Oh review
Skill open-horizon-labs/repo-native-alignment/.agents/skills/oh-review
Review a PR against its linked issue requirements, post structured feedbackFrom its SKILL.md
npx -y skills add open-horizon-labs/repo-native-alignment --skill oh-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 4 stars4 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.
SKILL.md
5.3 KB, as published. Nobody here has run it
oh-review
Review a PR, check it against the linked issue's requirements, and post structured feedback as a GitHub review. Read-only except for the review comment itself — no code changes, no commits.
Invocation
/oh-review <pr-number>
<pr-number>- the pull request number to review
Prerequisites
- Repo context: Run from the repo root where the PR exists
- GitHub CLI:
ghmust be authenticated
Flow
-
Load project background from
AGENTS.mdand relevant.oh/artifacts. Use RNArepo_mapandsearchfirst for repository exploration. If RNA is empty or incomplete, broaden queries and remove filters; before any Read/Grep/Bash fallback, document the reason and record a friction event in the repository session context. -
Fetch PR metadata:
gh pr view <pr-number> --json title,body,headRefName,baseRefName,additions,deletions,changedFiles,state,mergeableAbort if PR is not open.
-
Read the linked issue (CRITICAL) — the issue is the source of truth for requirements:
- Parse "Closes #N" / "Fixes #N" / "Resolves #N" from PR body
- Fetch the full issue:
gh issue view <N> --json title,body - Extract acceptance criteria, goal, context, constraints
- If no linked issue found, note "No linked issue — cannot verify requirements" and review code quality only
-
Fetch the PR diff:
gh pr diff <pr-number> -
Review the diff against:
- The linked issue requirements first — are acceptance criteria met? Does the implementation match the goal? Are constraints respected?
- Code quality (bugs, missing error handling, edge cases)
- Consistency with existing patterns in the repo
- Missing tests or documentation
-
Write a structured review to a temp file:
## Review: PR #<number> — <title> ### Requirements Check (from #<issue>) - [ ] <acceptance criterion 1> — met/not met/partially met - [ ] <acceptance criterion 2> — met/not met/partially met ### Blockers (P0-P1) - <issue with explanation and suggested fix> ### Improvements (P2-P3) - <issue with explanation> ### Clean - <what looks good> ### Follow-up Work - <concrete items that should be separate issues/PRs> -
Post the review on the PR:
-
If blockers found or requirements not met:
gh pr review <pr-number> --request-changes --body-file /tmp/review-<pr-number>.md -
If no blockers and requirements met:
gh pr review <pr-number> --approve --body-file /tmp/review-<pr-number>.md
-
-
Signal completion:
- Success with blockers:
signal_completion(status: "blocked", blocker: "PR has N blockers — see review comment") - Success clean:
signal_completion(status: "success", message: "Approved — N improvements noted") - Error:
signal_completion(status: "error", error: "<reason>")
- Success with blockers:
Review Principles
- The issue is the spec. The review must check the PR against the issue requirements, not just review code in isolation. A PR that passes code review but misses requirements is a failure.
- Be opinionated but structured — blockers vs improvements vs nits
- Read the full diff, not just file names — actually understand the changes
- Follow-up items should be concrete enough to become issues (not vague "consider refactoring")
- Do NOT auto-create follow-up issues — just post them as review comments. The human decides what to act on.
- Keep the skill read-only except for the review comment itself — no code changes, no commits
Exit Conditions
- Success: Review posted, PR approved (no blockers)
- Blocked: Review posted with request-changes (has blockers)
- Error: Could not fetch PR, diff, or post review
Completion Signaling (MANDATORY)
CRITICAL: You MUST signal completion when done. Call the signal_completion tool as your FINAL action.
Signal based on outcome:
| Outcome | Call |
|---|---|
| Review posted, PR approved | signal_completion(status: "success", message: "Approved — N improvements noted") |
| Review posted, changes requested | signal_completion(status: "blocked", blocker: "PR has N blockers — see review comment") |
| Unrecoverable failure | signal_completion(status: "error", error: "<reason>") |
If you do not signal, the orchestrator will not know you are done and the session becomes orphaned.
Fallback: If the signal_completion tool is not available, output your completion status as your final message in the format: COMPLETION: status=<status> message=<msg> or COMPLETION: status=<status> error=<reason>.
Example
$ /oh-review 99
Fetching PR #99...
PR: "Fix validation bug in auth module"
State: open, +147 -23, 4 files changed
Parsing linked issue...
Found: Closes #42
Fetching issue #42...
Issue: "Fix validation bug in auth module"
Acceptance criteria:
1. Empty string validation returns error
2. Tests cover edge cases
3. No changes to public API
Fetching PR diff...
Reviewing diff against requirements...
Posting review...
gh pr review 99 --approve --body-file /tmp/review.md
Review posted:
Requirements: 3/3 met
Blockers: 0
Improvements: 2
Follow-ups: 1
signal_completion(status: "success", message: "Approved — 2 improvements noted")
Done.
ARGUMENTS: