Inference worker

Configure an INFERENCE_WORKER for harness inference dispatch: declared operations, model and provider identity, modelLocality, timeout/resources, and INFERENCE_RUN workstation companionship.

How To Use

Author a worker with type INFERENCE_WORKER. Put model, modelProvider, and modelLocality on the worker when you need authored identity and locality; authored values always win over operator defaults. Declare operations with typed inputs and outputs for pre-dispatch compatibility. Keep operation context in the body. Bind the worker from an INFERENCE_RUN workstation that supplies slot bindings and the operation name used at dispatch. Inference workers are not AGENT_WORKER backends: they do not own agent loops or agentTools.policy.

Discriminator: type = INFERENCE_WORKER

INFERENCE_WORKER requires an INFERENCE_RUN workstation companion. The workstation binds the worker name and owns operation selection, slot bindings, limits, and outputs.

Inference-run workstation

Schema reference

The embed below shows the live Worker base definition with INFERENCE_WORKER overlay applicability. Selected fields are inference-capability exclusive. Excluded fields belong to other WorkerType families and must not appear on this variant. Open the full Factory schema reference for exhaustive field lookup rather than copying contract text into this page.

Worker

object

A reusable worker definition that tells the factory how a workstation should execute work, such as through a model-backed agent or a script.

additionalPropertiesfalse (closed)

Fields

agentToolsagentToolsOptionalAgentWorkerToolsConfigExcluded
Explicit agent-loop tool policy for AGENT_WORKER definitions. Omit or set policy DISABLED to run agent loops without advertising or executing tools.
argsargsOptionalarrayExcluded
Additional command arguments passed to the configured command.
authauthOptionalHostedWorkerAuthExcluded
Hosted-worker authentication contract. V1 hosted workers accept only auth.secretRef.
bodybodyOptionalstring
Inline worker instructions or script body when the worker is authored directly in factory config.
commandcommandOptionalstringExcluded
Command to execute when this worker runs through a command or script provider.
descriptiondescriptionOptionalNameValue
Optional localized customer-facing explanation of this worker.
executorProviderexecutorProviderOptionalWorkerProviderExcluded
Execution mechanism. Use `ACP` for ACP-backed workers and put the configured integration identity (for example `cursor-acp`) in modelProvider. `SCRIPT_WRAP` remains the command-wrapper compatibility value; legacy named executor identities remain accepted during migration.
ididOptionalstring
Optional durable public identifier for this worker. When present, graph and layout references should use this id instead of the mutable name.
linearlinearOptionalHostedLinearWorkerConfigExcluded
Provider-specific configuration for the built-in hosted LINEAR worker.
modelmodelOptionalstring
Model identifier to request from the configured model provider when this worker uses model execution.
modelLocalitymodelLocalityOptionalWorkerModelLocalitySelected
Provider locality for this model capability declaration. Use `LOCAL` for embedded or host-managed inference and `CLOUD` for remote provider execution.
modelProvidermodelProviderOptionaloneOf
Canonical provider identity used for model routing and provider diagnostics, or an exact invocation-parameter placeholder such as `${modelProvider}`. For `executorProvider: ACP`, this names the configured ACP integration, such as `cursor-acp`. Extension identities use lowercase standardized syntax; built-in values such as `CLAUDE` and `CODEX` remain compatibility conveniences.
namenameRequiredstring
Worker name referenced by Workstation.worker.
operationsoperationsOptionalModelOperation[]Selected
Provider-agnostic model operations that this worker can execute, including named input and output slots.
providerproviderOptionalHostedWorkerProviderExcluded
Built-in hosted provider identity when this worker uses repository-owned hosted execution.
reasoningEffortreasoningEffortOptionalReasoningEffort
resourcesresourcesOptionalResourceRequirement[]
Resource capacity this worker requires before it can be dispatched.
skipPermissionsskipPermissionsOptionalbooleanExcluded
When true, bypasses permission checks for providers that support permission gating.
stopTokenstopTokenOptionalstring
Marker that tells model-oriented workers where to stop generated output when the provider supports it.
timeouttimeoutOptionalstring
Optional Go duration that caps one worker execution attempt.
typetypeOptionalWorkerType
Worker implementation family to instantiate for this definition.

Full Factory schema reference

Examples

Use the minimal valid example when starting an INFERENCE_WORKER. The misuse example shows agentTools on an inference worker—agentTools belongs on AGENT_WORKER and fails validation on INFERENCE_WORKER.

Minimal valid INFERENCE_WORKER:

{
  "name": "tts-local",
  "type": "INFERENCE_WORKER",
  "model": "OMNIVOICE_Q4_K_M",
  "modelProvider": "CODEX",
  "modelLocality": "LOCAL",
  "resources": [
    {
      "name": "omnivoice-cache",
      "capacity": 1
    }
  ],
  "operations": [
    {
      "name": "TTS",
      "inputs": [
        {
          "name": "text",
          "required": true,
          "contentTypes": [
            "TEXT"
          ]
        },
        {
          "name": "voice",
          "contentTypes": [
            "JSON"
          ]
        }
      ],
      "outputs": [
        {
          "name": "audio",
          "contentTypes": [
            "AUDIO"
          ]
        }
      ]
    }
  ],
  "body": "Synthesize speech from resolved text content."
}

Incompatible misuse — agentTools on INFERENCE_WORKER (rejected):

{
  "name": "tts-local",
  "type": "INFERENCE_WORKER",
  "model": "OMNIVOICE_Q4_K_M",
  "modelProvider": "CODEX",
  "modelLocality": "LOCAL",
  "agentTools": {
    "policy": "ENABLED"
  },
  "operations": [
    {
      "name": "TTS",
      "inputs": [
        {
          "name": "text",
          "required": true,
          "contentTypes": [
            "TEXT"
          ]
        }
      ],
      "outputs": [
        {
          "name": "audio",
          "contentTypes": [
            "AUDIO"
          ]
        }
      ]
    }
  ],
  "body": "Synthesize speech from resolved text content."
}

agentTools is an agent-loop field. Validation rejects it on INFERENCE_WORKER. Use AGENT_WORKER for prompt-rendered agent loops.