Decision models

A decision model evaluates supplied evidence against a rubric and returns a typed answer: a choice, a score, or a Boolean probability. Use it for bounded judgments such as routing a request, scoring its urgency, or checking whether it meets a condition.

decisionModel is a model role with a shared API. Its providers can use different model architectures and inference backends. Sharing the interface does not make their reasoning ability or probabilities interchangeable.

The role and TypeSafe AI adapter were added after released OpenClaw 2026.9.5. These instructions apply to development checkouts containing those features and to later releases that include them. See each provider's setup page for its host requirements.

Decision models have a separate Decision picker in the Control UI. Selection chooses the provider for explicit evaluation and supported consumers. The core decision_evaluate tool follows that selection plus ordinary tool policy. Selection does not start background work or replace the chat model. Automatic experimental consumers additionally require explicit Decision assistance opt-in. The built-in OpenClaw runtime uses that opt-in for conversational tool filtering with bounded recent conversation. The selected Decision provider receives that secondary evidence; see the privacy and fallback contract. Explicit decision_evaluate remains independent of Labs.

Choose a provider and model

Configure the provider plugin before selecting its model:

• ONNX runs local CPU classifiers in a persistent subprocess. Follow its development-checkout or compatible-package setup, then explicitly download the model or prepare a local export. Inference needs no hosted API credential.
• TypeSafe AI connects to hosted Jev or a local System One server such as Kev. Install and enable the external plugin, then configure a protected hosted credential or an explicit loopback baseUrl. Hosted evaluations send the selected evidence to TypeSafe and incur its normal usage charges.

The current plugins declare these model references:

Both plugins are currently unpublished candidates. Their setup pages explain source-checkout use and the packaged host floor. The table describes the plugins' declared models, not which artifacts or credentials are ready on your machine.

After provider setup, merge the role selection into your configuration:

{
  agents: {
    ownership: "explicit",
    defaults: { decisionModel: "onnx/gliclass-edge-v3.0" },
    entries: {
      support: { decisionModel: "typesafe/jev-latest" },
      quiet: { decisionModel: "" },
    },
  },
}

An unset agent override inherits the global default. An empty agent override disables decisions for that agent. An unset or empty global default leaves the role off. There is no automatic fallback to a conversational model.

For local setup verification, openclaw onnx models lists the presets and openclaw onnx probe gliclass-edge-v3.0 runs a Choice, Score, and Boolean smoke evaluation after the model has been downloaded.

Define a decision

Each request contains shared state and a questions map. State accepts text, a JSON object or array, or null. Question keys identify answers; the instructions and criteria define the judgment. The rubric travels with each request, so there is no separate rubric-registration step.

Use descriptive alternatives and observable score anchors. Give Boolean questions both descriptions when targeting ONNX. Keep evidence focused on the question; ONNX's token budget includes the state, instructions, and rubric.

Agent evaluation tool

decision_evaluate is a core tool. An agent receives it when that agent has an effective decisionModel, subject to normal tool policy, explicit denies, and the active harness's capabilities. An unconfigured agent or one with an empty per-agent override does not receive the tool. The tool remains eligible whether or not other experimental consumers use Decision models. Provider plugins still need their own normal setup.

Call it with explicit shared state and a questions map:

{
  "state": { "message": "Checkout is failing for all customers." },
  "questions": {
    "escalate": {
      "type": "boolean",
      "instructions": "Does this need incident response?",
      "criteria": {
        "true": { "includes": ["Widespread service outages"] },
        "false": "A routine request can follow normal handling"
      }
    },
    "route": {
      "type": "choice",
      "criteria": { "support": "Service failures", "billing": "Payment questions" }
    },
    "urgency": {
      "type": "score",
      "criteria": ["No disruption", "One workflow blocked", "Widespread outage"]
    }
  }
}

State, instructions, and criterion descriptions accept text, JSON objects or arrays, or null. Each question sees the same state and cannot see another answer in the batch. Use a later call when one question depends on an earlier answer. The tool collects no ambient conversation. Its trusted calling-agent binding selects the provider and model; input cannot override that identity or selection.

Results preserve Boolean probabilities, fractional zero-based scores, original distributions and rounding, optional confidence and usage, and provider/model provenance. They do not generate explanations or grant permission to act. Evidence goes to the selected provider's configured endpoint or local runtime; only send data authorized for that destination. Provider plugins own credentials, model loading, transport, and wire-specific validation.

Provider capabilities are declared with their model metadata and exposed in models.list.decisionModels. Unsupported questions and bounds produce actionable, bounded guidance. Requests are rejected rather than silently truncated or split. Do not substitute shell or HTTP calls when a configured provider is unavailable, request credentials in chat, or treat a failure as a negative answer.

Missing credentials, rate limits, overload, and temporary provider errors leave the configured tool available and return an unavailable result. Configuration changes follow the existing tool/context refresh lifecycle; execution rechecks the effective selection and authority. Cancellation propagates to the shared Decision runtime and must not start fallback work.

Call from a plugin

Call from a live tool, hook, or other owned operation. Here api is the plugin API, agentId identifies the agent owning the work, and signal is that operation's cancellation signal. The consumer needs no provider SDK or API key.

const outcome = await api.runtime.decisions.evaluate(
  {
    state: { message: "Checkout is failing for all customers." },
    questions: {
      route: {
        type: "choice",
        instructions: "Which team should handle this?",
        criteria: {
          support: "Technical problems and service outages",
          billing: "Invoices, refunds, and incorrect charges",
          sales: "Pricing and purchasing questions",
        },
      },
      urgency: {
        type: "score",
        instructions: "Assess the operational impact.",
        criteria: [
          "No service disruption",
          "Minor inconvenience with a workaround",
          "An important workflow is blocked",
          "A widespread production outage",
        ],
      },
      escalate: {
        type: "boolean",
        instructions: "Does this require human incident response?",
        criteria: {
          true: "A serious service incident needs human attention",
          false: "A routine request can follow normal handling",
        },
      },
    },
  },
  {
    agentId,
    purpose: "support.triage",
    rubricVersion: "1",
    timeoutMs: 30000,
    signal,
  },
);

The host selects the provider and model from the owning agent's configuration. purpose identifies the consumer operation; rubricVersion identifies its rubric in result provenance. They do not replace question instructions. Change the rubric version when its meaning changes. Omit agentId only when intentionally using global-default selection.

An ok outcome contains outcome.result.answers, keyed by the submitted question IDs, plus the provider-reported model, optional usage, and provider/rubric/runtime provenance. A reported alias such as kev-latest does not identify a particular loaded checkpoint. Missing usage or confidence stays absent. For example, these are illustrative answers, not a promised model response:

{
  "route": {
    "type": "choice",
    "choice": "support",
    "probabilities": { "support": 0.97, "billing": 0.02, "sales": 0.01 }
  },
  "urgency": {
    "type": "score",
    "score": 2.85,
    "probabilities": [0, 0, 0.15, 0.85]
  },
  "escalate": { "type": "boolean", "probabilityTrue": 0.96 }
}

The host validates the complete batch before returning any answers. Narrow an answer by its type before accessing type-specific fields in TypeScript.

Interpret scores and probabilities

Scores are zero-based rubric positions. The four urgency descriptions above define a scale from 0 to 3, including fractions. A score of 2.85 lies near the highest-impact anchor. A display can rescale that position:

if (outcome.status === "ok") {
  const urgency = outcome.result.answers.urgency;
  if (urgency?.type === "score") {
    const scoreOutOf100 = (100 * urgency.score) / 3;
  }
}

The divisor is the last rubric index: criteria.length - 1. The display value is a position on the scale, not a confidence percentage or probability of being correct. Rubric wording determines what that scale means.

ONNX normalizes actual label logits with softmax. Choice uses the largest value; Score is the probability-weighted rubric index. TypeSafe preserves Jev's reported label, score, and probability estimates. Rounded vendor probabilities need not sum exactly to one, and its reported score need not equal an expectation computed from those rounded values.

A low Boolean probability favors false; a value near 0.5 gives similar weight to both outcomes. Missing evidence does not guarantee a value near 0.5. Include an explicit insufficient-evidence Choice alternative when that outcome matters. Choice alternatives compete; use separate Boolean questions for independent labels.

Your consumer chooses an action policy, such as escalating when probabilityTrue is at least 0.9. Validate that threshold on representative examples. Probabilities and optional provider-specific confidence values are estimates, not demonstrated accuracy guarantees. An answer does not grant permission to send a message or perform another effect.

Limits and unavailable results

For portable rubrics, use at most 32 questions, 2–64 Choice alternatives, 2–10 meaningful Score anchors, explicit true/false descriptions, and concise evidence. Provider limits differ:

The host bounds requests to one MiB, 20,000 JSON nodes, depth 32, and 256 questions. It admits at most four requests per provider and caps each deadline at 30 seconds. Provider-specific limits can be tighter. Unsupported input is rejected instead of silently truncated.

An unavailable outcome includes a reason such as disabled, not-configured, unsupported-input, overloaded, or deadline. The consumer decides whether to skip, defer, or use its existing fallback. Caller cancellation, closed consumer authority, and contract errors reject; do not turn cancellation into fallback work. See the SDK contract for the complete lifecycle and error behavior.

Consumers should build bounded useful evidence, evaluate it once, and retain their normal behavior when the provider returns unsupported-input, including context overflow. The runtime does not estimate model tokens, truncate evidence, retry, or choose another provider/model. Input rejection does not open the provider circuit; cancellation and closed authority remain terminal.

The decisions logger records DEBUG outcomes, dispatch status, elapsed time, question count, supplied JSON byte count, and actual provider usage when available. JSON bytes are not model tokens. Purpose/provider/model identifiers are hashed; existing ambient trace correlation is preserved. It logs no submitted text, rubrics, credentials, or provider error bodies, and does no extra input serialization when DEBUG is disabled. Generic input-rejection warnings are rate limited. Logs distinguish the provider result from the caller effect, which remains unobserved here.

For ONNX, cold-loading a large model can exceed the deadline on slower machines. Keep active models warm when memory permits, or choose a smaller model. See ONNX lifecycle and runtime for cache settings and the effect of active cancellation. For missing TypeSafe credentials, follow its setup instructions.

Provide models from a plugin

A provider plugin implements DecisionProviderV1 from openclaw/plugin-sdk/decisions and registers it with api.registerDecisionProvider(provider). Its id and contractVersion: 1 identify the contract; evaluate(batch, context) returns validated typed answers or a supported unavailable reason. The context includes the selected model, optional agent ID, composed cancellation signal, monotonic deadline, and an optional host-owned isAdmissible() predicate for automatic consumers. External transports check that predicate at final synchronous I/O, after awaited preparation; see the provider contract.

Declare provider ownership and static model metadata in the plugin manifest:

{
  "contracts": { "decisionProviders": ["example-decisions"] },
  "decisionModels": [{ "provider": "example-decisions", "id": "fast", "name": "Fast decisions" }]
}

This manifest fragment exposes the selector example-decisions/fast in the separate decision catalog. The consumer uses the same evaluation API for every provider. It does not register a provider itself.

Providers own model-specific input translation, transport or local inference, prepared credentials where needed, and physical cleanup on cancellation. Registration and optional isReady() must be synchronous and network-free; isReady() checks prepared credential availability, not model warmup. Return real model estimates rather than manufacturing certainty from a label-only response.

The provider contract owns the full SDK rules; the manifest reference owns discovery fields. For existing integrations, use the ONNX or TypeSafe AI setup pages.