j-review - SKILL.md Agent Skill

name: j-review description: Run pragmatic tests first, then get a second-opinion code review from another AI agent (claude, gemini, or codex) using the current git diff. Use when you want a fresh, correctness-focused review without heavy setup.

/j-review

Run pragmatic tests and then get a code review from a different AI agent for a fresh perspective.

Arguments

[file...] — (optional) specific files to review
--agent <name> — (optional) force a specific reviewer: claude, gemini, or codex

If no files are given, and we are on a branch different from main, review the complete branch. If we are on main, review the current diff (staged if any, otherwise unstaged).

Instructions

1. Run tests first (pragmatic)

Prefer a project test runner if present.
Be pragmatic: avoid heavy installs or services just to test a small change.
If tests are too heavy or unavailable, state what you did and why.

Suggested order:

If justfile exists: just test or a targeted recipe.
Else if package.json exists: pnpm test (or the repo's declared equivalent).
Else if pyproject.toml exists: just test or uv run pytest if already configured.
Else if go.mod exists: go test ./....
Else: run nothing and explain.

Capture test output in a variable or temp file.

2. Determine what to review

If file arguments provided: review those files
If there are staged changes (git diff --cached --stat): review the staged diff
If there are unstaged changes (git diff --stat): review the unstaged diff
If no changes: ask the user what to review

Capture the diff or file contents into a variable.

3. Pick the reviewer

If --agent specified, use that agent
Otherwise, pick a different agent than yourself:
- If you are Claude: use gemini
- If you are Gemini: use codex
- If you are Codex: use claude

Prefer the strongest available model for review if the CLI supports model selection.

4. Run the review

Include test output and the diff. Use non-interactive / print mode for each agent.

Unset ANTHROPIC_API_KEY and OPENAI_API_KEY when calling claude or codex so they use their own CLI auth instead of API-key mode.

Keep the review lean. Default codex exec is verbose — high reasoning, exploratory file reads via sed/nl/rg, and full test runs before writing a word. A 4-paragraph plan can produce 130KB of transcript and take 10 minutes. For a text-only review of piped content, scope it down.

Prompt directive (use for every agent):

"Review only the text shown. Do not read other files. Do not run commands, tests, or searches. Be terse — flag real issues, not style nits. Under 300 words."

Invocations:

PROMPT="Review only the text shown. Do not read other files. Do not run commands, tests, or searches. Be terse — flag real issues, not style nits. Under 300 words. Tests: $TEST_OUTPUT"

# Codex — read-only sandbox, low reasoning, capture final message only
OUT=$(mktemp)
printf '%s\n\n%s\n' "$PROMPT" "$content" | env -u ANTHROPIC_API_KEY -u OPENAI_API_KEY codex exec \
  -s read-only -c model_reasoning_effort=low --ephemeral -o "$OUT" - > /dev/null 2>&1
cat "$OUT"; rm -f "$OUT"

# Claude
echo "$content" | env -u ANTHROPIC_API_KEY -u OPENAI_API_KEY claude -p "$PROMPT"

# Gemini
echo "$content" | gemini -p "$PROMPT"

Use codex review --uncommitted (dedicated subcommand) only when you genuinely want codex to explore the whole working tree as part of the review. For a text-only "here's a diff, what's wrong" — use the lean codex exec invocation above. It keeps time and output size bounded.

5. Present the results

Print the reviewer's name and their feedback. Do not editorialize or filter the response.

Rules

Never review your own output — always use a different agent
Keep the review prompt focused on correctness, not style
Do not modify any files — this is read-only