Skill: Refactoring with Invariants Preserved

Name: Refactoring With Invariants Preserved
Rating: 62
Author: Kristopherlb

Intent

•Enable an agent to restructure systems while preserving semantic outputs: graph meaning, derived intents, policy outcomes, and plan artifacts.
•Make refactoring safe under strict determinism, provenance, and conformance constraints.

•In-scope: invariant identification, semantic equivalence reasoning, regression-risk assessment, and conformance-aligned change framing.
•Out-of-scope: “cleanup” that changes semantics without explicit intent and review; performance rewrites that weaken determinism.

•Invariant: property that must remain true (stable IDs, canonical ordering, least privilege, policy-driven compliance).
•Semantic equivalence: unchanged meaning of outputs even if internal structure changes.
•Regression surface: which contracts and golden cases could drift due to refactoring.

•Invariant catalog: the non-negotiable properties for the touched subsystems (versioned references).
•Golden cases + conformance baselines: expected outputs to preserve.
•Provenance/explain expectations: what must remain explainable and traceable after refactor.
•Change boundaries: what is allowed to change (structure) vs forbidden to change (semantics).

•
Structured outputs (minimum):
- •Refactor intent statement: what is being reorganized and what semantics must remain unchanged.
- •Invariant checklist: invariants asserted to be preserved, and any new invariants introduced (none weakened).
- •Conformance impact report: which golden cases/matrix areas might be impacted.
•
Human-readable outputs:
- •A concise “why this refactor exists” narrative tied to maintainability/clarity without semantic drift.

•No semantic drift: conformance suite outcomes remain unchanged unless explicitly intended and reviewed.
•Deterministic stability: stable ordering/IDs/serialization remain unchanged.
•Explainability preserved: provenance chains remain intact; no loss of traceability.

•Golden output stability: unchanged outputs across representative cases (deny on unexpected drift).
•Policy outcome stability: no unexplained changes in violations/actions.
•Diff quality: refactor changes do not introduce reorder-only churn (signals determinism risk).

•Forbidden: refactors that “simplify” by removing provenance or reducing auditability.
•Forbidden: refactors that widen default permissions/network posture.
•
Escalation (HITL required):
- •Any refactor that changes stable ID rules, canonical ordering, or policy enforcement semantics.