Specialized agents

Git Workflow includes eight specialized agents. The Markdown files in agents/ are canonical; setup installs generated Codex definitions into .codex/agents/ and Claude continues to discover the canonical files through its plugin package.

Agent	Purpose	Codex sandbox
`pr-reviewer`	Reviews code quality, security, performance, tests, and documentation	`read-only`
`silent-failure-hunter`	Finds swallowed errors and misleading fallbacks	`read-only`
`type-design-analyzer`	Evaluates invariants and type design	`read-only`
`pr-test-analyzer`	Identifies behavioral test-coverage gaps	`read-only`
`comment-analyzer`	Detects inaccurate comments and documentation	`read-only`
`version-delta-analyst`	Researches breaking changes and migration steps	`read-only`
`release-validator`	Runs pre-release checks and reports readiness	`workspace-write`
`qa-executor`	Executes QA plans and records results	`workspace-write`

Invocation

The workflow skills delegate to these agents when appropriate:

/review in Claude or $review in Codex fans out to the review specialists and waits for every report.
/release or $release delegates release validation and, when dependencies or APIs changed, version-delta analysis.
/start-qa or $start-qa delegates large QA plans to qa-executor.

You can also request an agent explicitly, for example: “Use the pr-reviewer agent to review these changes.” Routine edits do not automatically create agent sessions.

Maintaining agents

Edit agents/<name>.md, then regenerate Codex TOML files:

node scripts/generate-codex-agents.mjs
node scripts/generate-codex-agents.mjs --check

Generated definitions intentionally omit model and reasoning settings so they inherit the parent Codex session.