⚙️ FEATURE: TDD mode — auto-generate failing tests before code synthesis

[FernandoCelmer]

Problem

pycodeloop agents write code directly from a prompt with no enforced test-first discipline. The verification loop (#37) and regression detection (#38) can only catch failures after code is written. Research papers (TestART, RLTF, Klear-CodeTest) demonstrate that generating failing tests before implementation turn forces specification clarity and provides a tight feedback signal that guides the agent toward correct behavior faster.

Expected behavior

codeloop = CodeLoop(config=Config(
    tdd=True,
    check_cmd="pytest tests/ -q",
))

Turn sequence when tdd=True:

Turn 1 (forced): "Write failing tests for: <original prompt>"
  → agent writes test file
  → check_cmd runs → tests fail (expected)

Turn 2+: "<original prompt> — make the tests pass"
  → agent writes implementation
  → check_cmd runs → tests pass → done

Suggested implementation

• Add tdd: bool = False to Config
• In Agent.run(), when tdd=True, prepend a forced system turn before the user's prompt:
  "Before writing any implementation, first write failing unit tests that specify the expected behavior. Do not write implementation code in this turn."
• After this turn completes and tools are executed, inject the test results as context for the next turn
• Pairs with check_cmd from ⚙️ FEATURE: Execution-based verification loop (write → execute → verify → re-plan) #37/⚙️ FEATURE: Regression detection — capture lint/test baseline before and after each session #38 — tests written in TDD turn become the verification signal

References

• "Code as Agent Harness" — arXiv 2605.18747, Section 3.2 (Execution-based verification)
• TestART: Improving LLM-based Unit Test via Iterative Clarification
• RLTF: Reinforcement Learning from Unit Test Feedback