[aw-failures] P1: Smoke Gemini uses TTS preview model `gemini-3.1-flash-tts-preview` → unknown_model_ai_credits, 100% agent fail [Content truncated due to length]

[github-actions]

Recommendation

Fix the Smoke Gemini engine model — it is set to a non-agent TTS preview (gemini-3.1-flash-tts-preview) that the AWF api-proxy has no pricing for, so every inference call is rejected and the smoke canary is 100% red.

Problem statement

Smoke Gemini fails at "Execute Gemini CLI" with ApiError type unknown_model_ai_credits: model gemini-3.1-flash-tts-preview has no AI-credits pricing and no default pricing configured in the api-proxy (apiProxy.defaultAiCreditsPricing unset / model absent from the pricing table). The proxy rejects every generateContent / generateContentStream call, so 0 agent turns run and the agent job fails.

Affected workflow and run IDs

• Workflow: .github/workflows/smoke-gemini.lock.yml (Gemini engine)
• Run IDs: §27480363197, §27478319753 — identical signature, 100% failure this window.

Evidence

• Both GeminiChat.sendMessageStream (Turn.run) and BaseLlmClient.generateJson (classifier route) throw the same ApiError after retryWithBackoff exhausts retries.
• token_usage.jsonl is 0 bytes; agentic_fraction = 0 — no model call ever completed.
• Two gemini-client-error artifacts per run with byte-identical payloads.
• (Incidental, non-causal: 94% firewall block rate; the proxy-allowed inference endpoint itself returned the pricing error.)

Probable root cause

gemini-3.1-flash-tts-preview is a text-to-speech preview model, not an agent chat model, and is unknown to the proxy's AI-credits pricing table. This is a misconfigured/unintended model selection for an agent smoke test.

Proposed remediation

1. Point Smoke Gemini at a supported Gemini agent model (the model the canary is meant to validate), not a TTS-preview SKU.
2. Add the model (or a defaultAiCreditsPricing fallback) to the api-proxy pricing table so an unknown model degrades to a clear pricing error earlier, not deep in retry exhaustion.

Success criteria / verification

• Smoke Gemini "Execute Gemini CLI" completes with ≥1 successful agent turn; no unknown_model_ai_credits error over a 24h window.

Context

Filed by [aw] Failure Investigator (6h) — analyzed run https://github.com/github/gh-aw/actions/runs/27484878458. Parent: Workflow Health Manager Issue Group #29109.
Related to #29109
Related to #29109

Generated by 🔍 [aw] Failure Investigator (6h) · 421.4 AIC · ⌖ 12.8 AIC · ⊞ 4.5K · ◷

• expires on Jun 20, 2026, 5:52 PM UTC-08:00

[copilot-swe-agent]

The agent encountered an error and was unable to start working on this issue: Please try again later, or contact support if the problem persists. (Request id: 2040:20D086:9BCFE7:2447A60:6A2E1264)

[github-actions]

🍪 Issue Monster selected this for Copilot

I've identified this issue as a good candidate for automated resolution and requested assignment to the Copilot coding agent.

If assignment succeeds, the Copilot coding agent will analyze the issue and create a pull request with the fix.

Om nom nom! 🍪

🍪 Om nom nom by Issue Monster · 34.3 AIC · ⌖ 4.51 AIC · ⊞ 22.7K · ◷