A configuration of [
loop-controller]. That skill supplies the loop machinery — primitive selection, the full guardrail stack, state externalization. This skill supplies the two things specific to "build until the spec holds": a fresh-context evaluator as the proof (the [contract-auditor] graded against an authored contract, default-FAIL) and the generator/evaluator split that keeps the builder from grading its own work. Readloop-controllerfor the guardrails; they're inherited, not repeated here.Why
disable-model-invocation: this loop edits and commits code on its own and spends tokens spawning an evaluator subagent each round, until every contract criterion passes. You want to type/contract-conformance-loop(or have the orchestrator dispatch it) — not have Claude silently start an autonomous build loop because a contract happened to exist.
The 5-part contract
| Part | This loop |
|---|---|
| trigger | an authored contract exists (contract-author output in contracts/) plus an implementation to build or verify against it, or an explicit /contract-conformance-loop (optionally scoped to one contract) |
| action | ONE iteration: the generator (the builder / role agent) implements or fixes toward one not-yet-passing criterion → a fresh-context evaluator subagent (contract-auditor, spawned with no Write/Edit) checks ALL criteria against evidence and returns a default-FAIL verdict with per-criterion pass/fail + feedback → failing criteria + feedback feed the next iteration |
| proof | every entry in the criteria JSON has passing evidence confirmed by the fresh-context evaluator (not the builder) — default-FAIL: each criterion starts "passed": false and only the evaluator's verdict, citing evidence, flips it |
| memory | criteria.json (the default-FAIL criteria, one entry per contract criterion), PROGRESS.md (what's done / next), a git checkpoint per criterion turned green |
| stop | the evaluator confirms all criteria pass OR iteration cap OR no-progress for 3 rounds OR budget cap |
The proof: a fresh evaluator's verdict, default-FAIL
"Done" is not "the builder thinks the feature is built." It is every
criterion in criteria.json flipped to "passed": true by the fresh-context
evaluator, each citing concrete evidence — and re-confirmed against the whole
criteria set on the final round, not just the criterion last touched.
Two rules make this convergent rather than a rubber-stamp:
- The grader is not the doer. The builder is a pathological optimist about
its own work; a same-context critic approves mediocre output. The evaluator is
a fresh subagent that sees only the artifact + the contract, never the build
reasoning. This is the PGE / GAN pattern from
loop-controllerStep 2. - The evaluator cannot edit. It is spawned with no Write/Edit tools — so it can only report a failure, never "fix" one by quietly lowering the bar. It inspects (Read / Grep / Bash-to-run-tests) and returns a verdict.
This maps onto the existing pair: contract-author writes the spec the loop
builds against; contract-auditor is the evaluator. This loop does not
re-implement either — it sequences them in a bounded, default-FAIL loop. The full
generator/evaluator split, the criteria.json schema, and the evaluator dispatch
prompt are in references/pge-loop.md.
Step 1 — Build the default-FAIL criteria JSON
Read the authored contract from contracts/ (the contract-author output:
openapi.yaml, data-layer.yaml, types.<ext>, and the Domain Rules in
contracts/README.md). Decompose it into one testable criterion per checkable
fact and write each as a JSON entry with "passed": false. Store it at the
profile-defined path (default criteria.json). A criterion the evaluator cannot
check against evidence is not a criterion — split it until it is. Schema and a
worked example: references/pge-loop.md.
Step 2 — Generator: implement toward ONE failing criterion
Pick the highest-leverage "passed": false criterion and have the builder
implement or fix just that one. One criterion per iteration — don't batch
features; batching destroys the signal about which change moved which criterion
(loop-controller Step 5). Update PROGRESS.md with what was attempted.
Step 3 — Evaluator: grade ALL criteria from fresh context
Spawn the evaluator subagent (contract-auditor) fresh, with no Write/Edit
tools, blind to the builder's reasoning. Hand it only the contract + the
implementation + criteria.json. It re-checks every criterion against
evidence (re-verify the whole, not just the one touched — loop-controller
Step 5) and returns the criteria JSON with each entry either still false (plus
feedback) or flipped to true (plus the evidence that flipped it). The builder
never edits criteria.json; only the evaluator's verdict does.
Step 4 — Feed failures back, checkpoint, repeat
Merge the evaluator's verdict into criteria.json. On a green-ward round (a
criterion newly true, nothing regressed), commit a checkpoint naming the
criterion satisfied — the git trail is the loop's undo and post-mortem. Feed the
still-failing criteria + their feedback into the next generator pass. When the
evaluator confirms all criteria pass on a single whole-set re-check, the loop
is done; report the final criteria.json as evidence.
Guardrails specific to this loop
Inherits the full stack from loop-controller → references/safety.md. The
caps this loop sets:
- Iteration cap — default ~10–15 build/evaluate rounds (read from
.claude/profile.yamlif set). Hitting the cap is a stop-and-escalate, not a license to weaken the contract or have the builder self-pass a criterion. - No-progress detection — if the same set of criteria stays
falsefor 3 consecutive rounds, stop and escalate. Three rounds with the criterion count unchanged means the approach is wrong, or the criterion is ambiguous — surface it (it may be a contract gap, whichcontract-auditorflags rather than failing the implementation against). - Budget cap — the evaluator subagent costs a full fresh context per
round; a long run is materially more expensive than a self-graded loop.
Watch
/cost; terminate at the ceiling, don't just warn. - Never let the builder grade itself. Forbidden, and each is a finding:
the builder flipping a
criteria.jsonentry totrue, softening a criterion, or the evaluator being reused with the builder's context (it must be fresh and Write/Edit-free). A green that the builder wrote is not proof — only the evaluator's verdict is. - HITL before irreversible. If satisfying a criterion would touch something
irreversible (a migration against a real DB, an external API), that's an HITL
checkpoint — pause for the human (
loop-controllerguardrail 4).
Choosing the driver primitive
The canonical form is a skill orchestrating a generator + an evaluator
subagent — the evaluator subagent is the proof, so this loop always spawns it
(hence Agent in allowed-tools). Per loop-controller Step 1:
- Canonical — generator + fresh-context evaluator subagent. Use whenever any criterion is subjective or needs cross-file judgment (does the implementation import the shared types, do the domain rules actually hold). The evaluator's default-FAIL verdict is the stop signal.
/goalmay substitute only when every criterion is fully provable from command output (e.g. a contract reduced entirely to a test suite + schema-validation exit codes). Then/goal "every criterion in criteria.json is passed:true, confirmed by the contract's own checks — or stop after N turns", and the loop degenerates towardfix-until-green. The moment a criterion needs judgment, fall back to the fresh evaluator.
Using it under the orchestrator
This is the build-until-spec inner loop (archetype 2). The orchestrator
authors the contract via contract-author (its Phase 4), then dispatches this
loop with the builder role as generator and contract-auditor as the
evaluator. The orchestrator does not override a stuck loop — if the loop
escalates after no-progress, that's a real conformance blocker or a contract gap,
not a number to paper over. A satisfied criteria.json informs the build; the
qe-agent's qa-report.json still decides the gate (loop-controller's rule:
the loop informs, the gate decides).
How this differs from its neighbors
- vs.
contract-auditor— the auditor is a one-shot static audit that reports mismatches. This is the bounded iterative loop that drives them to zero, wrapping the auditor as its default-FAIL evaluator harness. - vs.
fix-until-green— that loop's proof is three exit codes; this loop's proof is per-criterion contract evidence graded by a fresh evaluator. Use fix-until-green when "done" is the test suite; use this when "done" is "the contract's criteria all hold," some of which no single exit code can prove.
Reference files
references/pge-loop.md— the generator/evaluator split in full, thecriteria.jsondefault-FAIL schema with a concrete JSON example, and the fresh-context evaluator dispatch (no Write/Edit, blind to the build reasoning, returns a per-criterion verdict).