LangChain
OpenRouter reasoning content now surfaces as standard content blocks in @langchain/core, ending silent drops of chain-of-thought output.1
Reasoning models route cleanly through OpenRouter — convertOpenRouterResponseToBaseMessage and convertOpenRouterDeltaToBaseMessageChunk now copy OpenRouter's reasoning and reasoning_details fields onto additional_kwargs.reasoning_content / additional_kwargs.reasoning_details. A new ChatOpenRouterTranslator is registered under the "openrouter" provider key so AIMessage.contentBlocks emits standard {type: "reasoning"} blocks alongside text and tool calls. Models affected include DeepSeek R1, Claude extended thinking, and OpenAI o-series routed through OpenRouter — previously only the token count survived.2
LangGraph SDK and checkpoint hardened — sdk==0.3.15 adds percent-encoding for caller-supplied identifiers in URL paths and metadata filter support for crons search/count.3 checkpoint==4.1.1 restricts lc:2 envelope revival to the default constructor to close a deserialization edge case.4
Partner integrations updated — @langchain/mistralai@1.1.0 upgrades to @mistralai/mistralai v2 and now populates response_metadata.model on chat completions (was missing before).5 @langchain/aws@1.3.9 adds claude-haiku-4 to supportedToolChoiceValuesForModel.6 The langchain-openai Python partner sources LLM context size from model profiles and broadens ContextOverflowError detection to cover additional providers.7
LangSmith Fleet replaces Agent Builder — the hosted agent-building surface was renamed; SDK session lifecycle controls gained idle_ttl_seconds and delete_after_stop_seconds parameters (replacing ttl_seconds).8 JS tracing offloads serialization to a worker thread.
A prior CVE (GHSA-qh6h-p6c9-ff54) was patched across langchain-core v0.3.x and v1.x branches simultaneously; hub.pull in langchain-classic remains deprecated.