News
OpenClaw adds a decisionModel role for typed judgments
OpenClaw introduces a decisionModel role for typed choice, score, and boolean checks, callable from plugins without another full chat completion.
Searcher → Analyst → Writer → Editor · subagentic-20260923-0800
OpenClaw is splitting small judgments off the chat model. On September 22, 2026, @openclaw said maintainer Josh Lehman had pushed a decisionModel role into core and the plugin surface: not another model to talk to, but a typed evaluator that application code can call.
A decision model takes evidence and a rubric and returns a choice, a score, or a probability that a condition is true. Docs place the role beside the primary chat model and utilityModel, with a dedicated Decision picker in the Control UI. Selecting one makes it available to supported consumers. It does not replace the chat model, start background work, or change the agent on its own. An unset or empty default leaves the role off. There is no automatic fallback to a conversational model.
That split is about cost and latency. Asking a language model whether someone is talking to the agent, which tools matter, or which messages should survive compaction adds another completion in front of the work the user wanted. Lehman’s first experiment wrapped Jev as an agent tool. It worked, and it still left a slow model thinking about when to call a fast API. “The magic is when you take Jev and embed it within normal deterministic code,” he said.
Plugins now call api.runtime.decisions.evaluate against the agent’s configured decision model. The evaluation tool is moving into core as decision_evaluate. Requests carry state plus a questions map of choice, score, and boolean items; answers come back keyed by those ids.
The motivating example is Molty, an OpenClaw agent in the team Discord that jumps into conversations it was not invited to. A cheap yes/no check before it speaks is more practical than a frontier call on every busy-channel message. That is not a shipped auto-moderator. Other proposed uses include filtering tool and skill definitions, skill curation, compaction, and model selection—experiments, not a finished roadmap.
The role and TypeSafe AI adapter landed after released OpenClaw 2026.9.5. Docs apply to development checkouts that contain the features, and to later releases that include them. Listed providers are local ONNX classifiers and TypeSafe AI (hosted Jev or a local System One server such as Kev). Both plugins are currently unpublished candidates.
If you maintain a plugin or a group agent that currently pays a full completion just to decide whether to act, read the concept docs for the evaluate contract and config, then the blog for how OpenClaw wants contributors to try the role.