--- name: skippy-correctness description: Use this skill when validating skippy staged execution against full-model execution, adding model families, changing split boundaries, testing activation wire dtypes, or diagnosing mismatch behavior. metadata: short-description: Validate staged execution exactness --- # skippy-correctness Use this skill when staged execution must be proven equivalent to full-model execution. ## What To Check - Single-stage direct GGUF parity. - Two-stage boundary parity for representative split points. - Multi-stage chain parity for package-backed serving. - Selected-device and pinned-device behavior. - Activation wire dtype exactness (`f16` by default, `q8` only with evidence). - Recurrent/hybrid family behavior and topology affinity. - Multimodal projector handling once native media execution is wired. When native code moves between Skippy modules, preserve the existing capability boundary during validation: - `execution` and `activation` changes require direct, two-stage, and multi-stage parity coverage. - `state` changes require KV import/export, checkpoint, trim, and exact-prefix cache coverage. - `sampling` and `speculative_decoding` changes require deterministic sampling, draft acceptance, rejection recovery, and checkpoint cleanup coverage. - `model_package` changes require inspection, tensor filtering, and artifact writer coverage. Do not put a test-only implementation hook back into `src/skippy.cpp`; keep the test with the capability that owns the behavior. ## Commands First check whether standalone correctness crates have been imported: ```bash cargo metadata --no-deps --format-version 1 | jq -r '.packages[].name' | sort ``` Current mesh-level checks: ```bash cargo test -p skippy-runtime --lib cargo test -p skippy-serving --lib cargo test -p mesh-llm-host-runtime --lib inference::skippy cargo test -p mesh-llm-host-runtime --lib ``` If `skippy-correctness` is imported later, prefer that harness for model-backed exactness gates instead of adding one-off tests.