uroboros-chunklisted

# Chunk a wave goal into loop-sized units Measured motivation (2026-08-31, one machine, one day): six wave-scale goals produced zero converged plans across three planner generations — the final judged-convergence build honestly returned pivot-conclude — while five small units shipped to completion in this repository the same day. Plan convergence decays with goal scope for context-free drafting seats. Decomposition belongs UPSTREAM of the loop, in the session that holds the project's context. ## Where this skill runs — and why that is the design This skill executes in the calling Claude Code session, not in a spawned seat. Two measured reasons: 1. A spawned seat is stateless; cold-reading a repository at wave scale is exactly the failure mode above. The caller holds the dissections, owner rulings, and memory that make chunk boundaries correct. 2. Spawned seats hit capability walls the caller does not (measured: Cursor's command runner refuses execution in -p mode — RAN_TESTS: no), which is the same lesson that made the harness the evidence runner. ## The contract: reasoning decides, determinism verifies and advises The MODEL decides — from understanding, never from rules: - chunk boundaries, chunk count, and granularity (no size caps, line budgets, or keyword splitters anywhere); - ordering and parallelism, from real dependency reasoning ("unit 2 edits the file unit 1 creates"); - whether the goal is even worth a campaign (one piece, or hand-build, are