LangChain (Python)

ToolRetryMiddleware now only retries retryable exceptions, and a new ToolErrorMiddleware is available.

Added a meta extra and langchain-meta support for init_chat_model, and explicit prompt caching for OpenAI.

Added explicit prompt caching support for OpenAI integrations.

Fixed cached prompt token usage reporting in Fireworks API calls. Also clarified prompt-cache session affinity guidance in documentation.

Suppressed a Pydantic serializer warning that appeared on structured output parsed fields. Also fixed asyncio.get_running_loop() usage in async contexts and bumped dependencies including langgraph-checkpoint to 4.1.1.

Fixed interrupt propagation through ToolRetryMiddleware and a shared process-group kill issue in shell middleware. Also sanitized Anthropic cache markers on fallback retries.

Fixed output parser bugs in xml.py and pydantic.py, and improved langsmith loader error messages. Also corrected docstring parsing for continuation lines with colons, avoided dict shadowing in language models, and hardened async context handling.

Chat responses from Mistral AI now surface citation metadata, and support for stop sequences has been added. Package version tracking is now included in tracing metadata.

Added support for default_headers to enable custom HTTP header injection in OpenRouter requests.

Fixed repeated finish metadata in responses and stripped reasoning IDs from Responses API output.

Fixed an issue where initial text was dropped on content_block_start events. Added missing messages to bare raise ValueError calls.

Changes since langchain-fireworks==1.4.2

Surface parallel_tool_calls on bind_tools and add cache_control passthrough on tool definitions. Minimum openrouter dependency bumped to 0.9.2, dropping a file workaround.

Fixed strict mode on tool schemas for OpenAI-compatible models in ProviderStrategy — strict=True is now set only on tools for models that support it.

Changes since langchain-anthropic==1.4.6

Tool schemas now default to strict=True for OpenAI-compatible models in ProviderStrategy, and response item IDs are dropped when store is false. Fixes include removing stop from the Responses API payload.

Fixed summary format switching and updated provider strategy detection for dated gpt-5.2/gpt-5.4 snapshots. Plus dependency bumps for cryptography, aiohttp, and pyjwt.

Preserved usage token details in v3 streaming events and memoized BaseTool.tool_call_schema for performance. Also includes various housekeeping changes.

Changes since langchain-openai==1.3.1

release(openai): 1.3.2 (#38130)

Fixed normalization for v1 streamed tool calls and updated tests to use gpt-4o for image token counting. Added package version tracking to tracing metadata.