack sandbox git/gh auth reliable end-to-end (`runtime`) (#132) - Forward LLM gateway base URLs in Harbor jobs (`eval`) (#133) - Run ci without omitting peer deps (`release`) - Skip connect prompts on unlinked thread replies (`runtime`) (#124) - Make github/slack/supabase UIs work on mobile (`examples`) (#125) - Update format (`test`) - Refresh stale seeds that teach /memories/AGENTS.md (`memory`) - Guide version mismatch toward project-local mda (`cli`) - Tighten Python runtime typing away from Any (`pypi`) (#112) - Harden release helper file handling - Patch Python dependency security alerts - Retry project deletes on LangSmith 429s (`integ`) (#98) - Cold vs hot memory (`runtime`) (#97) - Clean up context hub entries (`test`) (#96) - Require sandbox declaration and scaffold it in init (`cli`) (#95) - Require MDA_INGRESS_SECRET for trusted_backend identity (`deploy`) (#94) ## 0.3.1 - 2026-07-13 ### Changed - Bump the minor-and-patch group across 1 directory with 2 updates (`deps-dev`) (#87) - Update setuptools requirement in /packages/pypi (`deps-dev`) (#58) - Bump cliclack in the minor-and-patch group (`deps`) (#85) - Bump oxc from 0.138.0 to 0.139.0 in the major group (`deps`) (#86) ### Fixed - Lock language and runtime to install channel (`cli`) (#92) - Fix lock file (`npm`) - Refresh npm lockfile on release finalize (`ci`) - Update package-lock (`npm`) - Create git tag before GitHub Release in finalize (`ci`) ## 0.3.0 - 2026-07-12 ### Added - Richer version output and help UX polish (`cli`) - Managed identity, connectors, and Chat LangChain migration support (`identity`) (#83) - Restore --deployment-type flag for managed deep agents (`cli`) (#78) ### Fixed - Open studio from mda dev by default (#72) ## 0.2.0 - 2026-07-07 Published as `managed-deepagents==0.2.0` on PyPI. npm published the same CLI release train as `managed-deepagents@0.1.0` (manifests were not yet aligned). From the next release onward both packages share one version. ### Changed - Reoriented the project from language-specific API SDKs to a single `mda` CLI, written in Rust and cross-compiled for distribution to npm and PyPI (`cli`, `npm`, `pypi`). - Removed the previous Python (`python/`) and TypeScript (`js/`) SDKs. - `mda build` and `mda deploy` take the project path as a positional argument (defaults to `.`) instead of `--root` (`cli`). ### Added - `defineDeepAgent` / `define_deep_agent`: the v0 authoring contract. It accepts the full `createDeepAgent` surface minus the managed keys (`backend`, `store`, `checkpointer`) and returns a pre-runtime spec. Shipped from the TypeScript (`packages/npm`) and Python (`packages/pypi`) `managed-deepagents` packages, alongside a `runtime` helper (`compileManagedAgent` / `compile_managed_agent`) used by generated entry modules (`npm`, `pypi`). - `mda build ` / `mda deploy `: compiles a code-first Deep Agent project into a deployable managed LangGraph app. It copies your files verbatim, generates a managed entry module that consumes the `defineDeepAgent` definition and plugs it into an `agent(...)` graph factory, and generates `langgraph.json` (`cli`). - Managed Context Hub memory backends, MCP connectors, schedules, and scoped LangSmith sandbox backends (`cli`, `npm`, `pypi`). - Example TypeScript and Python Deep Agent projects under `examples/`. - npm and PyPI packages that ship the authoring interface plus the prebuilt CLI binary (`npm`, `pypi`). ### Removed - The AST transform of `agent.{ts,py}` (and the `oxc` / `rustpython-parser` dependencies). The CLI no longer rewrites user code; managed wiring lives in the generated entry module and the `defineDeepAgent` contract instead (`cli`). ## 0.1.0 - 2026-07-07 npm-only publish of the CLI-oriented `managed-deepagents` package (`managed-deepagents@0.1.0`). See [0.2.0](#020---2026-07-07) for the shared release notes for this train. ## 0.1.2 / 0.1.1 Earlier PyPI CLI packaging releases while npm was still on the `0.0.x` line. ## 0.0.2 Early npm packaging release of the CLI-oriented package (`managed-deepagents@0.0.2`), with matching platform optional dependencies. ## 0.1.0 (API SDK) Initial public beta SDK release (Python + TypeScript API clients). Superseded by the CLI direction above.