Examine: production-risk-first code review

Find what would break in production. Verify the change's claims instead of trusting them. Judge whether the solution is right-sized. Skip this skill for typo and doc-only changes.

Required references

Paths are relative to this file. <plugin root> is two directories above it.

Principles

• Trust nothing. The description, the diff, green tests, and your own candidates are claims. They become findings only after step 7.
• Judge against the project's rules, not generic practice. A finding cites its rule.
• No side effects. Read and run anything locally. Experiment in a detached worktree (git worktree add --detach) or outside the repo, never in the primary checkout. Never push, comment, trigger CI, or contact anyone unless the user asks during this review.

Target and mode

• PR: gh pr view <N>, gh pr diff <N>, and gh pr checks <N>.
• Branch or no argument: diff from the merge-base with the upstream, else with the local base branch. Add git diff HEAD when the tree is dirty.
• Range or path: exactly that. Read pre-change code with git show <merge-base>:<path>.
• Record the branch, HEAD, and git status --porcelain. The end check compares them.

/examine [quick|full] <target>. Default to full for migrations, auth, payments, personal data, dependencies, infra, or more than about 15 files. Otherwise use quick. State the mode and why. Quick runs steps 1 to 5 with axes 5a to 5d and 5i only, then steps 7 and 9. It skips host reviewers and the Occam pass, lists skipped axes under Not reviewed, and reports at most 8 Issues. Never pad toward the cap.

Workflow

Delegate independent passes to read-only subagents when the host has them. You own severity and synthesis.

1. Intent. Extract the problem, approach, constraints, and non-obvious decisions from the description, else from the commits. Read any linked ticket. A missing description is finding #1 (scope: PR). Mark inferred intent as inferred.
2. Rulebook. Read every agent instruction file that governs a changed file. Read in full the project docs the diff makes load-bearing: invariants, security, privacy, the touched spec, testing. Skim the rest.
3. Baseline, before opening the diff. Confirm the problem exists in the base snapshot. If it does not, or an existing utility already solves it, that is finding #1. Sketch two or three obvious approaches, one sentence each.
4. Approach gate. If the change matches a sketch, proceed. If it diverges, assume your sketch missed a constraint and look for it in the ticket, docs, history, and adjacent code. Found: note it under Verified. Not found: the approach question is the headline, and every line finding is provisional.
5. Audit. Read references/audit.md now and walk every axis that applies. Record each issue-shaped observation as a candidate with a one-line failure scenario. In full mode, also run the host's built-in review in a subagent on the reviewed checkout, as one more source of candidates. --no-host-reviewers skips it.
6. Occam pass (full mode, audit.md § 6). Ask whether the change is more solution than the problem needs.
7. Verify and sweep. Read references/verify.md now. Give every Medium or higher candidate an independent verdict from a subagent that sees only the diff, the files, and the candidate. Then make one fresh-eyes sweep for what the list missed.
8. Synthesize. Read ../../assets/writing.md and references/report.md now. Assign severities and build the report.
9. Report locally. Write the report to a file and run <plugin root>/scripts/prose-lint.py on it. Fix every finding in the text. If the host has a structured findings tool, call it once with the Issues. The terminal report is the deliverable. Post to the PR only if the user asks.

Done when

Every item in report.md § Definition of done holds, with evidence. If one does not, go back to the step that produces it.