Primitive of the day · 2026-10-03

chunk-refactor-trigger-thresholds

Switch from single-pass to chunk-and-checkpoint refactoring when the file exceeds ~200 lines, mixes multiple concerns, cannot be fully tested in under ~30 seconds, or a previous large agent change proved hard to debug.

When this applies

Deciding how to structure an upcoming agent-driven refactor of an existing file.

Preconditions

• The file's line count, concern mix, and test-cycle time are known or quickly measurable

Do this

Assess the four signals (size >200 lines, mixed concerns, test cycle >30s, prior hard-to-debug large change); if any holds, run the map/approve/plan/execute/checkpoint protocol instead of a one-shot refactor.

What you should see

Protocol overhead is spent only where single-pass failure (undiagnosable broken tests) is the realistic alternative.

How it fails if ignored

Applying the heavy protocol to trivial files (wasted cycles); skipping it on a 600-line file because it 'looks simple'.

Do not use when

Small, single-concern, fast-to-test files where a single reviewed pass is cheaper than the protocol's overhead.

Kind: calibration. Part of the skill Legacy refactor chunk checkpoint. Free to reuse in your own agent skills.