Stop trusting 'done'. Make AI coding agents prove their work — frozen specs, tamper-detected tests, an independent blind verifier, hard-blocking gates. Fails loudly instead of pretending. For Claude Code, Codex, OpenCode & any Agent Skills runtime. No unverified work passes.
Perfect health signals and solid docs offset by early-stage adoption (6 stars) and zero typed-language/dependency-manifest points.
Frozen specs, tamper-hashed tests, and an ed25519-signed manifest block agents from faking 'done'
Detects an agent weakening its own frozen test to fake a pass — halts with GAMING-DETECTED instead of retrying silently.
| Rating | Summary | |
|---|---|---|
| Quality | solid (17/24) | Strong health and documentation signals undercut by very early adoption and no typed-language/manifest engineering markers. |
| PAI Relevance | integrate (0.75) | Directly harvestable verification architecture that plugs cleanly into PAI's Skills/Hooks mechanism and fills a gap beyond the existing Evals skill. |
17/24 — actively-maintained / adequately-documented / early-or-minimal
Failed:
Passed:
.github/workflows/self-evals.ymlFailed:
Passed:
research/, .planning/, docs/evals/, CHANGELOG.md, and agentskills.ioFailed:
Passed:
| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 2 | Novel enforcement architecture — frozen specs, tamper-hashed tests, ed25519-signed manifest, independent blind verifier — directly applicable to PAI's Algorithm Verify phase and the Evals skill's design. |
| Integration Readiness | 2 | Distributed as Agent Skills (SKILL.md + shell hooks) with a one-command Claude Code plugin install; subprocess-executable, structured verdict output (DONE/STUCK/GAMING-DETECTED), matches PAI's CLI-first, skill/hook-based architecture. |
| Overlap Risk | 1 | Partially overlaps PAI's existing "Evals" skill (AI agent evaluation), though this repo's enforcement mechanics (tamper detection, independent verifier separation) go beyond what Evals currently covers. |
| Gap Fill | 1 | Addresses a functional area — hard-blocking proof-of-work enforcement against reward-hacking/gaming — where PAI's Evals skill has limited depth. |
Composite: 0.75
Conservancy editorial loop: Install make-no-mistakes as a Claude Code Agent Skill in .claude/skills/make-no-mistakes/ and wire its hard-blocking gate into the autonomous publish step — the editorial pipeline currently trusts the agent's own "done" signal before pushing to Cloudflare Pages; the frozen-spec + blind-verifier check stops a satirical post from going live on a self-graded pass.
Capture-to-Knowledge Pipeline: Replace the manual "clean-room Haiku validation" step with the skill's tamper-detected test + independent blind verifier mechanism — swaps an ad hoc second-model check for a checksummed, ed25519-signed manifest that proves validation actually ran against the frozen spec rather than a re-run that could silently pass.
Fabric Recommender (fab): Add the skill's frozen Definition-of-Done + checksummed test gate to the V2.1 release path so fab's test suite can't be edited to match failing code without invalidating the manifest hash — closes the gap where a shipped release currently relies on the agent's own "tests pass" claim.
Category: Developer Tools & CLIs
In this category: cheat--cheat, casey--just, mvanhorn--cli-printing-press, openclaw--gogcli, simonw--showboat, gastownhall--beads, ohmyzsh--ohmyzsh, tirth8205--code-review-graph, tw93--Mole, mvanhorn--printing-press-library, mattpocock--skills, atomantic--slashdo
Standing: Distinct from the category's CLI-utility norm — most peers ship task runners or code-generation tools, while this is the first entry that positions itself as a verification-enforcement layer against agent reward-hacking.
Density: 8/10 — README (partial, 8KB), release/star/fork/issue metadata, license, topics, and CI badge available; dependency manifest not available, full README (beyond 8KB) and commit history detail not confirmed.
Commit activity is concentrated in a 3-day burst (2026-07-02 to 2026-07-05) followed by 24 days of silence before appraisal — health probes pass on the letter of the rules, but the "actively-maintained" tag should be read with the caveat that development cadence has not been demonstrated over a longer window. Worth a re-check in a few months to confirm sustained maintenance beyond the initial release sprint.