fix(review-swarm): print why the swarm failed, not just that it did

[kjgbot]

The gate polls agent-relay cloud status --json, reads .status off the response and discards the rest. When a swarm fails, the only thing reaching the GitHub log is the word:

Review swarm did not complete successfully: failed

The reason is already in the payload the poll just fetched. For the five runs that failed on 2026-09-07 (34168392594, 34167663112, 34165035497, 34164872298, 34164770687) it was:

Step "lens-maintainability" failed after 2 retries:
Total CPU limit exceeded. Maximum allowed: 250.

Nothing in the log said "quota". Finding it meant pulling the run id out of the log and querying the run by hand — so the cause went uninvestigated for days while the failures were attributed to a guess. This is the observability gap, not the quota itself.

Change: in the wait step, when the terminal status is not completed, extract .result.error // .error from the response already in hand and print it to stderr and the step summary.

Verification — run against the real 04da7e48-87ec-4c7a-a1ee-22fd482e1cd1 payload:

• real payload → yields the quota text above
• {"status":"completed","result":{}} → empty string, not the literal null
• empty response (the first-call-failed path, where $response is unset) → empty, no error

Not merging this myself. The rule is a passing independent signoff at the exact head plus green CI, and the review swarm is exactly what cannot run right now — the failure this PR makes visible is the same one that blocks its own review. Wants a human or a swarm once capacity is back.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FtQSAcGDta5VH9xiZFT4sR

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Team

Run ID: c04df254-ca2a-40d8-9a06-1b1febfea1f2

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[cubic-dev-ai]

All reported issues were addressed across 1 file

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[github-actions]

Review swarm: maintainability

No fresh transcript was produced for run 0ded29b7-326f-47e8-a507-8e3bb5360850 (MISSING).

[github-actions]

Review swarm: history

No fresh transcript was produced for run 0ded29b7-326f-47e8-a507-8e3bb5360850 (MISSING).

[github-actions]

Review swarm: structure

No fresh transcript was produced for run 0ded29b7-326f-47e8-a507-8e3bb5360850 (MISSING).

[github-actions]

Review swarm: FAILED

• maintainability: MISSING
• history: MISSING
• structure: MISSING

Cloud run: 0ded29b7-326f-47e8-a507-8e3bb5360850

[kjgbot]

maintainability lens — PASS

Maintainability review — PR #235

Scope. Six lines of shell added after the poll loop in .github/workflows/review-swarm.yml:130 (new lines 130–147) to surface the failure reason from the last swarm status payload.

Concerns

1. response is reused from the polling loop and may not carry what the new code assumes. The comment claims "the reason … sits in the run payload we just fetched." That is only reliable for the failed/cancelled breaks. Two other paths through the loop reach this block:

   • status_error (review-swarm.yml:115-118, 119-122): the CLI or jq call errored, so response is either empty or the previous iteration's payload.
   • timed_out (deadline elapsed): the loop's last iteration observed a non-terminal state, so .result.error is legitimately empty and the swarm did not "fail" — it stalled.

   Today this degrades gracefully to no output because of // empty, so it is not silently misleading — but the comment asserts "print what we already have" as if we always have it, and a future edit that assumes the field is populated on failure will be wrong. Either tighten the guard to case "$status" in failed|cancelled) or update the comment to name the three ways response can arrive here.

2. Implicit contract with the CLI JSON schema, and no test would fail if it drifts. .result.error // .error // empty (review-swarm.yml:136) encodes two guessed paths into the agent-relay cloud status --json shape. If the schema renames to .failureReason or .result.failure.message, this falls through to empty and the exact problem the diff exists to solve returns — silently. There is no CI-testable pin on the CLI's failure contract. At minimum, note in the comment which CLI version/shape is expected, and echo "$response" (truncated) as a fallback so the reader is not left staring at an empty section.

3. Comment carries narrative that will not age well. Lines 131–135 are a six-line block containing a specific past-incident anecdote ("sent one reader inferring for days"). AGENTS.md endorses a Why: line, but the story-shaped rationale reads as commit-message material embedded in a workflow file — once the incident is forgotten, the paragraph becomes cryptic without adding rule-of-thumb value. Trim to one line stating what is being surfaced and why the poll loop can't just print it.

4. Markdown escaping in $GITHUB_STEP_SUMMARY. Fencing with plain triple backticks (review-swarm.yml:141-144) breaks if reason itself contains ```. Low probability, but the summary is human-facing and this is the kind of latent bug maintainability review is meant to catch. Consider ~~~ fences or a length cap.

Notes

• The added block belongs in scripts/swarm-post.sh (already invoked at line 137) rather than inline YAML; the diff would then be testable by a shell harness. Not a blocker for a six-line change, but the file trend is toward more inline shell.

Blockers

None. The change is additive, degrades to no-output on the schema/state edge cases, and does not affect the exit-code path.

REVIEW_PASSED

[kjgbot]

history lens — PASS

Blockers: none. PR #235’s commit 8f7a20f does not establish any of the three permitted grounds for rejection.

Notes:

• No recorded mistake is reintroduced. In .github/workflows/review-swarm.yml:130–147, the addition reads the existing response and prints its error to stderr and the step summary. It leaves the status output and final enforcement unchanged. DRIVE-LOG’s earlier failures involving lost transcripts, substring-based verdict parsing, and reviews using modified gates (ops/DRIVE-LOG.md:1152–1226, 1342–1364) are not reintroduced: this diff changes neither transcript handling nor the source or interpretation of verdicts.
• No new settled-decision contradiction. The diagnostic block at .github/workflows/review-swarm.yml:136–145 displays information; it does not make that information authoritative for completion or review approval. That fits RFC-0001 decisions flow/drive f59e279 08271341 #7 and drive: WP-11: repair PR #9 under review before anything else #11. It also introduces no new ability for reviewed agents to modify their judge under decision regressions: relaycast workspace-key repair answers an untyped 500 #6; the existing main-owned swarm definition and verdict scripts remain the review inputs.
• The commit’s scope description matches the diff. It claims to surface the failure reason, and .github/workflows/review-swarm.yml:136–145 does exactly that, including a top-level .error fallback. The commit does not claim to repair quota exhaustion, restore swarm capacity, or change product code.

Concerns: The commit and PR body describe verification against a real cloud payload, but do not include the literal verification command and its captured output. That falls short of the repository’s evidence-presentation standard. I did not independently retrieve that payload, so the historical execution claim remains uncorroborated; missing corroboration alone does not establish an untruth under this lens. Attaching the original command and output would resolve that concern.

This is a static HISTORY review; no runtime test result is claimed.

REVIEW_PASSED

[kjgbot]

structure lens — MISSING

[kjgbot]

🎯 review-swarm: FAILED (M:pass H:pass S:missing)

Lens transcripts posted as sibling comments above.