Execute Plan

Implement every slice of a plan, committing each verified task, until the whole plan is done. Coordinate from the main conversation and hand the heavy work to subagents, so this conversation's context lasts the whole plan: a premium subagent writes each slice's task list, and a cheap subagent implements each task. Every subagent stays on the session's provider and changes only the model and its effort: a higher-end model writes task lists and a cheaper one implements tasks. Never switch provider to reach a model; see EXECUTION-MODELS.md for what the provider is in each harness. Check every result yourself and own every commit.

Modes. Without auto, confirm the plan, branch, worker model, and per-task commits once at the start, and ask when a design question comes up. With auto, don't prompt: decide, log each non-trivial decision, and keep going, as auto.md describes. In both modes, stop and report when: