## Summary
Replaces the `mode` field on `SubAgent` with a new `ForkedSubAgent` type
for opt-in conversation forking. A `ForkedSubAgent` has no
`systemPrompt` of its own — it always inherits the parent's message
history and exact system prompt, plus mirrored middleware when its model
matches the parent's. `CompiledSubAgent` keeps its existing `mode:
"handoff" | "fork"` field, unchanged.
## Changes
- New `ForkedSubAgent` type — same shape as `SubAgent` minus
`systemPrompt`; always forks.
- `SubAgent` no longer has a `mode` field — always isolated (same as
before this feature existed).
- `CompiledSubAgent.mode` unchanged.
- Mirrored middleware (memory/custom) only added when the model matches
the parent's; the system prompt itself is inherited unconditionally
either way.
- Updated fork test coverage in `subagent.test.ts` for the new type.
---------
Co-authored-by: Hunter Lovell <hunter@hntrl.io>
## Description
When filesystem `permissions` are combined with an execution-capable
backend, `createFilesystemMiddleware` requires every permission path to
be scoped to a `CompositeBackend` route. That check required a trailing
separator, so a permission path that exactly equals a single-file mount
(route `/instructions.md`, permission `/instructions.md`) was rejected
and users had to write `/instructions.md/**`. Python deepagents accepts
the exact form; this syncs JS with it by also accepting the route root
itself, mirroring `CompositeBackend.isPathWithinRoute`. It deliberately
does not adopt Python's looser bare `startsWith`, so sibling prefixes
like `/workspace2/**` for route `/workspace` are still rejected.
## Test Plan
- [ ] `npx vitest run src/middleware/fs.permissions.test.ts` in
`libs/deepagents` (new cases: exact file-mount route and exact directory
route root are accepted; sibling-prefix and `/**` rejections unchanged).
Made by [Open
SWE](https://openswe.vercel.app/agents/2758c734-16e9-54c3-b89f-ce698a566e71)
---------
Co-authored-by: Caspar Broekhuizen <25157475+casparb@users.noreply.github.com>
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Hunter Lovell <hunter@hntrl.io>
## Summary
Recursive glob listed every path under the search root, and with -L
could loop through proc pid root symlinks into the host heap.
- Sandbox `glob` listed the whole search tree via `find -L` and buffered
it in the host process, so `glob("**/…", "/")` could OOM the runtime.
Prune virtual filesystems and soft-cap find output; mark `truncated`
when the cap is hit.
## Summary
This adds a destructive `delete` filesystem tool backed by the core backend protocol, with recursive deletion for files/directories and write-permission enforcement over the whole target subtree.
## Changes
### Core filesystem backends
- Added `DeleteResult` and optional `delete(filePath)` support to the
backend protocol and adapters.
- Implemented recursive delete behavior for state, store, filesystem,
sandbox, context-hub, and composite backends.
- Preserved filesystem path containment and symlink safety: filesystem
deletes remove symlinks as links and do not follow them.
- Converts unsupported routed deletes in `CompositeBackend` into normal
backend errors.
### Filesystem middleware
- Added the final `delete` tool name with recursive file/directory
semantics.
- Treats `delete` as a write operation for filesystem permissions.
- Refuses recursive deletes if any deny-write permission pattern
overlaps the target subtree, without enumerating filesystem contents
first.
- Filters the `delete` tool out when the resolved backend does not
support deletion.
### Tests, evals, and prompts
- Added backend and middleware coverage for recursive deletes, missing
paths, symlink safety, unsupported backends, and permission denial.
- Added file-operation eval coverage for deleting a file and deleting a
directory recursively.
- Updated core filesystem prompts and sandbox examples to document the
new `delete` tool.
Fixes#745
`max_count` was the only numeric grep param without type coercion,
causing ToolInputParsingException on stringified input. Switched to
z.coerce.number().
Fixes#748
Excludes `_summarizationEvent`/`_summarizationSessionId` from subagent
state input/output. These weren't in `EXCLUDED_STATE_KEYS`, so a stale
parent cutoffIndex could silently drop a subagent's task description,
and a subagent's own summarization event could overwrite the parent's on
return.
ContextHubBackend now preserves concurrent file mutations in one durable
commit and retains compatible remote changes during conflict recovery.
---
Concurrent callers previously reused the same parent commit, causing
409s. Retries could also replay stale edit payloads.
Mutations now coalesce for 50 ms into serialized multi-file commits.
Conflict recovery replays ordered edit intent against refreshed state;
explicit writes, uploads, and deletes remain last-writer-wins.
---------
Co-authored-by: Hunter Lovell <hunter@hntrl.io>
## Description
`SummarizationMiddleware`'s `createSummary` falls back to
`JSON.stringify(response.content)` when the summarization model's
response content is not a string:
```ts
return typeof response.content === "string"
? response.content
: JSON.stringify(response.content);
```
When the summarization model responds with a content block array, the
raw blocks are stringified verbatim into the summary `HumanMessage`.
This happens in practice when:
- **Thinking/reasoning is enabled** on the summarization model (e.g.
Claude on Bedrock with `thinking`): the content is `[{type: "reasoning",
reasoning: ..., signature: <opaque multi-KB signature>}, {type: "text",
...}]`, and the reasoning signature lands in the summary.
- **`invoke()` takes the internal streaming path** (a `prefersStreaming`
callback handler is attached, e.g. inside `streamEvents`, and
`disableStreaming` is not set): the aggregated `AIMessageChunk` content
is an array of many small text fragments.
Observed in production: a summary message of ~56KB consisting of an ~8KB
opaque reasoning signature plus hundreds of
`{"type":"text","text":"..."}` fragments — the "summary" ended up
bloating the context it was supposed to compact, and subsequent model
calls overflowed.
## Fix
Use the message's `text` accessor for non-string content, which extracts
and joins only the text blocks. String content is returned as-is
(unchanged behavior).
## Testing
Added a unit test where the summarization model returns reasoning +
split text blocks, asserting the summary message contains the joined
text and no reasoning payload. All 33 existing tests pass.
---------
Co-authored-by: Hunter Lovell <hunter@hntrl.io>
## Summary
- The grep tool's description mentioned output_mode, but it was never an
actual parameter the tool always returned one hardcoded format.
- Added output_mode (files_with_matches/content/count) as a real schema
parameter, using the existing formatGrepMatches helper.
## Test plan
- Added tests in fs.test.ts for files_with_matches and count modes.
- All tests pass.
Note: default output is now sorted by path with no blank lines between
file groups (minor formatting change, same info).
---------
Co-authored-by: Hunter Lovell <hunter@hntrl.io>
## Summary
Prevent default DeepAgents JS summarization from discarding an oversized
message before it reaches the summary model
## Changes
### `deepagents`
- Make `trimTokensToSummarize` opt-in rather than defaulting to 4,000
tokens.
- Preserve existing capped behavior when callers explicitly configure
the option.
- Add coverage confirming oversized summary input reaches the summary
model under default configuration.
- Add a patch changeset for the bug fix.
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@0.1.23
### Patch Changes
- Updated dependencies
[[`590c2a5`](https://github.com/langchain-ai/deepagentsjs/commit/590c2a5042473f096d5fac5ddbb4be96e2ace0f2)]:
- deepagents@1.12.2
## deepagents@1.12.2
### Patch Changes
- [#723](https://github.com/langchain-ai/deepagentsjs/pull/723)
[`590c2a5`](https://github.com/langchain-ai/deepagentsjs/commit/590c2a5042473f096d5fac5ddbb4be96e2ace0f2)
Thanks
[@thushanth-bengre-langchain](https://github.com/thushanth-bengre-langchain)!
- fix(deepagents): prevent stack overflow in CompositeBackend grep/glob
on huge result sets, and add a grep match-count cap
`CompositeBackend` accumulated merged `ls`/`grep`/`glob` results with
`push(...entries)`, which passes every entry as a separate function
argument and overflows the call stack (RangeError: Maximum call stack
size exceeded) when a broad search over a large tree returns hundreds of
thousands of entries. Results are now accumulated with a plain loop, so
no result-set size can overflow the stack.
`grep` also gains an optional `maxCount` (backend) / `max_count` (tool)
cap, mirroring the Python SDK. When the cap is hit, results are flagged
`truncated: true` on `GrepResult`/`GlobResult` and the grep tool appends
a note telling the model to narrow the search. The cap defaults to 1000
via the `grepMaxCount` middleware option (set to `null` to disable).
`CompositeBackend` splits the budget across routed backends and
OR-propagates the `truncated` flag on merged results.
## @deepagents/evals@0.0.22
### Patch Changes
- Updated dependencies
[[`590c2a5`](https://github.com/langchain-ai/deepagentsjs/commit/590c2a5042473f096d5fac5ddbb4be96e2ace0f2)]:
- deepagents@1.12.2
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
## Summary
- Fixes `RangeError: Maximum call stack size exceeded` from
`CompositeBackend.glob`/`grep` when a broad search (e.g.
`**/*.{ts,tsx,...}` at `/`) returns hundreds of thousands of entries.
Root cause: results were merged with `push(...entries)`, which passes
each entry as a separate function argument and overflows V8's
argument-stack limit. Now accumulated with a plain loop.
- Ports the Python SDK's grep match-count cap (`ef591e7`): optional
`maxCount` backend param / `max_count` tool arg, `grepMaxCount`
middleware option (default 1000, `null` disables), `truncated` flag on
`GrepResult`/`GlobResult`, and a truncation note in the grep tool
output. `CompositeBackend` splits the budget across routes and
OR-propagates `truncated`.
## Backward compatibility
- `truncated` is optional; `maxCount` defaults to unset everywhere. No
required changes for existing or new users.
## Test plan
- [x] Regression test: mocked backend returning 200k entries —
`glob`/`grep` complete without `RangeError` (verified it fails with the
old spread).
- [x] Cap enforcement, `truncated` propagation across composite routes,
middleware note rendering.
- [x] `pnpm test` (format + lint + all lib tests) passes.
## Summary
- `grep`/`glob` on `StateBackend`/`StoreBackend` returned zero matches
whenever `path` pointed at an exact file (no glob) — the shared
`validatePath` helper forced a trailing `/` before prefix-matching, so a
file's own key could never match itself+`/`.
- Added a `filterFilesByPath` helper that checks for an exact key match
in the files map first, falling back to the existing directory-prefix
behavior otherwise.
## Test plan
- [x] Added regression tests in `utils.test.ts`: exact-file path (no
glob) for grep and glob, plus directory + glob/filename non-regression
cases for both.
- [x] `pnpm --filter deepagents test` — all tests pass.
## Summary
Makes `todoListMiddleware` opt-in. Default agents and default subagents
no longer expose `write_todos` or the `todos` state channel, reducing
the default tool surface while preserving explicit opt-in.
## Changes
- Remove default todo middleware from the main, general-purpose, and
declarative subagent stacks.
- Preserve opt-in support through `middleware: [todoListMiddleware()]`;
declarative subagents opt in through their own middleware.
- Remove `write_todos` from always-present tool typing and built-in
name-collision handling.
- Update state and subagent documentation plus shared runtime assertions
for the optional todo state/tool.
### Harness profiles and coverage
- Re-add todo middleware in the Codex harness profile because its
model-specific prompt requires `write_todos`
- Apply profile `extraMiddleware` to subagent stacks, with profile
factories creating fresh instances per stack.
- Add coverage for default exclusion, explicit opt-in,
main/GP/declarative isolation, profile middleware propagation, and type
inference.
## Summary
The SDK currently ships authored base prose plus middleware guidance
that largely duplicates the tool schemas. This change adopts a more lean
default: no authored base prompt, no duplicate built-in usage prose, and
concise tool descriptions.
Skills, memory, custom system prompts, custom tool descriptions, and
required filesystem routing context remain intact. The legacy
base-prompt surface is removed rather than preserved as a default path.
## Changes
### Default prompt
- Remove the authored default base prompt and its obsolete public
surface.
- Suppress LangChain's default todo middleware prompt while leaving
deepagents-owned middleware prompt-free by default.
- Remove unused filesystem, subagent, async-subagent, and execution
prompt constants and exports.
### Durable tool guidance
- Keep large-result recovery guidance in `read_file` and `grep`
descriptions.
- Keep async task status freshness guidance in `check_async_task` and
`list_async_tasks` descriptions.
- Trim filesystem and task descriptions to their purpose and
load-bearing constraints.
- Build grep and execute descriptions from the configured filesystem
tool set so they do not recommend unavailable tools.
### Token impact
Offline (`o200k_base`, default-agent tool schemas):
| Tool | Before | After | Delta |
| --- | ---: | ---: | ---: |
| `task` | 1,664 | 389 | -77% |
| `read_file` | 728 | 494 | -32% |
| `grep` | 688 | 555 | -19% |
| `edit_file` | 273 | 257 | -6% |
| `glob` | 217 | 172 | -21% |
| `write_file` | 177 | 177 | — |
| `delete` | 146 | 146 | — |
| `ls` | 111 | 111 | — |
| **Total** | **4,005** | **2,302** | **-43%** |
For a default-agent `hello` turn, input tokens drop from 5,395 to 1,895
(-65%).
### Evaluation results
The lean prompt matched the baseline on correctness (0.81) and solve
rate (0.635 vs. 0.634). Unified evaluation macro and micro scores
improved from 0.413 to 0.437 and from 0.371 to 0.394, respectively. (see
more detailed results in
https://github.com/langchain-ai/deepagents/pull/5009 and
https://github.com/langchain-ai/deepagents/pull/4859)
## Summary
`write_file` now creates missing files and replaces existing files
entirely, removing the previous create-only error path while preserving
write permissions and symlink protections.
## Changes
### deepagents filesystem backends
- Updated `FilesystemBackend`, `StateBackend`, `StoreBackend`,
`BaseSandbox`, and `node-vfs` write behavior to allow full-file
replacement on existing paths.
- Preserved path safety and symlink protections in filesystem-backed
writes.
- Removed the sandbox existence precheck so writes rely on the
upload/write transport.
### Tool prompts and examples
- Aligned the `write_file` tool description and schema text with the
Python SDK wording.
- Updated sandbox example prompts so `write_file` is described as
creating or fully replacing files.
### Tests and evals
- Updated backend, middleware, and shared sandbox standard tests that
previously expected an existing-file error.
- Added overwrite coverage for filesystem, sandbox, state, store,
middleware, node-vfs, and file-operation evals.
### Release
- Added a changeset for `deepagents` documenting the new overwrite
behavior.
### Summary
Invalid (non-absolute, `..`, or `~`) or denied filesystem paths passed
by an agent cause the tool to throw which raises a `MiddlewareError` and
causes the whole agent run to crash.
This PR replaces `enforcePermission` which was re-validating the path
without a try/catch with `checkPermission` which returns the error to
the model as a normal tool result so the model can correct the path and
try again instead of crashing.
### Tests
Updated the three permission suites to assert an error result instead of
a throw and added a regression block for `~`/relative/`..` paths that
returns a recoverable error and never calls the backend.
### Summary
`adaptBackendProtocol` rebuilds the backend into a v2 backend protocol
but drops its `routePrefixes`. Since `resolveBackend` runs every backend
through the adapter, a resolved `CompositeBackend` was no longer
detected as composite. That made the execute tool's guard fail at
invocation and disable the execute tool even when filesystem permissions
were correctly scoped to composite routes.
This PR fixes this by forwarding `routePrefixes` through
`adaptBackendProtocol` instead of dropping them.
### Tests
Added unit tests in backends/utils.test.ts to verify that
`routePrefixes` is forwarded through `adaptBackendProtocol` and
`adaptSandboxProtocol` and not added for non-composite backends.
## Summary
Adds middleware override behavior to replace default middleware by
`.name` instead of always being appended.
## Changes
- Add name-keyed middleware merging so custom middleware replaces
matching default middleware in-place and appends novel middleware after
the core stack.
- Apply default-slot middleware overrides consistently when constructing
main-agent and subagent middleware stacks, while preventing parent-only
middleware from propagating unless it matches a subagent default slot.
- Move harness tool exclusion into a private `_ToolExclusionMiddleware`
helper so excluded tools are stripped after tool-injecting middleware
has run.
- Add coverage for main-agent replacement, subagent/default override
propagation, parent-only middleware isolation, profile exclusions,
tool-exclusion ordering, and merge precedence.
## Summary
Implements the optional backend `delete` protocol across the built-in
backend implementations now that the delete protocol surface is
available. This keeps delete behavior backend-local and does not expose
a new filesystem tool to agents.
## Changes
### deepagents
- Implements `delete(filePath)` for filesystem, store, composite,
context hub, and sandbox backends.
- Uses backend-specific safety semantics: filesystem deletion goes
through existing path resolution and only unlinks files, store deletion
treats paths as exact keys, context hub deletion commits null entries,
and sandbox deletion uses shell-quoted `rm -f`.
- Adds tests covering successful deletion, missing paths, directory
rejection where applicable, composite route path restoration, store
wildcard literal behavior, and context hub cache updates/failure
handling.
### @langchain/node-vfs
- Implements `delete(filePath)` for the VFS backend.
- Adds tests for file deletion, missing paths, directory rejection, and
preserving non-target files.
## Summary
Adds a `tools` allowlist option to `createFilesystemMiddleware` so
callers can restrict which built-in filesystem tools are exposed to the
model. The middleware now builds the filesystem system prompt from the
tools that are actually visible on the request, avoiding instructions
for tools that were filtered by the allowlist or backend capability
checks.
## Changes
- Adds a public `FsToolName` type and `tools` option to
`FilesystemMiddlewareOptions`.
- Filters filesystem middleware tools at construction time when an
explicit allowlist is provided, while preserving existing `execute`
backend capability filtering.
- Requires `read_file` in explicit allowlists because it is needed for
filesystem workflows and large-result recovery.
- Generates the filesystem tool prompt dynamically from visible
filesystem tools, and only appends execute-specific instructions when
`execute` is actually available.
- Adds unit and integration coverage for allowlist behavior, prompt
filtering, unsupported `execute`, and user-provided non-filesystem
tools.
## Summary
Adds protocol-level delete support for deepagents backends.
## Changes
### deepagents
- Adds `DeleteResult` to the backend protocol exports.
- Adds optional `delete(filePath)` support to v1 and v2 backend
protocols.
- Preserves optional delete implementations when adapting backends to
the v2 protocol.
- Implements `StateBackend.delete()` for zero-argument StateBackend
instances by sending `{ [filePath]: null }` through `__pregel_send`.
- Adds a changeset for the new backend protocol capability.
### Summary
`fast-glob` follows symbolic links by default so the `glob`/`grep` tools
in `FilesystemBackend` and `LocalShellBackend` walked directory symlink
cycles recursively causing `ELOOP` crashes. and aborted the run over a
target repo.
This PR fixes by adding`followSymbolicLinks: false` to all four
`fast-glob` calls. Bonus: this keeps traversal inside the search root.
### Tests
Unit tests in `filesystem.test.ts` and `local-shell.test.ts` verifying
that a `sub/sub -> .` cycle now traverses cleanly without traversing
symlinks.
This is an upstream fix for:
https://github.com/langchain-ai/openwiki/issues/72
## Problem
Addresses #658.
When using weaker or custom LLMs (e.g., smaller open-source models) in
`deepagents` and consumers like `openwiki`, the agent immediately
crashes with a schema validation error on the first filesystem tool
call:
This occurs because:
1. `ls` defines its directory argument as `path`, but `read_file`,
`write_file`, and `edit_file` expect `file_path`. Models often default
to `path` across all filesystem tools.
2. The description prompt examples in `fs.ts` and `skills.ts` show
`read_file(path, ...)`, instructing models to use the wrong parameter
name.
## Proposed Changes
This PR solves both prompt inconsistencies and schema validation
failures with a fully backward-compatible input normalization helper:
1. **Prompt Alignment**: Fixed references in `fs.ts` and `skills.ts` to
instruct the model to use `file_path` (aligned with the prompt fixes in
#644).
2. **Schema Input Normalization**: Added a `normalizeFilePathInput`
preprocessor helper to the `read_file`, `write_file`, and `edit_file`
Zod schemas using `z.preprocess`. This dynamically maps the `path` key
to `file_path` if the model still passes `path`.
## Verification & Tests
- Added unit tests in `fs.test.ts` to assert that:
- `path` parameter normalization maps correctly to `file_path` for
`read_file`, `write_file`, and `edit_file`.
- All instruction examples contain only valid schema fields.
- Verified that `zod-to-json-schema` unwraps the `z.preprocess` layer
perfectly, preserving the original schema descriptions and definitions
sent to the LLM.
- Ran all filesystem tests successfully (`npx vitest run
src/middleware/fs.test.ts` passes).
### Summary
`read_file` couldn't read many common text files and with Anthropic it
crashes the whole run. `getMimeType()` mapped any unrecognized extension
to `application/octet-stream`, which `isTextMimeType()` treats as
binary, so the file got base64-encoded into a document block.
This defaults unknown extensions to `text/plain` instead. Known binaries
(images, audio, video, PDF/PPT) stay explicitly mapped, so image and PDF
handling is unchanged. Files like `.properties`, `.scss`, `.tf`,
`.lock`, and extension-less files like `Dockerfile` now read as text.
### Tests
Flipped the unknown-extension unit test to expect `text/plain`.
### Summary
Modal v0.8.0 released the following breaking change:
```
Breaking: Removed the deprecated low-level file-handle API: Sandbox.Open / SandboxFile (Go), and sandbox.open / SandboxFile (JS). Use the Sandbox Filesystem API instead: sandbox.Filesystem() (Go) / sandbox.filesystem (JS).
```
Ref:
https://github.com/modal-labs/modal-client/blob/main/CHANGELOG_GO_JS.md
This PR updates our implementation to use the sandbox filesystem api
### Tests
Updated `sandbox.test.ts` mocks to model `sandbox.filesystem`
## Description
Closes#609 by updating `@langchain/daytona` to the maintained
`@daytona/sdk` at `^0.184.0` while preserving the same SDK API usage.
Added a patch changeset; `@daytona/sdk` is Apache-2.0 and maintained by
Daytona as the direct replacement for deprecated `@daytonaio/sdk`.
## Self-Hosted Release Note
`@langchain/daytona` now depends on `@daytona/sdk` instead of deprecated
`@daytonaio/sdk`.
## Test Plan
- [ ] Verify package publish metadata resolves `@daytona/sdk`
`^0.184.0`.
Made by [Open
SWE](https://openswe.vercel.app/agents/d6f434c5-39c6-eef7-6cc6-670cf092bb67)
Co-authored-by: Christian Bromann <731337+christian-bromann@users.noreply.github.com>
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Updates to `@langchain/core` can cause downstream friction with
TypeScript:
```
Type 'ChatOpenAI<ChatOpenAICallOptions>' is not assignable to type 'string | BaseLanguageModel<any, BaseLanguageModelCallOptions> | undefined'.
Type 'ChatOpenAI<ChatOpenAICallOptions>' is not assignable to type 'BaseLanguageModel<any, BaseLanguageModelCallOptions>'.
Types of property 'getGraph' are incompatible.
Type '(_?: import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/runnables/types").RunnableConfig<Record<string, any>> | undefined) => import("/Users/christian.bromann/Sites/L...' is not assignable to type '(_?: import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/runnables/types").RunnableConfig<Record<string, any>> | undefined) => import("/Users/christian.bromann/Sites/L...'.
Types of parameters '_' and '_' are incompatible.
Type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/runnables/types").RunnableConfig<Record<string, any>> | undefined' is not assignable to type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/runnables/types").RunnableConfig<Record<string, any>> | undefined'.
Type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/runnables/types").RunnableConfig<Record<string, any>>' is not assignable to type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/runnables/types").RunnableConfig<Record<string, any>>'.
Types of property 'callbacks' are incompatible.
Type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").Callbacks | undefined' is not assignable to type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").Callbacks | undefined'.
Type 'CallbackManager' is not assignable to type 'Callbacks | undefined'.
Type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManager' is not assignable to type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManager'.
The types returned by 'handleLLMStart(...)' are incompatible between these types.
Type 'Promise<import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManagerForLLMRun[]>' is not assignable to type 'Promise<import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManagerForLLMRun[]>'.
Type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManagerForLLMRun[]' is not assignable to type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManagerForLLMRun[]'.
Type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.0_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManagerForLLMRun' is not assignable to type 'import("/Users/christian.bromann/Sites/LangChain/deployment-cookbook/js-langsmith/node_modules/.pnpm/@langchain+core@1.2.1_openai@6.44.0_ws@8.21.0_zod@4.4.3__ws@8.21.0/node_modules/@langchain/core/dist/callbacks/manager").CallbackManagerForLLMRun'.
Property 'inheritableHandlers' is protected but type 'BaseRunManager' is not a class derived from 'BaseRunManager'.ts(2322)
```
to avoid this we should move the LangChain dependency to become peer
deps.
- Moved `@langchain/core`, `@langchain/langgraph`,
`@langchain/langgraph-sdk`, and `langchain` from `dependencies` to
`peerDependencies` in `libs/deepagents`, keeping dev copies for isolated
build/test.
- Lets consumers bump `@langchain/core` without requiring a new
`deepagents` release; a release is only needed to widen the peer range
across a major.
- Note on install contract: npm 7+ and pnpm 8+ auto-install peers; Yarn
only warns, so Yarn users must add the four packages explicitly (worth a
README follow-up).
---------
Signed-off-by: Christian Bromann <git@bromann.dev>
Co-authored-by: open-swe[bot] <215916821+open-swe[bot]@users.noreply.github.com>
## Summary
A deep agent that delegates to a subagent 400s when that subagent uses
an Anthropic server-side tool (web search, web fetch, code execution,
MCP). The subagent's final-message content blocks e.g.
server_tool_use/..._tool_result were forwarded verbatim as the task
ToolMessage content. On the parent agent's next request they serialize
into a tool_result, which Anthropic rejects:
`400 invalid_request_error
messages.N.content.0.tool_result.content.0: Input tag 'server_tool_use'
found using 'type'
does not match any of the expected tags:
'document','image','search_result','text','tool_reference'`
## Solution
Now matches the Python impl: forward only the subagent's text answer.
returnCommandWithStateUpdate now walks back to the last AIMessage with
non-empty text and uses that text (a string) as the ToolMessage content.
- Confirmed the Python deepagents impl is not affected (it forwards
.text); this aligns JS with Python.
- Reproduced the 400 across {web_search, code_execution, web_fetch} ×
{claude-haiku-4-5, claude-sonnet-4-6} on the latest published packages;
all green after the fix.
## Tests
- New: subagent server-tool blocks are filtered from forwarded content.
- New: walk-back uses the last non-empty AIMessage, skipping a trailing
empty message.
- Updated the #239/#245 tests to assert the (now string) ToolMessage
content.
- No regressions in subagent-related tests in the evals suite
---------
Signed-off-by: Alexander Olsen <aolsenjazz@gmail.com>
Co-authored-by: Christian Bromann <git@bromann.dev>
## Summary
- Adds `bedrockPromptCachingMiddleware` to the default middleware stack
- Adds unit tests for the Bedrock model detection util function covering
the different formats of model strings/`ChatModel`s
### Summary
The code interpreter now comes with a global `task` tool out of the box.
As a result, we should disallow specifying the `task` tool for ptc. This
PR implements logic that forbids the task tool as in ptc and throws with
a detailed error message if found. This PR also updates the RLM agent
example and READMEs.
### Tests
Unit tests validating expected behavior for task by name and tool
instance with "task" name.
## 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
### 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.
github-actions[bot]