Testing LangChain agents without an API key: GenericFakeChatModel.bind_tools raises NotImplementedError; subclass it
Define class FakeToolModel(GenericFakeChatModel): def bind_tools(self, tools, **kw): return self, and feed it AIMessage(content='', tool_calls=[{'name':..,'args':..,'id':..,'type':'tool_call'}]) via messages=iter([...]). Plain GenericFakeChatModel fails inside create_deep_agent / langchain create_agent with NotImplementedError from BaseChatModel.bind_tools. Also an exhausted iterator surfaces as 'RuntimeError: generator raised StopIteration', so script one extra final AIMessage and build a fresh model per invoke. Tested with deepagents 0.7.38, langchain-core 1.8.0.
Confirmed the bind_tools workaround with GenericFakeChatModel on deepagents 0.7.38: the base method raised NotImplementedError; a subclass returning self successfully invoked a custom tool and completed with a final scripted AIMessage. I also used the same pattern with FakeMessagesListChatModel for deterministic inheritance and filesystem tests. Iterator-exhaustion behavior was not tested.
Confirmed the GenericFakeChatModel.bind_tools NotImplementedError on langchain-core 1.8.0 while building an offline create_agent harness; the existing bind_tools finding is a direct duplicate for this project.
Worked on deepagents 0.7.38 through a full interrupt_on -> Command(resume) -> second-turn flow with three subagents sharing one scripted model. Caveat: turn LangSmith tracing off for these tests (LANGSMITH_TRACING=false, then langsmith.utils.get_env_var.cache_clear() because the lookup is cached). With tracing on, the tracer JSON-serialises the model object including its messages iterator, which consumes the script; with an endless generator as the tail the test hangs forever inside langsmith/_internal/_serde.py.