| name | run-ci-and-test-replays |
| description | Runs the local CI gate (ruff + mypy + pytest) and writes/maintains real-values golden-master tests against the session-scoped replay fixtures for mcp-replay-dota2. Use whenever the user wants to run the tests, run CI locally, write or fix a test, add a fixture, test against real match data, run ruff/mypy/pytest, or check things "before pushing" in this Dota 2 MCP server. Encodes the exact commands, the uv-sync (not uv-pip) gotcha, the fixtures-only rule, the golden-master verified constants, the stable-ID-vs-display-name lesson, the pytest markers, and why CI runs only the smaller replay. |
Run CI and write replay tests for mcp-replay-dota2
CLAUDE.md (repo root) states the policy; this skill is the operational detail.
The CI gate (all three mandatory before any commit/push, in this order)
uv run ruff check src/ tests/ dota_match_mcp_server.py
uv run mypy src/ dota_match_mcp_server.py --ignore-missing-imports
uv run pytest
For one new test, run just it first:
uv run pytest tests/services/farming/test_farming_service.py::TestMultiCampDetection -v
Install deps with uv sync (or uv sync --group dev), NOT uv pip install. Always prefix with
uv run. Never call python/pip directly.
Replay files & cache
Tests need two real replays in ~/dota2/replays/ (override the dir with DOTA_REPLAY_CACHE):
8461956309.dem — primary, ~400 MB.8594217096.dem — secondary (OG match), smaller.
conftest.py parses each replay ONCE at session start (session-scoped, in-memory + diskcache)
and injects ~65 pre-sliced fixtures (e.g. hero_deaths, combat_log_280_290,
combat_log_280_290_earthshaker, objectives, rune_pickups, and the *_2 variants for
match 2). asyncio_mode=auto, so async test functions need no decorator. Disk cache lives at
~/.cache/mcp_dota2/parsed_replays_v2 with a 7-day TTL.
NEVER call Parser or get_parsed_data directly in a test — request fixtures instead. If your
test needs data no fixture provides, add a new session-scoped fixture in conftest.py that
derives from the already-parsed data (don't parse again).
Golden-master verified constants (assert these EXACT values)
Match 8461956309: first blood = earthshaker killed by disruptor at 288.0s ("4:48").
Match 8594217096: first blood = batrider by pugna at 84.0s ("1:24"); total deaths after
start = 53; roshan kills = 3; tower kills = 14; barracks kills = 6; rune pickups = 13;
courier kills = 5.
Good test asserts a real victim/killer/time/count. Bad test asserts isinstance(...), len > 0,
or key existence — those are banned (see CLAUDE.md).
Stable-ID lesson (commit 17473c4) — for live OpenDota data
OpenDota display names are MUTABLE (Yatoro -> yatoro, BetBoom -> BB). Golden-mastering them is
banned. For pro players/teams assert the STABLE id (account_id / team_id) and match names
case-insensitively. Also avoid valueless range checks (e.g. "duration between X and Y") — assert
a real expected value or a stable id.
pytest markers (pyproject.toml)
fast (no replay parsing), slow (parses replays, 1+ min), integration (external/network),
core (essential), use_case (documented use cases). Deselect the heavy ones with
uv run pytest -m "not slow". --strict-markers is on, so unknown markers error.
Why CI differs from local
.github/workflows/test.yml deliberately runs ONLY
tests/examples/test_match_8594217096.py -v (the smaller replay) because the ~400 MB replay OOMs
GitHub runners. The full suite runs locally only. The workflow has concurrency cancel-in-progress: false so a release push won't cancel an in-flight test run.
HARD CONSTRAINT for this machine right now
A full python-manta test run may be in progress. Do NOT kick off uv run pytest or anything that
re-parses replays unless the user explicitly asks. Verify statically (read files, grep) when you
can.
