tdd

The agent codes better with a tight feedback loop than with a long specification. A failing test is the tightest loop there is: it states the target, and the target either goes green or it does not.

The loop

1. Red. Write one failing test that states the next slice of behavior. Run it. Confirm it fails for the right reason - a test that passes before you write the code is testing nothing. If it errors instead of failing on the assertion, fix the test setup first.
2. Green. Write the least code that makes the test pass. Not the general solution, not the abstraction - the minimal move. Run the test. Green.
3. Refactor. Now clean up with the test as a net: remove duplication, name things, simplify. Re-run after every edit. Behavior is frozen; only the shape changes.
4. Repeat for the next slice. Small slices. The rate of feedback is the speed limit - never take on a step too big to hold a single test.

What makes a test worth writing

• Tests behavior, not implementation. Assert on the observable result, not on which private method got called. A test that breaks on every refactor is a liability.
• One reason to fail. Each test pins one behavior. When it goes red you should know what broke without reading the body.
• Real seams, minimal mocks. Mock at the system boundary (network, clock, filesystem), not internal collaborators. A test that mocks the thing under test asserts nothing.
• No tautologies. A test that restates the implementation (expect(x).toBe(x) after setting x) verifies nothing. Assert the value the behavior should produce.

When to reach for it

Reach for TDD when the behavior is specifiable and a test can pin it: business logic, parsers, state machines, bug fixes (write the failing case first, then fix). Skip it for pure exploration, throwaway spikes, and layout-only UI where a test asserts nothing a human would not eyeball.

Output contract

Show each red-green transition, not just the final green. If a test is hard to write, say what the difficulty reveals about the design - untestable code is usually badly-seamed code, and that is a finding, not an obstacle.