ai · 2026
Whetstone
Reliability rails for AI coding agents — so they stop faking progress, quitting early, and shipping work that only looks done.
AI coding agents are confident liars. They declare a task finished, reference a file they never wrote, and stop at the first answer that looks plausible. Whetstone is the operating layer that holds them to a harder bar — prove it with evidence, never stop on a promise, verify with fresh eyes. In a blind test, a bare model fabricated an entire implementation for an app that didn't exist and called it done; the same model running Whetstone refused the fake and asked for the missing input instead. It's lean by default, portable to any model through a single WHETSTONE.md, and ships with a verified pipeline runner that halts the second a step produces a broken or empty artifact instead of cascading it forward. Built end-to-end solo and published on npm — it drops into Claude Code and Cowork in one command.
- Five laws, one dial — evidence over assertion, never stop on a promise, verify with fresh eyes, tuned from light to strict with a single knob
- Catches the fake "done" — in a blind eval a bare model invented a whole app and shipped it; with Whetstone loaded it refused and asked for what was missing
- Verified pipelines — every step must produce a real artifact (a non-blank image, a valid mesh) or the run halts, instead of passing garbage downstream
- Zero mandatory deps — one Node installer, a conflict gate so it never collides with your existing setup, portable to any model with no install at all
- Opt-in packs — security scans, local memory, multi-agent orchestration, and a clean-room ComfyUI content engine; add only what you need
- Node.js
- Claude Code
- npm
- GitHub Actions
- ComfyUI
Named Whetstone because it never does the work — it just sharpens whatever does. And it eats its own cooking: its test suite feeds it a booby-trapped save file that tries to delete outside its folder, and it refuses every time.
