Pr Classify

PR classify

Review the current branch (or a specified PR) and produce structured feedback in the user's canonical classification scheme. The user is a Staff Engineer who explicitly wants to filter out nitpicks and focus on meaningful engineering concerns.

Preconditions (before any review work)

1. Auth. Run gh auth status. If unauthenticated, stop and tell the user to run gh auth login before continuing. Do not fall back to a local-branch-only review silently.
2. PR target. A PR number/URL or a checked-out branch with an open PR is required. If invoked as a bare review with no PR context, ask for the target instead of guessing.
3. Open state. Verify the PR is still open: gh pr view <id> --json state,mergedAt. If it is merged or closed, surface that and confirm the user still wants the review before proceeding.

Steps

1. Identify the PR scope. If invoked on a branch, run gh pr view --json title,body,baseRefName,headRefName,files or git diff <base>...HEAD. If invoked with a PR number/URL, fetch via gh pr view <id>.
2. Read the tests first, then the diff fully. Start with the tests and the PR description to recover intended behavior — what the change is supposed to do — then read the implementation against that intent, not just changed-line context. Cross-reference touched files for callers and tests.
3. For each finding, decide the bucket BEFORE writing it. If it's borderline Important/Optional and you can't articulate the cost of ignoring it, drop it.
4. Group findings by bucket, not by file. Within each bucket, order by severity.
5. Cite file:line for every comment. Never describe an issue abstractly.
6. End with a summary verdict: approve / request changes / needs discussion. One sentence rationale.

Classification rules

Critical

Correctness, security, reliability. Examples:

• Logic bug that breaks the happy path or a documented edge case
• SQL injection, XSS, auth bypass, secret leak
• Race condition, unhandled promise rejection in hot path
• Data loss / migration that can't roll back
• Regression in existing test coverage

If you write more than ~3 Critical comments on a normal-sized PR, re-check — you're probably over-classifying.

Important

Maintainability, scalability, readability. Examples:

• Coupling that will hurt the next change in this area
• Missing test for non-trivial branch logic
• N+1 query, O(n²) where O(n) is trivial
• Naming that misleads future readers (not "I'd prefer X")
• Abstraction that leaks implementation across module boundary

Optional

Style, preferences, "would be nice". Examples:

• Could be more functional / could destructure here
• Comment phrasing
• Consistency with a pattern used elsewhere — only when the inconsistency genuinely costs something

If a finding is Optional and the cost of ignoring it is "none" — do not include it. This is the explicit anti-nitpick rule from persona.md.

Confidence

Classification is severity, not certainty — track the two separately. When you are not sure a finding is real (you could not fully trace the path, or it depends on context you cannot see), say so and phrase it as a question, not an assertion: "Is x guaranteed non-null here? If not, this throws." A confident wrong Critical costs more trust than a hedged correct one. Never inflate confidence to make a finding land; never drop a real concern because you are only 60% sure — ask.

Delegation

Delegate the reading, keep the ruling. Run the same gather/judge sequence every review, not only on large diffs — repeatability is the point:

1. Orient. A cheaper-tier subagent maps the changed surface (the codebase-map lens): entry points, touched invariants, risky areas. Compact map back, not file dumps.
2. Gather by angle, in parallel. One cheaper-tier subagent per angle — correctness, tests (the testing-checklist lens), security (the security-pass recon lens, only when the changed surface is sensitive), performance. Each returns compact candidate findings with file:line, not conclusions.
3. Judge. Classification, the confidence call, and the verdict stay on the main model. A candidate finding from a subagent is evidence, not a ruling.

Scale the number of gather subagents to the diff; never skip the orient-gather-judge split or fold judgment into a subagent.

Output format

## Summary
<one paragraph: scope, overall direction, verdict>

## Critical
- `path/to/file.ts:42` — <issue> — <why it's critical> — <suggested fix>

## Important
- ...

## Optional
- ...

## Verdict
<approve | request changes | needs discussion> — <one sentence>

If a bucket is empty, write _None._ rather than removing the heading.

Posting inline comments

The classified output above is for the chat. Posting it as inline PR comments is a separate, opt-in step.

• Draft first, post on confirmation only. Show the drafted inline comments and wait for the user to explicitly confirm. Never post to GitHub before that.
• Conversational English, plain prose. Write each comment the way a peer would leave it, not as a classification dump. English unless told otherwise.
• No symbols, no AI tells. Do not use arrows (→), tildes (~), or em-dashes. Strip AI-sounding phrasing — run the text through the humanizer skill's rules before posting.
• Anchor every comment to file:line via the gh review API; keep the Critical/Important/Optional context but phrase it naturally.