A layer-2 harness for coding agents

The quality gate for AI coding agents

AI agents will tell you they’re done. Chalk makes them prove it.

Ship agent-written code through gates, not vibes — a locked spec decides success, not the model.

npm install -g chalk-protocol

Get started GitHub

Works with any agent
  • Claude Code
  • Cursor
  • opencode
  • aider

Bring your own executor. Chalk is the referee — and it drives itself: this site ships through the same four gates.

The four levers

One rule: the agent never certifies itself

Each lever closes a way an agent can fake “done.” Together they make success an external fact.

Without / with Chalk

The difference a gate makes

The same agent, the same task — one path ships a bug, the other can’t.

Agent without Chalk

  • Agent marks the task done
  • The failing test was quietly deleted
  • Review never happened
  • The bug ships

Agent with Chalk

  • chalk done refuses — verify is RED
  • Locked tests can’t be touched
  • An adversary must sign off first
  • The gate closes only on real green

The loop

The loop — four gates every change must clear

A specced task enters the loop; nothing merges until all four gates say so.

  1. Locked test A real test is authored for the acceptance criteria, then locked read-only. The agent cannot weaken or delete it to pass.
  2. Verify The real toolchain plus a test-integrity check must print GREEN. The gate decides success — never the agent.
  3. Adversarial review A second agent tries to refute the change. Every blocking finding must be fixed before it passes.
  4. Merge gate `chalk done` only succeeds when verify is green, locked tests are untouched, and review passed.

Quickstart

The per-task loop, start to merge

  1. 01 chalk next find the one task to work on
  2. 02 chalk context <id> read the acceptance criteria and at-risk tests before coding
  3. 03 chalk start <id> begin — it refuses tasks with no acceptance criteria
  4. 04 write code satisfy the criteria, keep the diff small
  5. 05 chalk verify loop until the gate prints GREEN
  6. 06 chalk review <id> survive the adversarial reviewer
  7. 07 chalk done <id> the merge gate closes the task

Docs

Rendered from the source of truth

Rendered here at build time from the parent repo — the source of truth this site cannot drift from.