TDD Skill (Test-Only, Anti-Cheat, Minimal Version)

This skill defines a strict but lightweight TDD workflow designed for coding agents. It enforces honesty, prevents cheating, and minimizes token usage by focusing solely on tests.

Core Anti-Cheat Contract

(Always enforced in every state)

1. Never claim tests passed or failed unless you have just run them and shown the exact output.

2. If you cannot run tests, enter BLOCKED immediately. Do not guess or infer results.

3. Never modify test assertions just to make a failing test pass.

4. Only transition between TDD states when the required test evidence has been shown.

5. In RED, implement only the minimum code necessary to make the currently failing test pass.

6. Every message must begin with a TDD state prefix, and the prefix must match the actual test evidence.

7. Promote the Dependency Rule If making a failing test for the current SUT pass requires changing a dependency, you must stop and promote that dependency to be the SUT. Write a failing test for the dependency, make it pass, then return to the original SUT.

   Clarification:

   • Modifying a dependency counts as production code.
   • Production code may only be written to satisfy the currently failing test.
   • Therefore, any dependency change must be justified by its own failing test.

TDD Layer Discipline

When testing layer A that mocks layer B:

• Layer B only needs a method signature (stub returning null or throwing UnsupportedOperationException)
• Do NOT implement Layer B's logic until you write Layer B's tests
• Ask: "Is the service/dependency mocked in this test?" If yes, only a stub is needed.

Example:

• Controller test mocks CustomerService → CustomerService.createLocationForCustomer() should be throw new UnsupportedOperationException("implement me")
• CustomerService test (separate task) will drive the actual implementation

Outside-In Development

Use outside-in development: start from the external interface and work inward.

Why: Every piece of code should be immediately callable and tested from an external perspective. Avoid creating "infrastructure only" code to be connected later.

Correct order for implementing a feature:

1. Write test for the external interface (command handler, REST endpoint)
2. Implement the handler/controller, which calls the service method
3. Implement the service method, which uses infrastructure (events, repositories)
4. Create infrastructure as needed to satisfy the test

Anti-pattern: Creating event publishing code or domain methods without a caller (no REST endpoint or command handler to invoke them). Every method should have an entry point that exercises it.

Evidence Format (Required)

Whenever tests are mentioned:

Evidence:

• tests: pass | fail | not-run
• last_output: <exact test output snippet>

Rules:

• pass or fail may only be used if tests were just executed.
• not-run must be used if tests cannot be executed.
• Never infer test status from code.

TDD States

⚪ TDD: PLANNING

Goal:
Write a meaningful failing test that expresses the desired behavior.

Allowed actions:

• Ask clarification questions
• Write or update a test
• Run tests to show failure

Pre-conditions to enter RED:

• A test exists
• The test was just run
• A failing test is shown
• The failure reflects missing/incorrect behavior

Transition:

• If pre-conditions satisfied → RED
• Otherwise remain in PLANNING

🔴 TDD: RED

Goal:
Make the failing test pass using the minimum implementation.

Allowed actions:

• Analyze the failing test
• Implement the smallest possible change
• Run tests again
• Justify why the implementation is minimal

Pre-conditions to enter GREEN:

• The previously failing test now passes
• All tests pass
• Test output is shown

Transition:

• If all tests pass → GREEN
• If tests still fail → remain in RED
• If tests cannot run → BLOCKED

🟢 TDD: GREEN

Goal: The behavior is now correct; evaluate and prepare safe refactoring.

Allowed actions:

• Re-read the actual code changes made in this cycle
• Identify safe refactorings
• Explain improvement intent
• Run tests as needed

Required Evaluation (before any transition):

1. Re-read the actual code changes made in this cycle (use git diff HEAD, not memory)
2. List at least one observation about potential refactoring (even if "no duplication found")
3. If duplication exists, either refactor it OR note it as intentional with justification

Skipping directly from GREEN to VERIFY without this explicit evaluation is a TDD violation.

Pre-conditions to enter REFACTOR:

• All tests currently pass
• Recent test output shown
• A refactoring goal identified

Transition:

• If refactoring is needed → REFACTOR
• If no refactoring needed → VERIFY

🛠️ TDD: REFACTOR

Goal:
Improve internal code structure without changing behavior.

Allowed actions:

• Perform small safe refactorings
• Run tests immediately after each change

Pre-conditions to exit REFACTOR:

• Tests were run after the last change
• All tests pass
• Test output is shown

Transition:

• If tests pass → VERIFY
• If tests fail → return to RED
• If tests cannot run → BLOCKED

✅ TDD: VERIFY

Goal:
Confirm the entire TDD cycle is complete.

Allowed actions:

• Run the full test suite
• Show test results
• Summarize completed behavior

Pre-conditions to enter COMPLETE:

• All tests pass
• Test output is shown

Transition:

• If all tests pass → COMPLETE
• If any test fails → RED
• If tests cannot run → BLOCKED

🏁 TDD: COMPLETE

Goal:
The TDD cycle for this behavior is finished.

Allowed actions:

• Provide a concise summary
• Stop unless the user starts a new cycle

⚠️ TDD: BLOCKED

Goal:
Tests cannot be run; you cannot proceed honestly.

Actions:

• State explicitly why tests cannot run
• Show the exact command the user should run
• Ask for the output

Transition:

• When test results are provided → return to prior state
• Do not infer test status

🚨 TDD: VIOLATION

Goal:
The agent has broken the TDD rules.

Actions:

1. State what rule was violated
2. Explain the correct behavior
3. Undo or repair the incorrect action
4. Return to the correct state

End of Skill