Review

Review

Multi-agent code review that works on local diffs or open PRs.

Step 1: Determine the Review Target

If $ARGUMENTS contains a PR number, review that PR:

gh pr view {number} --json title,body,headRefName,baseRefName,labels,state
gh pr diff {number}

If no PR number, review the local diff against main:

git diff -w origin/main...HEAD

Step 2: Dispatch Parallel Review Agents

Launch each one as a background agent. Each agent gets the diff and reviews independently.

Agent: Bug Scan Shallow scan for obvious bugs in the changed code. Focus on logic errors, wrong file paths, incorrect commands, broken references. Ignore style issues, linting, and type errors (CI catches those). Only report issues a senior engineer would flag.

Agent: Git History Context Read git blame and history of modified files. Check for conflicts with recent changes, patterns from previous PRs, and whether the changes are consistent with the file's evolution.

Agent: Shared Code Impact For each modified file, check all callers and consumers. Flag if a shared file was changed without considering all code paths that use it. This is the #1 cause of errors that make it into production.

Agent: Project Convention Check Check CI-specific rules:

• Branch name matches ^[a-zA-Z0-9-]+\/(release|hotfix)$ (if reviewing a PR)
• PR title follows conventional commit format, under 65 chars

Agent: Security and Secrets Check

• Scan for hardcoded secrets, credentials, or sensitive information in the diff. Flag any findings with high confidence.
• Check for OWASP Top 10 vulnerabilities in the changed code, such as SQL injection, XSS, or insecure deserialization patterns.
• Ensure that any changes to authentication or authorization logic follow best practices and do not introduce security weaknesses.
• Flag any use of deprecated or insecure APIs that could lead to security issues.
• Run configured static analysis tools for security and report any findings that are relevant to the changed code.

Step 3: Score and Filter

For each issue found, score confidence 0-100:

• 0-25: Likely false positive
• 25-50: Might be real, could not verify
• 50-75: Verified issue, but minor or unlikely to hit in practice
• 75-100: Verified issue, will impact functionality or was explicitly called out in CLAUDE.md

Step 4: Produce the Review

## Code Review

Found {N} issues:

1. {description} ({evidence source})

   {link to file and line, or quote from diff}

2. ...

If no issues score 75+:

## Code Review

No issues found. Checked for bugs, shared code impact, project conventions, and security issues.

Step 5: Resolve Existing Threads (PR only)

If reviewing an open PR, check for existing review threads:

gh api graphql -f query='{
  repository(owner: "{orgName}", name: "{repositoryName}") {
    pullRequest(number: {N}) {
      reviewThreads(first: 50) {
        nodes { id isResolved comments(first: 1) { nodes { body } } }
      }
    }
  }
}'

For each unresolved thread, check if the issue has been fixed in the current diff. If fixed, resolve the thread:

gh api graphql -f query='mutation { resolveReviewThread(input: {threadId: "{id}"}) { thread { isResolved } } }'