Make Copilot Task-tool "No model available" failures non-fatal

[Copilot]

Copilot-engine workflows that spawn subagents via the Task tool fail 100% of the time: the SDK throws No model available. Check policy enablement under GitHub Settings > Copilot before any inference request is dispatched. The harness classified this as generic partial_execution, so it burned all 4 attempts reproducing the identical error and failed the job.

This implements remediation option 3 from the issue — a stopgap that keeps runs green while the policy/entitlement question is resolved separately.

Changes in actions/setup/js/copilot_harness.cjs

• Detection — NO_MODEL_AVAILABLE_PATTERN / isNoModelAvailableError() match the SDK's policy-gate message.
• Classification — new no_model_available failure class in classifyCopilotFailure, ranked below model_not_supported so the more specific configured-model error still wins. Surfaced in the per-attempt diagnostic line as isNoModelAvailableError=….
• No retry — the retry loop short-circuits on this signature, alongside the existing mcp_policy_blocked / model_not_supported persistent-error branches.
• Graceful degradation — subagent delegation is an optional capability:
  • if safe-outputs already contain a terminal output (e.g. the review was posted), exit 0 with a reason specific to blocked delegation;
  • otherwise emit a missing_tool signal for task so the unavailable capability is visible downstream instead of being lost in a red job.

[copilot-harness] attempt 1 failed: failureClass=no_model_available isNoModelAvailableError=true …
[copilot-harness] attempt 1: no model available for sub-agent delegation but safe-outputs already
  contain expected output — suppressing terminal verdict (delegation skipped, core work succeeded)

Tests

copilot_harness.test.cjs covers the detector (including negatives against the unrelated The requested model is not supported message), classification precedence, and both branches of the missing_tool emitter.

Not addressed here

Why subagent sessions resolve no model (options 1/2 in the issue — entitlement/allowlisting, or disabling Task for Copilot-engine workflows) remains open. This change only ensures the failure mode is cheap and non-fatal.

• Fixes [aw-failures] Fix Copilot CLI Task-tool subagent calls — 100% failure with No model available #51033

[pelikhan]

@copilot investigate why this is failing

[Copilot]

@copilot investigate why this is failing

Investigated — this isn't a test or build failure. Every workflow run on this branch has conclusion action_required, i.e. blocked awaiting maintainer approval, so none of them actually executed:

31215609704 CJS              | completed action_required | 45b6fda
31215650254 Agentic Commands | completed action_required | 45b6fda
31215541985 CJS              | completed action_required | 6b52d46
31215543793 Agentic Commands | completed action_required | 6b52d46
...

The only run that executed is Running Copilot cloud agent (31214768843), which succeeded. get_status on the PR head returns state: pending, total_count: 0 — no check has reported a result yet.

I re-ran every job the CJS workflow would run, locally, on 45b6fda:

Check	Command	Result
js-typecheck	npm run typecheck (tsc --noEmit)	pass
js-typecheck	make check-cjs-syntax	pass
js-tests	npm run test:js -- --no-file-parallelism	pass — 390 files, 12784 passed, 31 skipped
lint-js	make lint-cjs	pass — 0 errors (192 pre-existing warnings, none in the changed files)
lock freshness	make recompile	pass — clean tree, no diff

copilot_harness.test.cjs alone: 271/271 passing.

Two things worth noting from the investigation:

• make recompile produces no diff, which confirms copilot_harness.cjs is not inlined into .lock.yml — the lock files invoke it from ${RUNNER_TEMP}/gh-aw/actions/copilot_harness.cjs, so this change needs no workflow recompilation.
• The new NO_MODEL_AVAILABLE_PATTERN doesn't collide with detect_agent_errors.cjs, whose model_not_supported_error classifier matches only The requested model is not supported. A suppressed no_model_available run therefore won't be re-flagged downstream after the harness exits 0.

One false alarm on my side that's worth recording: the four copilot_requests_proxy_auth_403.md ENOENT failures I first hit were my own — promptsSourceDir is path.resolve("../md"), so the suite must be run from actions/setup/js, and CI additionally stages the templates into $RUNNER_TEMP/gh-aw/prompts. They pass when invoked correctly.

Approving the pending workflow runs should turn the PR green.

[github-actions]

Hey @Copilot 🎯 — thanks for the quick remediation on the Task-tool No model available failures! This PR looks great and ready for review.

Strengths:

• ✅ Clear problem statement — the issue ([aw-failures] Fix Copilot CLI Task-tool subagent calls — 100% failure with No model available #51033) documents 100% failure rate across multiple workflows, and your fix directly addresses the root cause (policy gate before model resolution).
• ✅ Focused scope — you're adding detection, classification, and graceful degradation for the specific error signature without changing unrelated logic.
• ✅ Comprehensive tests — copilot_harness.test.cjs covers the detector pattern (including negatives against similar-looking messages), classification precedence, and both output branches (missing_tool emission and early exit when safe-outputs exist).
• ✅ Clear explanation — the PR body walks through each change and explains why options 1/2 (entitlement/allowlisting or disabling Task) are left for follow-up.

Next steps:

• When ready, toggle from draft to ready for review
• Assign to pelikhan (already in requested reviewers) for final check
• After merge, monitor the next 5 runs of any Copilot-engine workflow using Task to confirm the stopgap is working

Generated by ✅ Contribution Check · auto · 78.4 AIC · ⊞ 8.7K · ◷