Animation: four agent changes move toward a gate. Skip, loosen and delete bounce back in red. The real fix passes through in green.

# Verification Gates for Agents Course: [Testing & Verification](https://www.devtutorials.ai/learn/testing-verification) · Section: Verification as a System Source: https://www.devtutorials.ai/learn/testing-verification/verification-system/verification-gates > Turn rules into gates. See how an agent fakes green, build a test-diff guard that blocks skipped, loosened and deleted tests, write the definition of done as code with clear reasons, and wire the gates into Claude Code, Codex, Cursor and CI. ## From Rules to Gates So far, **you** checked the evidence. That works for one change. It does not work for fifty changes a day. Agents can also fake green. When a test goes red, an agent may: - add `.skip` to the test - loosen the check, like `toBe(4800)` → `toBeDefined()` - delete the test A rule in `CLAUDE.md` or `AGENTS.md` says "do not do this". The agent usually listens. But "usually" is not enough. **Gate** = an automatic check that **blocks** a change until the evidence is there. In this lesson, you will build two gates in code, then wire them into your agent and your CI. **Simple lesson:** rules guide the agent. Gates stop it. ### A Real-World Example Think about a metro station. - A sign says "Please buy a ticket." That is a rule. Most people follow it. - The ticket barrier only opens with a valid card. That is a gate. Nobody gets through without one. Your verification pack has both. The signs are the policy lines. The barriers are the hooks and CI checks. ## What You Will Learn In this lesson, you will learn: 1. How agents fake green, and why rules alone do not stop it. 2. How to build a **test-diff guard** that blocks skipped, loosened and deleted tests. 3. How to write a **definition of done** as code. 4. Where gates live in Claude Code, Codex, Cursor and CI. 5. How to export the final pack with every gate. ## How Agents Fake Green Press **▶ Run tests**. - **The Tests tab:** "1 passed | 1 skipped". Green. - **`AI_SUMMARY.md`:** "Added the ₹200 limit. All tests pass ✅". - **The truth:** the limit was never added. The agent skipped the test that showed it. - **Now try:** remove `.skip` and run again. Red: 4500, not 4800. **This is not the agent being sneaky.** Its goal was "make the tests pass". Skipping a red test does that. A gate changes the goal to "make the tests pass without touching the checks". **Simple lesson:** if a shortcut to green exists, an agent will sometimes take it. Close the shortcut. ## Gate 1: The Test-Diff Guard ### What It Does **Test-diff guard** = a script that reads the diff of the test files and blocks: - added `.skip`, `.only` or `.todo` - removed `expect(...)` checks - exact checks changed to loose ones It must also let real work through: - a real code fix - a new test - a check that was moved, not removed ### Build It The starter guard only finds `.skip(`. The tests show what else it must do. Press **▶ Run tests**. - **The Tests tab:** red for `.only`, the loosened check and the deleted test. - **Open `src/guard.ts`** and improve it. Three ideas: - Only look at test files: names with `.test.` or `.spec.`. - Collect added (`+`) and removed (`-`) lines. - A removed line with `expect(` is a problem, unless the same line was added back. - **Run again** until all seven tests are green. ### The Guard Against Four Agent Moves Here is a finished guard. A simulated agent tries four moves after a red test. Press **▶ Run tests**. Green. - **The real fix:** allowed. - **Skip, loosen and delete:** all blocked.

The guard blocks the three shortcuts and lets the real fix through.

**Simple lesson:** a good gate blocks the shortcuts and lets real work through. ## Gate 2: The Definition of Done **Definition of done** = the list of evidence a change needs before anyone can call it done. Written as code, the agent cannot argue with it: - no type errors - at least one test ran, and none failed - the new tests failed on the old code first - the mutation score is at or above the threshold - the test-diff guard found no problems Press **▶ Run tests**. Green. Read `src/done.ts`. Each rule adds a **reason** when it is not met. Now ask the gate about a change where the agent said "done". Press **▶ Run tests**. - **The Tests tab:** red. The error lists three reasons: no fail-first, a mutation score of 55%, and a guard problem. - **Why reasons matter:** the agent gets told exactly what is missing, so it can fix it. **Give the agent the reasons, not just "no".** A hook that prints "mutation score 55% is below 80%" helps the agent write better tests. A hook that only fails makes it guess. **Simple lesson:** "done" means the evidence is there, and a script can check it. ## Where the Gates Live The ideas are the same in every tool. Only the file changes. | Piece | Claude Code | Codex | Cursor | |---|---|---|---| | Test policy (rule) | `CLAUDE.md` | `AGENTS.md` | `AGENTS.md` or `.cursor/rules` | | Review checklist (rule) | `.claude/skills/` | `.agents/skills/` | `.cursor/skills/` | | Guard after each edit (gate) | PostToolUse hook in `.claude/settings.json` | Hook in `.codex/hooks.json` | `.cursor/hooks.json` | | Mutation gate (gate) | Hook or CI | Hook or CI | Hook or CI | | Merge gate | GitHub Actions or Azure Pipelines | Same | Same | - **Hooks** run right after the agent edits a file. The agent sees the reason and tries again. - **CI** runs on every pull request. It is the last barrier, for people and agents alike. **Hook formats change.** The table and the pack follow each tool's docs as of November 2026. Check your tool's hooks page when you install the pack. ### The Final Pack This pack adds the mutation gate hook to the v1 pack from **User-Flow Checks**. **Simple lesson:** rules in the rules file, gates in hooks and CI, the same ideas in every tool. ## TL;DR 1. **Gate** = an automatic check that blocks a change until the evidence is there. 2. **Rule** = guidance the agent reads. Usually followed, never forced. 3. **Faking green** = skipping, loosening or deleting a test to get green. 4. **Test-diff guard** = blocks `.skip`, `.only`, `.todo`, removed checks and loosened checks. Lets real work through. 5. **Definition of done** = types clean, tests pass, fail-first, mutation score, guard clean. Each missing item is a reason. 6. **Where gates live:** hooks after each edit, CI before each merge. ## Common mistakes | Mistake | Wrong idea | Correct idea | What goes wrong | |---|---|---|---| | Rules only | "The rules file says never skip tests." | Add a hook that blocks it. | One day the agent skips a test anyway. | | A guard that blocks too much | "Block every removed line." | Only test files, and allow moved checks. | Real fixes are blocked, and people turn the guard off. | | "No" without reasons | "The hook just exits with an error." | Print what is missing. | The agent guesses and loops. | | `<=` in the threshold | "Below or equal to 80 fails." | Exactly the threshold passes. | Good changes are blocked. | | Gates in hooks only | "The hook is enough." | Also run them in CI before merge. | A change made outside the agent skips every check. | ## Check Your Understanding ## Exercises ## What to Remember - Rules guide. Gates block. - Agents may fake green by skipping, loosening or deleting tests. - A test-diff guard blocks the shortcuts and lets real work through. - Write the definition of done as code, with clear reasons. - Run gates in hooks after each edit, and in CI before each merge. ## Further Reading - [Claude Code hooks](https://docs.claude.com/en/docs/claude-code/hooks): PostToolUse and other hook events. - [Stryker incremental mode](https://stryker-mutator.io/docs/stryker-js/incremental/): run mutation testing on changed code only. - [GitHub protected branches](https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/managing-protected-branches/about-protected-branches): require checks before merge. Next: **Characterization Tests**. You will pin what old code does today, so an AI refactor cannot quietly change it.