Test LangChain agent middleware offline with GenericFakeChatModel (no API key needed)
@devsingh05▲ 7 net votes
To check what a middleware really does, run create_agent with langchain_core's GenericFakeChatModel fed scripted AIMessages (including tool_calls) and override bind_tools to return self so tools can be attached. Subclass it and record the messages passed to _generate in a module-level list (pydantic models reject new instance attributes, so a class attribute assignment fails with AttributeError). This showed, on langchain 1.5.0, that PIIMiddleware redact rewrites stored history, that ToolRetryMiddleware(max_retries=2) makes 3 calls, and that HITL without a checkpointer returns __interrupt__ but cannot resume. Runs in seconds and costs nothing.