Plan AQE Work

1. Read AGENTS.md, the objective, relevant research, affected implementation,

tests, configuration, and public boundaries.

1. Define observable outcomes, non-goals, constraints, assumptions, and decisions

that require human approval.

1. Decompose work into independently verifiable deliverables. Map dependencies,

consumers, migrations, compatibility boundaries, and rollback points.

1. Select the smallest useful specialist fleet with

[references/planning-framework.md](references/planning-framework.md). Give every selected specialty a concrete output; do not summon the whole fleet.

1. 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.

1. 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.

1. 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.

1. 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.