Code Review

Review changes independently and return only defects that the author would likely fix. Treat findings as claims that need evidence, not suggestions to improve code in general.

Choose the execution role

Determine which role this invocation has before doing anything else:

• Reviewer session: The current agent was asked to review the repository, or the prompt says this is a synthesized reviewer session. Perform the workflow below yourself. Do not launch another agent.
• Delegating session: The user asked to obtain an independent review from a named CLI agent. Run scripts/review-with-agent from this skill. The launcher uses native review commands where possible and explicitly invokes this skill for synthesized fallbacks.

Native review commands are process entry points, not tools a reviewer session must recursively call. Do not launch a second copy of the current agent merely because its CLI has native review support.

Delegate non-interactively

Run the launcher from the repository being reviewed:

scripts/review-with-agent --agent codex --base main
scripts/review-with-agent --agent claude --base main
scripts/review-with-agent --agent agy --commit HEAD
scripts/review-with-agent --agent agy --uncommitted

Paths beginning with scripts/ are relative to this skill directory. Use --dry-run to inspect the selected command without starting an agent. Use --synthesized to test the portable workflow even when a native route exists.

For a direct CLI invocation without the launcher, use the same distinction:

codex exec review --base main
claude -p "/code-review main...HEAD" --permission-mode dontAsk --no-session-persistence
agy --print "Use the installed agent-review skill to review the current branch against main. Perform the review yourself; do not delegate." --mode plan --sandbox

The launcher currently selects:

Agent	Direct route	Synthesized fallback
Codex	Native base-to-HEAD, commit, and uncommitted review	Explicit $agent-review prompt
Claude	Local /code-review for a base range or single commit	Explicit /agent-review prompt
agy	None	Explicit natural-language prompt

The agent-review name is intentionally distinct from Claude's bundled /code-review, so Claude can keep using that local reviewer as its direct route. Do not add ultra or substitute claude ultrareview: those launch a cloud-hosted multi-agent service with materially different execution and cost. agy has no documented forced skill-invocation syntax, so its fallback names the agent-review skill explicitly; verify skill loading in verbose logs when activation matters.

Perform a synthesized review

Read references/review-rubric.md completely before reviewing.

1. Confirm the repository and requested target. Support these target shapes:
   • Base branch: changes from the merge base of the requested base and HEAD through the current working tree.
   • Base plus head revision: committed changes from the merge base through the requested head revision.
   • Commit: changes introduced by one commit.
   • Uncommitted: staged, unstaged, and untracked changes.
2. Read every applicable repository instruction file before judging a changed file. More-specific instructions override broader ones.
3. Inspect repository state and the diff summary before reading the full diff. Review large diffs in bounded sections rather than truncating one large command result.
4. Read surrounding implementation, callers, tests, schema, and history as needed to prove or disprove each candidate finding. The diff identifies scope; it is not the only context available.
5. Run focused tests, type checks, linters, or small reproductions when they materially increase confidence. Do not edit source files. Normal disposable build and test outputs are acceptable.
6. Re-check every candidate against the rubric, deduplicate overlapping findings, and ensure each reported location overlaps changed lines.
7. Return the findings in the rubric's output format. If none qualify, say so explicitly.

Handle review results

Treat every finding from any reviewer as unverified until checked against the code. When the user asks to act on findings, record each disposition as accepted, rejected with evidence, or partially adopted. Re-run a fresh review after fixes instead of resuming the earlier reviewer session.

Reference Material

• Review Rubric: Finding thresholds, severity definitions, comment formats, and verdict reporting. See references/review-rubric.md.