Browse / General-purpose / keel

keel

visual-z

Project-agnostic workspace template: task cards as verifiable claims (RED re-verification, frozen done-definitions, test-weakening guard)

Open visual-z/keel → Checked against the four rules.

Does not say

no agent config and no instruction file at the root.

In its own words

README.mdvisual-z/keel

A language-agnostic workspace template, plus keel: a single static binary that checks what task cards claim.

keel treats each task card as a claim to be proven. The definition of done is the specification, the tests are the proof, and keel is the proof checker. keel re-derives every claim a card makes from git and from real exit codes, and never takes the card's word for it. Typical claims are "this test was red before the fix", "the definition of done was not relaxed" and "the gate passed".

Why

Letting a coding agent report its own "done" is unreliable. The usual shortcuts are:

  • deleting assertions or adding skips;
  • exiting the process before the asserts run;
  • loosening the acceptance criteria;
  • writing a test that was never red.

People under deadline pressure take the same shortcuts. keel's answer is to trust only the checker: a claim that a machine cannot re-verify does not count.

Your build tools (cargo, go, bun, tsc, pytest) check the code. keel checks claims about history, which no language toolchain checks:

  • "this test failed at commit X";
  • "this criterion is unchanged since the card was opened";
  • "this diff removed an assertion";
  • "this receipt covers every gate step".

That layer does not depend on the project's language, so keel is written once, in Go, and shipped as one binary with no runtime dependencies. Each project pins its version, the same way rust-toolchain.toml pins rustc.

Quoting 49% of the README. Read the rest on GitHub ↗

Read from the repository's git tree on 2026-08-17.