Testing

pip install -e . pytest pyflakes
python -m pyflakes src tests
python -m pytest -q

Layout

File	Covers
tests/test_unit.py	model.py structural validation (bounds, forbidden fields, enums, id patterns), fonts.py resolution, canonical.py hashing, the error-code table
tests/test_security.py	PathPolicy: traversal, prefix-collision (not containment), symlink escape, workspace escape, unsafe file names, NUL bytes; the forbidden-field boundary exercised through run_tool
tests/test_integration.py	Real Pillow rendering against real image fixtures: still-image thumbnails, PNG and JPEG output, z-order stacking, all four fit modes, multiline/Unicode (including Japanese) text, deterministic re-render (byte-identical), reuse (cache hit, cache-corruption recovery), provenance completeness, the run - stdin/stdout transport, and — when a real ffmpeg-skill checkout and ffmpeg/ffprobe are available — actual video_frame extraction at an exact timestamp and its INVALID_TIME_RANGE boundary
tests/test_contract.py	The contract and doctor report are valid JSON and match the implementation (tool ids, error codes, element types, output formats, font ids); no unimplemented operation (best-frame selection, auto layout, title generation, face detection, click-through prediction) is ever named in the contract

Fixtures

tests/fixtures/generate.py builds every fixture at test time (nothing binary is committed): solid-colour PNG/JPEG stills, an RGBA overlay image, a tall image (to exercise fit against a mismatched aspect ratio), a non-image text file (for "invalid image" cases), and — only when ffmpeg/ffprobe are on PATH — a short synthetic H.264 video. Tests that need the video fixture or an ffmpeg-skill checkout skip cleanly (pytest.skip) rather than failing when either is absent, the same convention audio-production-skill and transcription-skill use.

Real-media video E2E

CI clones a pinned ffmpeg-skill checkout into vendor/ffmpeg-skill and points THUMBNAIL_SKILL_FFMPEG_SKILL_DIR at it (see .github/workflows/tests.yml), the same pattern audio-production-skill uses. Locally:

git clone https://github.com/kajisho5/ffmpeg-skill vendor/ffmpeg-skill
THUMBNAIL_SKILL_FFMPEG_SKILL_DIR=$PWD/vendor/ffmpeg-skill python -m pytest -q

Deterministic / visual verification

test_deterministic_render_same_bytes renders the same document twice and asserts byte-identical output. test_reuse_hits_cache_on_second_render deletes the output and renders again, asserting reused: true and an identical sha256. test_reuse_rebuilds_when_cache_entry_is_corrupted corrupts the cache file directly and asserts the next render notices and rebuilds rather than serving a broken file as reused: true. test_z_order_stacking samples an actual output pixel to prove stacking order is respected, rather than trusting the code path alone. There is no subjective "is this a good thumbnail" check anywhere, deliberately: that judgement is not this skill's to make.