[aw-failures] Matt Pocock Skills Reviewer: 'Pre-fetch PR diff' step fails with SIGPIPE (exit 141) on large diffs #33769

Problem

The Pre-fetch PR diff step (step 19) in the Matt Pocock Skills Reviewer agent job fails with ##[error]Process completed with exit code 141. (SIGPIPE, 128+13). The failure occurs before agent activation — the agent never produces a turn, so there is no useful audit output beyond the step-exit code. 9 failures out of 30 runs in the last 30 days (~30% failure rate); 6 of those failures occurred in the last 11 hours alone.

Affected workflow and runs (last 6h window)

Additional same-signature failures in the prior 24h (same workflow, same step, same exit code):

• §26206536101 (copilot/refactor-oversized-functions-parser-workflow)
• §26203507064 / §26203506520 (copilot/decouple-engine-permission-mode)
• §26202121212 / §26202120930 (copilot/chore-bump-mcpg-to-v0316-and-firewall-to-v02550)

Root cause

The Pre-fetch PR diff step in .github/workflows/mattpocock-skills-reviewer.lock.yml (around line 539) runs this shape:

set -euo pipefail
mkdir -p /tmp/gh-aw/agent
gh pr diff "$PR_NUMBER" --repo $EXPR_GITHUB_REPOSITORY \
  --exclude '**/*.lock.yml' \
  --exclude '**/generated/**' \
  --exclude '**/dist/**' \
  --exclude '**/build/**' \
  | head -n 3000 \
  > /tmp/gh-aw/agent/pr-diff.patch
...

With set -o pipefail, when the PR diff exceeds 3000 lines, head -n 3000 closes its stdin and exits 0 after reading 3000 lines. The still-running gh pr diff then receives SIGPIPE (signal 13) on its next write, exits 141, and pipefail propagates that 141 as the step exit code. The step fails in ~1 second, before the agent runs.

This is a classic set -o pipefail + | head pitfall. The failure consistently affects PRs whose post-exclude diff is large — i.e. exactly the PRs that the truncation is designed to protect against. Workflows that DON'T have large diffs succeed; workflows that DO have large diffs always fail this step.

Proposed fix

The source is the .md workflow file (.github/workflows/mattpocock-skills-reviewer.md); the .lock.yml is generated. Update the Pre-fetch PR diff step body to one of:

Option A (preferred — fewest invariants changed): swallow SIGPIPE only for this command:

set -euo pipefail
mkdir -p /tmp/gh-aw/agent
{ gh pr diff "$PR_NUMBER" --repo $EXPR_GITHUB_REPOSITORY \
    --exclude '**/*.lock.yml' \
    --exclude '**/generated/**' \
    --exclude '**/dist/**' \
    --exclude '**/build/**' \
    || true; } | head -n 3000 > /tmp/gh-aw/agent/pr-diff.patch

Option B: write full diff to a temp file, then truncate (no pipe, no SIGPIPE):

gh pr diff "$PR_NUMBER" --repo $EXPR_GITHUB_REPOSITORY \
  --exclude '**/*.lock.yml' --exclude '**/generated/**' \
  --exclude '**/dist/**' --exclude '**/build/**' \
  > /tmp/gh-aw/agent/pr-diff.full
head -n 3000 /tmp/gh-aw/agent/pr-diff.full > /tmp/gh-aw/agent/pr-diff.patch

Option C: localize the pipefail relaxation: set +o pipefail; gh pr diff ... | head ...; set -o pipefail.

Option A is the minimal change and keeps the streaming-truncate property. The same fix pattern should be checked in other agent workflows that use gh ... | head shapes with set -o pipefail.

Success criteria / verification

• The Pre-fetch PR diff step exits 0 on a PR whose post-exclude diff exceeds 3000 lines (e.g. re-run on copilot/fix-github-actions-job-agent).
• pr-diff.patch contains exactly the first 3000 lines of the (excluded) diff.
• Across the next 24h of scheduled runs, the workflow's failure-on-step-19 rate drops to 0.
• No regression: small-diff PRs continue to succeed and pr-diff.patch still has fewer than 3000 lines for them.

Confidence / unknowns

• High confidence on root cause: exit code 141 is unambiguous (SIGPIPE), the step body matches the pipefail + head pattern exactly, all observed failures occurred at step 19 in ~1 second with no errors in any other step, and the failures correlate with branches likely to have large diffs (refactor / regex-cleanup / chore-bump-mcpg).
• Unknown: whether other workflows in this repo use the same pattern and silently mask similar SIGPIPE failures (e.g., not in step-fail mode). Worth a quick grep -n '| head -n' .github/workflows/*.lock.yml follow-up scan.

Parent report: #33620

References:

• §26222142424
• §26222141685
• §26206536101
  Related to [aw-failures] Agentic workflow failures (last 6h): MCP telemetry regression + codex model 404 + 4 smaller fixes #33620

Generated by 🔍 [aw] Failure Investigator (6h) · ● 12.2M · ◷

• expires on May 28, 2026, 2:03 PM UTC