langchain-ai/langchain - CodSpeed

langchain

Blog Docs Changelog

Performance History

Latest Results

feat(langchain): filter internal middleware model calls from `messages` projection SummarizationMiddleware, LLMToolEmulator, and LLMToolSelectorMiddleware call a model for their own bookkeeping inside the same graph node as the main agent turn, so stream_events(version="v3")'s messages projection couldn't tell their tokens apart from the agent's real answer. Tags those calls as internal via config metadata and adds an InternalCallTransformer (registered on every compiled agent) that keeps tagged runs out of run.messages and the raw event log, without adding a new projection. Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>

nishitha/langchain/filter-internal-model-calls-from-messages

5 hours ago

add on_parsing_failure and max_retires

nm/mw-streaming

5 hours ago

fix(langchain): clear stale `structured_response` between checkpointed turns (#39248) Closes #36957 --- With a checkpointer, `_build_commands` only wrote `structured_response` into state when the model call produced one, so a value persisted from a prior turn stuck around in the checkpoint. The routing edges (`_make_model_to_tools_edge`, `_make_model_to_model_edge`) only checked `"structured_response" in state` (key presence), so a turn whose first model call failed structured-output validation and needed a retry would exit the loop early using the previous turn's stale value instead of continuing to retry. `_build_commands` now explicitly clears `structured_response` to `None` when the agent has a `response_format` configured but the current call didn't produce one, and the routing edges check `state.get("structured_response") is not None` for freshness rather than mere key presence. --------- Co-authored-by: Kai Jiang <167662354+Kcstring@users.noreply.github.com>

master

6 hours ago

fix(langchain): stop HITL approval gates from silently failing open (#39247) Closes #38838 Related: #37093 --- Fixes an issue in `HumanInTheLoopMiddleware` where an invalid `allowed_decisions` config (missing, empty, or misspelled) silently disabled the approval gate. The constructor now raises `ValueError` with the offending tool and config keys. Also investigated #37093: rejected tool calls are already prevented from executing by `create_agent` routing. The rejected call must remain in the `AIMessage` so its rejection `ToolMessage` stays protocol-valid. Added an end-to-end regression test confirming the call is not executed and the message history remains correctly paired. Co-authored-by: Yigtwxx <yigiterdogan023@gmail.com>

master

6 hours ago

link to gateway docs

ccurme/langchain/langsmith-provider

8 hours ago

Merge branch 'master' into taehan79-kim/fix-nested-subgraph-rendering

taehan79-kim:taehan79-kim/fix-nested-subgraph-rendering

9 hours ago

chore: bump aiohttp from 3.14.1 to 3.14.3 in /libs/langchain_v1 --- updated-dependencies: - dependency-name: aiohttp dependency-version: 3.14.3 dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com>

dependabot/uv/libs/langchain_v1/aiohttp-3.14.3

10 hours ago

fix(langchain): type checkpointer thread config for mypy The regression test's `thread` config dict lacked a `RunnableConfig` annotation, which mypy's `agent.invoke` overload resolution needs to pick the right signature. CI's `lint_tests` job caught this. Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>

imnishitha/langchain/fix-stale-structured-response

13 hours ago

Latest Branches

0%

feat(langchain): filter internal middleware model calls from `messages` projection#39252

5 hours ago

e67b3d9

nishitha/langchain/filter-internal-model-calls-from-messages

0%

fix(langchain): prevent middleware model calls from leaking into the parent stream#39245

5 hours ago

74db80d

nm/mw-streaming

0%

feat(langchain): add LangSmith provider to `init_chat_model`#39224

8 hours ago

bf410cd

ccurme/langchain/langsmith-provider

© 2026 CodSpeed Technology

Home Terms Privacy Docs