code-decompose

code-decompose (Codex edition)

Steps 4-6 of THE PROCEDURE (~/.codex/AGENTS.md). Entry condition: codebase-first already returned a BUILD gate plus the residual gap. If it hasn't run, run it first — do not re-discover here.

The point: the change must be fully understood before any code is written — the code that exists and the code you intend to write — and that understanding must be written down precisely enough that it could be executed without further judgment. Decomposition is the expensive thinking step; execution is the cheap step done at volume.

Use one bounded Codex builder per unit and a distinct auditor against the same spec when native subagents are available. Never run sibling builders concurrently in a dirty checkout. If subagents are unavailable, preserve the separation in time: finish specing, build, then re-read the diff against the spec with independent-auditor eyes.

Phase 0 — FILE PLAN (mandatory, before any Unit)

Structure is the orchestrator's job (Ro 2026-10-03): builders lack the big picture, and "split it" yields unplanned files and functions cut midway. Before specing units:

1. wc -l every file the change will touch or create (signatures via tools/skeleton.py/graphify; no body reads).
2. Write the table:

path | one-line responsibility | public interface (signatures) | lines now -> budget after (<=200)

1. A touched file that would exceed 200 gets its split planned as its own first unit: a move-only refactor (behavior unchanged, same tests green before and after), cut only at function/module boundaries, never mid-function. Later units are after: U1.
2. Each unit's CHANGE names only files from the plan. A builder that needs another file, or would break a budget, stops and reports back; amend the plan first.

Phase 1 — Decompose (write the specs first, no code)

Start from the codebase-first discovery artifact (.scratch/discovery/<slug>.md if present) and carry its REUSE/ADAPT/REJECT verdicts and gate forward verbatim. Then produce: