LangChain

Recently shipped cached token reporting for Fireworks and content block filtering for OpenAI JS.

LangSmith SDK now traces voice agent runs — v0.9.6 added integrations for Pipecat, LiveKit, and OpenAI Realtime; v0.9.7 followed with Google ADK Live voice tracing. The SDK automatically sets ls_provider/ls_model_name on voice LLM spans for cost attribution.

Mistral chat responses expose citation metadata — langchain-mistralai==1.1.6 surfaces citation metadata from chat responses and adds support for stop sequences.

OpenRouter header injection and deduplication fixes — langchain-openrouter==0.2.6 accepts default_headers for custom HTTP header injection. v0.2.5 deduplicates repeated finish metadata and strips Responses reasoning IDs.

LangGraph emits valid UUIDs and fixes snapshot overwrites — v1.2.7 fixes DeltaChannel snapshot overwrites across supersteps and emits valid UUIDs for exit-mode delta task IDs for langgraph-api. v1.2.9 fixes updateState metadata and counters for delta channels.

Fireworks cached prompt token usage — langchain-fireworks==1.4.4 now reports cached prompt token usage, improving cost tracking for repeated prompts.

OpenAI JS filters rejected content blocks — @langchain/openai@1.5.5 filters out content blocks that the chat completions API rejects as input, and drops tool_call content blocks from chat completions input. Dependents (Fireworks, xAI, Together, DeepSeek, OpenRouter) also updated.

LangSmith SDK addresses URL path injection — v0.10.1 fixes a security vulnerability in the Python sandbox client where unencoded resource names could cause URL path injection. The JS SDK also exposes OpenAPI client error classes.