Skip to content

revert(ci): drop the automatic PR review trigger, keep the author gate - #191

Merged
jfrench9 merged 1 commit into
mainfrom
chore/revert-claude-auto-review
Aug 16, 2026
Merged

revert(ci): drop the automatic PR review trigger, keep the author gate#191
jfrench9 merged 1 commit into
mainfrom
chore/revert-claude-auto-review

Conversation

@jfrench9

Copy link
Copy Markdown
Member

Summary

Reverts the automatic PR review trigger added earlier today. It fired on every pull request and never once posted a review, while each run concluded success and cost $0.20–$1.12.

The comment-triggered path was never broken. /create-pr posts @claude please review this PR on every PR it opens, and that path demonstrably works. A working control invoked by the command beats a broken one wired to an event.

What went wrong

Three obstacles, hit in sequence:

  1. Anti-tampering guardclaude-code-action refuses to run when the workflow file on a PR branch differs from the default branch, so the rollout PRs could not test themselves.
  2. Token permissions — the job declared pull-requests: read; posting a comment is a write. Fixed, and it was not the cause.
  3. Tool permission denials — with the trigger live and write granted, a run still posted nothing: num_turns: 33, total_cost_usd: 1.12, permission_denials_count: 15. These are Claude Code's own tool-permission layer rather than GitHub rejections, and the root cause is unresolved.

Changes

.github/workflows/claude.yml

  • Removed the pull_request trigger.
  • Removed the concurrency group. It existed to de-duplicate draft/ready toggles on that trigger; on comment-driven runs it would cancel a legitimate in-progress review, so it is actively wrong without the trigger.
  • Removed the automation prompt, restoring the commented-out original.
  • Restored pull-requests / issues to read (robosystems only — the repo where they were bumped).

Kept deliberately

The author_association gate stays. Several of these workflows had no gate at all and would run for any commenter on a public repository — a real hole that predates today's work and is unrelated to the trigger. All four comment arms remain restricted to OWNER/MEMBER/COLLABORATOR.

Breaking Changes

None. CI configuration only.

Testing

Each workflow parsed with yaml.safe_load and asserted on: no pull_request trigger, no concurrency block, no pull_request arm in the job gate, no automation prompt, permissions at read, and the four author gates intact.

The pull_request trigger added earlier today fired on every PR and never
posted a single review. Three obstacles in sequence: the action's
anti-tampering guard, then insufficient token permissions, then 15 tool
permission denials that remain unexplained. Every run concluded "success"
while producing nothing, at $0.20-$1.12 each.

The comment-triggered path was never broken, and is what /create-pr uses. A
working control invoked by the command beats a broken one wired to an event.

Reverted: the pull_request trigger, the concurrency group (it existed to
de-dupe draft/ready toggles on that trigger, and on comment-driven runs it
would cancel a legitimate in-progress review), the automation prompt, and the
token permission bump that turned out not to be the fix.

KEPT deliberately: the author_association gate. Several of these workflows had
no gate at all and would run for any commenter on a public repository. That
hole was real and predates today's work.
@jfrench9
jfrench9 merged commit 1642239 into main Aug 16, 2026
1 check passed
@jfrench9
jfrench9 deleted the chore/revert-claude-auto-review branch August 16, 2026 00:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant