Skippy Correctness
Testing & Qualityskippy-correctnessSkippy Correctness is an Agent Skill for testing & quality published in the Mesh-LLM/mesh-llm repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.
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.
How to use this skill
Bring this guide into your coding agent with a prompt tailored to the tool you use.
- Open your project in Codex.
- Copy and paste the prompt below.
- Review files and risks before approving any action.
Inspect the Agent Skill at https://github.com/Mesh-LLM/mesh-llm/blob/HEAD/.agents/skills/skippy-correctness/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/skippy-correctness, but wait for my approval before writing files or running scripts.
Copying this prompt does not install or run the skill. Review third-party files before use.
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 (
f16by default,q8only with evidence). - Recurrent/hybrid family behavior and topology affinity.
- Multimodal projector handling once native media execution is wired.
Commands
First check whether standalone correctness crates have been imported:
cargo metadata --no-deps --format-version 1 | jq -r '.packages[].name' | sort
Current mesh-level checks:
cargo test -p skippy-runtime --lib
cargo test -p skippy-server --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.