llp-super-refine
The applying counterpart to llp-review: an orchestrating loop that revises the document between rounds and applies exactly one pre-authorized status transition. Invoke as /llp-super-refine <targets...> [--rounds N] [--reviser orchestrator|<family|model>|each] [--claude <model@effort>] [--external <runner:model@effort>]. Defaults: converge within the round budget; the orchestrating session revises. Default models: reviser Claude Fable 5.1 @ xhigh; reviewers codex gpt-6.1-sol @ ultra (codex exec -s read-only -C <capsule> --skip-git-repo-check -m gpt-6.1-sol -c model_reasoning_effort=ultra, prompt on stdin; confirm model and effort in the run log) and grok-4.7 @ xhigh (headless; the prompt MUST ban all shell commands or the permission prompt cancels the session).
Trigger
The document's author asks for the refine-until-both-families-ready loop on named Draft/Review documents.
Invariants (LLP 0010.001; LLP 0000 §Review honesty)
- Authorization is recorded per target and per author. Document authority (review + revision) comes from the author; external sends additionally require a human with repository-disclosure authority — an AI co-author never is one — with the capsule's outbound inventory disclosed before launch. Unsupported assertions of relayed authority are not accepted.
- MUST refuse
Accepted/Active/Superseded/Tombstonedtargets (addendum or new LLP is the path); already-Reviewtargets proceed unchanged;Drafttargets are set toReviewat loop start — the one authorized transition. Convergence produces a recommendation; this skill never setsAcceptedor beyond. - Target bytes are never redacted (redaction covers context only); a target that cannot be disclosed whole stops the loop.
- Each round: persist the round record (round number, target hashes, topology, runtimes, capsule + instruction hashes) before launching; one fresh review per family, mutually blind via a review capsule that structurally excludes the targets' review artifacts, run journal, and orchestrator notes. Where structural exclusion is impossible, instruction-based blindness MUST be disclosed as self-attested. A mid-round target change voids the round.
- MUST write each review verbatim to its family artifact (
llp/reviews/NNNN-slug.<family>.md, dated round sections, full provenance incl. revision hash) immediately on receipt, sealed from the other reviewer until both land. Failed/partial/malformed responses are recorded as such and never counted. Never fabricate, reconstruct, or overwrite a review. - Every review ends
READYorNOT READY. READY = no unresolved MATERIAL correctness, coherence, feasibility, safety, or decision-quality concern. Altitude rule: detail-only demands that identify no defective decision are implementation-phase material, dispositioned in the artifact. Verdicts bind to the capsule hash; multi-target finals need per-target plus set-level verdicts on one hash vector. - The reviser never reviews a revision it edited. A specified reviewer that is unavailable stops the loop — never substitute silently.
- Convergence is dual-READY on one revision or delta-round convergence (LLP 0010.001 §Delta rounds): a post-revision round whose instructions enumerate the delta and require every concern classified IN-DELTA vs PRE-EXISTING converges the loop when no family reports a MATERIAL concern in pre-existing text; remaining in-delta fixes get one more delta round or the final revision is labeled unreviewed. A family whose rounds are twice-consecutively convergence-clean — every remaining concern altitude-dispositionable detail or prior-delta churn, no defective decision identified — MAY be closed out via a detail-horizon stop: a normal terminal, not an escalation, with the family's NOT READY verdict recorded verbatim, a complete disposition ledger (never summarize a concern away), and the convergence claim scoped to say which family was cut off; MATERIAL defective-decision dissent still escalates. A round MAY run under a declared lens (e.g. implementer: draft a named contract from the document alone, log divergence decision points; UNDERDETERMINED = NOT READY) — recorded in provenance, all round invariants intact.
- The final report and close-out notes MUST state which revision hash the final verdicts bind to and whether later unreviewed revisions exist.
- Scope freeze + punch-list delta rounds (LLP 0010.001 §Scope freeze) — the default late-loop phase for decidable documents. For
Type: Spectargets, after the round-2 revision the document's normative surface is frozen by default (other revision-bearing types: at the author's authorization; never automatic for Principles/Explainers/Decisions/philosophical RFCs): subsequent revisions make corrections and restorations only, and any demand for new normative closure is recorded as an obligation (open-issue box or test obligation), never folded as new prose; the document carries a dated freeze declaration. Post-freeze rounds are punch-list delta rounds: the revision produces, per family, a punch list mapping that family's own findings to checkable dispositions (FOLDED at §N/REFUTED+ evidence /RECORDED-AS-OBLIGATION at §N); each reviewer receives the full document (context), the revision's unified diff, and its own family's punch list only (cross-family blindness preserved). Review scope: verify the dispositions; review the diff; outside findings admissible only as disposition errors or diff-created MATERIAL contradictions; new-closure demands are non-blocking obligations. READY = punch list correctly dispositioned + diff sound, modulo recorded obligations. From round 2 on, revisions classify each external finding (persistent / new-on-new-content / bar-inflation); the tallies go in the freeze-transition or terminal report so the choice between freeze, detail-horizon stop, and escalation rests on measured evidence. - Bounds: round budget (default 3/document), launch cap (2× budget per family per target, voids included), growth budget (~20% net over the entering revision; post-freeze, growth is corrections plus the mandatory checkable changelog only). Terminal escalations to the author, with every disposition enumerated: budget/cap exhaustion, material cross-family contradiction, sustained asymmetry (one family READY on two consecutive revisions, the other NOT READY on both), growth-budget exhaustion with open MATERIAL concerns. Each proposes stay-
Reviewor revert-Draft; the author applies — and may instead authorize extension rounds under the scope-freeze + delta protocol above. Fundamental-redesign feedback stops the loop for author authorization.
Workflow
Recipe (advisory) — read targets and governing LLPs fully; partition multi-target sets into clusters by coupling (
Related:, cross-references, sharedSystems:) and pick a topology per cluster — independent loops (loosely coupled; one sub-orchestrator each) or lockstep rounds (tightly coupled; reviewers see the whole cluster; the cluster review is preserved verbatim once, in the primary target's artifact, with per-target verdict sections) — stating the choice first. Run a context preflight; split oversized clusters rather than force-feeding. End with a cross-consistency pass: a sibling that converged before a referenced document changed gets one more round, until the hash vector passes a round unchanged. Build the capsule withgit archiveor an export excludingllp/reviews/<target-stems>*. Update**Revised:**on each revision (move long histories to a body Revision-history section). Prefer a delta round after any substantial revision — the loop's own revisions are a proven defect source — and an implementer-lens round before closing a protocol/contract-bearing document. After each revision, byte-verify security-critical literals on disk (transport layers decode backslash-u escapes).
Artifact
Per-family, per-target review files under llp/reviews/ with round records and verbatim bodies; a final report: rounds run, revisions made, verdicts, altitude-rule dispositions, and the proposed next step.
Hand-offs
- Feedback amounting to a new design →
llp-create, after author authorization. - Advisory review at any intensity, no revisions applied →
llp-review. - The author applies
Accepted; this skill never does.