Overview & Result
Canny is an open-source supervision layer for AI coding agents created by Qkal. It hooks into Claude Code and Codex CLI, maintaining an append-only ledger of actual agent actions, code diffs, and test runs. When an agent attempts to stop by claiming work is done, Jev evaluates whether the claim is supported by tangible test passes and diffs, blocking false completion claims.
How Jev fits in the loop
- Maintains an append-only ledger of agent file modifications, shell commands, and test outputs.
- Intercepts stop requests when the agent claims the task is complete.
- Jev evaluates the code diff against the prompt criteria and test evidence, rejecting unsubstantiated completion claims.
How to reproduce
- Clone https://github.com/qkal/canny and install dependencies with Node >= 22.
- Configure Canny hook in Claude Code or Codex CLI.
- Trigger agent task completion and inspect Canny's ledger check verifying diff and test evidence before exit.
Why this build matters
Solves one of the most frustrating coding agent failure modes: prematurely claiming success without running tests or producing real working code.
Limitations
- Requires Node.js >= 22.
- Zero runtime dependencies, but requires deterministic git repository state to compute diffs.