plan-executor

Plan executor

A plan of this kind — an architecture-research-report with a staged §6, a review with a recommendation, an ADR with a migration sequence — is already most of a machine-executable specification. It has ordered stages, an exit criterion per stage, and usually a section saying what not to do. What it lacks is a runtime that will not let itself be talked out of the exit criteria. That is what this skill adds.

The one rule everything else serves

Progress is derived by re-running acceptance, never claimed by whoever did the work.

A stage is done because a command exited as the contract said it would, in a clone that has never seen this run's working tree. Not because an agent said so, not because a prior session's notes say so. Everything else here — why the contract is a separate file, why the clean room is a fresh clone, why a fix that edits a test is rejected, why resuming is cheap — follows from that one rule.

It is also the rule these plans were written about. Day 1's finding was a repository that read green while a clone of it failed four checks, because the fix lived in an uncommitted working tree. Day 2's was a register that would happily accept a fabricated reduction, because the rule against it was prose. A harness that executes those plans and then reports success on its own say-so has learned nothing from either.

Phase 0 — Find the plan, or make one

You need a document with a staged recommendation. If the user named one, read it whole before anything else — the stages, the exit criteria, the "what not to do" section, and the falsifiers.

If there is no such document, stop and produce one with architecture-research-report first. It writes exactly the shape this skill consumes: a §6 that is staged, sequenced and costed, a §5 of prohibitions, a §7 of what would change the answer. Executing an unstaged wish is how a run ends up optimising something nobody agreed to.

If the document exists but its recommendation is one undifferentiated paragraph, say so and offer to stage it. Do not silently invent stage boundaries; the ordering in these plans is usually load-bearing.