langchain-ai/langchain

langchain-ai

langchain

Blog Docs Changelog

Performance History

Latest Results

skip async subagent e2e test on Python < 3.11 (config propagation)

nh/subagent-projection

22 hours ago

surface same-named/recursive subagents on run.subagents Relax the SubagentTransformer gate to 'lc_agent_name present' (was 'differs from parent'), matching the langgraph base discriminator, so a subagent that invokes itself (or any same-named nested agent) is surfaced. Trade-off: a non-agent subgraph that inherited the parent's name also surfaces. Tests: same-name nesting surfaces; unnamed inner agent surfaces under the inherited name. Bump langgraph pin to the matching #7928 commit.

nh/subagent-projection

22 hours ago

test(core): add regression tests for EvaluationResult.feedback_config (#31802)

begonia-474:fix/31802-feedback-config-preservation-v2

1 day ago

relock langgraph to #7928 reshaped HEAD (e3265ece4) Updates the lockfile so the branch builds against PR #7928's lc_agent_name reshape rather than the stale pre-reshape commit. The git+branch pin on langgraph remains temporary and must become a version constraint once langgraph releases with #7928.

nh/subagent-projection

1 day ago

fix(anthropic, openai): never close shared/cached httpx pools on aclose() The first cut of close()/aclose() unconditionally closed the SDK client's underlying httpx pool. But both integrations back their clients with a PROCESS-WIDE SHARED pool via @lru_cache (`_get_default_*httpx_client`): every model with the same base_url/timeout/proxy reuses one pool by design. Closing it from a single model's teardown broke every other live model in the process — observed in a long-lived worker as: RuntimeError: Cannot send a request, as the client has been closed. -> anthropic.APIConnectionError: Connection error. Fix: close()/aclose() now release the underlying httpx client ONLY when the model privately owns it; the shared cached pool and user-supplied clients are left intact. - anthropic: `ChatAnthropic` always wraps the shared cached pool (it has no http_client field), so close()/aclose() are effectively no-ops for the pool. An identity check against the lru-cache getter (`_wraps_shared_httpx`) guards a hypothetical future private-client path. `_http_client_params()` is factored out so the cached_property builders and the identity check stay in sync. - openai: ownership is computed in `validate_environment` and stored on `_owns_sync_http_client` / `_owns_async_http_client`. A client is owned iff the model built it privately — the unhashable-`httpx.Timeout` fresh-client path or an `openai_proxy` transport — and the user did not supply their own `http_client` / `http_async_client`. Default (shared cache) and user-supplied clients are never closed. Tests rewritten to pin the invariant: a regression test builds two default models, closes one, and asserts the other's shared pool is still open; plus owned-path and user-injected-not-closed cases. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

feat/chat-model-aclose

2 days ago

release(anthropic): 1.4.4 (#37757)

master

2 days ago

mdrxy/release-anthropic-1-4-4

2 days ago

Refresh lock for langgraph PR #7928 SHA bump (766a881c5)

nh/subagent-projection

2 days ago

Latest Branches

feat(langchain): project subagent runs onto typed run.subagents channel#37739

22 hours ago

126d114

nh/subagent-projection

test(core): add regression tests for EvaluationResult.feedback_config dict preservation (#31802)#37775

1 day ago

93bb96a

begonia-474:fix/31802-feedback-config-preservation-v2

feat(core, anthropic, openai): aclose() lifecycle + latched fallbacks#37718

2 days ago

e0ace82

feat/chat-model-aclose

Home Terms Privacy Docs