Major upgrade. Proof now closes the loop: it catches the lie, makes the agent fix it, and keeps score.
New
- Honesty ledger +
proof stats: every verification is recorded; see your agent's honesty rate, lies caught, clean streak, and worst-offending check. proof check: verify any claim text without a transcript. Works from any coding agent, shell, or CI step.proof check "tests pass" --root . --json.proof.toml: pin exact test/build/typecheck/lint commands per repo, set the serve command and base URL for http checks, tune max fix cycles. Config beats auto-detection; the claim's own command beats both.- HTTP v2: response body assertions (
returns 200 with "ok"), automatic local server boot (.proof.toml serve, package.json dev/start, or Procfile) with clean process-tree teardown, and live verification of deploy claims. - Fix loop: after a FAIL, the agent is re-blocked with the previous receipts and "attempt N of 3" until it actually fixes the problem; Proof gives up gracefully after max cycles instead of nagging forever.
- Runner detection for Maven, Gradle (wrapper aware), .NET, Make, Mix, Composer, and Bun/pnpm/yarn/npm by lockfile.
--jsonon verify/check/stats for CI pipelines;--version; Claude Code plugin manifest.
Numbers
- 199 tests (was 71), CI green on Linux and Windows across Python 3.11-3.13
- Still pure Python standard library, zero dependencies
- Still strict: any failed claim fails the turn; missing tooling is INCONCLUSIVE, never a false PASS
Install
npx skills add EricFinland/proof