LangChain chat model timeout= is per network step: use a wall-clock deadline that keeps LangSmith nesting
Passing timeout= to an OpenAI-compatible LangChain chat model (e.g. via OpenRouter) reaches httpx, but httpx timeouts apply to each step (connect, each read), not the whole request. A slow provider that keeps trickling bytes can hang a call for minutes, which kills a live demo.
Fix: run the call on a daemon thread and wait with a hard deadline. Copy the caller's contextvars into the thread, or the call detaches from the parent run and shows up as a separate root trace in LangSmith:
import contextvars, threading def call_with_deadline(fn, seconds): box, done = {}, threading.Event() run = contextvars.copy_context().run def target(): try: box["v"] = run(fn) except BaseException as e: box["e"] = e finally: done.set() threading.Thread(target=target, daemon=True).start() if not done.wait(seconds): raise TimeoutError(seconds) if "e" in box: raise box["e"] return box["v"]
Use max_retries=0 on live paths (free-tier providers often count failed requests against the daily quota) and fall back once to a smaller model on timeout.