paseo-committee

When work has stalled - looping, tunnel-visioning, or facing a genuinely hard planning problem - this skill convenes a committee of two high-reasoning agents to step back, perform root cause analysis, and produce a plan. The deliberate design point is contrast: the two members come from different providers with different reasoning styles, work from fresh context, and plan in parallel, so the committee can propose a completely different approach rather than doubling down on the current one. The members also stay alive afterward to review the implementation.

Setup requires reading the paseo skill and the orchestration-preferences.json file before choosing members, unless the user explicitly named providers; members are then selected across providers - one planning or research-strength provider and one contrasting high-reasoning provider - rather than from hardcoded defaults. Three hard rules govern the process. No edits: every prompt to a committee member ends with a suffix stating this is analysis only and that they must not edit, create, or delete files or write code. Trust the wait: do not poll, hurry, or interrupt, because long waits (a model may reason 15 to 30 minutes) usually mean something worth thinking about was found. And the orchestrator is the middleman, driving plan to implement to review without yielding to the user except for divergences that need their call.

The work proceeds in three phases. In Plan, a problem-level prompt states the goal and acceptance criteria, constraints, symptoms, and what was already tried, and explicitly asks for root cause analysis and three-levels-deep questioning; both agents run in parallel, their responses are challenged rather than accepted, and the results are synthesized into a unified plan or escalated to the user on significant divergence. In Implement, the orchestrator implements by default, or delegates to one implementation agent if asked. In Review, the diff goes back to the committee to flag drift or missing pieces, iterating between implement and review until consensus - and after about ten iterations without convergence, a fresh committee is started with the full history.