Test-Driven Development

Implement $ARGUMENTS using Test-Driven Development (TDD).

TDD Cycle

Repeat: Red → Green → Refactor

1. Red:    Write a failing test
2. Green:  Write minimal code to pass the test
3. Refactor: Clean up code (tests still pass)

Which test cases matter, and when to refactor, is judgment. The one thing that is not judgment is whether a run came out the way you expected: every test run in this skill goes through .claude/skills/_shared/run_tests.py, which turns "confirm failure" / "confirm success" into an exit code.

Who Does What (delegation-first)

Per .claude/rules/delegation.md, the lead does not write the cycles by hand. The split is fixed:

Work	Owner
Requirement clarification, test-case list, boundary values	Lead (judgment, Self-Handle List item 5)
Writing the tests and the production code, cycle by cycle	general-purpose-sonnet
Cycles with ambiguous design, cross-cutting invariants, or security / concurrency / data-integrity risk	general-purpose-opus
A Red that is red for the wrong reason, or a Green that will not go green after one retry	codex-debugger
Every run_tests.py / verify.sh gate and the final report	Lead — verification never delegates

Only a single trivial cycle on a file already open in context stays with the lead (Self-Handle List item 2). Two or more cycles, or a module the lead has not read, is delegated. Independent modules are delegated in parallel in one message; they share no test file, so nothing serializes them.

The Red/Green Invariant

python3 .claude/skills/_shared/run_tests.py \
  --target tests/test_{module}.py --expect fail --label red-1

--expect fail|pass (required) and --target PATH (repeatable, at least one) are the whole interface; --label names the log file under .claude/logs/ (default run-tests), so a later cycle does not overwrite an earlier log.

Read observed in the JSON — a non-zero exit is not all the same red:

observed	pytest exit	What it actually means
failed	1	The test ran and failed. This is the only valid Red.
passed	0	Nothing new is being asserted — the test cannot drive any code.
collection_error	2	The file did not import (syntax error, bad fixture, missing import): red for the wrong reason. Fix the test, not the production code.
no_tests_collected	5	Nothing was selected — the test was never written, or the name does not match the discovery pattern.

Exit codes: 0 observed matches --expect; 1 bad arguments — including a --target that does not exist (the mistyped path, caught before pytest runs) and a pytest usage error (exit 4, e.g. an unknown ::node id); 2 observed does not match --expect, or coverage is below --min-coverage; 3 external failure — no pytest runner available, a timeout (--timeout, default 600s), or a pytest internal error. On 1 and 3 the payload's observed is null: the script reports that it has no observation rather than guessing one.

Other payload fields: expected, runner, command, exit_code, summary, failed_tests (node ids from the short summary), coverage_percent, min_coverage, log_file, artifacts, error.

The runner is resolved, not hardcoded: uv run pytest when uv and a pyproject.toml are present, otherwise pytest from PATH, otherwise python -m pytest (probed for importability first). --runner pins one explicitly. If none is available the result is exit 3, never a silent pass.

Implementation Steps

Phase 1: Test Design

1. Confirm Requirements

   • What is the input
   • What is the output
   • What are the edge cases

2. List Test Cases — record them with TodoWrite (one todo per case), so the remaining cases survive a context reset instead of living only in this conversation:

   - [ ] Happy path: Basic functionality
   - [ ] Happy path: Boundary values
   - [ ] Error case: Invalid input
   - [ ] Error case: Error handling
   

Which cases to write, and which boundary values matter, is domain reasoning — never delegate it to a script.

Phase 2: Red-Green-Refactor

Step 0: Hand the Cycles Off

Delegate the Red-Green-Refactor loop with the six-element prompt contract from .claude/rules/delegation.md. One delegation per module, all launched together:

Task tool:
  subagent_type: "general-purpose-sonnet"   # or general-purpose-opus, see the table above
  prompt: |
    Objective: Implement {module} by strict TDD, one test case at a time.

    Scope:
    - Write only tests/test_{module}.py and src/{module}.py. Touch nothing else.
    - Do not modify existing tests, and never skip, delete, or weaken an assertion.

    Inputs:
    - Test cases to implement, in this order: {the Phase 1 list}
    - Read .claude/rules/coding-principles.md and .claude/rules/testing.md first.
    - Existing conventions to follow: {paths of comparable modules}

    Acceptance checks — run these yourself, per cycle, and never skip Red:
      python3 .claude/skills/_shared/run_tests.py \
        --target tests/test_{module}.py --expect fail --label red-{n}
      python3 .claude/skills/_shared/run_tests.py \
        --target tests/test_{module}.py --expect pass --label green-{n}
    A Red that exits 2 with observed=passed/collection_error/no_tests_collected
    is not a Red: fix the test, not the production code, and re-run.

    Output shape:
    ## Cycles completed (case -> red label -> green label)
    ## Files changed
    ## Deviations from the requested test cases, and why
    ## Anything that did not go green

    Context discipline: return this summary only; the run logs stay in
    .claude/logs/ and are referenced by label, not pasted.

Then verify rather than trust: re-run run_tests.py --expect pass yourself and read the diff. A reported cycle with no matching run_tests.py exit 0 did not happen. If a cycle came back unfinished, escalate it (general-purpose-opus or codex-debugger) instead of re-sending the same prompt to the same tier.

The steps below are the contract the delegate follows — and what the lead runs directly in the single-trivial-cycle case.

Step 1: Write First Test (Red)

# tests/test_{module}.py
def test_{function}_basic():
    """Test the most basic case"""
    result = function(input)
    assert result == expected

Confirm the test is red for the right reason:

python3 .claude/skills/_shared/run_tests.py \
  --target tests/test_{module}.py --expect fail --label red-{n}

Exit 0 means it genuinely failed. Do not proceed to Green on any other exit code: on exit 2 read observed (see the table above), on exit 1/3 read error and log_file.

Step 2: Implementation (Green)

Write minimal code to pass the test:

• Don't aim for perfection
• Hardcoding is OK
• Just make the test pass

Confirm success:

python3 .claude/skills/_shared/run_tests.py \
  --target tests/test_{module}.py --expect pass --label green-{n}

Exit 0 means the test now passes. Exit 2 with observed: failed means the implementation is not there yet — read failed_tests.

Step 3: Refactoring (Refactor)

Improve while tests still pass:

• Remove duplication
• Improve naming
• Clean up structure

python3 .claude/skills/_shared/run_tests.py \
  --target tests/test_{module}.py --expect pass --label refactor-{n}

When to refactor, and how far, stays a judgment call.

Step 4: Next Test

Return to Step 1 with the next test case from the Phase 1 list.

Phase 3: Completion Check

Run the full quality gates:

bash .claude/skills/_shared/verify.sh

Read the JSON: overall is pass / fail / no_gates. Exit 0 means overall: pass; exit 2 means a gate failed or no gate could run at all — inspect log_file and tools. no_gates is a failure by default because a code-editing session must not be able to declare done with zero checks executed; if the project genuinely has no configured gates, re-run with --allow-no-gates and verify manually with the project's own commands, and say in the report that you did so. Exit 1 is bad arguments, 3 the log file could not be written.

Then check coverage — with a threshold, so the answer is a gate and not a glance at term-missing output:

python3 .claude/skills/_shared/run_tests.py \
  --target tests/test_{module}.py --expect pass \
  --cov {module} --min-coverage {N} --label coverage

--cov MODULE (repeatable) requires the pytest-cov plugin; without it pytest reports a usage error and run_tests.py exits 1 pointing at log_file. Coverage below --min-coverage is exit 2; --cov with no coverage total in the output is exit 3, never a silent pass. --min-coverage requires --cov. Choosing the threshold is a project decision.

Report Format

## TDD Complete: {Feature Name}

### Test Cases
- [x] {test1}: {description}
- [x] {test2}: {description}
...

### Coverage
- {coverage_percent}% (threshold {min_coverage}%) — from run_tests.py, label `coverage`

### Quality Gates
- verify.sh: {overall} — `{log_file}`

### Implementation Files
- `src/{module}.py`: {description}
- `tests/test_{module}.py`: {N} tests

Every [x] above must correspond to a run_tests.py --expect pass run that exited 0; mark nothing complete on the strength of having written it.

Notes

• Write tests first (not after)
• Keep each cycle small
• Refactor after tests pass
• Prioritize working code over perfection
• Where the production code belongs follows the project's existing layout — that is a design decision, not a derived path