Review Fix Cascade Guard

Overview

A containment guard between a review tool's user-selected findings and the agent's Edit/Write calls. Line-scoped fixes recreate the cascade pattern: cycle 1 patches the reported line, cycle 2 raises a new valid finding the cycle 1 fix exposed. The recurring shapes the cascade takes are enumerated in §Phase 2 (13 archetypes).

The guard does not auto-fix. It restates the invariant in path-neutral terms, enumerates the sibling paths the same invariant applies to, picks the smallest envelope that closes the predicted next-cycle finding, and emits a gate_status enum. The caller (codex-review-cycle or a manual review/fix loop) reads gate_status and applies edits ONLY when it is closed or accepted-residual. The other three statuses block edits and force a deferral, split, or override-to-accepted-residual transition through AskUserQuestion.

After every selected finding has its per-finding envelope, a Phase 5.5 batch reconciliation pass runs over the combined fix set so envelopes that touch a shared surface (same file region, same caller schema, same doc section, same test) are ordered, split, doc-cascade deduped, or paused before the agent applies any edit. The batch produces its own <batch_reconciliation> carry record consumed by the next cycle.

If the skill is executed manually because the harness cannot invoke it, the manual fallback must still produce the Phase 3 sibling-path matrix and Phase 5 validation evidence. A recorded manual-fallback receipt without that evidence is a bypass, not a closed guard.

The skill's bias toward residual + deferral over scope expansion is documented in §Key Design Constraint at the bottom of the file.

Language

All user-facing output is rendered in the user's language (the language the user has been using in the conversation, or as configured in the Claude Code system-level language setting).

Translate into the user's language: