GeezerVison

Was running this against a flaky proxy the other day and hit something annoying: the agent would just sit there — no output, nothing. Turns out fetch() doesn't have a timeout. Node silently drops the timeout option if you pass it, so if a provider accepts the connection and then never answers, you're stuck until you hit Ctrl+C.

So the LLM client now takes a timeoutMs (default 5 min, overridable with CORECODER_TIMEOUT_MS) and every request races against a timer. If the timer wins, it aborts the fetch and the retry loop treats it as a transient error and retries. Your own Ctrl+C still comes through as an AbortError — the timeout can't swallow it.

Three tests cover it:

• a provider that hangs forever → request times out, retry loop gives up with a "timed out" error
• a pre-aborted signal → still surfaces as AbortError, not a timeout
• first attempt times out, second succeeds → retry works

Writing the tests caught something fun: my fetch stub didn't reject when handed an already-aborted signal, so the second test hung forever. Real fetch rejects immediately in that case, so I fixed the stub, not the client.

Design note: the timeout covers getting a response, not the streaming body — a model that's actively streaming can legitimately take a while, and that's Ctrl+C territory. Say the word if you'd rather have a whole-call timeout instead.