# Bug Receipt

> A portable Agent Skill that prevents AI coding agents from declaring software bugs fixed without reproducible evidence.

## Agent entry points

- [SKILL.md](https://raw.githubusercontent.com/lMysticl/bug-receipt/main/skills/bug-receipt/SKILL.md): activation metadata and executable workflow
- [Receipt contract](https://raw.githubusercontent.com/lMysticl/bug-receipt/main/skills/bug-receipt/references/receipt-contract.md): machine-readable status invariants
- [JSON Schema](https://lmysticl.github.io/bug-receipt/receipt.schema.json): validation contract
- [Benchmarks](https://github.com/lMysticl/bug-receipt/blob/main/benchmarks/RESULTS.md): exact model methodology, raw telemetry, deterministic invariants, failed holdout, and limits
- [Repository](https://github.com/lMysticl/bug-receipt): source, tests, releases, and installation

## Discover and install

Search: `gh skill search "bug fix evidence"`

Install: `gh skill install lMysticl/bug-receipt bug-receipt --agent codex --scope user`

The GitHub CLI installer supports Codex, Claude Code, Cursor, GitHub Copilot, Gemini CLI, and other Agent Skills clients.

## Activate

Use for bug fixes, regressions, flaky or failing tests, runtime errors, broken UI/API/database behavior, incident follow-ups, proposed-fix verification, and requests to prove a fix or find the root cause. Skip feature-only work and generic review.

## Required result

Return exactly one evidence-backed status: VERIFIED, PARTIAL, or BLOCKED. Include Problem, Baseline, Root cause, Change, Proof, Gaps, and Source. VERIFIED requires an observed baseline, concrete root-cause evidence, an identified responsible change, passing direct verification, and no material proof gaps. Redact secrets and sensitive payloads.

## Measured evidence

- Deterministic validator: 22/22 invariant cases passed.
- Fresh v1.4 robustness routing: Bug Receipt selected in 4/4 bounded probes.
- Pre-registered robustness Skills ON: 18/20 blind-judged assertions passed (90%).
- Robustness Skills OFF: 5/20 assertions passed (25%); measured lift +65 percentage points, exact McNemar p = 0.000244.
- Correct status and complete receipt: 4/4 ON cases; leaked supplied secrets: 0; evidence-safety regressions: 0; candidate actions: 0.
- Earlier audit gate: 18/20 ON versus 7/20 OFF, corroborated at 19/20 versus 7/20. See the raw reports and limitations before generalizing.