Subagent Review Orchestrator

Purpose

Create narrow project-scoped subagent definitions that can be explicitly spawned for review, research, QA, and trajectory auditing.

Inspect First

• existing .codex/agents
• review docs, security docs, browser QA docs, trajectory risk artifacts, package manifests, architecture docs, and validation scripts
• whether each proposed agent has real repo artifacts to inspect

Procedure

1. Suggest useful subagents.

   • Run scripts/suggest_project_subagents.py <repo> --json.
   • Recommend only agents backed by project artifacts.
   • Do not generate personal ~/.codex/agents entries or mutate global Codex config.

2. Generate project agents.

   • Run scripts/generate_project_subagents.py <repo>.
   • Write standalone TOML files under .codex/agents/.
   • Default generated agents are read-only unless a repo explicitly chooses a broader execution agent.

3. Keep agents narrow.

   • repo-explorer maps affected code paths and avoids fixes.
   • reviewer leads with correctness, security, tests, and regressions.
   • security-reviewer focuses on auth, secrets, data, dependencies, and migrations.
   • docs-researcher verifies APIs and version-specific behavior.
   • browser-qa-reviewer inspects UI paths, accessibility, console/network, and Playwright gaps.
   • trajectory-auditor audits task traces, skipped evidence, and risky transitions.

4. Validate agent files.

   • Run scripts/check_project_subagents.py <repo>.
   • Confirm required TOML fields exist and that read-heavy agents use read-only sandbox mode.

5. Explain explicit usage.

   • Subagents only run when the user or parent agent explicitly asks Codex to spawn them.
   • Do not treat generated subagents as hooks or daemons.

Validation

• Run scripts/suggest_project_subagents.py <repo> --json.
• Run scripts/check_project_subagents.py <repo>.
• Confirm generated agents preserve existing user-edited files unless overwrite is explicitly requested.

Completion Criteria

• The repo has only useful project-scoped subagents backed by real artifacts.
• Generated subagents are narrow, mostly read-only, and documented as explicit-use specialists.