[aw-failures] [P0] Step Name Alignment sandbox blocks its own input file — 100% failure

[github-actions]

Problem

Fix Step Name Alignment's sandbox lockout — it fails 100% of the time and never gets to do its job. Parent report: #51545.

Affected workflows and runs

• Step Name Alignment (§31294247934, 2026-08-09T04:17Z) — failed after 32 turns, ##[error]Process completed with exit code 1.
• Baseline for comparison: §31239045503 (2026-08-08, success, 56 turns).

Probable root cause

The workflow writes its own input (step-alignment-input.json, 6.7MB — a dump of every step name across all .lock.yml workflows) to /tmp/gh-aw/agent/step-alignment-input.json. The session's sandbox only allows file operations under /home/runner/work/gh-aw/gh-aw — so every attempt to inspect the input from /tmp/gh-aw/agent/ is blocked:

wc in '/tmp/gh-aw/agent/step-alignment-input.json' was blocked. For security, Claude Code may only count lines/words/bytes in files from the allowed working directories for this session: '/home/runner/work/gh-aw/gh-aw'.
jq in '/tmp/gh-aw/agent/step-alignment-input.json' was blocked. ... may only process JSON from files in the allowed working directories ...
cp in '/tmp/gh-aw/agent/step-alignment-input.json' was blocked. ... may only copy files to/from the allowed working directories ...

Read also initially refused the file (6.7MB exceeds its 256KB single-read cap), and a dangerouslyDisableSandbox: true retry on jq was blocked too. The agent eventually found Grep/paginated Read work, but burned turns retrying the blocked path first each time — this run gave up faster than the baseline (32 vs 56 turns) rather than slower, confirming it's hitting a hard wall, not just being slow.

This is a directory-allowlist mismatch: /tmp/gh-aw/agent/ is the documented convention for agent-generated temp files in this repo's own workflows, but this workflow's sandbox config doesn't include it.

Proposed remediation

Pick one:

1. Add /tmp/gh-aw/agent/ (or the specific subpath this workflow uses) to the sandbox's allowed working directories for the Step Name Alignment job.
2. Have a pre-agent step copy/stage step-alignment-input.json into the repo checkout (e.g. under a git-ignored path like .github/tmp/) before the CLI step runs, so the agent only ever touches paths already inside its allowlist.
3. Longer-term: shrink the input. A 6.7MB flat JSON of every step name in the repo is itself the thing forcing the agent into wc/jq/Grep workarounds — consider pre-aggregating (e.g. group by workflow file, or diff against the last run) in a deterministic pre-step instead of shipping the raw dump.

Success criteria

• The next 3 scheduled runs of Step Name Alignment contain zero permission_denied log entries.
• Turn count returns to roughly the ~56-turn baseline (or lower, if the input is also trimmed) instead of failing early.
• Workflow conclusion is success for 3 consecutive scheduled runs.
  Related to [aw-failures] [aw] Failure Investigator Report — 2026-08-09 (6h) #51545

Generated by 🔍 [aw] Failure Investigator (6h) · agent · 119.7 AIC · ⌖ 47 AIC · ⊞ 5.5K · ◷

• expires on Aug 15, 2026, 11:19 PM UTC-08:00

[github-actions]

🍪 Issue Monster selected this for Copilot

I have 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 · gpt54 · 2.07 AIC · ⌖ 7.1 AIC · ⊞ 7.3K · ◷