Add provider-neutral explicit Decision evaluation through the existing runtime, declared provider capabilities, and a default-off Labs consent foundation. Keep explicit evaluation independent of Labs and retire the unreleased TypeSafe-specific tool. No automatic consumer or public selection/local-availability inspection API is added. Verified the registered core tool with real ONNX CPU inference, host-bound rejection before dispatch, and agent disablement. Preserve the contributor implementation and the reviewed Labs UI/config behavior. Related: #155115, #155314, #155317 Co-authored-by: Jacqueline Henriksen <jjjhenriksen@gmail.com> Co-authored-by: jalehman <550978+jalehman@users.noreply.github.com> Co-authored-by: Josh Lehman <550978+jalehman@users.noreply.github.com> |
||
|---|---|---|
| .. | ||
| assets | ||
| scripts | ||
| src | ||
| index.ts | ||
| openclaw.plugin.json | ||
| package.json | ||
| README.md | ||
@openclaw/onnx
Official local ONNX decision-model plugin for OpenClaw. It evaluates Choice, Score, and Boolean rubrics with GLiClass, GLiNER2.5, and DeBERTa classifiers in a persistent CPU inference process.
Setup
Released OpenClaw 2026.9.5 lacks the decision-provider API. Packaged installs
require a host and plugin API of at least 2026.9.6.
For development, use an OpenClaw checkout containing both the decision-provider
API and this plugin. Run pnpm install --frozen-lockfile and pnpm build, enable
plugins.entries.onnx, and run these commands from the checkout:
pnpm openclaw onnx models
pnpm openclaw onnx download gliclass-edge-v3.0
pnpm openclaw onnx probe gliclass-edge-v3.0
Select onnx/gliclass-edge-v3.0 as agents.defaults.decisionModel or as an
agent's override. Inference stays local; model downloads use pinned revisions,
sizes, and SHA256 hashes. Large models may need preloading to meet the host's
30-second decision deadline.
On a compatible packaged host, install the local candidate with
openclaw plugins install npm-pack:/path/to/openclaw-onnx.tgz. A development
checkout's co-versioned source loading does not grant compatibility to an older
packaged host.
See Local ONNX decision models for the model catalog, configuration, local exports, and runtime limits. The Decision models guide explains rubric definitions, score semantics, and the plugin API shared with TypeSafe AI.