From the PR #422 Copilot re-review (suppressed comment, src/aorta/chat/inference/providers/local_vllm.py:44). Verified live on 1090cd26.
url = settings.vllm_base_url.rstrip("/").replace("/v1", "") + "/health"str.replace removes every occurrence. A valid base such as https://host/v1-gateway/v1 becomes https://host-gateway/health, so preflight polls a URL that does not exist and reports a healthy server unreachable after the full timeout.
Fix: strip only a trailing /v1 --
base = settings.vllm_base_url.rstrip("/")
url = (base[: -len("/v1")] if base.endswith("/v1") else base) + "/health"Small, but it wants a couple of table-driven cases (plain host, trailing /v1, /v1 in a path segment, no /v1 at all) rather than a one-line change, and it is unrelated to the review classes #422 was addressing.
Also worth checking in the same pass: remote_openai / remote_litellm do not do this, so this is the only site.
From the PR #422 Copilot re-review (suppressed comment,
src/aorta/chat/inference/providers/local_vllm.py:44). Verified live on1090cd26.str.replaceremoves every occurrence. A valid base such ashttps://host/v1-gateway/v1becomeshttps://host-gateway/health, so preflight polls a URL that does not exist and reports a healthy server unreachable after the full timeout.Fix: strip only a trailing
/v1--Small, but it wants a couple of table-driven cases (plain host, trailing
/v1,/v1in a path segment, no/v1at all) rather than a one-line change, and it is unrelated to the review classes #422 was addressing.Also worth checking in the same pass:
remote_openai/remote_litellmdo not do this, so this is the only site.