The methodology loop, not a gate loop. Iteration zero writes CHECKPOINT.md: the explicit, checkable success criteria as unchecked boxes plus every assumption the task rests on marked UNVERIFIED. Each subsequent pass does exactly one step — resolving an assumption by reading the artifact, or advancing one criterion — then appends a checkpoint stating what was done, what was verified (with the command output or file:line that proves it), and what remains, and refuses to proceed if it can't describe the current state back. Distinct from spec-first-ship (implements an external spec.md of feature requirements) and pr-self-review (reviews a finished diff): here the artifact is the agent's own evolving plan-and-evidence ledger.
Loop
- Goal
- every success criterion and assumption in CHECKPOINT.md is resolved — each checked off with a verification note (command output or file:line) proving it, or struck out as false with the correction noted
- Exit when
- CHECKPOINT.md exists with zero unchecked '- [ ]' boxes and zero UNVERIFIED entries, and every checked item carries a 'verified:' note
Check between iterations