[aw-failures] [aw-fix] P1: Daily BYOK Ollama Test — self-hosted `qwen2.5:0.5b` has no AI-credits pricing; api-proxy returns 400 [Content truncated due to length]

[github-actions]

Fix the AI-credits pricing gate for the BYOK path — add apiProxy.defaultAiCreditsPricing so self-hosted qwen2.5:0.5b stops getting rejected with HTTP 400.

The Daily BYOK Ollama Test has failed every daily run for 8+ consecutive days (100% AR). The Ollama setup succeeds (model pulls, /v1/models ready); the agent dies at Execute GitHub Copilot CLI because the api-proxy blocks the model on pricing, not connectivity.

Sub-issue of #46846 (same unknown_model_ai_credits 400 mechanism, different workflow/model).

Problem statement

All three copilot-harness attempts fail with an identical HTTP 400 from the api-proxy:

400 400 400 Model "qwen2.5:0.5b" has no AI credits pricing and no default pricing is configured.
Set apiProxy.defaultAiCreditsPricing in the AWF config (e.g. {"input": 3.0, "output": 15.0})
to provide a fallback rate, or add the model to the pricing table.

Exit code 1 → agent job failure. Output flag unknown_model_ai_credits is set; isHTTP400ResponseError=false, failureClass=partial_execution.

Affected workflow and run IDs

• Workflow: Daily BYOK Ollama Test (.github/workflows/daily-byok-ollama-test.md / .lock.yml)
• Representative run: §30050463506 (2026-07-23)
• Persistence (nearest prior failures, all same signature by run history): 29963539174, 29874295528, 29784505320, 29618245341 — failing daily back to at least 2026-07-14.

Probable root cause

The workflow declares max-daily-ai-credits: 10000, which activates the proxy's maxAiCredits accounting. The BYOK model qwen2.5:0.5b (set via COPILOT_MODEL) is a self-hosted Ollama model and will never appear in the built-in pricing table. Per pkg/workflow/schemas/awf-config.schema.json, when maxAiCredits is active and the model is unrecognized and no apiProxy.defaultAiCreditsPricing fallback is set, the proxy rejects with HTTP 400 unknown_model_ai_credits. No fallback is configured, so the request is rejected before any inference — a config gap, not an Ollama/network fault.

Evidence (agent-stdio.log, run 30050463506)

• Ollama install/serve/pull//v1/models readiness steps all pass.
• 3× copilot-harness attempt N failed: exitCode=1 failureClass=partial_execution each with the qwen2.5:0.5b ... no AI credits pricing 400 body.
• Trailing `503 for (apiproxy/redacted) lines are post-failure shutdown noise, not the cause.

Proposed remediation

1. Primary: Add an apiProxy.defaultAiCreditsPricing fallback to the workflow's AWF config. Since Ollama is self-hosted with no real per-token cost, {"input": 0, "output": 0} is the natural rate; any nonzero rate also works if credit accounting should still be exercised.
2. Alternative: Register qwen2.5:0.5b in the pricing table used by the proxy.
3. Consider making the BYOK path skip the AI-credits pricing gate entirely when COPILOT_PROVIDER_BASE_URL points at a self-hosted endpoint — this generalizes the fix to any BYOK model (ties to [aw-failures] [aw-fix] P1: Copilot auto model has no AI-credits pricing — api-proxy returns 400 and fails Smoke Copilot / AI-c [Content truncated due to length] #46846).

Success criteria / verification

• Next scheduled Daily BYOK Ollama Test run completes with agent job success (no unknown_model_ai_credits 400).
• copilot-harness reaches inference (tokenCount > 0) instead of failing all retries at pricing.
• Re-run on demand confirms the agent produces output against qwen2.5:0.5b.
  Related to [aw-failures] [aw-fix] P1: Copilot auto model has no AI-credits pricing — api-proxy returns 400 and fails Smoke Copilot / AI-c [Content truncated due to length] #46846

Generated by 🔍 [aw] Failure Investigator (6h) · age00 · 176.2 AIC · ⌖ 34.9 AIC · ⊞ 5.2K · ◷

• expires on Jul 30, 2026, 5:23 PM UTC-08:00

[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 · gpt54 · 2.98 AIC · ⌖ 6.73 AIC · ⊞ 8.1K · ◷