Multi-Task Orchestrator

Overview

The orchestrator lets an agent break a complex request into multiple independent sub-tasks that run in parallel, with results fed into dependent tasks as they complete.

User: "Research 3 competitors, write a comparison table, draft an executive summary"
         ↓
Planner LLM → [subtask 0: research A] [subtask 1: research B] [subtask 2: research C]
                           ↓                    ↓                    ↓
                     [subtask 3: comparison table] (depends on 0,1,2)
                           ↓
                     [subtask 4: executive summary] (depends on 3)
                           ↓
                     Synthesise → final response in parent session

Usage

From the TUI

/orchestrate Research the top 3 AI coding tools, write a comparison, and draft a recommendation memo.

Check progress:

/plan-status <plan-id>

Via the agent tool

The plan_tasks tool is available to every agent:

Use plan_tasks to: research the latest EU AI Act developments, identify 5 affected industries, and draft a compliance checklist for each.

Via REST API

POST /v1/orchestrate
Authorization: Bearer <token>
Content-Type: application/json

{
  "taskDescription": "...",
  "maxSubtasks": 5
}

Response:

{ "planId": "uuid-..." }

How It Works

1. Planner Session

A temporary opencode session is created with the "task decomposition specialist" system prompt. It receives the task description and outputs a JSON array:

[
  { "description": "Research competitor A", "agent": "build", "dependencies": [] },
  { "description": "Research competitor B", "agent": "build", "dependencies": [] },
  { "description": "Write comparison table", "agent": "build", "dependencies": [0, 1] }
]

2. Sub-task Execution

Each sub-task with no unmet dependencies is enqueued as an orchestrator-subtask job immediately. Tasks with dependencies re-enqueue themselves every 5 seconds until all deps complete.

Each sub-task:

• Creates its own opencode session
• Receives completed dependency results as context: "Context from previous steps:\nSubtask 1: ...\nResult:\n..."
• Polls for completion (10-minute timeout)

3. Synthesis

When all sub-tasks reach terminal status (completed or failed), an orchestrator-synthesize job fires. It prompts the original parent session to combine all results into a coherent response.

Checking Status

GET /v1/orchestrate/:planId

Returns:

{
  "plan": { "id": "...", "status": "executing", "taskDescription": "..." },
  "subtasks": [
    { "index": 0, "status": "completed", "description": "...", "result": "..." },
    { "index": 1, "status": "running",   "description": "..." },
    { "index": 2, "status": "pending",   "description": "..." }
  ]
}

Limits

Parameter	Default	Max
maxSubtasks	5	10
Sub-task timeout	10 minutes	—
Planner timeout	3 minutes	—