Break infinite AI repair loops with bounded verification

Break AI Fix Loops replaces patch-and-retry behavior with a bounded, evidence-producing repair protocol. It treats a changed patch as progress only when an observable state changes, using stable failure fingerprints to detect when an AI coding agent cycles through similar patches without changing the observed failure. The skill enforces a three-attempt budget per acceptance claim and requires direct observation of real execution paths - CLI, API, UI, persistence, or deployment - rather than accepting unit tests or mocks as proof.

Use this skill when an AI coding agent cycles through similar patches without changing the observed failure, when a focused test passes but the installed, deployed, UI, API, persistence, or other real execution path still fails, or when a repair claim needs a verifier that can reject a known-bad state and a rollback that has actually restored the baseline on a disposable copy.

Do not use for a one-shot, already-understood edit whose acceptance check directly exercises the complete claimed behavior.

You provide the exact defect and the behavior that would disprove it, the revision, configuration, input, and execution path under test, the baseline command with literal result and exit status, the strongest check that directly observes the claimed behavior, and the rollback command and the state it must restore.

You receive a structured evidence ledger documenting each repair attempt with hypothesis, prediction, change, focused check, real-path check, symptom fingerprint, and decision (ADVANCE, SHIFT_CAUSE, PROVEN, or STOP). The skill documents failure fingerprints with fields including command, input_digest, exit_code, failure_class, stable_excerpt, and real_path_state.

Compatible with claude-code, codex-cli, copilot, cursor, and gemini-cli AI coding tools.

This skill is for developers and teams working with AI coding agents who need to prevent unproductive repair loops and establish verifiable proof that fixes work in real execution paths. It is particularly valuable for critical-risk code quality scenarios where false-green tests or incomplete verification can mask persistent defects. The skill enforces a rigorous evidence-based approach: when a symptom fingerprint repeats, the patch changes but the decisive state does not, or a focused test passes while the real path still fails, it immediately triggers a SHIFT_CAUSE decision that stops editing and requires new discriminating evidence at the next unobserved owner boundary. The protocol includes negative controls on disposable copies to prove verifiers can actually fail, and tested rollback procedures to ensure baseline restoration works before claiming a fix is proven.