code-review

Files (3)

• SKILL.md
• openai.yaml · agents/
• smell-baseline.md · references/

SKILL.md (first lines)

# Code Review

Review the diff from a fixed point to `HEAD` without letting standards findings influence spec findings or vice versa.

## Pin the review range

1. Use the fixed point supplied by the user. Ask for it only when none was provided.
2. Resolve it before delegation and record the immutable commit SHA.
3. Capture these commands once:
   - `git diff <fixed-point>...HEAD`
   - `git log <fixed-point>..HEAD --oneline`
4. Stop early when the ref is invalid or the diff is empty.

Keep unrelated working-tree changes outside the review unless the user explicitly includes them.

## Find the spec source

Search in this order:

1. Issue, PRD, or pull request references in the review commits.
2. A path or tracker item supplied by the user.
3. A branch-matching file under `docs/`, `specs/`, or `.scratch/`.
4. The user's explicit statement that no spec exists.

Follow a repository tracker guide such as `docs/agents/issue-tracker.md` when present. Otherwise confirm the tracker and required profile or repository before fetching a referenced item. Run `gh auth status` before every GitHub lookup.

If no source is discoverable, ask where it is. If the user confirms none exists, skip the Spec subagent and report `No spec available`.

## Find the standards sources

Discover repository instructions and standards files that govern the changed paths, including applicable `AGENTS.md`, `CLAUDE.md`, `CONTRIBUTING.md`, and coding-standard documents.

Read [references/smell-baseline.md](references/smell-baseline.md) for the fixed judgement-call baseline. Repository rules override that baseline. Exclude rules already enforced deterministically by checks unless the diff shows the enforcement is bypassed.

## Run independent reviews

Launch the Standards and Spec subagents concurrently. Give both the pinned diff command and commit list, but do not give either subagent the other axis's prompt or findings.

### Standards brief

Give the subagent:

- Applicable standards-source paths and contents.
- The complete smell baseline.
- This task:

> Report every documented-standard violation by file and hunk, citing the source file and rule. Separately report baseline smells by name with the relevant hunk. Treat documented breaches as hard findings and baseline smells only as judgement calls. Let repository rules override the baseline. Skip deterministically enforced rules. Stay under 400 words.

### Spec brief

Give the subagent the spec path or fetched contents and this task:

> Report requirements that are missing or partial, behavior not requested by the spec, and implemented requirements whose behavior appears incorrect. Cite the relevant spec line for every finding. Stay under 400 words.

Each subagent is complete only when it has inspected the full pinned diff and either reported every finding in its axis or explicitly reported no findings.

## Aggregate without reranking

Present the reports under:

```markdown
## Standards

## Spec
```

Preserve axis-local severity and lightly clean formatting only. Do not merge, deduplicate across axes, or choose one overall worst finding.

End with one line containing:

- Finding count for each axis.
- Worst finding within Standards, if any.
- Worst finding within Spec, if any.

From deepaktiwari09/team-agent-skills, MIT.

More from deepaktiwari09/team-agent-skills