[aw-failures] [aw] Failure Report 2026-05-03: 3 failures (1 P0 chronic + 2 P1 regressions)

[github-actions]

Overview

Failure investigation covering the 6h window ending 2026-05-03T07:27Z. Found 3 failed runs out of 34 total (8.8% failure rate). Two workflows are P1 regressions; one is a P0 chronic failure spanning 7 consecutive days.

⚠️ GitHub issue correlation was skipped — gh CLI is not authenticated in this environment. New issues may duplicate existing tracking; please cross-check the agentic-workflows open issues before acting on sub-issues.

Failure Clusters

#	Workflow	Run ID	Engine	Duration	Severity	Root Cause
1	GitHub Remote MCP Authentication Test	§25271765723	Copilot / gpt-4.1-mini	2.2m	P0 chronic	Unsupported model — 7 consecutive failures
2	Schema Consistency Checker	§25271731933	Claude Code	7.5m	P1 regression	max_turns (60) hit; playwright CLI migration expanded scope
3	Artifacts Summary	§25271895822	Copilot / claude-sonnet-4.6	17m	P1 first-run	Infeasible pagination strategy; never reached safeoutputs

Evidence

Cluster 1 — GitHub Remote MCP Authentication Test (P0 chronic)

7 consecutive failures (2026-04-27 → 2026-05-03); no successful run in history.

Error from agent-stdio.log:

● Request failed (transient_bad_request). Retrying...
● Request failed (transient_bad_request). Retrying...
400 The requested model is not supported.

• Engine: Copilot CLI v1.0.40, model gpt-4.1-mini
• Turns: 0 (agent exits immediately before any work)
• Firewall: 2 requests blocked from (unknown) domain (likely the model endpoint)
• All runs: Apr 27, Apr 28, Apr 29, Apr 30, May 2, May 3

Cluster 2 — Schema Consistency Checker (P1 regression)

First failure after 6 consecutive successful runs. Triggered same day as commit 15b7dd5 (playwright CLI migration).

Error from agent-stdio.log:

{"subtype":"error_max_turns","stop_reason":"tool_use",
 "total_cost_usd":2.0151049,"terminal_reason":"max_turns"}

Audit-diff vs last success (run 25245322929 → 25271731933):

• Removed: safeoutputs.create_discussion (was called 1× in successful run — never reached in failed run)
• 85 new bash investigation tools (all # Check playwright..., # Check mcp..., schema inspection) that weren't in the previous run
• GitHub API core consumption: +1,671% (21 → 372 quota units)
• Turns: 61 (hit the 60-turn cap) vs unknown baseline
• Cost: $2.02 for a failed run

Root cause: The playwright CLI migration commit added new schema fields/patterns that Schema Consistency Checker tried to fully validate, causing an exploratory bash loop that consumed all 60 turns before reaching the reporting step.

Cluster 3 — Artifacts Summary (P1 first-run)

First-ever run of this workflow; immediately failed.

Agent approach observed in logs:

1. Tried gh api repos/github/gh-aw/actions/artifacts?per_page=100 — got partial data
2. Attempted to paginate 20+ pages of artifacts to build 30-day summary
3. Tried to build run_id → workflow_name mapping by paginating 500+ runs
4. Recognized "at 500 runs per ~4 hours, it's ~90k runs for 30 days — This isn't feasible"
5. Tried alternative strategies but never produced output; safe_outputs job was skipped

Key metrics: 17 minutes, 20 turns, 843K tokens (842K cache reads = 48.5% cache hit rate)

Proposed Fix Roadmap

Priority	Issue	Workflow	Action
P0	Unsupported model gpt-4.1-mini	GitHub Remote MCP Authentication Test	Replace model with supported alternative
P1	max_turns regression from playwright migration	Schema Consistency Checker	Add turn budget or pre-agent deterministic data collection
P1	Infeasible pagination strategy	Artifacts Summary	Rewrite with deterministic pre-agent artifact collection

References:

• §25271765723 — GitHub Remote MCP Authentication Test failure
• §25271731933 — Schema Consistency Checker failure
• §25271895822 — Artifacts Summary failure

Generated by [aw] Failure Investigator (6h) · ● 401.4K · ◷

• expires on May 10, 2026, 7:37 AM UTC