On-device LiteRT model knowledge
OKF v0.2 bundle for Kai’s pinned on-device model downloads. Agents use this folder for pin policy, digest provenance, and freshness when checking or bumping the runtime catalog.
Runtime vs knowledge
| Layer | Path | Role |
|---|---|---|
| Runtime (shipped) | composeApp/src/commonMain/kotlin/com/inspiredandroid/kai/inference/LocalModelCatalog.kt |
Source of truth for download URL, SHA-256, size, and context defaults |
| Knowledge (this bundle) | docs/knowledge/litert/ |
Pin policy, attested digest snapshot, sources, stale dates |
| Refresh skill | .grok/skills/update-litert-models/SKILL.md |
Checks HuggingFace pins, updates Kotlin + this bundle only on an explicit bump |
The app never reads this directory at runtime. Production downloads use the pinned URL in Kotlin only.
Scope
- In: Catalog
.litertlmfiles from HuggingFacelitert-community: pinned commit, SHA-256, exact byte size. - Out: User imports (no digest). GPU-memory heuristics, tool allowlist, and engine lifecycle — those stay in
docs/features/on-device-inference.md. - Out: Cloud model Elo / free-tier flags.
A refresh verifies pins. It does not silently move a pin to main. A bump (new commit + digest + size) is an explicit product decision because users re-download gigabytes.
Catalogs
- Pinned LiteRT models — current pins vs HuggingFace
Policy
- Pin and bump policy — never
main, commit + digest + size together
Playbooks
- Refresh LiteRT pins — check or bump (also
/update-litert-models)