ci: let the Claude review fan out, and make an unfinished review visible - #1439
Merged
Conversation
The claude-review action's allowlist had no `Task`, so on a large diff the reviewer fails silently. On #1437 (~3800 lines across 6 files) it announced "dispatched parallel sub-reviews of the three logical chunks", had all 12 of those calls denied, then exited with is_error: false and 17 turns — leaving a sticky comment that still read "Review in progress" with an unticked checklist and no findings. The check went SUCCESS regardless, because it reports that the action ran, not that a review happened. So the PR presented as reviewed-and-clean when nothing had been reviewed. Two runs failed identically, ~2 minutes apart; it is not a flake. Smaller diffs are unaffected — #1433 reviewed fine because it never chose to fan out. Two changes: - Allow `Task`, so the fan-out the reviewer already attempts can actually run. - Tell it, in the prompt, to say so in the tracking comment when it cannot finish. Its system prompt already asks it to report missing permissions and it did not, and a silent stall is the expensive failure here: a green check over a half-written comment is worse than a red one. Cost note recorded inline: fan-out means several sub-agents per review, and a single non-fanned-out run on that PR cost ~$1. Drop `Task` again if that proves too expensive, accepting that large PRs then need splitting to be reviewable. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019iCzZAuCMsZS7UJYfy3htd
|
This was referenced Sep 5, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Based on
master, independent of stack #1438, so it can merge first — the othertwo PRs then rebase onto it and get a review that can actually complete.
The bug
The
claude-reviewallowlist had noTask, so on a large diff the reviewerfails silently.
On #1437 (~3800 lines across 6 files) it announced "Dispatched parallel
sub-reviews of the three logical chunks", had all 12 of those calls denied
(
permission_denials_count: 12), then exitedis_error: falseafter 17 turns —leaving the sticky comment reading:
…and no findings, ever.
The check went
SUCCESSregardless, because it reports that the actionran, not that a review happened. So the PR presented as reviewed-and-clean
while nothing had been reviewed. Two runs failed identically ~2 minutes apart
(2m20s and 1m49s), so it is not a flake.
Smaller diffs are unaffected: #1433 reviewed fine because it never chose to fan
out. That is what makes this nasty — it only bites the PRs most in need of
review, and it looks like success.
The fix
1. Allow
Task. The reviewer already tries to fan out on large diffs; thislets it. Effective allowlist was
Glob, Grep, LS, Read, mcp__github_comment__update_claude_comment, Bash(git …)plus the workflow'sghcommands — no
Task.2. Tell it to report an unfinished review. Its system prompt already asks it
to explain missing permissions ("so that the user can update your
--allowedTools") and it did not. The prompt now says explicitly: if you cannotfinish, name what blocked you in the tracking comment, and do not leave an
unticked "Review in progress" checklist.
That second change is the more general one.
Taskfixes this instance; a greencheck over a half-written comment is the failure class, and the next cause
won't be permissions.
Cost — worth a conscious decision
Fan-out means several sub-agents per review. The single non-fanned-out run on
#1437 cost ~$1, so budget a multiple of that on large diffs. Recorded in an
inline comment next to the setting, with the fallback stated: drop
Taskagainif it proves too expensive, accepting that large PRs then need splitting to be
reviewable at all.
Verification
claude_argsand theon.pull_request.types: [opened, synchronize]triggers confirmed unchanged viayq.review should complete instead of stalling. Worth watching that specific PR
rather than assuming.
Follow-up
reopenedis not in the trigger types, which is why closing/reopening a PRre-runs
Testsbut not the review. Not changed here — adding it would re-run abilled review on every reopen. Noting it because it cost me a confusing round.
This PR was generated with the help of AI, and reviewed by a Human
🤖 Generated with Claude Code
https://claude.ai/code/session_019iCzZAuCMsZS7UJYfy3htd