openccu-loom: Skill for Claude Code

Cross-stack model-snapshot verification

An end-to-end regression check, not a parity mandate: it catches unintended model regressions when both stacks load the same wire data. The common-schema definition is notes/parity/model_snapshot_schema.md.

Both stacks run against the same godevccu release (the Python side builds the binary pinned in go.mod), so the wire data is identical by construction.

Run the three scripts in this order:

# 1. Dump OpenCCU-Loom's model against godevccu (~80k DPs, 60+ MB JSON).
go test -tags=integration -timeout=300s \
    -run TestModelSnapshotDumpAgainstGodevccu ./tests/integration/...

# 2. Dump aiohomematic's model against godevccu (~8k DPs, ~8 MB JSON).
#    The script builds the go.mod-pinned godevccu (needs `go` on PATH, or
#    GODEVCCU_BIN) and re-execs in the aiohomematic venv when openccu_data is
#    not on sys.path — without that it silently emits empty parameter labels
#    and masks real drift.
python3 script/aiohomematic_snapshot.py

# 3. Per-field diff (tolerated: `profile`, `wrapped_dps`; paramsets-channel
#    field excluded). Exit 0 means full intersection parity.
python3 script/model_snapshot_diff.py

OPENCCU_LOOM_SNAPSHOT_DEVICES=A,B,C scopes both sides to a smoke subset for fast iteration; the default loads the whole embedded fleet. The two snapshot JSONs under tests/integration/testdata/ (~70 MB) are gitignored — produced on demand, never committed.

Reading the result

The baseline is ~270 architecturally-accepted drifts. Growth beyond that in your area, without a matching entry in notes/parity/by_design.md, is a regression — not something to re-baseline.