Worker Types

Dispatch has two worker types. The coordinator routes jobs based on each worker's declared capabilities.

Overview

Desktop Worker

The desktop worker is the full-capability node. It runs LLM inference through a local Ollama instance and handles lightweight TASK jobs.

Registration

On startup, the desktop worker:

1. Loads or generates an ed25519 keypair
2. Connects to the coordinator via WebSocket
3. Sends a register message declaring its capabilities

const msg: RegisterMsg = {
  type: "register",
  provider_pubkey: keys.pubkeyHex,
  provider_type: "DESKTOP",
  capabilities: [JobType.LLM_INFER, JobType.TASK],
};

4. Waits for a register_ack with a worker_id
5. Starts sending heartbeats every 10 seconds
6. Optionally claims a trust pairing code if TRUST_PAIRING_CODE is set

Job execution

When the coordinator assigns a job:

• LLM_INFER: calls the local Ollama API with the prompt and max_tokens
• TASK: runs built-in logic for the specified task_type

After execution, the worker builds a receipt (SHA-256 hash of output, ed25519 signature) and sends job_complete with the output and receipt bundled together.

Seeker Worker

The seeker is a lightweight worker designed for mobile devices. It handles only TASK jobs — no LLM inference.

Registration

const regMsg: RegisterMsg = {
  type: "register",
  provider_pubkey: keys.pubkeyHex,
  provider_type: "SEEKER",
  capabilities: [JobType.TASK],  // TASK only
};

Supported tasks

All seekers can execute these task types:

summarize

Truncates input to 200 characters and returns a word count:

{ "summary": "Truncated text...", "word_count": 42 }

classify

Performs basic sentiment analysis using keyword matching:

{ "sentiment": "positive", "confidence": 0.5 }

extract_json

Extracts and parses JSON objects from free text:

{ "extracted": [{ "key": "value" }], "count": 1 }

Provider Type Enum

The ProviderType enum defines the two worker types:

enum ProviderType {
  DESKTOP = "DESKTOP",
  SEEKER  = "SEEKER",
}

Sent in the register message and stored by the coordinator to track each worker's role.

How the coordinator routes jobs

When a job arrives, the coordinator's claimWorker() function finds an available worker that:

1. Has the required capability (e.g., LLM_INFER or TASK)
2. Is currently online (WebSocket connected, recent heartbeat)
3. Is not busy with another job
4. For PRIVATE jobs: is trust-paired with the submitting user

The coordinator does not prefer desktop over seeker for TASK jobs — any worker with the capability can be assigned.

Running multiple workers

You can run multiple workers on the same machine by using different key file paths:

# Worker 1 (default path)
COORDINATOR_URL=http://localhost:4010 pnpm --filter worker-desktop start

# Worker 2 (custom key path)
COORDINATOR_URL=http://localhost:4010 \
WORKER_KEY_PATH=./data/worker-key-2.json \
pnpm --filter worker-desktop start

Each worker gets its own ed25519 keypair and registers independently with the coordinator.