OpenClaw Adds Dedicated Decision Model Support to Core and Plugins via TypeSafe Jev

OpenClaw has added official support for dedicated decision models across its core framework and plugin ecosystem, integrating TypeSafe AI's Jev models for fast,

tau · September 23, 2026

#OpenClaw #Jev #TypeSafeAI #DecisionModel #Agents #DevTools

OpenClaw Adds Dedicated Decision Model Support to Core and Plugins via TypeSafe Jev

Open-source agent framework OpenClaw has officially introduced a dedicated 'Decision Model' role across its core engine and plugin architecture. Announced in September 2026, this capability allows developers to offload discrete routing, conditional branching, and scoring decisions from heavy conversational LLMs to lightweight decision-specialized models.

OpenClaw official announcement graphic: How OpenClaw is using decision models

Image source: @openclaw on X

In complex multi-step agent workflows, relying on general-purpose chat models to make binary choices or evaluate structured probabilities creates unnecessary token costs and inference latency. Through plugin PR #152298 (which bundled the TypeSafe decision provider) and core PR #155134 (which introduced the shared decision_evaluate tool), OpenClaw now isolates micro-decisions into a separate execution pathway tailored for high-speed deterministic evaluation.

Dedicated Decision Picker and the decision_evaluate Tool

The defining structural shift in OpenClaw's update is the rigorous separation between decision models and conversational chat engines.

Within the Control UI, OpenClaw introduces a standalone 'Decision picker'. Models registered under this provider role do not appear in the primary chat, fallback, or general utility model pickers, preventing operational confusion and accidental model substitution.

  • Specialized Decision Primitives: Rather than generating unstructured free-form text, decision models focus entirely on choices, numerical scoring, and boolean probability calculations.
  • Narrow Evaluation Interface: The accompanying decision_evaluate tool accepts structured runtime states, validates their format and boundary ranges, and returns verified judgment values.
  • Host-Governed Model Routing: Native evaluations enforce host-configured model policies rather than allowing downstream plugin callers to invent arbitrary routing pathways.
  • Strictly Non-Autonomous Execution: Configuring a decision model does not replace conversational agents or initiate background loops; it functions purely as an explicit, on-demand evaluation backend.

TypeSafe Plugin Integration and Jev Model Configuration

To operationalize decision models, OpenClaw provides the official typesafe external plugin (extensions/typesafe).

Operators can connect OpenClaw to TypeSafe AI's hosted cloud models, such as typesafe/jev-latest or the pinned typesafe/jev-1.13.0, or point to a self-hosted System One server like Kev. When using hosted Jev models, operators store their API credentials securely via Settings → Secrets under the TYPESAFE_API_KEY identifier.

The integration is configured in OpenClaw's JSON5 configuration file:

{
  plugins: {
    allow: ["typesafe"],
    entries: {
      typesafe: {
        enabled: true,
        config: {
          apiKey: { source: "store", provider: "default", id: "TYPESAFE_API_KEY" },
        },
      },
    },
  },
  agents: {
    ownership: "explicit",
    defaults: { decisionModel: "typesafe/jev-latest" },
    entries: {
      research: { decisionModel: "typesafe/jev-1.13.0" },
    },
  },
}

This structure allows operators to define a global default decision model while pinning specific agents, such as research workers, to immutable versioned checkpoints for deterministic behavior.

Compatibility Requirements and Deployment Considerations

Engineering teams planning to deploy OpenClaw's decision model pipeline should take note of several version and runtime requirements.

The underlying adapter and decision-provider abstractions were merged after OpenClaw 2026.9.5. Consequently, packaged installations mandate an OpenClaw host and plugin API version of at least 2026.9.6. The package installer automatically verifies this condition and rejects older host versions prior to initialization.

  • Source Checkout Build Path: Pending initial packaging and publishing on npm and ClawHub, teams must build the plugin from a source checkout containing the decision API and extensions/typesafe using pnpm install --frozen-lockfile and pnpm build.
  • Safe Opt-In Defaults: The agents.defaults.decisionModel setting remains unconfigured by default, ensuring that existing OpenClaw deployments experience zero unexpected behavioral changes until explicitly enabled.
  • Transparent Billing and Telemetry: Hosted Jev evaluations consume standard TypeSafe API usage, while plugin lifecycle states and provider health checks integrate directly with OpenClaw's host monitoring.

Sources