[aw-failures] [P1] Fix Go Logger Enhancement's 15-minute timeout loop on oversized manifest.json #52865

Problem

Stop letting the agent burn its entire 15-minute budget re-reading a file it already knows is too big — cap the retries and fail fast instead. Go Logger Enhancement has now failed 3 of its last 4 daily runs (2026-08-11, 2026-08-12, 2026-08-15), and this run's failure was a hard step timeout with zero forward progress.

Affected workflow and runs

• Workflow: Go Logger Enhancement (.github/workflows/go-logger.lock.yml)
• Failed run: §31861677993 (2026-08-15 03:23 UTC), job 94956121140, step Execute Claude Code CLI
• Comparator (last success): §31768069259 (2026-08-14)
• Prior failures same pattern likely: §31665298469 (2026-08-13), §31561379584 (2026-08-12) — not yet confirmed to share this exact root cause, worth a quick check.

Probable root cause

The agent tried to inspect /tmp/gh-aw/agent/go-logger/manifest.json (763.6KB / ~35K tokens — both over the 256KB file-size and 25K-token read limits) and, instead of adapting, looped through six different failing approaches in ~40 seconds: Read (twice, hitting token limit both times), cd into the directory (blocked — outside allowed working directory), jq (blocked — same reason), cp (blocked), wc (blocked). None of these succeeded, and the agent never fell back to offset/limit on Read as the error message explicitly suggests. The harness itself never logged a claude-harness exit/closed event — the job just idled until GitHub Actions killed the step after the hard 15-minute limit (##[error]The action 'Execute Claude Code CLI' has timed out after 15 minutes.).

audit-diff against the last successful run (§31768069259) shows no firewall/network drift (0 anomalies) and far fewer GitHub API rate-limit points consumed (18 vs 693 core points) — confirming this run did negligible real work before stalling, not an infra/connectivity regression.

Proposed remediation

1. Have the go-logger workflow pre-summarize or pre-chunk manifest.json before handing it to the agent (e.g., emit a smaller index/summary file under the token limit), so the agent never needs to read the raw 763KB file directly.
2. Add a prompt instruction telling the agent to use offset/limit on the first Read failure instead of retrying with other tools.
3. Add a harness-level guard: if the same tool-call pattern repeats N times with the same size-limit error, fail fast with a clear error instead of idling to the 15-minute timeout — this makes future occurrences (in any workflow) diagnosable in seconds instead of burning the full step budget.

Success criteria / verification

• Go Logger Enhancement completes Execute Claude Code CLI well under 15 minutes on its next 3 scheduled runs.
• No manifest.json-size tool-result errors appear in the transcript.

Parent report: #52570
Analyzed run: §31861677993
Related to #52570

Generated by 🔍 [aw] Failure Investigator (6h) · agent · 210.5 AIC · ⌖ 9.45 AIC · ⊞ 5.6K · ◷

• expires on Aug 21, 2026, 11:15 PM UTC-08:00