A reminder can be forgotten by an agent; a supported hook can run a deterministic check. Neither replaces CI, and a hook configuration is only tested when the real client dispatches it.

Tool event → Hook handler → Exit status → Continue or stop
  1. 1Tool event
  2. 2Hook handler
  3. 3Exit status
  4. 4Continue or stop

Work through the example

Use a synthetic protected file to test both denial and allowance. Then deliberately fail verification and inspect whether the client stops.

Start with a disposable branch and synthetic data. Write the expected outcome before changing the implementation, then keep the first failing result. This prevents a later repair from quietly redefining the task. The procedure below is grounded in the repository reference; its examples must still be checked against your project and installed toolchain.

Implementation reference

The following focused section is adapted from the maintained project guide. It preserves the source’s examples and limitations.

1. Hook vs. CI vs. reviewer

Three enforcement layers. Use the cheapest one that can decide the question.

Layer Decides Cost Feedback speed
Hook Rules a script can evaluate ~free Immediate — model self-corrects mid-turn
CI Same, plus full build/test minutes After push
Reviewer subagent Judgment: is this correct, does it match intent tokens End of task

The ordering matters. Spending a reviewer subagent on "did you use DispatchQueue.main.async" is waste — a grep answers that for free, instantly, and cannot be argued with. Reserve model judgment for what rules cannot express.

Rule of thumb: if you can write the check as a grep or an exit code, it is a hook. If it needs to understand intent, it is a reviewer.

Acceptance and failure review

Checkpoint What to inspect If it does not match
Tool event Confirm the input and environment Preserve the failure and return to this step
Hook handler Inspect the intermediate artifact Preserve the failure and return to this step
Exit status Run the focused check Preserve the failure and return to this step
Continue or stop Record the observed result Preserve the failure and return to this step

Ask the agent to explain the smallest change that resolves the observed mismatch. Keep unrelated refactors out of the repair. A change that makes a warning disappear is not enough if the behavior or ownership contract has changed. Re-run the same acceptance check so the before and after results are comparable.

Evidence and limits

Claude script-level checks and Muse Stop discovery have different scopes. Full cross-client enforcement remains unverified.

This is an educational guide. Its presence in the series does not certify a completed client-specific lab. The series evidence record separates executed checks from exercises and blocked environments.

Inspect the source used in this lesson.

What to do next

Next: Making the agent prove it: screenshots as evidence, not "done"