Fixed
- A model answer has a deadline now, not just a socket timeout.
AI_PROVIDER_TIMEOUT
was handed to the HTTP client, whose read timeout measures the gap between two
chunks — so a model that keeps generating never trips it, and a review of a
nineteen-file pull request sat for minutes with its headers already in the log, looking
answered. A second review in the same log finished in seventy-eight seconds, which is
what a slow model looks like when it does finish. The whole request is now under that
timeout, generation included: a model that will not stop is cut off and asked again,
which also makes "the attempt count times the timeout" the real worst case rather than
a number nothing enforced. - Exceeding that deadline is retried once, not to the full count. Running out of
time is about the request rather than about luck: the same prompt on the same model
takes the same minutes again. Five attempts of five minutes is twenty-five minutes
spent on a review that was not going to arrive, at the price of the whole prompt each
time, so a deadline gets one more attempt. The fullAI_PROVIDER_RETRIESstill applies
to failures that carried no verdict at all — a broken connection, a rate limit, a
server error.
Changed
- A review says when it is waiting for the model. The request line in the log arrives
with the response headers, not with the answer — httpx logs before it reads the body —
so a log can show a200in under a second and then look dead for minutes while the
model is still generating. That is exactly how a fifty-eight file pull request read as
stuck when nothing at all was wrong with it. The line before the call names the pull
request, the connection, the mode and the size of the prompt, so the wait has a
beginning and a number attached to it.