Consolidate AI PR review into single Claude Code workflow

[DiamondDagger590]

Summary

Replaces the two-stage AI PR review pipeline (pr-review.yml + ai-persona-review.yml) with a unified claude-review.yml workflow powered by the Anthropic Claude Code Action. The new workflow posts a single sticky comment per review instead of multiple top-level comments, reduces per-push review spam, and enables on-demand re-review via @claude mentions.

Key Changes

• Removed: pr-review.yml (detect-changes + prepare-review jobs) and ai-persona-review.yml (persona orchestration)
• Added: claude-review.yml with two jobs:
  • auto-review: runs once when a PR opens non-draft or leaves draft (no per-push reviews)
  • on-demand: triggered by @claude mentions in PR comments or review comments
• Added: .github/claude-review-prompt.md — orchestration protocol defining how the review orchestrator routes to persona subagents, consolidates findings, and posts a single sticky comment with inline review comments for Important findings only
• Added: .claude/agents/review-*.md — four isolated persona subagents (extensibility, security, testing, and a placeholder for future lenses) that return structured findings to the orchestrator instead of posting comments directly
• Updated: .coderabbit.yaml to:
  • Switch profile from "assertive" to "quiet" (reduce noise)
  • Enable auto_incremental_review for incremental feedback
  • Add [skip-review] title keyword and skip-review label support (mirrors Claude workflow)
  • Update base branches to mainline + develop (remove recode)
  • Clarify that Claude Code review handles API extension points and test structure, so CodeRabbit should avoid duplicating those domains
• Updated: CLAUDE.md to document the new review workflow and persona agent definitions

Notable Implementation Details

• Single sticky comment: All findings (Important + nits) go into one tracking comment, with nits collapsed in a <details> block. Inline review comments are created only for Important findings.
• Re-review convergence: On @claude re-review, the orchestrator reads the previous sticky comment, passes previously-reported findings to each lens, and suppresses new nits to avoid review spirals on unchanged code.
• Concurrency control: New pushes or @claude mentions on the same PR cancel in-flight reviews via a concurrency group.
• No per-push reviews: Auto-review runs only on PR open/ready-for-review, not on every push. Re-review is on-demand via @claude mention.
• Persona isolation: Each subagent (extensibility, security, testing) runs in its own context with its own checklist, verifies findings against the actual checkout, and returns only structured blocks — no comments posted by subagents.
• Deduplication & capping: The orchestrator dedupes findings across lenses, ranks by severity, and caps at 10 total findings (5 nits max) to keep reviews focused.

Rationale

The old pipeline posted a fresh top-level comment per persona on every push, creating review spam and making it hard to track which findings were resolved. The new workflow:

• Posts once per review cycle (auto on open, on-demand via @claude)
• Consolidates all findings into a single sticky comment for easy tracking
• Reduces noise by suppressing new nits on re-review
• Enables concurrent persona analysis via Claude Code Action's Task tool
• Maintains the same review rigor (extensibility, security, testing lenses) with cleaner output

https://claude.ai/code/session_017RS1pjrpm7c5AdirmmAtwX

[github-actions]

Claude finished @DiamondDagger590's task in 4m 11s —— View job

---

McCore PR Review — claude-review.yml consolidation

• Gather context (PR diff via gh pr diff, changed files)
• Read .github/claude-review-prompt.md orchestration protocol
• Route to lenses per protocol
• Manual review of workflow/config changes (no lens matched)
• Post final review

Verdict: 1 important, 1 nit

No src/main/**/*.java (or any Java) files are touched by this PR, so per the routing table in .github/claude-review-prompt.md §2, none of review-security/review-testing/review-extensibility match — this PR only changes CI workflows, agent/prompt definitions, and CodeRabbit config. Per §2 that would normally mean "no reviewable code changes," but since the changed files are the review pipeline's own security boundary, I read them directly rather than skip straight to a one-liner.

Important findings:

• .github/workflows/claude-review.yml:74-80 — The new on-demand job (triggered by any @claude mention in a PR comment or review comment) has no fork/author-trust gate, unlike auto-review which explicitly checks github.event.pull_request.head.repo.full_name == github.repository (line 39). On a public repo, anyone who can comment on a PR — including a PR opened from a fork — can trigger this job. claude-code-action's tag mode operates on the PR's actual head commit, so it runs Bash/Read/Grep against that (potentially attacker-controlled) checkout with ANTHROPIC_API_KEY in the environment and pull-requests: write/issues: write permissions available — a classic "pwn request" pattern. The deleted ai-persona-review.yml had explicit comments guarding against exactly this (never checking out PR-head code from untrusted branches); that protection wasn't carried forward for the new on-demand path. Suggested fix: gate on github.event.comment.author_association being OWNER/MEMBER/COLLABORATOR, posted as an inline suggestion on that line. Fix this →

Nit (1)

• .github/workflows/claude-review.yml:9 — auto-review triggers on [opened, ready_for_review] only (intentionally dropping synchronize to avoid per-push spam). It also drops reopened, so a closed-then-reopened PR won't get an initial automated review and will need an @claude mention to kick one off. Likely fine given the on-demand path exists, but worth confirming it's intentional.

[coderabbitai]

Warning

Review limit reached

@DiamondDagger590, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 50 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: f707e2d3-b3e5-42c5-ad85-85d63d9165c7

📥 Commits

Reviewing files that changed from the base of the PR and between d07cc96 and cf64dde.

📒 Files selected for processing (9)

• .claude/agents/review-extensibility.md
• .claude/agents/review-security.md
• .claude/agents/review-testing.md
• .coderabbit.yaml
• .github/claude-review-prompt.md
• .github/workflows/ai-persona-review.yml
• .github/workflows/claude-review.yml
• .github/workflows/pr-review.yml
• CLAUDE.md

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/review-workflow-cleanup-g4z1dp

---

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.}

[github-actions]

IMPORTANT (security): on-demand has no fork/author gate, unlike auto-review which explicitly checks head.repo.full_name == github.repository (line 39).

issue_comment/pull_request_review_comment containing @claude can be posted by anyone who can comment on a public repo's PR — including on a PR opened from a fork. claude-code-action's tag mode checks out and operates on the PR's actual head commit, so this job would run Bash/Read/Grep against attacker-controlled code with ANTHROPIC_API_KEY in the environment and pull-requests: write / issues: write permissions available — a classic "pwn request" pattern. The deleted ai-persona-review.yml had explicit comments guarding against exactly this (never checking out PR-head code from untrusted sources); that protection wasn't carried over here.

Suggest gating on commenter trust, mirroring GitHub's own recommended pattern:

Suggested change

	on-demand:
	if: >-
	(github.event_name == 'issue_comment' &&
	github.event.issue.pull_request != null &&
	contains(github.event.comment.body, '@claude')) ||
	(github.event_name == 'pull_request_review_comment' &&
	contains(github.event.comment.body, '@claude'))
	on-demand:
	if: >-
	((github.event_name == 'issue_comment' &&
	github.event.issue.pull_request != null &&
	contains(github.event.comment.body, '@claude')) ||
	(github.event_name == 'pull_request_review_comment' &&
	contains(github.event.comment.body, '@claude'))) &&
	contains(fromJson('["OWNER","MEMBER","COLLABORATOR"]'), github.event.comment.author_association)