Tao Tdd Pipeline

tao-tdd-pipeline

Drives a TDD discipline on top of tao_loop.run_until_done. The user provides a goal that names the test scenarios first, the implementation last; the worker writes failing tests, makes them pass, refactors. The pipeline gates "done" on three conjunctive conditions:

1. The standard tao_judge.judge returns done=True (GOAL_MET).
2. git diff --name-only HEAD includes at least one file matching pytest test conventions (test_*.py, *_test.py, tests/*.py, conftest.py).
3. python -m pytest -q tests/ exits 0 with a passed summary line.

If the loop reports GOAL_MET but #2 or #3 fail, the result is returned with done=False, reason='TESTS_NOT_GREEN' or 'NO_TEST_FILES_MODIFIED', and discipline_violations populated for the caller to act on.

When to trigger

• A user issues a TDD-style brief — explicitly mentions tests, scenarios, red/green, or asks for "test-first" / "TDD" / "test-driven".
• The orchestrator wants stronger guarantees than vanilla tao-loop for ambiguous correctness goals (e.g. "fix this bug" → "add a regression test, then fix").

Public API

from app.server.tao_tdd_pipeline import run_tdd, TddResult

result = await run_tdd(
    goal="implement a hex-string parser; tests cover empty / non-hex / "
         "uppercase / mixed-case / leading-0x cases",
    workspace="/path/to/repo",
    max_iters=15,
    max_cost_usd=2.00,
    timeout_per_iter_s=600,
    evt: ...,
    session_id="tdd-hex-parser-001",
)

# result.done                   — True iff loop AND tests AND test-files
# result.reason                 — composite (see TddResult.reason)
# result.loop                   — full LoopResult from tao_loop
# result.test_files_modified    — list of paths matching test convention
# result.final_pytest_passed    — pytest exit-code 0 + passed summary
# result.final_pytest_summary   — last 1500 chars of pytest combined output
# result.discipline_violations  — strings explaining loop.done but result.done=False

Autoresearch envelope

CLI

python scripts/run_tao_tdd.py \
    --goal "..." \
    --workspace /path/to/repo \
    --max-iters 10 \
    --max-cost 1.50 \
    --judge-every 1

Exit codes: 0 done, 1 loop met but discipline violated, 2 kill-switched / never met goal.

Composition with other Wave 1+2 modules

• tao-context-prune (RA-1990): runs upstream of the SDK call so the worker's transcript stays small across iters.
• tao-context-mode (RA-1969): sharper file-expand decisions during the implementation phase.
• tao-judge (RA-1970): unchanged — the TDD discipline is added as a post-loop gate so the existing judge contract isn't fractured.

Behaviour invariants

• Goal is enriched with a _TDD_PREAMBLE describing the red→green→refactor order. Idempotent — if "test-first discipline" already appears in the goal we don't double-add.
• The worker is NOT prevented from deciding to skip tests; the gate is retrospective. This keeps the existing tao_loop contract clean and surfaces discipline violations as data the caller can route on (e.g. flag-as-needs-review rather than auto-merge).
• Pytest invocation uses -q --tb=line so summary lines parse reliably.
• Test-file regex is liberal: test_*.py, *_test.py, conftest.py, tests/*.py. Conservative regression risk if a project uses a non-pytest convention (e.g. nose2); document and override per repo if that turns up.