Skippy Family Certification
Testing & Qualityskippy-family-certificationSkippy Family Certification 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 certifying a GGUF model family for skippy stage-split serving, reviewing capability data, promoting family evidence into topology policy, or updating staged split certification docs.
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-family-certification/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/skippy-family-certification, 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-family-certification
Use this skill for end-to-end family certification, not a one-off correctness smoke. Certification means collecting evidence for full-model parity, staged activation handoff, recurrent/hybrid state behavior, topology constraints, selected-device behavior, and package materialization.
Workflow
-
Inspect the model with
skippy-runtime::ModelInfoor the model-package helpers before choosing split points. Keep topology policy incrates/skippy-topology. -
Prefer reviewed capability data in
crates/skippy-topology/capabilities/reviewed-family-capabilities.json. Do not enable default staged splits for a family without reviewed evidence. -
For dense models, validate at least one representative two-stage boundary and one multi-stage boundary. For recurrent or hybrid families, validate recurrent ranges explicitly and treat recurrent owners as topology-affinity constraints.
-
Compare staged output against full-model execution with the correctness harness when it is present. In this mesh repo, some standalone skippy harness crates may still be migration candidates; do not invent replacement commands without checking
cargo metadata.
Decision Rules
Default activation wire dtype is f16. Treat q8 as per-family and per-split
opt-in only after exactness evidence exists.
Do not recommend transferring recurrent state during normal decode unless the family has explicit reviewed evidence for it. Prefer sticky recurrent ownership and route future tokens for the same sequence back to those owners.
Keep lifecycle phases separate for large models: inspect/materialize, drop any full source model, then launch staged serving. Avoid holding a full source GGUF resident while testing staged servers.