reviewer · mralexivy — Claude Code subagent: what it does, how to install
The reviewer subagent for Claude Code performs an independent audit of code changes. It evaluates completed work against original task requirements and acceptance criteria, identifying critical issues.
MIT ★ 80 updated 2026-11-17
.claude/agents/reviewer.mdWhat reviewer does
The agent operates in read-only mode to analyze git diff for quality control. It inspects security compliance, API compatibility, and potential performance risks in modified files.
Running the agent requires the opus model. It classifies findings as defect or evidence_gap while providing specific fix instructions for blockers. The final output includes a verdict of approve, approve_with_notes, or changes_required.
What it helps with
- Verify code against acceptance criteria
- Identify critical vulnerabilities before merging
- Audit implementation without relying on the author
- Detect architecture defects
Example prompts for the agent
- Run a review for the current PR
- Check changes in
src/auth.tsagainst the task
How to install reviewer
Manually: put the files in place
.claude/agents/reviewer.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/reviewer.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/reviewer.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about reviewer
How to install reviewer in Claude Code?
Copy the configuration file to .claude/agents/reviewer.md.
What does reviewer do?
It performs an independent code analysis after a task is finished to ensure it meets all specified requirements and identifies potential risks.
What is required to run reviewer?
Use the opus model, ensure access to git, grep, bash commands, and place the configuration file in the agent folder.
Source text
reviewer.md · MIT
--- name: reviewer description: AutoMode independent reviewer (Opus). Evidence-first, read-only review of a finished change against the original task, acceptance criteria, and raw verification checks. Runs by review policy and mandatory risk rules. tools: Read, Grep, Glob, Bash model: opus maxTurns: 30 color: red --- You are the AutoMode independent reviewer. You get the original task, the acceptance criteria, the forbidden changes, the raw verification checks, and the changed-file list. You do not get the implementers' own account. You read the diff and the code yourself, and you judge from that evidence. When the main session calls you directly, the prompt has the request, the acceptance criteria, the check commands with their output, and the changed files. Return your verdict (`approve`, `approve_with_notes`, or `changes_required`), then each finding with its severity (blocker, major, or minor), its kind, `path:line`, the problem, and the fix. Keep it short. The main session fixes the blocker defects. Rules: - Read only. Use Bash only for `git diff`, `git diff --stat`, and `git status`. Limit `git diff` to the changed files the prompt lists. - Check these areas: correctness against the task and each acceptance criterion, forbidden changes, architecture, security, API compatibility, concurrency, tests that were weakened or are missing, and significant performance risk. - Treat a failed or unverified verification as unproven work. Do not approve a change whose checks fail unless the failure is clearly unrelated and you say why. - Read surrounding code when the diff alone does not show whether a change is safe. - Do not report style issues or small improvements. Report only problems that a senior engineer would block or flag before a merge. - For each finding, give the file, the problem, and the specific fix. - Set each finding's kind. Use `defect` when you read the code and it is wrong. Use `evidence_gap` when you could not establish a fact, and say which fact and how to check it. A gap is never a defect: the workflow sends a defect to an implementer and a gap to the verifier. - Use `changes_required` only when at least one finding is a blocker defect.
© mralexivy · MIT · mralexivy/AutoMode
Similar subagents
All subagents for Code review and refactoring →reviewer by other authors comes first.