[aw-failures] 6h Failure Report: 3 claude-engine runs failed (2× no safe output, 1× cache memory blocked)

[github-actions]

Executive Summary

In the 6 hours ending 2026-05-04T07:41Z, 3 of 29 runs failed (10.3% failure rate, all claude engine). Two distinct failure clusters identified. No blocked firewall domains contributed to Step Name Alignment or Schema Consistency Checker; Multi-Device Docs Tester had 2 firewall blocks.

Failure Clusters

#	Workflow	Run	Turns	Cost	Root Cause
1	Schema Consistency Checker	§25304247413	61	$1.91	Agent exited without calling any safeoutputs tool
2	Multi-Device Docs Tester	§25304397726	81	$1.93	Agent exited without calling any safeoutputs tool; multiple permission blocks (playwright, /tmp writes, settings.json)
3	Step Name Alignment	§25301975942	31	$1.10	Cache memory file blocked by security policy (/tmp/gh-aw/cache-memory/ not in allowed dirs); yq glob commands blocked; exhausted max turns

Evidence

Cluster A: Missing Safe Output (Schema Consistency Checker + Multi-Device Docs Tester)

Both workflows completed their full analysis work but never called noop, create_discussion, or any safeoutputs tool. This caused:

1. output_types output → null (empty)
2. detection job skipped (condition: output_types != '' → false)
3. safe_outputs job skipped (condition: detection.result == 'success' → false)
4. Workflow conclusion: failure

Key logs from detection system evaluation:

Evaluating: (always() && (needs.agent.result != 'skipped') && (((needs.agent.outputs.output_types != '') || (needs.agent.outputs.has_patch == 'true'))))
Expanded: (true && ('failure' != 'skipped') && ((null != '') || ('false' == 'true')))
Result: false

For Multi-Device Docs Tester, the agent also encountered:

• "Output redirection to '/tmp/preview.log' was blocked" — dev server log writes
• "Contains simple_expansion" — bash security policy block
• "Execute skill: playwright-cli" — playwright skill blocked
• "Claude requested permissions to write to /home/runner/.claude/settings.json" — settings write blocked
• "Unhandled node type: string/array" — parser errors in tool results

Despite 81 turns of browser testing attempts (curl checks returned 200 for 8 pages), the agent never called noop before exiting.

Cluster B: Cache Memory Access Blocked (Step Name Alignment)

The agent tried to read its cache from /tmp/gh-aw/cache-memory/step-name-alignment/patterns.json but Claude Code's security policy restricts file access to /home/runner/work/gh-aw/gh-aw:

"cat in '/tmp/gh-aw/cache-memory/step-name-alignment/patterns.json' was blocked.
For security, Claude Code may only concatenate files from the allowed working directories
for this session: '/home/runner/work/gh-aw/gh-aw'."

Additionally, yq commands using glob patterns (.github/workflows/*.lock.yml) were blocked as "multiple operations requiring approval". The agent correctly called missing_data(data_type=cache_memory, reason=cache_memory_miss) but then hit error_max_turns after 31 turns.

The audit reported: "evidence": "plan tool_call tool_call tool_call tool_call tool_call tool_call tool_call tool_call tool_call tool_call tool_call tool_call tool_call error finish" confirming the max-turns exhaustion pattern.

Existing Issue Correlation

Note: GitHub issue read API is returning 403 in this context; could not verify existing open issues. Sub-issues should be inspected for duplicates before acting.

Proposed Fix Roadmap

Priority	Issue	Action
P1	Schema Consistency Checker never calls safeoutputs	Workflow definition needs explicit safeoutputs call in all exit paths
P1	Multi-Device Docs Tester never calls safeoutputs + permission failures	Add noop call; expand allowed file paths in settings.json for /tmp writes
P2	Step Name Alignment cache memory blocked	Grant /tmp/gh-aw/cache-memory/ read access or rewrite cache lookup to use allowed path

References:

• §25304247413 Schema Consistency Checker failure
• §25304397726 Multi-Device Docs Tester failure
• §25301975942 Step Name Alignment failure

Generated by [aw] Failure Investigator (6h) · ● 545.5K · ◷

• expires on May 11, 2026, 7:51 AM UTC

[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 · ● 22.2K · ◷