agentsclimarketplace

Pr review

Skill sujeet-pro/agents-devkit/plugins/adk/skills/pr-review

A plugin system for Claude Code that adds skills, agents, guidelines, repo-aware configurations, and MCP integrations to supercharge your development workflow.

Install
npx -y skills add sujeet-pro/agents-devkit --skill pr-review

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

HEAVYWEIGHT, GitHub-only deep PR review. Takes a GitHub PR URL, prepares a read-only git worktree at the PR head (SSH clone only), and reviews the diff with full cross-file context via Read/Grep/Glob. Fans out one agent per dimension (correctness, security, tests, performance, api, docs, observability, concurrency, feature-flow) through the Workflow tool, then adversarially verifies every finding before it survives. Traces feature flags / experiments via the statsig MCP and pulls linked Jira/Confluence via the atlassian MCP. Posts inline review comments, a review summary, and appreciations through the gh CLI after confirmation. NEVER merges, never force-pushes, never edits the PR's code. For a quick read-only pass with no worktree, use /adk:review.

SKILL.md

4.6 KB, as published. Nobody here has run it

pr-review — deep PR review (heavyweight, GitHub-only)

A real reviewer's pass on a GitHub pull request: a checked-out worktree, cross-file context, a multi-dimension fan-out with adversarial verification, feature-flag tracing, and inline comments posted back to the PR. This is the heavyweight path. For a quick read-only skim with no worktree, use /adk:review.

Scope is a GitHub PR URL only. https://github.com/<owner>/<repo>/pull/<n>. Bitbucket, GitLab, and self-hosted forges are explicitly out of scope — see rules.md. All GitHub access is the gh CLI; all git is git directly; cloning is SSH only.

The full operating contract lives in this skill folder — read these as you need them:

AspectFile
How you review (voice, severity rubric, evidence rules)persona.md
The phased process + the Workflow orchestrationworkflow.md
Hard rules + refusals + safetyrules.md
Per-dimension review guide + feature-flow tracingdimensions.md
Classifying pre-existing PR review threadscomment-resolution.md

Quick start

  1. Phase 0 — parse + auth. Parse the URL → (owner, repo, number). Verify gh auth status. (workflow.md Phase 0.)
  2. Phase 1 — checkout. Ensure an SSH clone exists, git fetch, add a detached git worktree at the PR head SHA. The worktree is read-only — you never edit the PR's code.
  3. Phase 2 — gather. PR metadata + diff + existing review threads via gh pr view --json / gh pr diff / gh api; linked Jira/Confluence via the atlassian MCP.
  4. Phase 3 — the Workflow. Fan out one agent per dimension over diff + worktree, trace feature flags via the statsig MCP, then adversarially verify and synthesize. (dimensions.md, workflow.md Phase 3.)
  5. Phase 4 — threads. Classify each existing review thread resolve / reopen / leave-as-is by re-checking the worktree at its anchor (comment-resolution.md).
  6. Phase 5–6 — triage + post. Triage findings, then render inline comments + a summary + appreciations via the gh CLI. Confirm before transmitting.

Workflow is the default

Every real PR gets the parallel-dimension Workflow in workflow.md: review fans out one dimension per agent → each finding is adversarially refuted by an independent skeptic → only survivors are reported. This catches what a linear pass misses and kills plausible-but-wrong findings. There is no "skip the workflow" fast path here — that's what /adk:review is for.

Posting policy + merge policy

  • Posting inline review comments is the purpose of this skill, so it happens by default in non-interactive mode — but you always summarize what will post ("about to post N inline comments + a summary on PR #X — proceed?") before transmitting, per rules.md. --no-post produces the report and posts nothing.
  • The verdict is approve when there are no surviving blocker/critical findings and no thread needs reopening; otherwise request changes or comment.
  • NEVER merge a PR. Even if asked, print the merge link and exit (rules.md). Recommend merge; the human clicks. Never force-push.

Modes

  • default — full review, post inline comments + summary + appreciations after a one-line confirmation.
  • -i / --interactive — walk each finding (accept / reject / edit) before anything is posted.
  • --scope security|correctness|tests|all — restrict the fan-out to one dimension family (default all).
  • --no-post — review and report only; transmit nothing to the PR.
  • --deep — stronger reasoning profile; auto-selected for security-sensitive, cross-module, or feature-flag-bearing diffs.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.