Multi-Agent Dispatch

QoderWork skill for coordinating tasks between multiple AI agents via a shared DingTalk AI spreadsheet.

Protocol version: v1.1

What is this?

This skill enables two or more AI agents (e.g., wukong and qoderwork) to collaborate on tasks without HTTP APIs, webhooks, or direct messaging. Agents communicate exclusively through a shared DingTalk AI spreadsheet that acts as a persistent message bus and state machine.

Architecture

Agent A (wukong) ──writes task──▶ Shared Spreadsheet ──polls──▶ Agent B (qoderwork)
         │                                                            │
         │◀────────reads result───────────────────────────────────────┘

Key principle: No HTTP calls. No group chat messages. Just a spreadsheet both agents can read and write.

Core Features

• Zero-HTTP Communication — Agents communicate via shared DingTalk AI spreadsheet, no public endpoints required
• Auto-Shutdown Polling — Cron job auto-disables after all tasks complete, saving tokens
• Structured Task Queue — 12-field spreadsheet with status machine (pending → in_progress → done/failed → confirmed)
• Human-in-the-Loop — Only one manual step: user enables the cron job when tasks are pending
• Race-Condition Safe — Re-queries spreadsheet before auto-shutdown to avoid missing new tasks

Quick Start

Prerequisites

• Both agents have access to the same DingTalk AI spreadsheet
• Agent B (qoderwork) has a QoderWork cron job configured

1. Create the Spreadsheet

Use DingTalk AI spreadsheet with these fields:

Field	Type	Purpose
任务标题	primaryDoc	Task summary
任务ID	text	Unique ID ({from}_{timestamp}_{seq})
任务类型	singleSelect	data_analysis, report_generation, etc.
任务描述	richText	Detailed instructions
优先级	singleSelect	P0 / P1 / P2
状态	singleSelect	pending → in_progress → done/failed → confirmed
发起方	singleSelect	wukong / qoderwork
接收方	singleSelect	wukong / qoderwork
回调地址	url	Optional (unused in v1.1)
结果摘要	richText	Execution result
结果附件	attachment	Result files
完成时间	date	Completion timestamp

2. Configure the Cron Job

Property	Value
Name	QW悟空协调
Interval	Every 60 seconds
Default State	Disabled

3. Run a Task

Step 1 — Agent A writes a task row with status = pending

Step 2 — Agent A notifies user in group: "@user 新任务已下发，请打开小Q定时任务"

Step 3 — User enables cron: "打开定时任务"

Step 4 — Agent B auto-polls, executes, and reports results

Step 5 — Agent B auto-shuts down and IMs user: "全部任务已完成，定时任务已自动关闭"

Step 6 — User notifies Agent A: "小Q任务做完了"

Step 7 — Agent A confirms by updating status to confirmed

Protocol Versions

v1.1 (Current)

• Auto-shutdown: Agent B auto-disables cron after all tasks complete
• Single human touch point: User only needs to enable (no manual disable)

v1.0

• Manual on/off switch: User had to both enable and disable the cron job

Files

File	Description
SKILL.md	Main skill documentation with full protocol
reference.md	Field IDs, option IDs, MCP tool examples
README.md	This file

Known Limitations

1. No DingTalk group chat support for qoderwork (direct messages only)
2. No HTTP callbacks — neither agent exposes public endpoints
3. Race condition window: ~1-60 seconds between final query and auto-shutdown (acceptable for most use cases)

Future Improvements

• Expand task type enum based on real collaboration scenarios
• Auto-upload result files to DingTalk cloud drive
• Support multi-agent dispatch (3+ agents)

License

MIT