[aw-failures] [P1] Daily BYOK Ollama Test — local Ollama backend returns 503 on real inference despite passing warm-up

[github-actions]

Fix Ollama backend availability for the BYOK test — warm-up passes, then real inference gets HTTP 503.

§31542632254 — Daily BYOK Ollama Test — Execute GitHub Copilot CLI step failed: every chat-completion request to host.docker.internal:11434 (proxied) returned 503 with content_type: text/html (a generic gateway error page, not an Ollama JSON error) across all 4 harness retry attempts over ~4 minutes — even though the preceding "Install Ollama", "Pull model", "Warm up model", and "Verify Ollama BYOK readiness" steps all succeeded.

Probable root cause: the readiness check doesn't reflect the backend's real state once inference traffic starts. Either the Ollama server crashes/OOMs between warm-up and the first real request, or the proxy route to host.docker.internal:11434 drops after warm-up — the html 503 body points to a gateway/proxy-level failure rather than an Ollama-level error.

Proposed remediation:

1. Capture Ollama server logs/exit status for 2026-08-11 ~22:30–22:40 UTC to check for a crash or OOM between warm-up and the first real request.
2. Make the readiness check send a real (non-trivial) prompt through the same proxy path the CLI uses, not just a lightweight health ping.
3. If resource-constrained, reduce the model size or add 503-specific retry-with-backoff before failing the step.

Success criteria: Daily BYOK Ollama Test completes 3 consecutive scheduled runs without a 503 during real inference.

Analyzed run IDs: 31542632254 (audit). Parent: #52061.
Related to #52061

Generated by 🔍 [aw] Failure Investigator (6h) · agent · 221.7 AIC · ⌖ 47.4 AIC · ⊞ 5.3K · ◷

• expires on Aug 18, 2026, 5:28 PM UTC-08:00

[pelikhan]

@copilot add a step to test ollama outside the awf first, then investigate failure