what stackpulse tracks
LangChain.js releases from GitHub
StackPulse watches LangChain.js release notes and keeps the original source link close to every summary.
Build context-aware reasoning applications in JavaScript StackPulse turns upstream changelogs into scannable summaries with risky changes, deprecations, migration notes, and source links.
what stackpulse tracks
StackPulse watches LangChain.js release notes and keeps the original source link close to every summary.
upgrade risk
Risky changes are separated from normal feature notes so you can scan upgrade impact before changing production dependencies.
migration notes
Migration steps and recommended actions are only shown when the upstream release notes support them.
This patch release adds Gemini 4 Hydrogen and Carbon model capability profiles. It also adds support for the XHIGH thinking level, clamping it to HIGH on models that do not support it and passing through future thinking levels on level-based models.
LangChain.js users working with Google Gemini models can use the new Gemini 4 profiles and the expanded thinking-level handling.
This patch updates `AnthropicToolsOutputParser` to fall back to `tool_calls`. This makes `withStructuredOutput` work when streamed content is normalized to `tool_call` blocks.
Users of `withStructuredOutput` with streamed Anthropic content normalized to `tool_call` blocks are affected.
This patch prevents failed tool calls or subagents from crashing the process when their `output` promise is never consumed. It also ensures a `Command` returned from `wrapModelCall` is applied once when passed through by outer middleware.
Applications using tool calls, subagents, or middleware that passes through `wrapModelCall` may benefit from these reliability fixes.
This patch changes how streamed `input_json_delta` blocks are handled. They are merged into the preceding `tool_use` or `server_tool_use` block at the same index.
Users processing streamed tool-use blocks with `@langchain/core` are affected.
This patch release updates the `@langchain/openai` dependency to version 1.8.4. The release notes describe no other changes.
Users of `@langchain/xai` receive the updated `@langchain/openai` dependency.
`TypeSafeClassifier` is now a decision model extending `BaseDecisionModel` from `@langchain/core/decisions`, with updated request and response types. This release also requires Zod 4.6.6 or newer to avoid versions affected by GHSA-mrhx-c7xc-m3w5, while retaining support for Zod v3 schemas through the bundled `zod/v3` entry point.
Users of `@langchain/typesafe`, especially those using `TypeSafeClassifier` or its experimental middleware, must adapt to breaking API changes and the Zod dependency requirement.
Update classifier calls and response handling, switch deprecated middleware imports to `langchain`, and ensure installed Zod is version 4.6.6 or newer.
This patch release updates the `@langchain/openai` dependency to version 1.8.4. The release notes describe no other changes.
Users of `@langchain/together-ai` receive the updated `@langchain/openai` dependency.
This patch requires installed dependencies to use Zod 4.6.6 or newer, excluding versions affected by GHSA-mrhx-c7xc-m3w5. Zod v3 schemas remain supported through the bundled `zod/v3` entry point.
Users of @langchain/tavily with installed Zod dependencies older than 4.6.6 are affected; Zod v3 schemas remain supported via the bundled `zod/v3` entry point.
Ensure installed Zod dependencies meet the 4.6.6 minimum; continue using the bundled `zod/v3` entry point for Zod v3 schemas.
This patch release adds middleware for routing an agent to a model selected by a decision model and for blocking tool calls judged risky. It also raises the minimum installed Zod version to 4.6.6 to exclude versions affected by GHSA-mrhx-c7xc-m3w5; Zod v3 schemas remain supported via the bundled `zod/v3` entry point.
LangChain.js users gain model-routing and risky-tool-blocking middleware, while users with installed Zod dependencies must use version 4.6.6 or newer.
Ensure installed Zod dependencies are at least 4.6.6; use the bundled `zod/v3` entry point if you need Zod v3 schemas.
This patch release updates the package’s dependency on @langchain/openai to version 1.8.3. The release notes provide no further details about behavior changes.
Users of @langchain/xai receive the updated @langchain/openai dependency.
This patch release updates the `@langchain/openai` dependency to version 1.8.3. The release notes do not describe any other changes.
Users of `@langchain/together-ai` receive the updated `@langchain/openai` dependency.
This patch release updates the `@langchain/openai` dependency to version 1.8.3. No other changes are described in the release notes.
Users of `@langchain/openrouter` receive the updated `@langchain/openai` dependency.
This patch surfaces model refusals as `ModelRefusalError`. Chat Completions retains refusal information, refusals are passed back natively in assistant messages (including `v1` messages), and `withStructuredOutput` using `jsonSchema` throws the error when a refusal occurs.
Applications using LangChain.js OpenAI Chat Completions or structured output may now receive `ModelRefusalError` when the model refuses a request.
This patch adds `modelRefusalMiddleware`, which throws `ModelRefusalError` when a model refuses. Middleware also passes `ModelRefusalError` through unwrapped rather than wrapping it as a `MiddlewareError`.
Users handling model refusals through LangChain middleware are affected; refusals now surface as an unwrapped `ModelRefusalError`.
This patch changes routing for Google chat models: requests use the LangSmith gateway only when no credentials are provided. When credentials are explicitly configured, requests go directly to Google. The release also updates the `@langchain/google-webauth` dependency to 2.3.6.
Users of Google chat models who pass credentials explicitly are affected; those requests now bypass the LangSmith gateway and go directly to Google.
If you rely on the LangSmith gateway, avoid passing `apiKey`, `authOptions`, `credentials`, or `googleAuthOptions`.
Google chat models now use the LangSmith gateway only when no credentials are provided. When explicit authentication options are supplied, requests go directly to Google.
Users configuring explicit Google credentials or authentication options are affected because those requests now bypass the LangSmith gateway.
This patch release updates the `@langchain/google-common` dependency to version 2.3.6. The release notes describe no other changes.
Users of `@langchain/google-webauth` receive the updated `@langchain/google-common` dependency.
The release notes state that there are no changes in this release. No features, fixes, or migration details are described.
view details view source on github->Adds experimental support for Google’s Interactions API on AI Studio and Vertex AI, including invocation, streaming, event streaming, and API converters. Google chat models with explicitly supplied credentials now connect directly to Google instead of being routed through the LangSmith gateway.
LangChain.js users integrating Google chat models can opt into the experimental Interactions API, while users providing Google credentials will have requests routed directly to Google.
Use `apiHandler: "interactions"` to try the experimental Interactions API; no migration is required for existing API usage.
This patch release updates the `@langchain/openai` dependency to version 1.8.2. No other changes are described in the release notes.
Users of `@langchain/xai` receive the updated `@langchain/openai` dependency.