mirror of
https://github.com/langchain-ai/deepagentsjs.git
synced 2026-07-21 03:45:22 -04:00
@langchain/quickjs@0.5.1
177 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 github-actions[bot]
|
a9e1ba1b89 |
chore: version packages (#605)
This PR was opened by the [Changesets release](https://github.com/changesets/action) GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated. # Releases ## deepagents-acp@0.1.15 ### Patch Changes - Updated dependencies [[`7c4a11e`](https://github.com/langchain-ai/deepagentsjs/commit/7c4a11eacc11c3720b70d802068300ac3b4d8651)]: - deepagents@1.10.5 ## deepagents@1.10.5 ### Patch Changes - [#598](https://github.com/langchain-ai/deepagentsjs/pull/598) [`7c4a11e`](https://github.com/langchain-ai/deepagentsjs/commit/7c4a11eacc11c3720b70d802068300ac3b4d8651) Thanks [@christian-bromann](https://github.com/christian-bromann)! - refactor(stream): use langchain `run.subagents` instead of bespoke transformer Remove deepagents' custom `createSubagentTransformer` and rely on the native subagent stream that `createAgent` registers (langchain#37739). Keep `DeepAgentRunStream` as a compile-time overlay that narrows `run.subagents` to declared subagent specs. Update streaming tests for `cause` and per-subagent message coverage. ## @langchain/quickjs@0.5.1 ### Patch Changes - [#602](https://github.com/langchain-ai/deepagentsjs/pull/602) [`204cb27`](https://github.com/langchain-ai/deepagentsjs/commit/204cb27414c82c34a0c681c7e5a5336e1834f058) Thanks [@colifran](https://github.com/colifran)! - chore(quickjs): refine dynamic subagent prompt to trigger on workflow keyword and to improve iterative eval behavior - [#604](https://github.com/langchain-ai/deepagentsjs/pull/604) [`0971007`](https://github.com/langchain-ai/deepagentsjs/commit/0971007ab2491e73eb1a78c5426ab934470d5620) Thanks [@colifran](https://github.com/colifran)! - fix(quickjs): unwrap Command/ToolMessage envelopes from tool and subagent results ## @deepagents/evals@0.0.14 ### Patch Changes - Updated dependencies [[`7c4a11e`](https://github.com/langchain-ai/deepagentsjs/commit/7c4a11eacc11c3720b70d802068300ac3b4d8651)]: - deepagents@1.10.5 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> |
||
|
Christian Bromann
|
7c4a11eacc |
fix(deepagents): use langchain run.subagents instead of bespoke transformer (#598)
## Summary - Remove deepagents' bespoke `createSubagentTransformer` (~460 lines) and wire `createDeepAgent` to langchain's native `run.subagents` projection from `createAgent` (langchain#37739). - Slim `stream.ts` to type-only exports: re-export `SubagentRunStream` from langchain and keep `DeepAgentRunStream` as a compile-time narrowing over declared subagent specs. - Update streaming tests for `sub.cause` (replacing `taskInput`), add single-subagent `sub.messages` coverage, and document parallel-subagent streaming behavior. - Pin `langchain@1.4.6-dev-1781497519109` and `@langchain/langgraph` / `@langchain/langgraph-checkpoint` to `d2312cf` dev builds (root `pnpm.overrides`) for end-to-end verification until published releases are available. - Clarify `streamTransformers` docs: custom transformers surface on `run.extensions`; built-in streams like `run.subagents` and `run.toolCalls` are separate. fixes #560 |
||
|
Colin Francis
|
0971007ab2 |
fix(quickjs): unwrap Command/ToolMessage envelopes from tool and subagent results (#604)
### Summary The deepagents `task` tool resolves to a LangGraph `Command` (and some tools return a `ToolMessage` or a list of messages), but the quickjs bridges passed these envelopes straight to the REPL so `task(...)` and PTC `tools.*` calls handed the model the wrapper instead of the subagent's actual output, breaking the contract that a `responseSchema` dispatch resolves to a typed value. This PR adds `unwrapToolEnvelope`: a duck-typed unwrap of `Command` / `ToolMessage` / message-list shapes that reverse-scans `update.messages` for the last message with content and reads both live and serialized forms. Non-envelope values pass through unchanged. `unwrapToolEnvelope` is wired into both bridges, i.e., the `task()` dispatch and the PTC `tools.*` bridge. Inspired by the Python's [`coerce_tool_output_for_ptc`](https://github.com/langchain-ai/deepagents/blob/4c347603162e181c721e5e2a5ca8bbff6828e23c/libs/partners/quickjs/langchain_quickjs/_format.py#L93-L119). ### Tests - `coerce.test.ts` provide unit test coverage for `unwrapToolEnvelope`: string/null/undefined passthrough, `Command` (single + multi-message), reverse-scan past empty trailing messages, serialized `kwargs.content`, bare `ToolMessage`, message lists, content-block arrays and plain `{ content }` objects left untouched. - `middleware.test.ts` provides an integration test that the `task()` bridge returns the parsed structured value from a `Command` result. |
||
|
Colin Francis
|
204cb27414 |
chore(quickjs): refine dynamic subagent prompt to trigger on workflow keyword and to improve iterative eval behavior (#602)
### Summary Prompt change only, no runtime behavior change. Reworks the quickjs code-interpreter subagent prompt to steer the orchestrator toward an explore-then-distribute workflow. It explores with its own tools first, hands subagents file paths instead of pasted contents, reuses persisted REPL variables rather than re-pasting prior eval output, and returns results via the last expression instead of `console.log`. Also gates the REPL sandbox note so it only mentions the `tools.*` namespace when PTC is actually exposed. --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
github-actions[bot]
|
750d69c024 |
chore: version packages (#568)
This PR was opened by the [Changesets release](https://github.com/changesets/action) GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated. # Releases ## @langchain/node-vfs@0.2.0 ### Minor Changes - [#575](https://github.com/langchain-ai/deepagentsjs/pull/575) [`5d266a2`](https://github.com/langchain-ai/deepagentsjs/commit/5d266a213d8d3d15c2281e2e00ab71ee30ca3ffe) Thanks [@hntrl](https://github.com/hntrl)! - BREAKING: remove shell execution from the VFS provider `VfsBackend` now operates as a filesystem-only `BackendProtocolV2` implementation and no longer exposes command execution. The provider now implements `read`, `ls`, `grep`, and `glob` directly against the in-memory VFS, path resolution is confined to the virtual workspace root, and legacy `VfsSandbox` / `createVfsSandboxFactory*` aliases are removed. ## @langchain/quickjs@0.5.0 ### Minor Changes - [#590](https://github.com/langchain-ai/deepagentsjs/pull/590) [`7ee3feb`](https://github.com/langchain-ai/deepagentsjs/commit/7ee3feb674c361305b2a031db9132253e3d0543b) Thanks [@colifran](https://github.com/colifran)! - chore(quickjs): remove skills backend and associated plumbing - [#592](https://github.com/langchain-ai/deepagentsjs/pull/592) [`72cfb0c`](https://github.com/langchain-ai/deepagentsjs/commit/72cfb0c0384b30059b5e8028139a2e167c1be882) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): implement default subagent primitive in code interpreter for programmatic subagent calling ### Patch Changes - [#594](https://github.com/langchain-ai/deepagentsjs/pull/594) [`61db938`](https://github.com/langchain-ai/deepagentsjs/commit/61db9381494f61b4548cd80b715e5af72fdde381) Thanks [@colifran](https://github.com/colifran)! - chore(quickjs): remove configurable concurrency and add subagents option with true default - [#546](https://github.com/langchain-ai/deepagentsjs/pull/546) [`b7a5f26`](https://github.com/langchain-ai/deepagentsjs/commit/b7a5f26fd9cba992b11e1a4cf18607d64533503b) Thanks [@open-swe](https://github.com/apps/open-swe)! - fix(quickjs): scope `CodeInterpreterMiddleware` REPL prompt sandbox bullet to the runtime ## deepagents-acp@0.1.14 ### Patch Changes - Updated dependencies [[`18557db`](https://github.com/langchain-ai/deepagentsjs/commit/18557db7bbdf92052ed5f994512fb70e11989e69), [`773cac5`](https://github.com/langchain-ai/deepagentsjs/commit/773cac5dc7efc7843dd882642d91f7d64d6fde81), [`1ca6dc9`](https://github.com/langchain-ai/deepagentsjs/commit/1ca6dc92fd40a6d845d24b95ba14b8f2643db394), [`03df237`](https://github.com/langchain-ai/deepagentsjs/commit/03df237385fbdfefd862076c5588eb39cb6e43c3), [`84f3c0c`](https://github.com/langchain-ai/deepagentsjs/commit/84f3c0c2f1cad271191bcc138b84ba5b9c9205c9), [`72cfb0c`](https://github.com/langchain-ai/deepagentsjs/commit/72cfb0c0384b30059b5e8028139a2e167c1be882), [`04cc3fc`](https://github.com/langchain-ai/deepagentsjs/commit/04cc3fc26001ee566ed94de44c2dda2cf6adecc4), [`18fbb48`](https://github.com/langchain-ai/deepagentsjs/commit/18fbb4839050e98ae3cfd36ec69b11f0725ad6d6), [`e3d4b53`](https://github.com/langchain-ai/deepagentsjs/commit/e3d4b5367b1825df56c919b483ec4a3e117d631f)]: - deepagents@1.10.4 ## deepagents@1.10.4 ### Patch Changes - [#551](https://github.com/langchain-ai/deepagentsjs/pull/551) [`18557db`](https://github.com/langchain-ai/deepagentsjs/commit/18557db7bbdf92052ed5f994512fb70e11989e69) Thanks [@antonnak](https://github.com/antonnak)! - fix(deepagents): gate cache_control writes on per-call request.model `createCacheBreakpointMiddleware` and `createMemoryMiddleware` were gating the Anthropic-specific `cache_control` write at agent-creation time only. When `modelFallbackMiddleware` swapped `request.model` to a non-Anthropic provider mid-flight (e.g. on Anthropic 5xx), the marker leaked through and the fallback provider rejected the request with `400 Unknown parameter: 'cache_control'`. Both middlewares now also check `isAnthropicModel(request.model)` inside `wrapModelCall`. Fixes [#550](https://github.com/langchain-ai/deepagentsjs/issues/550). - [#591](https://github.com/langchain-ai/deepagentsjs/pull/591) [`773cac5`](https://github.com/langchain-ai/deepagentsjs/commit/773cac5dc7efc7843dd882642d91f7d64d6fde81) Thanks [@colifran](https://github.com/colifran)! - chore(deepagents): expose createSubAgent - [#541](https://github.com/langchain-ai/deepagentsjs/pull/541) [`1ca6dc9`](https://github.com/langchain-ai/deepagentsjs/commit/1ca6dc92fd40a6d845d24b95ba14b8f2643db394) Thanks [@ixchio](https://github.com/ixchio)! - fix getMimeType to return application/octet-stream for unknown file extensions instead of text/plain - [#572](https://github.com/langchain-ai/deepagentsjs/pull/572) [`03df237`](https://github.com/langchain-ai/deepagentsjs/commit/03df237385fbdfefd862076c5588eb39cb6e43c3) Thanks [@hntrl](https://github.com/hntrl)! - fix: scope CompositeBackend grep/glob route fanout by search path CompositeBackend now limits fallback route fanout to routes mounted under the requested search path, instead of querying all routed backends unconditionally. This avoids unrelated routed backend calls (and side-effect errors) for scoped searches like `path="/workspace"`, while preserving full fanout behavior at root (`path="/"`). - [#574](https://github.com/langchain-ai/deepagentsjs/pull/574) [`84f3c0c`](https://github.com/langchain-ai/deepagentsjs/commit/84f3c0c2f1cad271191bcc138b84ba5b9c9205c9) Thanks [@hntrl](https://github.com/hntrl)! - fix(deepagents): add explicit browser and node entrypoints - add `deepagents/browser` and `deepagents/node` subpath exports - route browser bundlers to the browser-safe bundle via the root `browser` export condition - avoid named Node builtin imports in backend utils that can break browser builds - document browser guidance to import from `deepagents/browser` - [#592](https://github.com/langchain-ai/deepagentsjs/pull/592) [`72cfb0c`](https://github.com/langchain-ai/deepagentsjs/commit/72cfb0c0384b30059b5e8028139a2e167c1be882) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): implement default subagent primitive in code interpreter for programmatic subagent calling - [#566](https://github.com/langchain-ai/deepagentsjs/pull/566) [`04cc3fc`](https://github.com/langchain-ai/deepagentsjs/commit/04cc3fc26001ee566ed94de44c2dda2cf6adecc4) Thanks [@hntrl](https://github.com/hntrl)! - fix(deepagents): propagate subagent `lc_agent_name` during task delegation - Ensure `task` tool subagent invocations override `metadata.lc_agent_name` with the selected `subagent_type`. - Add regression coverage for both compiled subagents (`runnable`) and standard subagent specs to verify tool-time metadata reflects the active subagent. - Update the `langsmith` peer dependency range in `deepagents` to `^0.7.1`. - [#595](https://github.com/langchain-ai/deepagentsjs/pull/595) [`18fbb48`](https://github.com/langchain-ai/deepagentsjs/commit/18fbb4839050e98ae3cfd36ec69b11f0725ad6d6) Thanks [@christian-bromann](https://github.com/christian-bromann)! - fix(deepagents): count tokens once per model call in summarization middleware `createSummarizationMiddleware` counted tokens twice on every model call—once inside `truncateArgs` and again for the should-summarize check—even when nothing was truncated or summarized. Count once and pass the total into `truncateArgs`; recount only when truncation actually modifies messages. - [#242](https://github.com/langchain-ai/deepagentsjs/pull/242) [`e3d4b53`](https://github.com/langchain-ai/deepagentsjs/commit/e3d4b5367b1825df56c919b483ec4a3e117d631f) Thanks [@alvedder](https://github.com/alvedder)! - feat(deepagents): support direct skill paths as sources in createSkillsMiddleware ## @deepagents/evals@0.0.13 ### Patch Changes - Updated dependencies [[`18557db`](https://github.com/langchain-ai/deepagentsjs/commit/18557db7bbdf92052ed5f994512fb70e11989e69), [`773cac5`](https://github.com/langchain-ai/deepagentsjs/commit/773cac5dc7efc7843dd882642d91f7d64d6fde81), [`1ca6dc9`](https://github.com/langchain-ai/deepagentsjs/commit/1ca6dc92fd40a6d845d24b95ba14b8f2643db394), [`03df237`](https://github.com/langchain-ai/deepagentsjs/commit/03df237385fbdfefd862076c5588eb39cb6e43c3), [`84f3c0c`](https://github.com/langchain-ai/deepagentsjs/commit/84f3c0c2f1cad271191bcc138b84ba5b9c9205c9), [`72cfb0c`](https://github.com/langchain-ai/deepagentsjs/commit/72cfb0c0384b30059b5e8028139a2e167c1be882), [`04cc3fc`](https://github.com/langchain-ai/deepagentsjs/commit/04cc3fc26001ee566ed94de44c2dda2cf6adecc4), [`18fbb48`](https://github.com/langchain-ai/deepagentsjs/commit/18fbb4839050e98ae3cfd36ec69b11f0725ad6d6), [`e3d4b53`](https://github.com/langchain-ai/deepagentsjs/commit/e3d4b5367b1825df56c919b483ec4a3e117d631f)]: - deepagents@1.10.4 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> |
||
|
Christian Bromann
|
18fbb48390 |
fix(deepagents): count tokens once per model call in summarization middleware (#595)
## Summary - Port [langchain-ai/deepagents#3877](https://github.com/langchain-ai/deepagents/pull/3877): `createSummarizationMiddleware` now counts tokens once per model call instead of twice on the common pass-through path. - Pass the precomputed total into `truncateArgs` via an optional `totalTokens` parameter; recount only when truncation modifies messages. - Add a regression test asserting `countTokensApproximately` runs exactly once when nothing is truncated or summarized. |
||
|
Colin Francis
|
61db938149 |
chore(quickjs): remove configurable concurrency and add subagents option with true default (#594)
### Summary This PR remove the configurable concurrency option from code interpreter middleware options and adds a subagents option that is true by default |
||
|
Colin Francis
|
72cfb0c038 |
feat(quickjs): implement default task primitive in code interpreter for programmatic subagent calling (#592)
### Summary This PR adds a default task primitive to the code interpreter that allows the agent to natively write code to programmatically dispatch subagents. The task primitive supports passing response schemas for dynamic structured output that helps with more complex orchestration patterns. This PR also removes the swarm task tool in favor of the task primitive. --------- Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Colin Francis
|
773cac5dc7 |
chore(deepagents): expose createSubAgent (#591)
### Summary Exposes `createSubAgent` as a dedicated function which allows us to reuse the inner subagent factory outside of the middleware (e.g. using it as the single source of truth for defining subagents in quickjs). Port of similar implementation from https://github.com/langchain-ai/deepagents/pull/3846 |
||
|
Colin Francis
|
7ee3feb674 |
chore(quickjs): remove skills backend and associated plumbing (#590)
### Summary Removes skills backend and associated plumbing from code interpreter middleware and options |
||
|
Alexei Vedernikov
|
e3d4b5367b |
feat(deepagents): support direct skill paths as sources in createSkillsMiddleware (#242)
## Description
Previously, sources had to be parent directories (e.g. `/skills/`) whose
subdirectories each contained a SKILL.md. Passing a direct skill path
(e.g. `/skills/my-skill/`) had no effect because `listSkillsFromBackend`
scanned for subdirectories inside the source, found none, and returned
an empty list.
## Real case
The motivating scenario is when a curated list of skills organised as
individual directories under a shared directory. Each skill has a clear
responsibility, and different skills are intentionally assigned to the
orchestrator agent vs. subagents:
```typescript
// Desired: pass individual skill paths — one per capability
const SHARED_SKILLS = ['/skills/web-research/'];
const ORCHESTRATOR_SKILLS = [
'/skills/requirement-completeness-validator/',
'/skills/diff-risk-analyzer/',
];
createDeepAgent({
skills: [...SHARED_SKILLS, ...ORCHESTRATOR_SKILLS],
subagents: subagentsConfig.map(agentConfig => ({
skills: [
...SHARED_SKILLS,
...agentConfig.skills,
]
}))
});
```
The workaround would be to create nested directories like this:
```
skills/
web-research/
web-research/
SKILL.md
equirement-completeness-validator/
equirement-completeness-validator/
SKILL.md
...and so on
```
## What's done
Detection is done by inspecting the lsInfo() result of the source
directory: if a SKILL.md file appears directly in the listing, the
source itself is treated as the skill directory. Otherwise the existing
parent-directory scan runs unchanged.
The two modes can be freely mixed in the sources array:
```ts
sources: [
'/skills/', // parent dir: all subdirs with SKILL.md
'/skills/my-skill/', // direct path: SKILL.md at its root
]
```
---------
Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com>
|
||
|
Anton Nak
|
18557db7bb |
fix(deepagents): gate cache_control writes on per-call request.model (#551)
Fixes #550. ## Problem `createCacheBreakpointMiddleware` and `createMemoryMiddleware` write Anthropic-specific `cache_control: { type: "ephemeral" }` markers into the request payload. Both gate the write at **agent-creation time** in `createDeepAgent` (whether the *primary* model is Anthropic). When `modelFallbackMiddleware` swaps `request.model` to a non-Anthropic provider at request time, the markers leak through and the fallback provider rejects the request: ``` 400 Unknown parameter: 'messages[0].content[N].cache_control' ``` Full root-cause writeup, production trace, and reproduction in #550. ## Fix Move the model classification from boot-time to **per-call**, using the existing `isAnthropicModel` helper (`src/utils.ts:9`). - **`src/middleware/cache.ts`** — bail at the top of `wrapModelCall` when `request.model` isn't Anthropic. - **`src/middleware/memory.ts`** — AND the existing `addCacheControl` flag with the same per-call check. `addCacheControl` keeps its current public meaning (opt-in switch); the per-call gate is an additional safety net. - **`src/agent.ts`** — left untouched. The boot-time install gate stays (small perf win when primary is non-Anthropic — middleware isn't installed at all). The new per-call check handles the fallback case where the middleware *is* installed but the model has been swapped. ## Tests Each middleware's `.test.ts` got a `per-call provider gating` block. They build a request with a `ChatAnthropic`-shaped stub vs a `ChatOpenAI`-shaped stub on `request.model`, run `wrapModelCall`, and assert `cache_control` is/isn't written. Also covers `ConfigurableModel` with `modelProvider: "anthropic"` and the string forms (`"anthropic:claude-..."`, `"openai:gpt-5"`). Existing fixtures that didn't set `request.model` now pass a `ChatAnthropic` stub so they keep exercising the write path — they were previously relying on the unconditional behavior. No behavioral assertions changed; only the fixture got more explicit. 31/31 tests pass on `libs/deepagents` (`cache.test.ts` + `memory.test.ts`). ## Compatibility - **No public API change.** `addCacheControl?: boolean` on `MemoryMiddlewareOptions` keeps its existing semantics. - **Anthropic-primary happy path unchanged** — same writes, same cache hits. - **Per-call cost:** one `getName()` / string check inside `wrapModelCall`. Sub-microsecond. - **Failure mode improvement:** users running `createDeepAgent` with an Anthropic primary + non-Anthropic fallback no longer get a 400 from the fallback provider on swap. ## Changeset `patch` bump on `deepagents`. ## Out of scope `request.modelSettings.cache_control` (written by LangChain core's `anthropicPromptCachingMiddleware`) can also leak across the swap. Mentioned in #550 for tracking — that's a LangChain-side fix, not deepagents. --------- Co-authored-by: Anton Nakaliuzhnyi <anakaliuzhnyi@gipartners.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
Aman Pandey
|
1ca6dc92fd |
fix(deepagents): return application/octet-stream for unknown MIME types (#541)
Fixes #542. `getMimeType()` falls back to `"text/plain"` for any extension not in the `MIME_TYPES` lookup table. Since `isTextMimeType("text/plain")` returns `true`, this means binary files (.zip, .exe, .wasm, .sqlite, .pyc, .class, .jar, etc.) get treated as text everywhere downstream — grep searches through them and returns garbage, read_file splits their content on newlines and paginates it, and StateBackend can silently corrupt binary data stored via createFileData(). The fix is straightforward: change the default to `"application/octet-stream"`, which is what unknown binary data should be. `isTextMimeType()` already returns `false` for it, so binary files stop getting mangled. To avoid breaking existing behavior for actual text files that were previously relying on the fallback (like .txt, .ts, .py, .json, .js — none of which had explicit entries), this also adds those to the MIME_TYPES map directly. Covers ~67 common text/code extensions. All 386 backend tests pass (10 files). The updated test expectations in utils.test.ts verify both sides: known text extensions still resolve correctly, and unknown/binary extensions now get `application/octet-stream` instead of being misidentified as text. --------- Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Hunter Lovell
|
5d266a213d |
refactor(node-vfs): make VfsSandbox filesystem-only (#575)
## Summary This PR removes command-execution semantics from `@langchain/node-vfs` and aligns `VfsSandbox` with `BackendProtocolV2` as a filesystem-only backend. The provider no longer uses sandbox execution abstractions and now performs core file operations directly against the in-memory VFS. ## Changes ### Protocol alignment - `VfsSandbox` now implements `BackendProtocolV2` directly instead of extending sandbox/base-execution abstractions. - Removed command execution from the provider surface (`execute` is no longer part of this backend implementation). - Removed `timeout` from `VfsSandboxOptions` and corresponding docs. ### Runtime behavior - Added native backend implementations for: - `read` - `readRaw` - `write` - `edit` - `ls` - `grep` - `glob` - Kept upload/download support and added path confinement handling for invalid traversal paths. - Continued in-memory VFS-only operation (no temp-dir shell sync path). ### Tests and docs - Updated unit and integration tests to remove execute-focused assertions and validate filesystem behavior. - Updated package docs to describe `BackendProtocolV2` usage and removed execute/timeout references. - Added changeset for `@langchain/node-vfs`. |
||
|
open-swe[bot]
|
b7a5f26fd9 |
fix(quickjs): scope REPL prompt sandbox bullet to the runtime (#546)
Ports langchain-ai/deepagents#3528 to deepagentsjs. The base `CodeInterpreterMiddleware` REPL prompt asserted "no filesystem, no stdlib, no network, no real clock, no `fetch`, no `require`". When `ptc` is configured, the same system prompt also exposes a `tools.*` namespace whose members can include things like web fetch, file I/O, or shell access — directly contradicting the sandbox bullet. A model reading both sections sees "no network" *and* `await tools.webFetch({...})` in the same prompt. With PTC disabled it's even worse: exposed tools are still reachable through the normal tool-calling path *outside* `eval`, and the model could interpret the bullet as a statement about the agent's capabilities rather than the REPL's. This rewords the bullet to scope it to the QuickJS runtime, and notes that external side effects from inside the REPL flow through `tools.*` when it's exposed (otherwise the REPL is pure computation). `@langchain/quickjs` middleware unit tests, lint, and format all pass locally. _Opened collaboratively by Mason Daugherty and open-swe._ Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com> Co-authored-by: Mason Daugherty <61371264+mdrxy@users.noreply.github.com> |
||
|
Hunter Lovell
|
84f3c0c2f1 |
fix(deepagents): add browser and node entrypoints (#574)
## Summary Fixes #292, #543 This updates `deepagents` packaging to split environment-specific exports and avoid browser build failures caused by Node-only imports in the default bundle. Browser consumers now have an explicit `deepagents/browser` entrypoint, while Node consumers can keep using `deepagents` (or optionally `deepagents/node`) for the full runtime API. ## Changes ### `deepagents` package exports and entrypoints - Added `src/browser.ts` as a browser-safe public barrel that excludes Node-only APIs. - Added `src/node.ts` as an explicit Node entrypoint that re-exports the full API. - Updated `package.json` exports to: - route the root `browser` condition to `./dist/browser.js` - expose `./browser` and `./node` subpath exports (ESM + CJS + types). - Updated `tsdown` config to build `index`, `browser`, and `node` outputs. ### Runtime import graph hardening - Updated `createDeepAgent` to import `StateBackend` directly from `backends/state` so it no longer pulls in the full Node-oriented backends barrel. - Updated `types.ts` to import `AnyBackendProtocol` from `backends/protocol` instead of `backends/index`. - Removed named Node builtin path imports from `backends/utils.ts` by replacing `path.basename`/`path.extname` with local helpers. ### Documentation and release notes - Added a README section documenting import guidance: browser apps should use `deepagents/browser`, and Node apps can continue to use `deepagents`. - Added a patch changeset for `deepagents` describing the entrypoint and packaging fix. |
||
|
Hunter Lovell
|
03df237385 |
fix(deepagents): scope composite route search fanout by path (#572)
## Summary Fixes #241. This updates `CompositeBackend` search routing so scoped `grep`/`glob` requests only fan out to routed backends mounted under the requested path. Previously, non-route paths (for example `/workspace`) still queried all routes, which could trigger unrelated backend errors. ## Changes ### `libs/deepagents/src/backends/composite.ts` - Added route/path helper predicates to distinguish: - when a search path is inside a route, - when a route is a descendant of the requested search path. - Updated `grep` fallback behavior to query the default backend plus only descendant routes (instead of unconditional all-route fanout). - Updated `glob` fallback behavior to use the same descendant-route filtering. - Normalized `grep` path handling to accept `null` and treat it as root (`/`). - Reused the same route-matching helper in `ls` for consistent prefix matching semantics. ### `libs/deepagents/src/backends/composite.test.ts` - Added regression test to ensure `grep("...", "/workspace")` does not call unrelated routes (for example `/skills/`) while still searching relevant descendant routes. - Added regression test for equivalent `glob` behavior. Co-authored-by: Christian Bromann <731337+christian-bromann@users.noreply.github.com> |
||
|
github-actions[bot]
|
c524561ad0 |
chore: version packages (#538)
This PR was opened by the [Changesets release](https://github.com/changesets/action) GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated. # Releases ## deepagents-acp@0.1.13 ### Patch Changes - Updated dependencies \[[`bfb6eec`](https://github.com/langchain-ai/deepagentsjs/commit/bfb6eecdfe617645b3bdebf9a60d4b08e575cef7), [`9c666ba`](https://github.com/langchain-ai/deepagentsjs/commit/9c666ba44adc1f8b428546c2191ea71d88b03998), [`04cc3fc`](https://github.com/langchain-ai/deepagentsjs/commit/04cc3fc26001ee566ed94de44c2dda2cf6adecc4), [`9221c8a`](https://github.com/langchain-ai/deepagentsjs/commit/9221c8a2b5236954f674e30f3ef0e2962f54fb56), [`f6d3f13`](https://github.com/langchain-ai/deepagentsjs/commit/f6d3f13559bd2d9dfad63e379b59ad7577e828be)]: - deepagents@1.10.3 ## deepagents@1.10.3 ### Patch Changes - [#500](https://github.com/langchain-ai/deepagentsjs/pull/500) [`bfb6eec`](https://github.com/langchain-ai/deepagentsjs/commit/bfb6eecdfe617645b3bdebf9a60d4b08e575cef7) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): add swarm task tool - [#288](https://github.com/langchain-ai/deepagentsjs/pull/288) [`9c666ba`](https://github.com/langchain-ai/deepagentsjs/commit/9c666ba44adc1f8b428546c2191ea71d88b03998) Thanks [@christian-bromann](https://github.com/christian-bromann)! - fix(deepagents): handle non-string content blocks in tool result sizechecking - [#566](https://github.com/langchain-ai/deepagentsjs/pull/566) [`04cc3fc`](https://github.com/langchain-ai/deepagentsjs/commit/04cc3fc26001ee566ed94de44c2dda2cf6adecc4) Thanks [@hntrl](https://github.com/hntrl)! - fix(deepagents): propagate subagent `lc_agent_name` during task delegation - Ensure `task` tool subagent invocations override `metadata.lc_agent_name` with the selected `subagent_type`. - Add regression coverage for both compiled subagents (`runnable`) and standard subagent specs to verify tool-time metadata reflects the active subagent. - Update the `langsmith` peer dependency range in `deepagents` to `^0.7.1`. - [#549](https://github.com/langchain-ai/deepagentsjs/pull/549) [`9221c8a`](https://github.com/langchain-ai/deepagentsjs/commit/9221c8a2b5236954f674e30f3ef0e2962f54fb56) Thanks [@colifran](https://github.com/colifran)! - chore(deepagents): move required ptc tools to metadata - [#537](https://github.com/langchain-ai/deepagentsjs/pull/537) [`f6d3f13`](https://github.com/langchain-ai/deepagentsjs/commit/f6d3f13559bd2d9dfad63e379b59ad7577e828be) Thanks [@hntrl](https://github.com/hntrl)! - bump langsmith sdk version ## @langchain/quickjs@0.4.1 ### Patch Changes - [#545](https://github.com/langchain-ai/deepagentsjs/pull/545) [`b3f0d6e`](https://github.com/langchain-ai/deepagentsjs/commit/b3f0d6e8fa2980652123e0053c116fc9945720ac) Thanks [@colifran](https://github.com/colifran)! - chore(quickjs): update swarm task tool to use withStructuredOutput - [#500](https://github.com/langchain-ai/deepagentsjs/pull/500) [`bfb6eec`](https://github.com/langchain-ai/deepagentsjs/commit/bfb6eecdfe617645b3bdebf9a60d4b08e575cef7) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): add swarm task tool - [#549](https://github.com/langchain-ai/deepagentsjs/pull/549) [`9221c8a`](https://github.com/langchain-ai/deepagentsjs/commit/9221c8a2b5236954f674e30f3ef0e2962f54fb56) Thanks [@colifran](https://github.com/colifran)! - chore(deepagents): move required ptc tools to metadata ## @deepagents/evals@0.0.12 ### Patch Changes - Updated dependencies \[[`bfb6eec`](https://github.com/langchain-ai/deepagentsjs/commit/bfb6eecdfe617645b3bdebf9a60d4b08e575cef7), [`9c666ba`](https://github.com/langchain-ai/deepagentsjs/commit/9c666ba44adc1f8b428546c2191ea71d88b03998), [`04cc3fc`](https://github.com/langchain-ai/deepagentsjs/commit/04cc3fc26001ee566ed94de44c2dda2cf6adecc4), [`9221c8a`](https://github.com/langchain-ai/deepagentsjs/commit/9221c8a2b5236954f674e30f3ef0e2962f54fb56), [`f6d3f13`](https://github.com/langchain-ai/deepagentsjs/commit/f6d3f13559bd2d9dfad63e379b59ad7577e828be)]: - deepagents@1.10.3 --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Colin Francis <colin.francis@langchain.dev> |
||
|
Hunter Lovell
|
04cc3fc260 |
fix(deepagents): propagate subagent lc_agent_name for delegated tasks (#566)
## Summary Fixes #206. This updates deepagents subagent delegation so tool executions can reliably identify the active subagent via `config.metadata.lc_agent_name` instead of inheriting the parent agent name. It also adds focused regression tests that validate both compiled subagents and standard subagent specs follow the same metadata behavior. ## Changes ### `libs/deepagents` subagent metadata propagation - Updated `createTaskTool` subagent invocation config to explicitly set: - `metadata.lc_agent_name = subagent_type` - existing `configurable.ls_agent_type = "subagent"` behavior remains unchanged - Added new regression coverage in `subagent.test.ts`: - compiled subagent (`runnable`) path - standard subagent spec (`systemPrompt/tools/model`) path - Tests assert tool-time metadata receives the delegated subagent name (`worker`), preventing parent-name leakage. ### `libs/deepagents` dependency range update - Bumped `langsmith` peer dependency range in `libs/deepagents/package.json` from `>=0.6.0 <1.0.0` to `^0.7.1`. - Updated `pnpm-lock.yaml` accordingly. |
||
|
Christian Bromann
|
9c666ba44a |
fix(deepagents): handle non-string content blocks in tool result sizechecking (#288)
Fixes #277 The tool result eviction logic in `processToolMessage` only checked `content` when it was a plain string (`typeof msg.content === "string"`), which meant array-based content blocks (e.g. `[{ type: "text", text: "..." }]`) bypassed the size limit entirely and were sent back to the agent untruncated. The fix uses the existing `BaseMessage.text` getter from `@langchain/core`, which already handles both content formats: - **String content**: returned as-is - **Array content blocks**: joins all blocks with `type === "text"` into a single string This avoids introducing any new utility functions — the upstream API already solves this. ## Changes - Updated `processToolMessage()` in `fs.ts` to use `msg.text` instead of directly checking `msg.content` - Added 2 unit tests covering array-based content block eviction (large and small) |
||
|
Colin Francis
|
9221c8a2b5 |
chore(deepagents): move required ptc tools to metadata (#549)
This PR aligns SkillMetadata with the agent skills spec by removing the top level required ptc tools field and keeps it contained within metadata |
||
|
Colin Francis
|
b3f0d6e8fa |
chore(quickjs): update swarm task tool to use withStructuredOutput (#545)
### Summary Switch JS swarm task tool to use withStructuredOutput instead of hand-rolled bindTools + tool_calls extraction. ### Tests - All 202 unit tests pass - Removed normalizeSchema tests and bindTools-specific tests, replaced with withStructuredOutput test using FakeToolCallingModel |
||
|
Colin Francis
|
bfb6eecdfe |
feat(quickjs): add swarm task tool (#500)
### Summary
Adds the swarm task tool to @langchain/quickjs and fixes skill module
loading for the agentskills.io spec, enabling parallel subagent dispatch
from inside the QuickJS code interpreter.
Swarm task tool (tools/swarm-task.ts) is a PTC-only tool that bridges
the code interpreter to the host runtime for subagent dispatch. It
supports two modes:
1. Direct model calls with structured output (invoke mode) for
classification/extraction tasks
2. Full agentic loops with tools (agent mode) for multi-step reasoning.
It handles schema normalization, model variant caching with TTL-based
sweeping, and structured output binding.
Spec-compliant entrypoint resolution — loadSkill() now resolves the
entrypoint from metadata.entrypoint (the agentskills.io extension point)
first, falling back to the legacy top-level module field. This lets
skills use spec-compliant frontmatter without a top-level module key.
Subdirectory entrypoint support — the QuickJS module normalizer now uses
the entrypoint's directory (not the skill root) when resolving relative
imports from a bare specifier. This fixes import { ... } from
"./sibling.js" when the entrypoint lives in a subdirectory like
scripts/index.ts.
Supporting changes include wiring skills metadata through the middleware
and session layers, extending the transform pipeline to strip import
type/export type declarations, adding requiredPtcTools validation, and
session improvements (extractToolText, stripLineNumbers, __sessionId__).
The swarm skill itself (table management, batching, filtering,
interpolation) is distributed separately and not included here.
### Tests
- 41 tests for the swarm task tool covering both dispatch modes, schema
normalization, variant caching and TTL sweeping, error handling, and
edge cases around model binding
- 3 new tests for entrypoint resolution (metadata.entrypoint preferred
over module, subdirectory relative imports)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Hunter Lovell
|
f6d3f13559 | chore(deepagents): bump langsmith sdk version (#537) | ||
|
github-actions[bot]
|
163ee4949e |
chore: version packages (#532)
This PR was opened by the [Changesets release](https://github.com/changesets/action) GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated. # Releases ## deepagents-acp@0.1.12 ### Patch Changes - Updated dependencies \[[`f088089`](https://github.com/langchain-ai/deepagentsjs/commit/f0880899ea6726b7320b0888d0f6a10a7749e1bf), [`7c33a86`](https://github.com/langchain-ai/deepagentsjs/commit/7c33a8695f2e16217779bef5c6fca28230f18815)]: - deepagents@1.10.2 ## deepagents@1.10.2 ### Patch Changes - [#533](https://github.com/langchain-ai/deepagentsjs/pull/533) [`f088089`](https://github.com/langchain-ai/deepagentsjs/commit/f0880899ea6726b7320b0888d0f6a10a7749e1bf) Thanks [@vishnu-ssuresh](https://github.com/vishnu-ssuresh)! - feat(deepagents): add `ContextHubBackend` for LangSmith Hub agent repos - [#526](https://github.com/langchain-ai/deepagentsjs/pull/526) [`7c33a86`](https://github.com/langchain-ai/deepagentsjs/commit/7c33a8695f2e16217779bef5c6fca28230f18815) Thanks [@colifran](https://github.com/colifran)! - feat(deepagents): implement harness profiles ## @deepagents/evals@0.0.11 ### Patch Changes - Updated dependencies \[[`f088089`](https://github.com/langchain-ai/deepagentsjs/commit/f0880899ea6726b7320b0888d0f6a10a7749e1bf), [`7c33a86`](https://github.com/langchain-ai/deepagentsjs/commit/7c33a8695f2e16217779bef5c6fca28230f18815)]: - deepagents@1.10.2 --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Colin Francis <colin.francis@langchain.dev> |
||
|
Vishnu Suresh
|
f0880899ea |
feat(deepagents): add ContextHubBackend backend (#533)
This ports Python deepagents PR #3338 behavior into the JS SDK by introducing a new first-class backend: `ContextHubBackend`. Why: - Context Hub repo persistence is now available as an SDK backend in JS, matching the Python backend architecture. - This gives JS users a canonical LangSmith Hub-backed file backend with the same operational semantics (lazy pull, commit chaining, cache invalidation, batch upload behavior). What changed: - Added `ContextHubBackend` in `libs/deepagents/src/backends/context-hub.ts`. - Exported it from backend and root package exports. - Added parity-focused unit tests in `context-hub.test.ts`. - Added integration contract tests in `context-hub.int.test.ts` (gated by `LANGSMITH_API_KEY`). - Updated `deepagents` peer dependency on `langsmith` from `>=0.5.20` to `>=0.6.0` to rely on Hub directory APIs (`pullAgent` / `pushAgent`). - Added a changeset for release notes. Behavior highlights: - Missing repo on pull is treated as empty and lazily created on first write. - In-memory file cache with linked-entry tracking (`getLinkedEntries`). - Commit hash tracking with optimistic `parentCommit` chaining. - Cache invalidation on mutating Hub failures. - `uploadFiles` commits all valid UTF-8 files in a single batch commit and returns per-file success/error. Validation run: - `pnpm install` - `pnpm exec oxfmt --write ...` (touched files) - `pnpm --filter deepagents exec oxlint ...` (touched files) - `pnpm --filter deepagents exec vitest run src/backends/context-hub.test.ts` (33 passed) - `pnpm --filter deepagents exec vitest run --mode int src/backends/context-hub.int.test.ts` (skipped without `LANGSMITH_API_KEY`) |
||
|
Colin Francis
|
7c33a8695f |
feat(deepagents): implement harness profiles (#526)
### Summary
Introduces harness profiles which is a way to declaratively control how
agents behave at runtime without changing which model is used. A profile
can override the system prompt, hide tools, add or remove middleware,
and configure the general-purpose subagent. Profiles are registered by
provider or model key ("anthropic", "openai:gpt-5.4") and looked up
automatically when `createDeepAgent` is called.
The system ships with built-in profiles for Anthropic Opus 4.7, Sonnet
4.6, Haiku 4.5, and OpenAI Codex that include vendor-recommended
prompting guidance. Users can register their own profiles on top and
registrations merge additively so provider-wide defaults compose with
model-specific overrides.
Profiles are plain frozen objects with Zod schemas for parsing from
JSON/YAML config files, including prototype-pollution protection for
untrusted input.
### Tests
Unit tests cover key validation, profile construction and freezing,
serialization round-trips, poison-key rejection, merge semantics across
all field types, registry lookup with provider fallback, and prompt
overlay behavior. Built-in profiles are tested for correct registration
and expected content.
Also verified end-to-end via a standalone script against LangSmith and
confirmed the custom suffix appears in the system prompt and excluded
tools are absent from the model call.
### Usage Example
```ts
import { HumanMessage } from "@langchain/core/messages";
import { createDeepAgent, registerHarnessProfile } from "deepagents";
registerHarnessProfile("anthropic:claude-sonnet-4-6", {
systemPromptSuffix: "Always respond in exactly one sentence.",
excludedTools: ["grep"],
generalPurposeSubagent: {
description: "A customized GP subagent via harness profile.",
},
});
const agent = createDeepAgent({
model: "anthropic:claude-sonnet-4-6",
});
const result = await agent.invoke({
messages: [new HumanMessage("What is the capital of France?")],
});
const lastMessage = result.messages[result.messages.length - 1];
console.log("Response:", lastMessage.content);
```
|
||
|
github-actions[bot]
|
c231aed3ee |
chore: version packages (#525)
This PR was opened by the [Changesets release](https://github.com/changesets/action) GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated. # Releases ## @langchain/quickjs@0.4.0 ### Minor Changes - [#531](https://github.com/langchain-ai/deepagentsjs/pull/531) [`a76b7df`](https://github.com/langchain-ai/deepagentsjs/commit/a76b7df62310e7f2dd49bb1ea5f1b3ee6c8590b6) Thanks [@colifran](https://github.com/colifran)! - chore(quickjs): update `REPLMiddleware` to be named `CodeInterpreterMiddleware` ### Patch Changes - [#524](https://github.com/langchain-ai/deepagentsjs/pull/524) [`2cbd524`](https://github.com/langchain-ai/deepagentsjs/commit/2cbd5245a43fb1ba97fa532c1942a8903e090cfa) Thanks [@colifran](https://github.com/colifran)! - fix(quickjs): individual repl sessions use individual wasm module causing inefficient memory usage ## deepagents-acp@0.1.11 ### Patch Changes - Updated dependencies \[[`f164f99`](https://github.com/langchain-ai/deepagentsjs/commit/f164f992e06a157573612fb2640232f44d9daa18)]: - deepagents@1.10.1 ## deepagents@1.10.1 ### Patch Changes - [#479](https://github.com/langchain-ai/deepagentsjs/pull/479) [`f164f99`](https://github.com/langchain-ai/deepagentsjs/commit/f164f992e06a157573612fb2640232f44d9daa18) Thanks [@ramon-langchain](https://github.com/ramon-langchain)! - feat(deepagents): add snapshot/start/stop lifecycle to LangSmithSandbox ## @deepagents/evals@0.0.10 ### Patch Changes - Updated dependencies \[[`f164f99`](https://github.com/langchain-ai/deepagentsjs/commit/f164f992e06a157573612fb2640232f44d9daa18)]: - deepagents@1.10.1 --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Colin Francis <colin.francis@langchain.dev> |
||
|
Colin Francis
|
a76b7df623 |
chore(quickjs): update REPLMiddleware to be named CodeInterpreterMiddleware (#531)
### Summary This PR updates `REPLMiddleware` to `CodeInterpreterMiddleware` |
||
|
Ramon Nogueira
|
f164f992e0 |
feat(deepagents): add snapshot/start/stop lifecycle to LangSmithSandbox (#479)
## Summary Expose the new LangSmith sandbox snapshot and lifecycle APIs (landed in `langsmith@0.5.20`) through the `LangSmithSandbox` wrapper. - New methods on `LangSmithSandbox`: `start()`, `stop()`, `captureSnapshot()` - `LangSmithSandbox.create()` now accepts `snapshotId` (preferred) or `templateName` (deprecated); throws if neither is provided - Re-export upstream types as `LangSmithSnapshot`, `LangSmithCaptureSnapshotOptions`, `LangSmithStartSandboxOptions` - Bump `langsmith` peer dep to `>=0.5.20` - Update the example to boot from `LANGSMITH_SANDBOX_SNAPSHOT_ID` ## Test plan - [x] Unit tests: added coverage for `start()`, `stop()`, `captureSnapshot()`; 41/41 pass - [x] Type-check clean - [x] Dev smoke test against `dev.api.smith.langchain.com` — create from snapshot, execute, stop, start, execute after restart all working --------- Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com> Co-authored-by: Ramon Nogueira <270434257+ramon-langchain@users.noreply.github.com> Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Colin Francis
|
2cbd5245a4 |
fix(quickjs): individual repl sessions use individual wasm module causing inefficient memory usage (#524)
### Summary
This PR optimizes memory usage by allowing us to use a single WASM
module. To do this we needed to fix two failure modes that occur when
multiple sessions share a single WASM module. Both are caused by the
async module loader creating asyncify suspensions (unwind/rewind cycles
on a module-global stack buffer) during import resolution.
__Failure 1__: Concurrent eval crash. The asyncify variant allows only
one async operation per module at a time. If two sessions eval
simultaneously and both hit an import, the second eval tries to suspend
asyncify while the first is already suspended. This crashes with
QuickJSAsyncifySuspended: Already suspended.
Fix: AsyncEvalQueue queues evalCodeAsync calls so only one runs at a
time.
__Failure 2__: Silent disposal corruption. When a session imports a
multi-file skill (e.g. an entry point that imports greet.js and
shout.js), each import causes an asyncify unwind/rewind cycle. The eval
succeeds. But when the session's runtime is disposed, the disposal does
not fully reset the module's asyncify stack buffer and the unwind/rewind
cycles left residual state (saved stack frames, rewind pointers) that
emscripten manages at the module level, not the runtime level. The next
session creates a fresh runtime on the same module, but the corrupted
asyncify state causes module loader callbacks to silently never fire.
Fix: synchronous module loader. Skill files are preloaded into a Map
before eval starts. The loader callback does a plain map.get(). No
await, no asyncify suspension, no residual state on disposal.
Note: The eval queue alone does not fix failure 2. The corruption
happens at runtime disposal, which runs outside the queue. Both fixes
are required.
#### What This PR Adds
- Replaces per-session WASM module instantiation with a process-global
singleton shared by all sessions. Each session still gets its own
runtime and context (full isolation); only the underlying WASM module is
shared. Reduces WASM memory from ~16 MB per session to ~16 MB total.
- Adds AsyncEvalQueue to serialize evalCodeAsync calls across sessions,
enforcing the asyncify one-concurrent-async-call-per-module constraint.
- Makes the module loader synchronous by preloading skill files into an
in-memory cache before eval (preloadReferencedSkills). Why is this
needed? An async module loader causes asyncify suspensions on each
import. Disposing a runtime after multi-file skill imports corrupts the
shared module's asyncify state and silently breaks the loader for all
subsequent sessions. The sync loader eliminates asyncify suspensions
from imports entirely.
#### Synchronous Module Loading
As stated above, synchronous module loading is a necessary change due to
async module loading causing asyncify suspensions on each import which
corrupts the singleton WASM module's asyncify state which silently
breaks the loader for all subsequent sessions. The following
demonstrates the before and after regarding this change:
__Before__:
User code: `import { run } from "@/skills/demo"`
1. eval() is called
2. evalCodeAsync() runs the code
3. QuickJS hits the import, calls the module loader callback
4. The callback is async: await ensureSkillLoaded("demo")
→ fetches /skills/demo/index.js from the backend (network I/O)
→ this await causes the first asyncify suspension in the WASM module
5. index.js has: import { greet } from "./greet.js"
6. QuickJS calls the loader again for greet.js
→ await ensureSkillLoaded() → network fetch
→ The second asyncify suspension in the WASM module is hit
7. index.js has: import { shout } from "./shout.js"
8. QuickJS calls the loader again for shout.js
→ await ensureSkillLoaded() → network fetch
→ third asyncify suspension in the WASM module is hit
9. eval completes, runtime is disposed
10. runtime.dispose() is called to clean up the session
11. The runtime disposal does NOT fully reset the module's asyncify
stack buffer. The 3 unwind/rewind cycles left residual state (saved
stack frames, rewind pointers, etc.) that the dispose path doesn't know
to clean up, because asyncify bookkeeping is managed by emscripten at
the module level, not by QuickJS at the runtime level.
12. Next session creates a fresh runtime on the same module, but the
module's asyncify machinery is corrupted. When QuickJS tries to call the
new session's module loader, the stale asyncify state causes the
callback to silently not fire. No crash, no error. Imports just resolve
to nothing.
__After__:
User code: `import { run } from "@/skills/demo"`
1. eval() is called
2. preloadReferencedSkills() runs BEFORE evalCodeAsync():
→ scans source code for "@/skills/*" specifiers
→ finds "demo"
→ fetches all 3 files from the backend into a Map:
"@/skills/demo" → source code
"@/skills/demo/greet.js" → source code
"@/skills/demo/shout.js" → source code
→ (this is async but happens OUTSIDE the WASM module so no asyncify)
3. evalCodeAsync() runs the code
4. QuickJS hits the import, calls the module loader callback
5. The callback is sync: map.get("@/skills/demo") → returns source
immediately
→ NO asyncify suspension
6. QuickJS resolves greet.js → map.get("@/skills/demo/greet.js") →
immediate
→ NO asyncify suspension
7. QuickJS resolves shout.js → map.get("@/skills/demo/shout.js") →
immediate
→ NO asyncify suspension
8. eval completes, runtime is disposed
9. Zero asyncify suspensions occurred → no residual state → module is
clean
10. Next session works perfectly
### Tests
- 4 new AsyncEvalQueue unit tests: serialization (no interleaving of
concurrent ops), return values, error isolation (rejected op doesn't
break queue), insertion order preservation
- All 160 existing tests pass (session, skills, middleware, transform,
utils, eval queue)
### Memory Benchmarking
#### Per-session module (before)
| Sessions | RSS Delta (MB) | Per-Session RSS (MB) | WASM Memory (MB) |
Module Init (ms) | Setup (ms) |
|----------|----------------|----------------------|------------------|------------------|------------|
| 1 | 0.53 | 0.53 | 16 | 11.4 | 15.9 |
| 5 | 2.13 | 0.43 | 80 | 30.2 | 37.9 |
| 10 | 5.28 | 0.53 | 160 | 31.1 | 42.8 |
| 15 | 5.53 | 0.37 | 240 | 57.4 | 72.1 |
| 20 | 6.80 | 0.34 | 320 | 73.6 | 91.1 |
---
#### Singleton (after)
| Sessions | RSS Delta (MB) | Per-Session RSS (MB) | WASM Memory (MB) |
Module Init (ms) | Setup (ms) |
|----------|----------------|----------------------|------------------|------------------|------------|
| 1 | 0.61 | 0.61 | 16 | 6.7 | 9.1 |
| 5 | 0.59 | 0.12 | 16 | 12.5 | 28.5 |
| 10 | 0.91 | 0.09 | 16 | 12.2 | 37.0 |
| 15 | 1.81 | 0.12 | 16 | 12.0 | 45.1 |
| 20 | 1.70 | 0.09 | 16 | 10.4 | 50.4 |
---
#### Comparison at 20 sessions (realistic workload)
| Metric | Before | After | Improvement |
|------------------------|---------|--------|-----------------|
| WASM virtual memory | 320 MB | 16 MB | 20x reduction |
| RSS delta | 6.80 MB | 1.70 MB| 4x reduction |
| Module instantiation | 73.6 ms | 10.4 ms| 7x reduction |
| Setup time | 91.1 ms | 50.4 ms| 1.8x reduction |
|
||
|
github-actions[bot]
|
6da7007e52 |
chore: version packages (#521)
This PR was opened by the [Changesets release](https://github.com/changesets/action) GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated. # Releases ## deepagents@1.10.0 ### Minor Changes - [#458](https://github.com/langchain-ai/deepagentsjs/pull/458) [`b1e1b7b`](https://github.com/langchain-ai/deepagentsjs/commit/b1e1b7bd3bcc3bd5b03dd461e72559ed69c77e22) Thanks [@christian-bromann](https://github.com/christian-bromann)! - feat(deepagents): new events streaming ## @langchain/quickjs@0.3.0 ### Minor Changes - [#520](https://github.com/langchain-ai/deepagentsjs/pull/520) [`2548954`](https://github.com/langchain-ai/deepagentsjs/commit/254895457fce2de7df4547ebcf825d60868bacce) Thanks [@colifran](https://github.com/colifran)! - chore(quickjs): rename to REPLMiddleware and adjust defaults ## deepagents-acp@0.1.10 ### Patch Changes - Updated dependencies \[[`b1e1b7b`](https://github.com/langchain-ai/deepagentsjs/commit/b1e1b7bd3bcc3bd5b03dd461e72559ed69c77e22)]: - deepagents@1.10.0 ## @deepagents/evals@0.0.9 ### Patch Changes - Updated dependencies \[[`b1e1b7b`](https://github.com/langchain-ai/deepagentsjs/commit/b1e1b7bd3bcc3bd5b03dd461e72559ed69c77e22)]: - deepagents@1.10.0 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> |
||
|
Christian Bromann
|
b1e1b7bd3b |
feat(deepagents): support new streaming features (currently experimental in LG.js) (#458)
This patch adopts the new streaming primitives we have implemented in LangGraph.js for Deep Agent. LangGraph.js PR: https://github.com/langchain-ai/langgraphjs/pull/2312 LangChain.js PR: https://github.com/langchain-ai/langchainjs/pull/10697 |
||
|
Colin Francis
|
254895457f |
chore(quickjs): rename to REPLMiddleware and adjust defaults (#520)
### Summary - Rename `createQuickJSMiddleware` → `createREPLMiddleware` and `QuickJSMiddlewareOptions` → `REPLMiddlewareOptions` to match the Python `REPLMiddleware` naming - Rename middleware registration from `"QuickJSMiddleware"` → `"REPLMiddleware"` - Rename default tool name from `"js_eval"` → `"eval"`, now configurable via `toolName` option - Add `captureConsole` option (default `true`) to toggle console output capture - Align default memory limit to 64 MiB (was 50 MiB) and timeout to 5s (was 30s) to match Python ### Tests - All 156 existing unit tests pass with updated references - Updated `middleware.test.ts`, `middleware.int.test.ts`, and example files to use new names --------- Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
github-actions[bot]
|
9a8ff1941a |
chore: version packages (#444)
This PR was opened by the [Changesets release](https://github.com/changesets/action) GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated. # Releases ## deepagents-acp@0.1.9 ### Patch Changes - [#501](https://github.com/langchain-ai/deepagentsjs/pull/501) [`5b0eaea`](https://github.com/langchain-ai/deepagentsjs/commit/5b0eaea7b20461414983b71ba08d26d078b49214) Thanks [@sukhmanghotraa](https://github.com/sukhmanghotraa)! - fix: bump @langchain/core to ^1.1.42 across all workspace packages - Updated dependencies \[[`5b0eaea`](https://github.com/langchain-ai/deepagentsjs/commit/5b0eaea7b20461414983b71ba08d26d078b49214), [`e90171a`](https://github.com/langchain-ai/deepagentsjs/commit/e90171abe4bcc76767246be470a7b17b94692f41), [`8fd575f`](https://github.com/langchain-ai/deepagentsjs/commit/8fd575f06ca27cb0bef1a649aa34124a2c04ddd3), [`3657941`](https://github.com/langchain-ai/deepagentsjs/commit/3657941ea36b21b9b512c1eb68a250ae79124383), [`998d772`](https://github.com/langchain-ai/deepagentsjs/commit/998d772a07acc76fcc0d419e65b3c74a64d9ac52), [`55f3bd8`](https://github.com/langchain-ai/deepagentsjs/commit/55f3bd8d74cac22d124fd6d1b11538dc2c2c2aec), [`79e20e1`](https://github.com/langchain-ai/deepagentsjs/commit/79e20e18082a19b65094b953cd857908a7525801), [`2442d7d`](https://github.com/langchain-ai/deepagentsjs/commit/2442d7d080c8a1008197eda526de52400303dd72), [`43cd121`](https://github.com/langchain-ai/deepagentsjs/commit/43cd121133562abf0dee76c6db01f2bde0eb3fd3), [`2994444`](https://github.com/langchain-ai/deepagentsjs/commit/2994444f32a6c0503defa6157652e742361abb00)]: - deepagents@1.9.1 ## deepagents@1.9.1 ### Patch Changes - [#501](https://github.com/langchain-ai/deepagentsjs/pull/501) [`5b0eaea`](https://github.com/langchain-ai/deepagentsjs/commit/5b0eaea7b20461414983b71ba08d26d078b49214) Thanks [@sukhmanghotraa](https://github.com/sukhmanghotraa)! - fix: bump @langchain/core to ^1.1.42 across all workspace packages - [#442](https://github.com/langchain-ai/deepagentsjs/pull/442) [`e90171a`](https://github.com/langchain-ai/deepagentsjs/commit/e90171abe4bcc76767246be470a7b17b94692f41) Thanks [@christian-bromann](https://github.com/christian-bromann)! - fix(deepagents): closer align store backend with Python - [#496](https://github.com/langchain-ai/deepagentsjs/pull/496) [`8fd575f`](https://github.com/langchain-ai/deepagentsjs/commit/8fd575f06ca27cb0bef1a649aa34124a2c04ddd3) Thanks [@colifran](https://github.com/colifran)! - feat(deepagents): implement functional skills for quickjs middleware - [#448](https://github.com/langchain-ai/deepagentsjs/pull/448) [`3657941`](https://github.com/langchain-ai/deepagentsjs/commit/3657941ea36b21b9b512c1eb68a250ae79124383) Thanks [@ItayCoCo](https://github.com/ItayCoCo)! - fix: follow symlinks in sandbox find commands by adding -L flag to find invocations in buildLsCommand, buildFindCommand, and buildGrepCommand - [#486](https://github.com/langchain-ai/deepagentsjs/pull/486) [`998d772`](https://github.com/langchain-ai/deepagentsjs/commit/998d772a07acc76fcc0d419e65b3c74a64d9ac52) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): remove built-in VFS globals, add PTC instance injection and StateBackend read-your-writes - [#470](https://github.com/langchain-ai/deepagentsjs/pull/470) [`55f3bd8`](https://github.com/langchain-ai/deepagentsjs/commit/55f3bd8d74cac22d124fd6d1b11538dc2c2c2aec) Thanks [@jacoblee93](https://github.com/jacoblee93)! - Adds agent type metadata prop to configurable - [#451](https://github.com/langchain-ai/deepagentsjs/pull/451) [`79e20e1`](https://github.com/langchain-ai/deepagentsjs/commit/79e20e18082a19b65094b953cd857908a7525801) Thanks [@JadenKim-dev](https://github.com/JadenKim-dev)! - remove unconditional @langchain/anthropic import - [#465](https://github.com/langchain-ai/deepagentsjs/pull/465) [`2442d7d`](https://github.com/langchain-ai/deepagentsjs/commit/2442d7d080c8a1008197eda526de52400303dd72) Thanks [@hntrl](https://github.com/hntrl)! - fix(deepagents): route summarization through active request model - [#492](https://github.com/langchain-ai/deepagentsjs/pull/492) [`43cd121`](https://github.com/langchain-ai/deepagentsjs/commit/43cd121133562abf0dee76c6db01f2bde0eb3fd3) Thanks [@colifran](https://github.com/colifran)! - implement file system permissions for fs middleware tools - [#459](https://github.com/langchain-ai/deepagentsjs/pull/459) [`2994444`](https://github.com/langchain-ai/deepagentsjs/commit/2994444f32a6c0503defa6157652e742361abb00) Thanks [@open-swe](https://github.com/apps/open-swe)! - fix(deepagents): skill loading should default to 1000 lines ## @langchain/quickjs@0.2.6 ### Patch Changes - [#501](https://github.com/langchain-ai/deepagentsjs/pull/501) [`5b0eaea`](https://github.com/langchain-ai/deepagentsjs/commit/5b0eaea7b20461414983b71ba08d26d078b49214) Thanks [@sukhmanghotraa](https://github.com/sukhmanghotraa)! - fix: bump @langchain/core to ^1.1.42 across all workspace packages - [#496](https://github.com/langchain-ai/deepagentsjs/pull/496) [`8fd575f`](https://github.com/langchain-ai/deepagentsjs/commit/8fd575f06ca27cb0bef1a649aa34124a2c04ddd3) Thanks [@colifran](https://github.com/colifran)! - feat(deepagents): implement functional skills for quickjs middleware - [#486](https://github.com/langchain-ai/deepagentsjs/pull/486) [`998d772`](https://github.com/langchain-ai/deepagentsjs/commit/998d772a07acc76fcc0d419e65b3c74a64d9ac52) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): remove built-in VFS globals, add PTC instance injection and StateBackend read-your-writes - [#498](https://github.com/langchain-ai/deepagentsjs/pull/498) [`192bbd7`](https://github.com/langchain-ai/deepagentsjs/commit/192bbd7109817a826aa6bd982e16b54b5a5c8cb5) Thanks [@colifran](https://github.com/colifran)! - fix(quickjs): bound console buffering at capture time - [#497](https://github.com/langchain-ai/deepagentsjs/pull/497) [`79dcf9a`](https://github.com/langchain-ai/deepagentsjs/commit/79dcf9ae4567fcce71ea7e529f332baf6a3d6dcd) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): add maxPtcCallsbudget for ptc calls - [#492](https://github.com/langchain-ai/deepagentsjs/pull/492) [`43cd121`](https://github.com/langchain-ai/deepagentsjs/commit/43cd121133562abf0dee76c6db01f2bde0eb3fd3) Thanks [@colifran](https://github.com/colifran)! - implement file system permissions for fs middleware tools - [#494](https://github.com/langchain-ai/deepagentsjs/pull/494) [`e64c7e3`](https://github.com/langchain-ai/deepagentsjs/commit/e64c7e31a7e62929d5d2a9048ad7df3315a546cd) Thanks [@colifran](https://github.com/colifran)! - feat(quickjs): implement delete session in after agent hook to clean up completed repl sessions - [#499](https://github.com/langchain-ai/deepagentsjs/pull/499) [`352e487`](https://github.com/langchain-ai/deepagentsjs/commit/352e4876679fb708d9690ffeab82363ba02677dc) Thanks [@hntrl](https://github.com/hntrl)! - fix(quickjs): add ls_code_input_language metadata to js_eval tool ## @deepagents/evals@0.0.8 ### Patch Changes - Updated dependencies \[[`5b0eaea`](https://github.com/langchain-ai/deepagentsjs/commit/5b0eaea7b20461414983b71ba08d26d078b49214), [`e90171a`](https://github.com/langchain-ai/deepagentsjs/commit/e90171abe4bcc76767246be470a7b17b94692f41), [`8fd575f`](https://github.com/langchain-ai/deepagentsjs/commit/8fd575f06ca27cb0bef1a649aa34124a2c04ddd3), [`3657941`](https://github.com/langchain-ai/deepagentsjs/commit/3657941ea36b21b9b512c1eb68a250ae79124383), [`998d772`](https://github.com/langchain-ai/deepagentsjs/commit/998d772a07acc76fcc0d419e65b3c74a64d9ac52), [`55f3bd8`](https://github.com/langchain-ai/deepagentsjs/commit/55f3bd8d74cac22d124fd6d1b11538dc2c2c2aec), [`79e20e1`](https://github.com/langchain-ai/deepagentsjs/commit/79e20e18082a19b65094b953cd857908a7525801), [`2442d7d`](https://github.com/langchain-ai/deepagentsjs/commit/2442d7d080c8a1008197eda526de52400303dd72), [`43cd121`](https://github.com/langchain-ai/deepagentsjs/commit/43cd121133562abf0dee76c6db01f2bde0eb3fd3), [`2994444`](https://github.com/langchain-ai/deepagentsjs/commit/2994444f32a6c0503defa6157652e742361abb00)]: - deepagents@1.9.1 --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Hunter Lovell <hunter@hntrl.io> |
||
|
sukh
|
5b0eaea7b2 |
fix: bump @langchain/core to ^1.1.42 across all workspace packages (#501)
## Problem
When using `@langchain/google` as a model provider in the examples
workspace, the server crashes immediately with:
1. Runtime crash on startup:
```bash
ERR_PACKAGE_PATH_NOT_EXPORTED: Package subpath './utils/uuid' is not
defined by "exports" in @langchain/core/package.json
```
`@langchain/google` imports `@langchain/core/utils/uuid`, a subpath
added in `@langchain/core@1.1.42`. Both the examples workspace and
`libs/deepagents` were pinned to `^1.1.40` which does not export this
path.
2. TypeScript type error after fixing the runtime crash:
```bash
Type 'ChatGoogle' is not assignable to type 'BaseLanguageModel<any,
BaseLanguageModelCallOptions>'
```
Even after bumping `@langchain/core` in `examples/` only, pnpm kept two
separate copies in its store - `1.1.40` for `deepagents` and `1.1.42`
for `@langchain/google`. TypeScript treats them as structurally
incompatible types because the type import paths differ.
## Fix
- Bump `@langchain/core` from `^1.1.40` → `^1.1.42` in all
`../package.json` files
## Testing
```bash
cd examples/research
pnpm dlx @langchain/langgraph-cli dev
# -> No ERR_PACKAGE_PATH_NOT_EXPORTED and agent starts correctly
---------
Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com>
|
||
|
Hunter Lovell
|
352e487667 |
fix(quickjs): add ls_code_input_language metadata to js_eval tool (#499)
## Summary
Adds LangSmith code-language metadata to the QuickJS REPL evaluation
tool so tool inputs are correctly interpreted as JavaScript in traces.
## Changes
### @langchain/quickjs
- Added `metadata: { ls_code_input_language: "javascript" }` to the
`js_eval` tool definition in `createQuickJSMiddleware`.
- Extended middleware tool-registration tests to assert that `js_eval`
exposes `ls_code_input_language` metadata.
- Verified behavior with quickjs unit tests (`src/middleware.test.ts`).
|
||
|
Colin Francis
|
192bbd7109 |
fix(quickjs): bound console buffering at capture time (#498)
Port of https://github.com/langchain-ai/deepagents/pull/2999 |
||
|
Colin Francis
|
79dcf9ae45 |
feat(quickjs): add maxPtcCallsbudget for ptc calls (#497)
Port of https://github.com/langchain-ai/deepagents/pull/2994 |
||
|
Colin Francis
|
8fd575f06c |
feat(deepagents): implement functional skills for quickjs middleware (#496)
### Summary
Adds functional skill modules to the QuickJS REPL. Skills can now ship
JS/TS entrypoints via a module: frontmatter key on SKILL.md, which the
model dynamically imports at eval time with await
import("@/skills/<name>").
When skillsBackend is passed to createQuickJSMiddleware, the middleware
reads skillsMetadata from the current task state before each eval, scans
the source for skill references, and pushes the metadata + backend into
the session. The session wires a QuickJS module loader that lazily
fetches skill files from the backend, strips TypeScript syntax, and
caches the result for the lifetime of the session. Skills that are
referenced but not available on the agent short-circuit with a clear
error before evaluation starts.
On the deepagents side, SkillMetadata gains an optional module field
(validated and normalized from frontmatter), and formatSkillsList now
advertises importable skills in the system prompt with an await
import(...) hint.
```ts
const backend = new FilesystemBackend({ rootDir: "/project", virtualMode: true });
const agent = createDeepAgent({
model: "claude-sonnet-4-6",
backend,
skills: ["/skills/"],
middleware: [
createQuickJSMiddleware({ skillsBackend: backend }),
],
});
```
### Tests
- Unit tests cover module path validation, TS syntax stripping, skill
loading (single/multi-file, error cases, bundle size cap), skill
reference scanning, module loader resolution and caching, and system
prompt formatting
- Ran an E2E eval that exercised the full pipeline from SKILL.md parse
through dynamic import and function call in the REPL, with a LangSmith
trace
### Benchmarks
#### memory: console_log concurrent
- Total time: 15820ms
| name | hz | min | max | mean | p75 | p99 | p995 | p999 | rme | samples
|
|--------------------------|---------|--------|--------|--------|--------|--------|--------|--------|--------|---------|
| 1 concurrent sessions | 145.91 | 1.8263 | 27.7702| 6.8536 | 9.1753 |
21.1492| 25.0420| 27.7702| ±5.30% | 730 |
| 8 concurrent sessions | 39.1773 | 15.5854| 61.5015| 25.5250| 27.8461|
49.0536| 61.5015| 61.5015| ±3.58% | 196 |
| 32 concurrent sessions | 13.9753 | 48.5504| 110.40 | 71.5547| 76.6394|
110.40 | 110.40 | 110.40 | ±3.72% | 70 |
---
#### memory: ptc_tools concurrent
- Total time: 16755ms
| name | hz | min | max | mean | p75 | p99 | p995 | p999 | rme | samples
|
|--------------------------|---------|--------|--------|--------|--------|--------|--------|--------|--------|---------|
| 1 concurrent sessions | 86.9205 | 4.1809 | 32.1833| 11.5048| 15.3222|
25.2676| 28.2697| 32.1833| ±4.76% | 435 |
| 8 concurrent sessions | 21.8310 | 33.5598| 69.0041| 45.8065| 50.8740|
60.2001| 69.0041| 69.0041| ±2.77% | 110 |
| 32 concurrent sessions | 7.0269 | 137.88 | 148.90 | 142.31 | 144.05 |
148.90 | 148.90 | 148.90 | ±0.70% | 36 |
---
#### throughput: single thread many iterations
- Total time: 31656ms
| name | hz | min | max | mean | p75 | p99 | p995 | p999 | rme | samples
|
|-------------------------------------------|--------|--------|--------|--------|--------|--------|--------|--------|--------|---------|
| 200 sequential evals (ptc + console.log) | 3.5897 | 240.92 | 347.19 |
278.58 | 289.81 | 341.02 | 347.19 | 347.19 | ±1.44% | 108 |
|
||
|
Colin Francis
|
e64c7e31a7 |
feat(quickjs): implement delete session in after agent hook to clean up completed repl sessions (#494)
### Summary ReplSession uses a static map to preserve QuickJS WASM runtime state across js_eval calls within a run. Previously sessions were never cleaned up, causing unbounded memory growth as each thread_id accumulated a live WASM runtime for the lifetime of the process. - Added ReplSession.deleteSession(key) to dispose of the QuickJS runtime and removes the entry from the session map - Added an afterAgent hook in createQuickJSMiddleware that calls deleteSession for the current thread's session key at the end of each agent run Sessions still persist across all js_eval calls within a single run. The map remains for correct concurrent thread isolation. Without it, simultaneous runs from different threads would share a runtime. ### Tests Unit tests covering the following scenarios: - ReplSession.deleteSession disposes and removes an existing session - ReplSession.deleteSession is a no-op for a key that does not exist - afterAgent disposes the session for the finished thread - afterAgent on a thread with no session does not throw - afterAgent only removes the session for the finished thread, leaving other active threads intact |
||
|
Colin Francis
|
998d772a07 |
feat(quickjs): remove built-in VFS globals, add PTC instance injection and StateBackend read-your-writes (#486)
### Summary - Removed built-in readFile/writeFile globals from the QuickJS REPL. File I/O is now exclusively through PTC tools, keeping the REPL a pure JS sandbox with no implicit side effects - ptc now accepts StructuredToolInterface instances directly alongside tool name strings. Custom tools can be injected into the REPL without being registered on the agent - StateBackend now provides read-your-writes semantics within a single js_eval superstep by reading via __pregel_read with fresh=true. Pregel already tracks pending sends in task.writes; asking for a fresh read applies them through the reducer, removing the need for a manual write buffer - Removed ptc: boolean shorthand from QuickJSMiddlewareOptions; use the array or object forms instead ### Tests - Added 12 unit tests to state.test.ts covering write-then-read, chained edits, ls/grep/glob/readRaw visibility, upload/download, duplicate write error, edit-nonexistent, and superstep boundary behavior (commitSends/clearPending) - Expanded middleware.test.ts with PTC instance injection coverage - Verified end-to-end with createDeepAgent + createQuickJSMiddleware across 7 scenarios (write-read, write-edit-read, write-ls, write-grep, write-glob, cross-eval, upload-download); traces confirmed in LangSmith --------- Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Colin Francis
|
43cd121133 |
feat(deepagents): implement file system permissions for fs middleware tools (#492)
### Summary - Adds a FilesystemPermission class and a permissions parameter to createDeepAgent and SubAgent that controls which filesystem operations each agent is allowed to perform - Rules are evaluated first-match-wins with a permissive default; mode: "deny" blocks the operation before any backend call - Permissions are closed over at tool-creation time — no runtime config threading, no ToolPolicy in langchain-core - ls, glob, and grep enforce permissions in two passes: the base path check throws early; results are post-filtered to strip entries the caller can't read - Subagents receive the parent agent's permissions by default; setting permissions on a SubAgent is a full replacement, not a merge - FilesystemPermission, FilesystemPermissionOptions, FilesystemOperation, and PermissionMode are exported from the top-level package entry point ### Tests - permissions/types.test.ts — FilesystemPermission construction, validation (rejects relative paths, .., ~), defaults - permissions/enforce.test.ts — validatePath, globMatch (wildcards, brace expansion, dotfiles), decidePathAccess (first-match-wins, operation mismatch, multiple ops/paths) - middleware/fs.permissions.test.ts — each fs tool (read, write, edit, ls, glob, grep) with allow/deny rules and mock backends; verifies backend is never called on a denied path; verifies execute is unaffected - permissions/agent.permissions.test.ts — CreateDeepAgentParams.permissions type and default; wiring to createFilesystemMiddleware - middleware/subagents.permissions.test.ts — SubAgent.permissions type; ?? resolution (inherit vs override vs own deny); createFilesystemMiddleware behavior for each resolved-permissions case --------- Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Jacob Lee
|
55f3bd8d74 |
feat(deepagents): Adds agent type metadata prop to configurable (#470)
Will be useful for trace display features Python: https://github.com/langchain-ai/deepagents/pull/2788 --------- Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Hunter Lovell
|
2442d7d080 |
fix(deepagents): align summarization with active request model (#465)
## Summary Fixes #398 Align deepagents summarization with the active request-time model path instead of relying on a separate middleware-specific model resolution path. This keeps summarization behavior consistent with the main agent model selection flow and avoids divergence in provider/model handling during middleware execution. The update also simplifies runtime behavior by removing unnecessary explicit config threading for summary calls. ## Changes ### `deepagents` middleware/model routing - Updated summarization middleware to prefer `request.model` as the resolved model used for profile-derived limits and summary generation. - Kept fallback model resolution support for cases where a request model is not present. - Removed explicit runtime config plumb-through in summary invocation, relying on LangChain runnable context propagation. ### `deepagents` agent wiring - Updated `createDeepAgent` built-in and subagent middleware wiring to construct summarization middleware without explicitly passing a separate model option, so summarization follows the active request model path. ### `deepagents` tests - Added/updated summarization tests to cover request-model-based summarization behavior and ensure middleware behavior remains stable with the new routing path. - Simplified test request construction to keep request-model assumptions explicit. |
||
|
ItayCoCo
|
3657941ea3 |
fix: follow symlinks in sandbox find commands (#448)
## Summary - Add `-L` flag to `find` in `buildLsCommand`, `buildFindCommand`, and `buildGrepCommand` so that symlinks are followed - Fixes #447 ## Problem `BaseSandbox` helper functions use `find` without `-L`, which means symbolic links are not dereferenced. On Modal (and potentially other providers), volumes are mounted as symlinks, so the `ls`, `glob`, and `grep` sandbox tools all return empty results when pointed at symlinked paths. ## Fix The `-L` flag tells `find` to follow symbolic links. This is a one-line change in each of the three helper functions: ```diff - find ${quotedPath} -maxdepth 1 ... + find -L ${quotedPath} -maxdepth 1 ... ``` This matches user expectations — filesystem exploration tools should transparently follow symlinks, just like `ls` and `find -L` do in a normal shell. ## Test plan - [ ] Verify `ls` tool returns files when the target path is a symlink - [ ] Verify `glob` tool traverses symlinked directories - [ ] Verify `grep` tool searches inside symlinked directories - [ ] Confirm no regression on non-symlinked paths (behavior is identical when no symlinks are present) --------- Co-authored-by: Itay Cohen <itaycohen@Itays-MacBook-Pro.local> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
open-swe[bot]
|
2994444f32 |
fix: skill loading should default to 1000 lines (#459)
## Description Ports the skill loading fix from the Python SDK (langchain-ai/deepagents#2721) to JS. The system prompt now instructs agents to pass `limit=1000` when reading skill files via `read_file`, since the default of 100 lines is too small for most skill files. ## Test Plan - [ ] Verify skill loading prompt includes `limit=1000` instruction _Opened collaboratively by Sydney Runkle and open-swe._ --------- Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com> Co-authored-by: Sydney Runkle <54324534+sydney-runkle@users.noreply.github.com> Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Youngho Kim
|
79e20e1808 |
fix(deepagents): remove unconditional @langchain/anthropic import (#451)
## Summary Fixes #443 - Removed static `import { ChatAnthropic } from "@langchain/anthropic"` at module level - Changed default `model` parameter from `new ChatAnthropic("claude-sonnet-4-6")` to the string `"claude-sonnet-4-6"` ## Problem `deepagents@1.9.0` unconditionally imported `@langchain/anthropic` at the top of `agent.ts`. Bundlers follow static imports regardless of which model the user actually uses, causing builds to fail for users who only use other providers (e.g. `ChatOpenAI`) and don't have `@langchain/anthropic` installed. ## Solution `createAgent` from `langchain` natively accepts a model name string (`string | AgentLanguageModelLike`), so no `ChatAnthropic` instantiation is needed at the library level. With a string default, bundlers have no static reference to `@langchain/anthropic` and can tree-shake it entirely. Users who want Anthropic models can either: - Pass `new ChatAnthropic(...)` explicitly (requires `@langchain/anthropic` in their own `package.json`) - Rely on the `"claude-sonnet-4-6"` string default (also requires `@langchain/anthropic` at runtime, but no longer breaks builds for non-Anthropic users) --------- Co-authored-by: Hunter Lovell <40191806+hntrl@users.noreply.github.com> |
||
|
Christian Bromann
|
e90171abe4 |
fix(deepagents): closer align store backend with Python (#442)
Aligned `StoreBackend` in `libs/deepagents/src/backends/store.ts` more closely with the Python version while keeping the current TS behavior working: - `namespace` now supports either a static `string[]` or a dynamic factory, so callers can derive store namespaces from runtime state/config instead of hardcoding them. - Legacy namespace fallback now checks config metadata for `assistant_id` / `assistantId` first, then falls back to the existing injected `assistantId`, and finally to `["filesystem"]`. - Added exported types for the new namespace factory context in `libs/deepagents/src/backends/index.ts` and `libs/deepagents/src/index.ts`. |
||
|
github-actions[bot]
|
02c1c7fa38 |
chore: version packages (#433)
* chore: version packages * cr --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Hunter Lovell <hunter@hntrl.io> |