iii

name	iii
description	WebSocket-routed worker mesh — the engine's Function/Trigger/Worker model and the iii-sdk surface for authoring them. Teaches the ordered way to gain a capability before writing code — (1) check functions already registered in the engine, (2) search the public registry via iii-directory, (3) build a worker. Single self-contained skill — meant for system-prompt injection; do not re-fetch.

iii

iii is a WebSocket-routed worker mesh. One engine process (default port 49134) holds a live registry of every connected worker, every function those workers expose, and every trigger bound to them. Workers are independent OS processes that open a WebSocket to the engine and register Functions (service::name handlers) and Triggers (the events that invoke those Functions). There is no direct worker-to-worker traffic — every call routes through the engine, which makes the language, runtime, and physical location of any worker invisible to its callers.

When to Use

Use this skill to discover live iii capabilities, call functions, author SDK workers, bind trigger types, or operate project workers through Compose.

You extend yourself by writing iii workers. A few lines get you on the bus:

import { registerWorker } from 'iii-sdk'

const iii = registerWorker(process.env.III_ENGINE_URL!, { workerName: 'demo' })

iii.registerFunction('demo::add', async (payload: { a: number; b: number }) => {
  return { c: payload.a + payload.b }
})

The instant the handshake completes, demo::add is callable from any worker (and the harness itself) via iii.trigger({ function_id: 'demo::add', payload: { a: 2, b: 3 } }). No restart, no registration with the harness — the engine routes it automatically.

The four primitives

Primitive	What it is	Owned by
Engine	One coordinator process. Routes every invocation.	The operator
Worker	A process that opens a WebSocket to the engine.	Anyone who writes one
Function	A named handler inside a worker, id service::name. Stable across worker restarts.	The registering worker
Trigger	A (type, config, function_id) triple. Causes a function to run when an event fires.	A worker (the type-publisher) + a caller (the binding)

Three consequences worth internalising:

1. No worker-to-worker traffic. Every call is worker → engine → worker. Workers never address each other directly. Location and language are invisible.
2. No restart coordination. Restarting a worker is invisible to callers as long as it re-registers the same function ids. Two workers registering the same function id = automatic load-balance.
3. No polling unless you opt in. Triggers are the engine's push channel. The engine fans events out to bound functions when the underlying source fires.

The function id is the only contract between any two workers.

graph TD
  Harness["harness (LLM worker)"] <-->|"WS"| Engine["iii engine :49134 (registry + router)"]
  Engine <-->|"WS"| WorkerA["your authored worker my::fn"]
  Engine <-->|"WS"| WorkerB["installed registry worker"]
  Engine <-->|"WS"| Provider["trigger-type provider"]
  External["external event (request, timer, queue, ...)"] -->|"native protocol"| Provider

Every edge to the engine is a WebSocket. A trigger-type provider terminates some native protocol — an inbound request, a timer, a queue message — and translates it into engine traffic.

Need a capability? Discover before you build — in this order

The most common harness mistake is reimplementing something that already exists, or hardwiring one worker out of habit. Work the steps in order; stop at the first that satisfies the need.

1. Look at what is already registered in the engine. The capability may be one call away.

// engine::functions::list   — every function on this engine, across all workers.
//   Filter with { prefix: 'svc::' } or { search: 'resize' }.
// engine::workers::list      — every connected worker.

If a registered function fits, just call it: iii.trigger({ function_id, payload }).

2. Search the public registry. If nothing registered fits, look for a worker to install. This goes through the iii-directory worker:

// directory::registry::workers::list { search: 'image resize' }
//   → published workers matching the query.
// directory::registry::workers::info { name: '<worker>' }
//   → that worker's README, config keys, API reference, and skills.

When one fits, add it through the project's Compose daemon:

iii trigger -n <compose-namespace> compose::add worker=<worker>

iii-directory is itself a registry worker, so confirm it is connected before calling directory::*:

// engine::functions::list { prefix: 'directory::' }
//   → empty? add it to worker-compose.yaml through compose::add first.

3. Build a worker. Only when steps 1 and 2 both come up empty. Author it with the SDK (below), then deploy it. Discover the deployment/runtime surface the same way as any other capability — directory::registry::workers::list / ::info and its skill — rather than assuming a worker name. Add local code as a path:// container in worker-compose.yaml; relative paths resolve on the Compose daemon host.

Discover in order. Don't jump to a worker you remember; the registry may hold a better fit, and the right surface is whatever the live engine and registry report — not training-data recall.

The TypeScript SDK in brief

import {
  registerWorker, // factory; opens the WS from your code's perspective synchronously
  TriggerAction, // .Void() | .Enqueue({ queue })
  InvocationError, // typed error thrown by iii.trigger()
} from 'iii-sdk'
import { Logger } from '@iii-dev/helpers/observability' // OTel-aware structured logger; falls back to console.*

const iii = registerWorker(process.env.III_ENGINE_URL!, {
  workerName: 'my-worker', // appears in engine::workers::list
  invocationTimeoutMs: 30_000,
  reconnectionConfig: { maxRetries: -1 }, // -1 = infinite (the default)
})

// Publish a function. Same handler shape regardless of how the invocation arrives.
const ref = iii.registerFunction('svc::do-thing', async (payload) => ({ ok: true }), {
  description, // JSON-Schema-shaped metadata
  request_format,
  response_format,
})
ref.id // 'svc::do-thing'
ref.unregister() // drop just this function, keep the WS open

// Invoke. Three modes — same method, different `action`.
await iii.trigger({ function_id, payload, timeoutMs })
await iii.trigger({ function_id, payload, : .() })
 iii.({ function_id, payload, : .({ queue }) })


iii.({ , function_id, config })


iii.({ id, description }, { registerTrigger, unregisterTrigger })

 iii.() 

registerWorker(url, options?) opens the WebSocket synchronously from your code's perspective — there is no separate await connect(). The handle queues calls until the handshake lands.

Schemas in registerFunction (description, request_format, response_format) are JSON-Schema-shaped metadata — the engine does not validate payloads against them today. Declare them anyway: they surface in engine::functions::info, document the contract for the next caller, and reserve a slot for future runtime validation.

The three invocation modes

action	Caller blocks?	Retries?	Returns	Use when
(omitted)	yes	no	the function's result	you need the value to continue
TriggerAction.Void()	no	no	null	one-way notification, no result needed
TriggerAction.Enqueue({ queue })	no	yes	{ messageReceiptId }	slow/unreliable work; the queue handles retry + back-pressure

Errors

• Throw inside the handler → propagates to the caller as InvocationError (carries code, function_id, stacktrace). Use for unexpected failures a retry might fix.
• Return a structured value ({ ok: false, reason }) → the call succeeds; the caller branches on the shape. Use for expected failures (validation, not-found, business rules).

Rule of thumb: if a retry might succeed, throw; if it will fail the same way, return a value.

Lifecycle

• iii.shutdown() flushes pending traffic and closes the WS; the engine evicts the worker's functions immediately and resolves in-flight calls to it as invocation_stopped.
• ref.unregister() removes one registration (FunctionRef, Trigger, or TriggerTypeRef) without touching the others.
• The SDK reconnects automatically with backoff and replays registrations verbatim — never re-register manually. Callers see invocation_stopped during the disconnect window; treat it as cancellation, not transient failure.

Triggers — discover the type, don't hardcode it

A trigger's type is a literal string published by some worker, and its config shape is defined by that worker. There is no fixed catalogue — discover what is available rather than assuming:

// engine::triggers::list             → every trigger TYPE currently published (legal `type:` values).
// engine::triggers::info { id }       → that type's config + return JSON Schema, and its provider.
// directory::registry::workers::info  → the provider's README, when you need prose + examples.

Then bind, passing the literal type string and the config its schema requires:

iii.registerTrigger({
  type: '<type from the list above>',
  function_id: 'svc::handler',
  config: {
    /* keys per the type's schema */
  },
})

Two cautions that apply to every trigger type:

• registerTrigger succeeds at the engine even when the type provider is not connected or the config keys are wrong — the binding lands but never fires. Confirm the provider is up (engine::triggers::list shows the type) and copy config keys from the type's schema, not from memory.
• The bound function receives whatever payload that trigger type delivers (its request schema in engine::triggers::info) and must return whatever shape that type expects. The handler contract is the trigger type's, not a generic one — check the schema before writing the handler.

Custom trigger types — the deepest leverage

registerTriggerType turns your worker's native event source (a webhook hit, a file change, a row update) into something the whole bus can react to without polling. Keep a { trigger_id → { function_id, config } } table in memory and walk it when the source fires:

type FsWatchConfig = { path: string; recursive?: boolean }
const bindings = new Map<string, { function_id: string; config: FsWatchConfig }>()

iii.registerTriggerType<FsWatchConfig>(
  { id: 'fs::watch', description: 'Fires when a file under `path` changes.' },
  {
    async registerTrigger({ id, function_id, config }) {
      bindings.set(id, { function_id, config })
      startWatching(id, config)
    },
    async unregisterTrigger({ id }) {
      stopWatching(id)
      bindings.delete(id)
    },
  },
)

function onChange(triggerId: string, path: string) {
  const binding = bindings.get(triggerId)
  if (!binding) return
  iii.trigger({ function_id: binding.function_id, payload: { path }, action: TriggerAction.Void() })
}

From the caller's side, your custom type is indistinguishable from any built-in one.

Worker lifecycle — compose::*

Project workers live in worker-compose.yaml. The Compose daemon owns their install, startup, restart, update, and shutdown lifecycle:

iii trigger -n dev compose::add worker=state
iii trigger -n dev compose::status file=worker-compose.yaml
iii trigger -n dev compose::logs file=worker-compose.yaml worker=state tail=100
iii trigger -n dev compose::restart file=worker-compose.yaml worker=state
iii trigger -n dev compose::update file=worker-compose.yaml worker=state
iii trigger -n dev compose::down file=worker-compose.yaml

iii worker and worker::* were removed. Use engine::workers::list for live registrations and compose::status for the supervisor's process state. Use compose::logs for raw worker stdout and stderr.

Discovery surface

Call	Returns
engine::functions::list	Every function across all workers. Filter prefix / search.
engine::functions::info { function_id }	One function's schemas, description, owning worker.
engine::workers::list	Every WS-connected worker.
engine::triggers::list	Every trigger TYPE published (legal type: values).
engine::triggers::info { id }	One trigger type's config / return schema + provider.
engine::registered-triggers::list	Every trigger INSTANCE bound. Filter function_id / worker.
compose::status	Declared containers, process state, PID, ownership, and last error.
compose::logs	Bounded worker stdout/stderr entries and continuation cursors.
compose::list	Projects held by one Compose daemon.
directory::registry::workers::list	Workers published in the public registry. Filter search.
directory::registry::workers::info { name }	A registry worker's README, config, API reference, and skills.
directory::skills::list / directory::skills::get { id }	The markdown how-to a worker shipped — deeper than engine::functions::info.

engine::workers::list is the live registration view. compose::status is the process-supervisor view; consult both when diagnosing a declared container that did not register.

Trust runtime probes over introspection

engine::*::list reads can come back empty for blurred reasons: an older engine that lacks the surface, a store that lags live state, or genuinely nothing registered. Disambiguate with a runtime probe — call the function with iii.trigger(...). If the probe succeeds, the registration is live regardless of what *::list reported. Don't unbind or re-register on the strength of an empty list alone; you'll churn a working worker.

Anti-patterns

• Polling instead of a trigger type. A function on a timer reading a queue / file / table every N seconds is almost always wrappable as a custom trigger type. See Custom trigger types.
• Reinventing what exists. Run the discovery steps (engine::functions::list, then directory::registry::workers::list) before authoring anything.
• Hardwiring a remembered worker. Pick the capability the live engine and registry surface, in order — not the one you reached for last time.
• Side-channel state between workers. Don't have workers read each other's files or hit each other's endpoints; route every cross-worker call through iii.trigger, and use a shared-state worker (discover one in the registry) for shared key/value.
• Catching the wrong error type. iii.trigger() throws InvocationError; catch that specifically or you lose code / function_id / stacktrace.
• Trusting introspection over runtime probes. An empty *::list can mean lag, not absence — a successful iii.trigger() is the authoritative signal.

Boundaries

• Do not put project workers in config.yaml; use worker-compose.yaml.
• Do not call removed iii worker commands or worker::* functions.
• Do not add registry packages of kind engine as Compose roots; the engine supplies them.
• Treat function ids and trigger schemas from the live engine as authoritative.