[aw-failures] P1: Copilot CLI aborts at CAPI layer — model-not-supported (4 daily workflows, regression) + max-LLM-invocations ( [Content truncated due to length]

[github-actions]

Problem statement

Five agentic workflows on the GitHub Copilot CLI engine aborted at step "Execute GitHub Copilot CLI" (agent job, exitCode=1) during a single 6h window. audit shows the aborts originate at the model-inference / CAPI layer, not from tool/permission denials — permissionDeniedCount=0, guard.tool_denials_exceeded absent, isMCPPolicyError=false, isAuthError=false, 0 firewall blocks. This is a distinct root cause from #39667 (which is guard.tool_denials_exceeded).

Two sub-signatures:

• (A1) CAPIError 400: The requested model is not supported (isModelNotSupportedError=true) — 4 of 5 runs, each aborting on attempt 1 in ~14–17s with no retry and 0 tokens. The configured model is unavailable for the current subscription tier.
• (A2) CAPIError 429: Maximum LLM invocations exceeded (50/50) — 1 run (Code Simplifier), hit the per-run LLM-invocation / max-turns cap (~51 turns, ~995k tokens, ~14.5m) after exhausting all harness retries.

Affected workflows and run IDs

Workflow	Run	Signature
Daily Compiler Quality Check	§27664093833	A1 model-not-supported
Daily Firewall Logs Collector and Reporter	§27663612723	A1 model-not-supported
Metrics Collector - Infrastructure Agent	§27664135353	A1 model-not-supported
PR Description Updater	§27664727661	A1 model-not-supported
Code Simplifier	§27666488465	A2 max-LLM-invocations (50/50)

Evidence — confirmed regression (A1)

audit-diff on Daily Compiler Quality Check comparing the last healthy run §27592420137 (2026-06-16) vs the failed run §27664093833 (2026-06-17):

• failed run: turns=1, token_usage=0 (aborted before any model work)
• no firewall anomaly (has_anomalies=false; only a benign sentry-domain volume shift)
• GitHub API usage normal

The workflow ran cleanly on this schedule for ≥5 consecutive days and began failing on 2026-06-17 — i.e. an environment/model-availability regression, not a workflow-content change.

Per-run audit details

• A1 runs: abort at attempt 1, ~14–17s, isModelNotSupportedError=true, isCAPIQuotaExceededError=false, timeout_detected=false, permissionDeniedCount=0, firewall blocked_requests=0.
• A2 run (27666488465): CAPIError 429 Maximum LLM invocations exceeded (50/50), ran ~11–14m across 4 retry attempts (each 5 internal retries) before giving up; not a billing/quota denial (isCAPIQuotaExceededError=false).

Probable root cause

• A1: the model id these workflows request is no longer served for the account/tier on the Copilot CLI engine, so CAPI returns 400 model not supported immediately. Because the harness treats this as fatal (no retry), every scheduled run fails. Likely a model-availability/deprecation change or a model-id drift in the shared engine config that landed on/around 2026-06-17.
• A2: Code Simplifier's task genuinely exceeds the 50-invocation cap (large diff / exploratory behaviour), so it exhausts the per-run budget rather than completing.

Proposed remediation

1. A1 — pin a supported model: verify the configured Copilot model id against the tier's currently-served models and update the shared engine config to a supported id (or fall back automatically). Add a fast preflight that fails the compile/lock step with a clear message when the configured model is not in the served set.
2. A1 — surface model-not-supported distinctly: the model_not_supported_error output is already set on the agent job — wire it into the failure summary so this stops being misattributed to generic agent failure.
3. A2 — bound the work or raise the cap: scope Code Simplifier to a smaller diff/file budget, or raise its per-run LLM-invocation limit, and have it emit partial progress before hitting 50/50.

Success criteria / verification

• Re-run the 4 A1 workflows: agent job completes without 400 model not supported; isModelNotSupportedError=false.
• A preflight rejects an unsupported model id at compile/lock time rather than at run time.
• Code Simplifier completes (or cleanly partial-completes) without 429 Maximum LLM invocations exceeded.
• No CAPI-layer aborts across these 5 workflows for 48h.

Scope note

This is not #39667 (guard.tool_denials_exceeded): that signature did not recur in this window (tool denials = 0 on every run here).

Parent: #29109. Filed by the [aw] Failure Investigator (6h), lookback 2026-06-17 ~02:37–08:37Z.
Related to #29109

Generated by 🔍 [aw] Failure Investigator (6h) · ◷

• expires on Jun 24, 2026, 12:52 AM UTC-08:00