planner

mode: subagent — dispatched after @literature_scout, before any execution step.

What it does

Decomposes the user's task into an ordered, numbered plan. Returns a plan only — it never executes anything. @cosmotron writes the result with write_plan, which opens the plan-approval human gate; the runtime confirmation dialogue is the approval, and no execution step can run before it resolves.

The plan gate is human, not automated

There is no automated plan-review agent — @planner self-validates its own plan, @cosmotron does a sanity check, and the human is the actual gate (one dialogue click per plan draft). An automated review step running on a small local model risks stalling the whole session on its own judgement call; asking the human once is more reliable and no slower in practice.