Worked example: manually walked /plan-rollout against
garrytan/gstack#1241 (fix(ask-user): keep question payloads compact)
— 41 files, +661 / -282, regeneration-heavy.
Verdict the skill should emit: "one PR." 39 of 41 files are
deterministic regenerations of one source change in
scripts/resolvers/preamble/generate-ask-user-format.ts; slicing them
off would leave dependent fragments. Reader time: ~21 min (under cap).
Findings (v1.1 todos) surfaced by the dogfood:
1. Deterministic-regeneration detector — merge build-output slices
into their source slice automatically.
2. Regen-multiplier on reader-time so skim-only output isn't
over-counted.
3. --explain mode — when verdict is "one PR," print the rejected
slicing alternatives and the signals that rejected them.
4. Calibration loop — predicted vs actual reader-time on first ~10
real invocations to ground v2 heuristics in data.
SYSTEM.md is explicitly called out as the WRONG primitive for catching
build-output coupling — that needs a Makefile-style dependency, not a
contract graph.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Manas Srivastava