AI Moderator agent job fails on every run: pi + copilot/auto gets 400 model_not_supported, misreported as generic infrastructure_error #63113

Summary

Since ai-moderator.md moved to engine: pi / model: copilot/auto (144a253, #59695, 2026-09-09), AI Moderator's agent job has not succeeded in the run history I could see. The Pi provider sends model: "auto" to the api-proxy (POST http://api-proxy:10002/chat/completions), which rejects it:

provider_error provider=aw-gateway model=auto api=openai-completions ... error="400: {"message":"The requested model is not supported.","code":"model_not_supported","param":"model","type":"invalid_request_error"}"

The agent then emits report_incomplete / infrastructure_error ("All 1 Pi provider requests failed before safe outputs were emitted").

Numbers (AI Moderator, last 1,000 runs, 2026-09-15 → 2026-09-24)

• 260 failure, 107 success, 374 skipped, 259 action_required.
• 0 of the 107 success runs executed the agent job (all skipped at activation), so every run that reached the agent failed.
• model_not_supported found in the failed-job log for 5 of 7 sampled failures (including the newest and oldest in the window); the other 2 had no retrievable failed-step log.
• Same symptom in other engine: pi + model: copilot/auto workflows, e.g. duplicate-code-detector (17 of its last 20 runs failed; sampled failure 35924730146 shows model_not_supported).

Example run: https://github.com/github/gh-aw/actions/runs/35917931650

Why it isn't getting fixed: detection gap

The conclusion job's classifier logs Model not supported error: false for these runs even though the agent log contains model_not_supported. So they're filed as generic "[aw] AI Moderator reported incomplete result" issues (#62169, #62364, #62543, #62698, #62912, now #63068). Those accumulate bot-only comments, expire, and close as NOT_PLANNED, and no human sees the actual cause.

Prior art

#60867 found the same model_not_supported spike on Issue Monster (Pi resolving to aw-gateway/auto, 60 consecutive failures). It recommended an engine-level fix (fail fast / never send bare auto to the gateway) to be filed against the Pi model-resolution path. I couldn't find that follow-up issue.

Possibly separate

#63068's last agent output also shows Extension error (.../pi_steering_extension.cjs): Cannot read properties of undefined (reading 'steer').

Ask

1. Pin AI Moderator (and the other pi + copilot/auto workflows) to a concrete model, or make the gateway accept/resolve auto for the Pi provider.
2. Have the conclusion classifier recognize model_not_supported in Pi provider errors, so these runs are reported as model errors rather than generic incompletes.