ralplan

Ralplan runs a structured planning loop before execution begins. A Planner agent drafts the initial plan, an Architect reviews it for soundness, and a Critic challenges it for quality and testability. If either reviewer finds issues, the loop repeats with a revised plan until the Critic returns an approval or five iterations are exhausted. The result is a plan that three independent perspectives have signed off on, stored on disk for $ralph or $team to execute.

When to use

• The task is large enough that a wrong approach would waste significant execution time
• You say "ralplan", "consensus plan", "plan this before coding", or "plan the"
• You want planner, architect, and critic to agree before any implementation starts
• You are following $deep-interview and need architectural validation before handing off to $ralph or $team
• The request is underspecified and you want the gate to redirect to planning automatically

How to invoke

Natural language triggers: "ralplan", "consensus plan", "plan this", "plan the".

Explicit slash: $ralplan

Flags:

• --interactive — pause for user review after the draft and after approval
• --deliberate — add a pre-mortem (3 failure scenarios) and expanded test planning; auto-enables for auth, migrations, destructive changes, and public API work

codex
> ralplan "add OAuth2 login to the existing auth module"

codex
> ralplan --deliberate "migrate the users table to a new schema"

What happens

Ralplan starts with a pre-context intake that loads or creates a grounded snapshot at .omx/context/{slug}-{timestamp}.md. The Planner agent then drafts a plan and a compact decision-record summary covering principles, decision drivers, and at least two viable options with tradeoff notes. The Architect reviews the draft next, providing a steelman counterargument and at least one concrete tradeoff tension. The Critic evaluates the plan against quality criteria — principle consistency, fair alternatives, risk mitigation, testable acceptance criteria, and concrete verification steps — and returns either APPROVE or ITERATE. Any non-approval verdict sends the plan back through the full Architect and Critic loop; this repeats up to five times. When the Critic approves, the final plan is written to .omx/plans/ with an Architecture Decision Record, staffing guidance, and explicit launch hints for both $ralph and $team. If --interactive is set, you choose the execution path; otherwise the plan is output and ralplan stops.

Outputs

• .omx/plans/{date}-{topic}.md — consensus-approved implementation plan with ADR
• .omx/context/{slug}-{timestamp}.md — grounded context snapshot
• Explicit staffing roster and reasoning-level guidance per execution lane
• Launch hints for $ralph (sequential) and $team (parallel) execution

Related skills

• $deep-interview — clarify requirements before ralplan runs
• $ralph — sequential persistent execution of the approved plan
• $team — parallel coordinated execution of the approved plan