Unslop code

Orchestrate only. Do not implement, grep the tree yourself, or commit. Spawn subagents with the Agent tool.

Goal: a clean new state of the app, not compatibility layers left behind by refactors.

Models

Workers are the agents from model-policy: Haiku high, compacting at 100k, so pass no model or effort for them. If the Agent tool lists no implementer, model-policy is not installed: pass model: haiku, effort: high on every worker call (this skill explicitly asks for that effort), including Explore — the built-in one otherwise runs on the session model — and use general-purpose in place of implementer and verifier (verifier: no edits).

Keep each worker brief small enough that the worker stays under ~100k tokens of context; split a phase by folder before it gets bigger. When a Haiku worker returns wrong or shallow work, re-run that one brief as general-purpose, model: sonnet, effort: high and say so in chat. A worker that went over 100k but did the work right needs no re-run. The rest of the model and effort policy is the model-policy rule.