Plan AQE Work
- Read
AGENTS.md, the objective, relevant research, affected implementation, tests, configuration, and public boundaries. - Define observable outcomes, non-goals, constraints, assumptions, and decisions that require human approval.
- Decompose work into independently verifiable deliverables. Map dependencies, consumers, migrations, compatibility boundaries, and rollback points.
- Select the smallest useful specialist fleet with references/planning-framework.md. Give every selected specialty a concrete output; do not summon the whole fleet.
- Arrange phases so contracts, safety controls, and durable tests precede or accompany implementation. Check CLI/MCP and Claude/Codex parity when shared capabilities cross those surfaces.
- Identify work that can run in parallel and a clear critical path. Do not use unsupported precision for time estimates; size work by risk and dependency.
- Attach acceptance criteria and the narrowest verification command to each deliverable. Include security, resilience, performance, accessibility, or data-integrity gates only when triggered by the actual scope.
- End with prioritized milestones, risks and mitigations, open decisions, and a recommended first vertical slice.
Use Ruflo for persistent memory, learned routing, hooks, or multi-agent coordination only when its operational benefit exceeds its setup overhead. Treat Claude agent definitions as domain references; translate their workflows into available Codex actions rather than copying Claude-only tool syntax.