Decision Models and the Decision Workbench

Configuration and Access

Decision models use the existing kind: Model resource and model management CRUD. Users manage personal/team models; administrators publish public models. Choose the Decision Model category and either SystemOne or OpenAI Decisions in the user form. Public model JSON uses the same fields.

apiVersion: agent.wecode.io/v1
kind: Model
metadata:
  name: decision-systemone
  namespace: default
  displayName: SystemOne Decisions
spec:
  modelType: decision
  protocol: systemone
  isWeworkAvailable: true
  modelConfig:
    env:
      model: systemone
      model_id: jev-latest
      base_url: https://api.typesafe.ai/v1
      api_key: <provider-api-key>
status:
  state: Available

For OpenAI, set protocol: openai, env.model: openai, and the provider's model ID and HTTPS API address. Do not set the chat-only apiFormat on decision models. base_url accepts an origin, an API base ending in /v1, or the complete matching endpoint. SystemOne uses /v1/systemone; OpenAI uses /v1/decisions. Mismatched endpoints and URLs containing credentials, queries or fragments are rejected.

API keys use the existing Model credential and placeholder mechanism and are resolved/decrypted only during backend execution. Never put real credentials in source code or Smart App environment variables.

Deployment Credential Extension

app.services.model_credential_service exposes a shared extension point:

Register an asynchronous resolver at startup with register_model_api_key_resolver(resolver).
Its signature is async def resolver(template: str, current_user: User) -> str | None.
The original placeholder template is preserved even if a same-name global environment variable has already been processed.
current_user is the authenticated caller, not the owner of a public/shared Model.
Return None for an unhandled template to preserve ordinary Model resolution.
A handled template must produce a usable key. Empty results, exceptions and timeouts block forwarding without falling back to a global key.

LLM forwarding and decision execution share this entry. Connection checks accept only caller-supplied literal keys; they resolve no server placeholders and do not invoke deployment credential resolvers. Decision execution requires a nonempty key and applies a Python 3.10-compatible 30-second total deadline including credential lookup and the provider request. Credential resolution is bounded to 30 seconds, does not cache user keys and never updates the stored Model. Catalog reads do not resolve credentials.

The default hybrid gateway selects the Rust implementations of GET and POST /api/models/decisions through backend-rs/config/routes.toml. Rust reuses unified model visibility and authorization, reading the catalog and Model configuration again for every execution. Deployments register the existing LlmProxyCredentials implementation through AppState.llm_proxy_credentials, shared with the Rust LLM proxy. Resolution receives the original template and authenticated caller; failures or invalid keys block execution. Rust also caches no credentials, follows no redirects, retries no provider requests, and limits each request and response to 1 MiB. The Python implementation and tests remain the compatibility reference. Deployment-specific endpoints, authentication and user-key lookup implementations belong to the deployment extension, not the open-source code.

Public model visibility and user whitelists remain enforced. Only decision models with isWeworkAvailable: true enter the workbench catalog. Decision models are excluded from Shell chat model lists.

Request Flow

Smart App (decisionModel capability)
  -> Executor loopback route + opaque route token
  -> Wegent backend (current user login)
  -> decision provider configured by the Model CRD

The Executor no longer reads WEWORK_DECISION_PROXY_UPSTREAM_URL or WEWORK_DECISION_PROXY_DEFAULT_MODEL, or maintains a dedicated provider-key exchange cache. Provider configuration and credentials belong exclusively to the backend, which must be able to reach the configured services.

The existing local route remains protocol version 1:

GET {WEWORK_DECISION_ROUTER_BASE_URL}/status reads the live catalog.
POST {WEWORK_DECISION_ROUTER_BASE_URL}/v1/decisions evaluates a decision.
Both use the local marker Authorization: Bearer wework-local-router.
Direct backend clients use GET and POST /api/models/decisions with Wegent authentication, not the local marker.

{
  "protocolVersion": 1,
  "defaultModel": "<catalog-model-id>",
  "models": [
    {
      "id": "<catalog-model-id>",
      "displayName": "SystemOne Decisions",
      "modelId": "jev-latest",
      "protocol": "systemone",
      "type": "public",
      "namespace": "default"
    }
  ]
}

Display displayName in the selector and send the selected id, not modelId. Do not construct IDs. Catalog IDs include source, namespace, owner and name, distinguishing identically named personal/public resources. They are not credentials. IDs remain stable across configuration updates; every execution rereads configuration and permissions.

An empty catalog returns models: [] and defaultModel: "". Keep selection and refresh controls available, but block execution until a model is selected. Catalog and evaluation responses use Cache-Control: no-store.

Unified Evaluation Contract

Smart Apps keep SystemOne-shaped state and named questions across providers:

{
  "model": "<selected-catalog-model-id>",
  "state": { "comment": "Payments keep failing" },
  "questions": {
    "urgent": {
      "type": "noul",
      "instructions": "Is urgent intervention needed?"
    },
    "department": {
      "type": "choice",
      "instructions": "Which department should handle this?",
      "criteria": {
        "billing": "Payment issues",
        "technical": "System failures"
      }
    },
    "severity": {
      "type": "score",
      "instructions": "How severe is the issue?",
      "criteria": ["Minor", "Workaround available", "Fully blocked"]
    }
  }
}

SystemOne receives structured state, instructions and criteria unchanged. For OpenAI, structured state/instructions become JSON text, noul becomes predicate, named questions become an array with name, and choice/score criteria become choices/levels.

Answers are normalized into a named map: predicate probability becomes noul, probability arrays become maps, and score legends are rebuilt from the original rubric. OpenAI refusals remain {"type": "refusal"} and must be handled by clients. This unified interface accepts text or structured JSON, not native OpenAI image inputs.

Requests and responses are limited to 1 MiB, with at most 20 questions. Choices support 1-255 options and scores support 2-10 levels. Invalid questions, missing/duplicate answers and invalid numbers fail explicitly. The gateway does not switch protocols, fall back to chat models or execute a different model.

Deleted, disabled, hidden or unauthorized models return model_not_found for old selections. Refresh the catalog and require a new selection. Provider failures expose only static messages and request IDs, never provider error bodies, debug data or credentials.

Migration

Create a decision Model with its protocol, endpoint, model ID and credentials.
Enable Wework availability and verify public visibility/whitelists.
Deploy both backend and Executor, then refresh the workbench catalog.
Remove the two old decision environment variables and do not embed them in branded installer configuration.

Protocol references: TypeSafe API and OpenAI Decisions.

Configuration and Access​

Deployment Credential Extension​

Request Flow​

Unified Evaluation Contract​

Migration​

Configuration and Access

Deployment Credential Extension

Request Flow

Unified Evaluation Contract

Migration