| name | review |
| description | Run the project's two reviewer agents in parallel against the current diff and report findings in chat. Never posts to GitHub. Use when Alex says "review my changes", "/review", or otherwise wants a quality check before commit, push, or merge. |
Review Skill
Run the two reviewer agents in parallel against the current branch's
diff vs main, plus any uncommitted edits in the working tree. Collate
their findings into a single chat report. Do not post to GitHub.
The reviewers are:
code-reviewer — module boundaries, naming, imports, configuration
access, logging, error handling, idempotency, type discipline (parse
at boundaries, discriminated unions, stringly-typed narrowing), and
test coverage (error cases, naming, no-skipIf, body-visible fixtures,
full-pipeline integration tests, parallel-LLM pattern, FP/FN
threshold direction).
llm-quality-reviewer — N-shot prompts, fixture content,
FP-over-FN bias, mandatory routing through services/llm.ts, token
tracking, failure modes, concurrency stacking.
Step 1: gather the diff
git diff --name-only main...HEAD
git diff --name-only
git diff --name-only --staged
Deduplicate into one list of changed paths. If the list is empty, say
so and stop — there is nothing to review.
Step 2: spawn both agents in parallel
Each agent already knows its own scope. Send both tool calls in a
single message so they run concurrently. Each prompt should give the
agent:
- The diff source — "the diff
main...HEAD plus any uncommitted
edits in the working tree".
- The list of changed paths.
- A reminder that its output is collated back into a single chat
report, so it should be concrete (file path, line number, fix) and
skip restating its own scope.
Don't try to pre-filter which agents to run based on which files
changed. If llm-quality-reviewer has no relevant files (no prompts
or fixtures touched), it will return "no issues found" — that's the
expected outcome and takes seconds.
Step 3: collate the report
Once both agents return, assemble a single chat report:
# Review of <diff source>
Changed files: <count> across <areas>.
## code-reviewer
<findings, or "No issues found.">
## llm-quality-reviewer
<findings, or "No issues found.">
## Summary
- <N issues found total, or "Clean across both reviewers.">
- Top priorities: <bullet the most impactful items, if any>.
Keep it tight. Pull out the concrete findings — don't paste raw agent
transcripts.
Notes
- Never post GitHub review comments. This skill produces a chat report
only.
- If an agent errors out or returns nothing, note that under its
section rather than silently dropping it.
- Re-running this skill is cheap and expected — Alex iterates.