[aw-failures] Fix: 'general-purpose' review subagent 'No model available' hard-fails whole job (regression of #49022)

[github-actions]

Warning

Threat Detection Engine Failure — The analysis engine could not complete. This is a tooling failure, not a security finding.

What happened

The threat detection engine failed to produce results.

Review the workflow run logs for details.

Stop letting a spawned general-purpose review subagent take down the whole run when Copilot model allocation fails — this is a confirmed regression of #49022, hitting 5 runs across 2 workflows in one window

Problem

The task tool spawns a background general-purpose subagent (used by PR Code Quality Reviewer for parallel/first-pass review, and by daily-experiment-report). When the Copilot harness can't allocate a model for that subagent, it logs [copilot-sdk-driver] error: Execution failed: Error: No model available. Check policy enablement under GitHub Settings > Copilot and the entire job fails — either by exhausting all 4 harness retries, or by burning the full step timeout while retrying.

This is item #2 from #49022 ("Fleet-wide: Copilot CLI subagent model allocation fails with 'No model available'"), which was closed as fixed. It is recurring with the identical signature.

Affected workflows and runs (2026-08-06, 08:43–12:59 UTC)

• PR Code Quality Reviewer:
  • §31102988000 — 4/4 retries fail identically, failureClass=partial_execution, job fails.
  • §31102532937 — same, 4/4 retries fail identically.
  • §31100482108 — same, 4/4 retries fail identically.
  • §31102469614 — same signature on attempt 1, but the step then hits ##[error]The action 'Execute GitHub Copilot CLI' has timed out after 15 minutes instead of exhausting retries cleanly — a second observable failure mode of the same root cause.
• daily-experiment-report:
  • §31086066320 — same general-purpose / No model available signature; attempt 1 alone takes 3m35s before the harness stops (not retrying).

Probable root cause

The background subagent's model resolution path (through the task tool → general-purpose agent type → Copilot model routing) resolves to a model class that GitHub Copilot policy doesn't currently authorize for this context — most likely a routing-table entry (see sonnet-6x/agent/small/large aliases in the AWF config) that doesn't match whatever the subagent actually requests, or an org Copilot policy gap for that specific model. Because this fails deterministically on every attempt (not intermittently), it is a resolution/policy bug, not transient flakiness — consistent with #49022's original diagnosis, meaning the prior fix likely did not durably close this path (only the sibling idle-timeout failure mode may have been addressed).

Proposed remediation

1. Re-open or re-diagnose why [aw-failures] Fleet-wide: Copilot CLI subagent model allocation fails with 'No model available' (hit PR Code Quality Reviewer, L [Content truncated due to length] #49022's fix didn't hold for the model-allocation failure mode specifically (as opposed to the idle-timeout failure mode it also described).
2. Verify GitHub Copilot org policy enablement for whatever model alias the general-purpose subagent resolves to when spawned via the task tool.
3. Make background review-subagent failures non-fatal: if the task-tool subagent can't get a model after 1–2 quick retries, degrade to skipping that sub-review (or falling back to inline review) rather than failing the whole job / burning the full step timeout.

Success criteria / verification

• Next 20 PR Code Quality Reviewer and daily-experiment-report runs show zero No model available / failureClass=partial_execution job failures originating from a task-tool subagent spawn.
• If the underlying model-allocation issue can't be eliminated immediately, verify instead that a subagent failure of this kind no longer fails the parent job (graceful degrade), and reference this issue + [aw-failures] Fleet-wide: Copilot CLI subagent model allocation fails with 'No model available' (hit PR Code Quality Reviewer, L [Content truncated due to length] #49022 in the fix commit.
  Related to [aw-failures] Failure Investigator Report — 2026-08-06 11:50–13:15 UTC (6h) #50887

Generated by 🔍 [aw] Failure Investigator (6h) · agent · 177.3 AIC · ⊞ 5.2K · ◷

• expires on Aug 13, 2026, 5:34 AM UTC-08:00