LangChain

Recently shipped mid-conversation model support in the Anthropic and OpenAI integrations, plus a broad LangGraph checkpoint-replay fix pass.

Model integrations gained mid-conversation support.

• langchain-anthropic 1.7.6 and langchain-openai 1.7.1 declare mid-conversation support in their model profiles, alongside a refresh of the profile data.
• SummarizationMiddleware now accepts a custom summarizer, so summarization no longer depends on the built-in model call.

MCP adapters 2.0 is a breaking upgrade. The adapter moved to MCP SDK 2 and serves modern and legacy MCP servers from one place, letting servers ask users for input through LangGraph interrupts. Tool names, configuration, and tool results changed: MultiServerMCPClient becomes MCPAdapter, and server maps move under { servers: { ... } }. A migration guide covers the rest. A subsequent patch keeps the SSE attempt's SDK error as cause after a failed Streamable HTTP connection, so OAuth login handlers checking UnauthorizedError work against SSE-only servers.

LangSmith's MCP server now queries runs the way the SDK does.

• query_runs takes the SDK's client.runs.query() parameters, including typed selects for costs, token breakdowns, and full inputs/outputs with cursor pagination; fetch_runs is deprecated in its favor.
• Costs resolve for OpenAI models under "All models" and for the open models Vertex AI hosts: DeepSeek, Qwen, GLM, Kimi, MiniMax, and GPT-OSS.
• Public trace and thread share links now require read permission.
• LangSmith Chat offers Claude Sonnet 5.5 (Sonnet 5 stays default), takes roughly 4x more steps per turn, and completes multi-tool turns on Gemini 3.x through Vertex AI.

Feedback ingest no longer drops runs over bad scores. A score that violates the workspace feedback-key configuration is accepted and recorded as error feedback on the run with the rejection reason in the comment, instead of a 400 that discarded the run and every other score with it. Annotation-queue feedback is attributed by whichever identity the request resolves.

LangGraph checkpointing got a correctness pass.

• An empty checkpoint_id now resolves to the latest entry in delta history.
• Delta replay is ordered by task path, and every message written to a messages DeltaChannel gets an id.
• Namespace labels are validated before store dispatch, and ToolNode no longer swallows invalid resume values.
• Exit runs and update_state writes stay on the checkpoints they belong to instead of leaking into newer branches.

Smaller items. LangSmith sandboxes accept any JOSE asymmetric algorithm the JWKS key chooses and support customer-managed KMS keys for S3 mounts; @langchain/core preserves runtime context across tool calls, nested runnables, and streaming; langgraph-sdk percent-encodes thread_id and assistant_id in thread stream requests.