mcp-replay-dota2: Skill for Claude Code

Add a replay collector service to mcp-replay-dota2

Read CLAUDE.md (repo root) first for python-manta attribute/enum rules, lane naming, and the tests+docs policy. This skill covers the service-layer architecture those rules sit inside.

A "collector" is a service: business logic that takes a parsed replay and returns models. Tools live one layer up. To then expose the service as a tool, use the add-mcp-tool skill.

Layer order (never bypass)

external (tools/CLI/web) -> services -> python-manta. Tools must hold no extraction logic; services must hold no MCP code. Adding a service is purely a services-layer change until step 7.

The input contract

Every public service method takes a ParsedReplayData (from src.services.models.replay_data). It is produced ONCE by await ReplayService.get_parsed_data(match_id) and cached on disk. Services never call python-manta Parser and never re-parse — they read everything from the single passed-in ParsedReplayData. ReplayService does the one-and-only single-pass parser.parse(...) with the combat-log types and entity interval_ticks requested together.

Workflow

1. Create the package: src/services/<domain>/__init__.py and src/services/<domain>/<domain>_service.py. The __init__.py exports the class.
2. Class shape: a class <Domain>Service: whose public methods take ParsedReplayData (plus optional filters / a GameContext) and return Pydantic models.
3. Constructor injection for cross-service dependencies — match the existing pattern:

   # src/services/rotation/rotation_service.py
   def __init__(self, combat_service=None, fight_service=None):
       self._combat = combat_service or CombatService()
       self._fight = fight_service or FightService()
   

   FightService(combat_service=...) and RotationService(combat_service=..., fight_service=...) are the live examples. The entry point wires the real singletons; the or Default() fallback keeps the service usable standalone (e.g. in tests).
4. NO MCP imports. Add NO MCP DEPENDENCIES. to the module docstring (every existing service does). Importing fastmcp here breaks the CLI/web reusability the whole layer exists for.
5. Service-layer model: add src/services/models/<domain>_data.py with the Pydantic return types. These are SEPARATE from the MCP response models in src/models/ — keep them distinct. Existing ones: combat_data, farming_data, jungle_data, lane_data, rotation_data, seek_data, replay_data.
6. GameContext when you need map geometry / team mapping: build it once via GameContext.from_parsed_data(data) (src/models/game_context.py) and pass it in. Services that consume it import it under TYPE_CHECKING to avoid hard coupling.
7. Wire the singleton in dota_match_mcp_server.py: instantiate _<domain>_service = ... alongside the others (~line 86) and add "<domain>_service": _<domain>_service to the services dict (~line 97) so tools can reach it.
8. Optional export from src/services/__init__.py __all__. Note it currently lists replay/cache/combat/fight/analyzers/jungle/lane/seek but NOT farming/rotation — exporting is optional, dict-wiring in the entry point is what actually matters.
9. Real-values tests under tests/services/<domain>/ using conftest fixtures only — see the run-ci-and-test-replays skill. Never instantiate Parser/get_parsed_data in a test.
10. CI gate (also in CLAUDE.md):

    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