tests/audio_cli/ mirrors implementation ownership under src/audio_cli/: dsp/denoise/, environments/, export/, media/, packages/, pipeline/, and transcribe/ with its adapter, orchestrator, planner, refusal, result, stage, and transport subpackages. Tests for root composition modules and shared services remain directly under tests/audio_cli/. CLI tests belong there even when they exercise a feature command; provider workflow tests belong to the orchestrator, including tests of the frozen native compatibility entry point.
tests/architecture/ enforces the package graph, isolated stage imports, source layout, and file size budget. tests/docs/ checks specification consistency and shipped command payloads. tests/benchmark/ checks research runner behavior. Shared helpers live beside their owning tests and use explicit package imports; cross-owner helpers are imported from that owner rather than through path injection. tests/fixtures/ contains tracked, compact recorded fixtures and pinned upstream source samples.
Run the full suite with uv run --extra dev pytest, or select an owner, for example uv run --extra dev pytest tests/audio_cli/transcribe/adapters. Install hooks with uv run --extra dev pre-commit install; run the deterministic lint and format gate with uv run --extra dev pre-commit run --all-files.
For a user-facing feature, finish with fresh-agent acceptance. It defines the operator isolation, original-media cases, required evidence, component-level outcomes and final feature verdict. Green implementation tests alone do not establish that an agent can complete the user's workflow; documentation-only changes follow its narrower verification rule.
Put private recordings and generated acceptance output under tests/artifacts/. The directory is ignored except for its placeholder and excluded from source distributions. The local samples are tests/artifacts/media/autio-test-sample.m4a, tests/artifacts/media/test-sample-multispeaker.m4a, and tests/artifacts/media/demo-video-audio-to-improve.mp4. They are optional local inputs and are never modified in place. Write generated output to a separate path, such as tests/artifacts/runs/.
Benchmark evidence keeps its existing convention: runners under model_tests/benchmark/, tracked compact results under model_tests/benchmark/results/, and ignored raw artifacts under model_tests/benchmark_runs/. Historical result JSON retains its original paths and hashes; only current runnable examples and fixture discovery use the relocated media paths.