About full-pr-review
Trigger only when the user or another skill explicitly invokes this skill
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ sztadii/ agent-skills --skill full-pr-reviewAfter installing, type/full-pr-reviewin Claude Code to confirm it loaded.
Bundled filesagents/openai.yaml
Details
From the project READMEFull PR Review
Scope
Run one review of the PR inferred from context and publish, at most once, only the comments the user accepts. This workflow requires a PR, not a standalone diff or commit range; no prior skill run is required.
Review bugs, regressions, unnecessary complexity, excessive tests and unused additions. Preserve author and unrelated work; isolate probes. Do not fix code, push, merge, approve, request changes, monitor the PR or start follow-up rounds; the author owns delivery.
Dependencies
Required: $proof, $council and $leaving-pr-comments.
Before starting, find each required skill in the agent's catalog or installed locations, read its SKILL.md and complete its dependency check. Repository links alone are insufficient. If any skill is missing or unreadable, stop this workflow and its caller; name the affected skills without substituting or installing.
Workflow
Review
- Identify the PR from context; ask if missing or ambiguous. If it is merged or closed, report that and stop. Read intent, repository rules, discussion and current reviews.
- Fetch its branch and base and record the head SHA. Use a separate review checkout when needed to preserve local work; exclude unpublished changes.
- Start
$proofand$councilin parallel on that snapshot and wait for both:$proof: acceptance criteria and required checks; run e2e tests only when the PR CI pipeline runs them. Reproduce suspected failures without fixing code. If execution is blocked, record the gap.$council: intent, scope, repository rules and user-selected reviewers/settings. Ask it to trace seemingly unused new files, helpers and dependencies through framework discovery, configuration and runtime loading; missing imports alone do not prove disuse. Tell it$proofruns in parallel so it does not rerun those checks. It owns independent review and finding validation; do not launch separate reviewers.
- Combine both results; validate and deduplicate findings. Separate introduced defects from existing issues and material findings from style preferences.
Comment
If there are no comment-worthy findings, report that and stop without publishing.
Otherwise show each proposed comment with a short ID (C1, C2, …), its exact
text, location and evidence, formatted per $leaving-pr-comments without
invoking it yet. Ask the user to accept all, some or none, and wait.
Publish only the accepted comments, with any user edits, in one pass through
$leaving-pr-comments as a single COMMENT review, passing the reviewed SHA and
text approval. Never publish rejected or unanswered drafts. Verify the published
URLs, then stop.
Result
Report the reviewed SHA, $proof/$council evidence and coverage gaps, findings,
accepted and rejected comments and published URLs. State that the PR is not
monitored and no approval was submitted. Explain blocked outcomes without invented
success.