Defining operators

An operator is a named AI worker with a goal, a fixed set of capabilities, and a three-phase loop: it senses through connectors, reasons with a base model, and proposes a validated ExecutionPlan that a deterministic executor disposes. This page is the author's reference for defineOperator(): subscriptions, the reasoning contract and the plan it must produce, the action allowlist, guardrail binding, and what state an operator may keep. The concept page covers the why; your first operator is the tutorial.

#Anatomy of an operator

import { defineOperator } from '@fibric/sdk';

export default defineOperator({
  name: 'ship-risk',
  goal: 'Catch orders that will miss their promised ship date, and hold the ones a human should look at first.',
  requires: ['orders.read', 'orders.hold', 'notify.send'],
  model: 'router:reasoning',            // resolved by the model-router seam
  trigger: { on: 'order.*' },           // or { every: '15m' } / { on: 'ask' }

  async run(ctx) {
    const open = await ctx.sense('orders.read', { status: 'open' });
    const plan = await ctx.reason(open);   // model proposes; returns an ExecutionPlan
    return plan;                           // the executor disposes. the operator stops here.
  },
});

#Senses: event subscriptions

The trigger decides when run() executes. Event subscriptions are the primary form: the kernel's router matches the trigger pattern against every envelope's event_type and hands matching envelopes to the operator.

A subscription narrows attention; it does not deliver state. Inside run(), the operator reads current state through ctx.sense(), which resolves a capability to whichever connector the tenant bound it to. Sense from the world, not from the triggering payload, whenever the decision matters: the envelope says something happened, and time has passed since.

async run(ctx) {
  // the envelope that woke us: identity of the entity, not the source of truth
  const orderId = ctx.event.payload.entity?.id;

  // current state, through the capability binding (Magento today, anything tomorrow)
  const order = await ctx.sense('orders.read', { order_id: orderId });
  const thread = await ctx.sense('conversations.read', { about: orderId });
  // ...
}

#The reasoning contract

The platform tenet, stated once more because everything here follows from it: the LLM proposes a validated ExecutionPlan; a deterministic executor disposes. ctx.reason() is the only model call in the loop. It receives what the operator sensed, reasons inside the goal, and must return a plan that validates against the kernel type; anything else, malformed JSON, an unknown tool, a missing key, is rejected before it goes anywhere near the executor. A model cannot talk its way past a schema.

#The plan an operator returns

export interface PlannedAction {
  connector: string;
  tool: string;
  args: Record<string, unknown>;
  value?: number;          // e.g. a refund amount, for max_value policies
  entity_key: string;      // single-flight key: serialize side-effects per entity
  idempotency_key: string; // dedup key for the side-effect
}

export interface ExecutionPlan {
  reasoning?: string;
  actions: PlannedAction[];
}

An empty plan is a first-class outcome. An operator that senses, reasons, and concludes nothing should be done returns { actions: [] }, ideally with its reasoning filled in; proposing something because the loop ran is exactly the failure mode the contract exists to prevent. Populate reasoning in every plan: it is stored with the plan, shown in review queues, and is what a human reads when deciding whether to trust the operator with more.

#Constructing the keys

The two keys on every action are the operator's responsibility, built deterministically in code after reasoning, never left to the model to invent:

const verdicts = await ctx.reason(risky);   // model decides WHAT; code decides the keys

return {
  reasoning: verdicts.summary,
  actions: verdicts.holds.map((h) => ({
    connector: ctx.binding('orders.hold'),        // the bound connector's id
    tool: 'order.hold',
    args: { order_id: h.order_id, reason: h.reason },
    entity_key: `order:${h.order_id}`,
    idempotency_key: `ship-risk:${h.order_id}:hold`,
  })),
};

#The action allowlist

The requires array is a hard boundary, enforced in three places, none of which is the model's good behavior:

1. At sense time: ctx.sense() resolves only capabilities in requires. There is no call that reaches an unbound capability.
2. At validation time: a returned plan whose action names a tool outside requires fails validation; it is never submitted.
3. At disposition time: the executor evaluates the tenant's trust policy default-closed. Even an allowlisted tool is blocked if no policy rule allows it.

Keep the list minimal and by intent, never by vendor: orders.hold, not cn-magento. The tenant's capability bindings decide which connector answers, which is what makes swapping the vendor a configuration change under a running operator.

#Guardrail binding

Guardrails do not live in the operator; they live in the tenant's policy, expressed in the kernel's TrustPolicy shape, and the executor consults them on every side-effecting action. The operator author's obligations are narrower but real:

• Set value on every action a value ceiling could apply to. A refund action without value cannot be bounded by a maxValue rule, and review treats an unbounded money action as a defect.
• Expect BLOCK and design for it. A blocked action is a receipt, not an exception. The next run should sense the unchanged world and may propose again; it must not escalate around the refusal.
• Ship recommended defaults if the operator will be packaged. A pack carries a suggested TrustPolicy[] the installer may accept; the tenant policy always has the last word. See the pack manifest.

const policy: TrustPolicy[] = [
  { tool: 'order.hold',  decision: 'ALLOW' },
  { tool: 'notify.send', decision: 'ALERT' },              // act, and page a human
  // no rule for order.refund: proposing it yields BLOCK. fail closed.
];

#Memory and state

Runs are stateless by default, and most operators should stay that way: the world is the state, and sensing it fresh each run is what keeps behavior auditable. What continuity exists is explicit and bounded:

#A complete operator

import { defineOperator } from '@fibric/sdk';

export default defineOperator({
  name: 'ship-risk',
  goal: 'Catch orders that will miss their promised ship date, and hold the ones a human should look at first.',
  requires: ['orders.read', 'orders.hold', 'notify.send'],
  model: 'router:reasoning',
  trigger: { every: '15m' },

  async run(ctx) {
    // SENSE
    const open = await ctx.sense('orders.read', { status: 'open' });

    // REASON: the model decides which orders are at risk and why
    const verdicts = await ctx.reason({
      input: open,
      instruction: 'Identify orders that will miss their promised ship date. For each, give a hold reason and a one-line explanation.',
    });

    if (verdicts.holds.length === 0) {
      return { reasoning: 'No orders at risk this run.', actions: [] };
    }

    // PROPOSE: deterministic keys, allowlisted tools, value-free actions
    return {
      reasoning: verdicts.summary,
      actions: verdicts.holds.flatMap((h) => [
        {
          connector: ctx.binding('orders.hold'),
          tool: 'order.hold',
          args: { order_id: h.order_id, reason: h.reason },
          entity_key: `order:${h.order_id}`,
          idempotency_key: `ship-risk:${h.order_id}:hold`,
        },
        {
          connector: ctx.binding('notify.send'),
          tool: 'notify.send',
          args: { channel: 'ops', text: `held ${h.order_id}: ${h.explanation}` },
          entity_key: `order:${h.order_id}`,
          idempotency_key: `ship-risk:${h.order_id}:notify`,
        },
      ]),
    };
  },
});

Nothing in this file can reach the order system. The only exit to the world is the returned plan, and the executor stands behind it with the tenant's policy, per-order single-flight, and dedup on every key. That split, intelligence in the operator, guarantees in the kernel, is the entire design.

#Keep going

• Operator pack manifest: packaging this shape for others to install.
• Exposing actions: the contract on the other side of every proposed tool.
• Governance & trust: the disposition machinery in full.
• Example projects: complete operators, including MQTT and scheduled reporting.