Team Tasks — Multi-Agent Pipeline Coordination

Overview

Coordinate dev team agents through shared JSON task files + AGI dispatch. AGI is the command center — agents never talk to each other directly.

Two modes:

• Mode A (linear): Fixed pipeline order code → test → docs → monitor
• Mode B (dag): Tasks declare dependencies, parallel dispatch when deps are met

Task Manager CLI

All commands use: python3 <skill-dir>/scripts/task_manager.py <command> [args]

Where <skill-dir> is the directory containing this SKILL.md.

Quick Reference

Status Values

• pending — waiting for dispatch
• in-progress — agent is working
• done — stage completed
• failed — stage failed (pipeline blocks)
• skipped — intentionally skipped

Pipeline Workflow (Mode A: Linear)

Step 1: Initialize Project

python3 scripts/task_manager.py init my-project \
  -g "Build a REST API with tests and docs" \
  -p "code-agent,test-agent,docs-agent,monitor-bot"

Default pipeline order: code-agent → test-agent → docs-agent → monitor-bot

Step 2: Assign Tasks to All Stages

python3 scripts/task_manager.py assign my-project code-agent "Implement REST API with Flask: GET/POST/DELETE /items"
python3 scripts/task_manager.py assign my-project test-agent "Write pytest tests for all endpoints, target 90%+ coverage"
python3 scripts/task_manager.py assign my-project docs-agent "Write README.md with API docs, setup guide, examples"
python3 scripts/task_manager.py assign my-project monitor-bot "Verify code quality, check for security issues, validate deployment readiness"

Step 3: Dispatch Agents Sequentially

For each stage, AGI follows this loop:

1. Check next stage:   task_manager.py next <project> --json
2. Mark in-progress:   task_manager.py update <project> <agent> in-progress
3. Dispatch agent:     sessions_send(sessionKey="agent:<agent>:telegram:group:<id>", message=<task>)
4. Wait for reply      (sessions_send returns the agent's response)
5. Save result:        task_manager.py result <project> <agent> "<summary>"
6. Mark done:          task_manager.py update <project> <agent> done
7. Repeat from 1       (currentStage auto-advances)

Step 4: Handle Failures

If an agent fails:

python3 scripts/task_manager.py update my-project code-agent failed
python3 scripts/task_manager.py log my-project code-agent "Failed: syntax error in main.py"

To retry:

python3 scripts/task_manager.py reset my-project code-agent
python3 scripts/task_manager.py update my-project code-agent in-progress
# Re-dispatch...

Step 5: Check Progress Anytime

python3 scripts/task_manager.py status my-project

Output:

📋 Project: my-project
🎯 Goal: Build a REST API with tests and docs
📊 Status: active
▶️  Current: test-agent

  ✅ code-agent: done
     Task: Implement REST API with Flask
     Output: Created /home/ubuntu/projects/my-project/app.py
  🔄 test-agent: in-progress
     Task: Write pytest tests for all endpoints
  ⬜ docs-agent: pending
  ⬜ monitor-bot: pending

  Progress: [██░░] 2/4

Agent Dispatch Details

Session Keys (Dev Team)

Dispatch Template

When dispatching to an agent, include:

1. Project context — what the project is about
2. Specific task — what this agent should do
3. Working directory — where to create/find files
4. Previous stage output — if relevant (e.g., test-agent needs to know what code-agent built)

Example dispatch message:

Project: my-project
Goal: Build a REST API with tests and docs
Your task: Write pytest tests for all endpoints in /home/ubuntu/projects/my-project/app.py
Target: 90%+ coverage, test GET/POST/DELETE /items
Working directory: /home/ubuntu/projects/my-project/
Previous stage (code-agent) output: Created app.py with Flask REST API, 3 endpoints

Delivery Context Fix

⚠️ If an agent's session was first created via sessions_send, its deliveryContext is webchat, not telegram. Agent replies won't appear in the Telegram group.

Workaround: After getting the agent's reply via sessions_send, use the message tool to relay key results to the group:

message(action="send", channel="telegram", target="-5189558203", message="✅ code-agent 完成: Created app.py")

Mode B: DAG Workflow (Parallel Dependencies)

Step 1: Initialize DAG Project

python3 scripts/task_manager.py init my-project -m dag -g "Build REST API with parallel workstreams"

Step 2: Add Tasks with Dependencies

TM="python3 scripts/task_manager.py"
# Root tasks (no deps — can run in parallel)
$TM add my-project design     -a docs-agent  --desc "Write API spec"
$TM add my-project scaffold   -a code-agent  --desc "Create project skeleton"

# Tasks with dependencies (blocked until deps are done)
$TM add my-project implement  -a code-agent  -d "design,scaffold" --desc "Implement API"
$TM add my-project write-tests -a test-agent -d "design"          --desc "Write test cases from spec"

# Fan-in: depends on multiple tasks
$TM add my-project run-tests  -a test-agent  -d "implement,write-tests" --desc "Run all tests"
$TM add my-project write-docs -a docs-agent  -d "implement"             --desc "Write final docs"

# Final gate
$TM add my-project review     -a monitor-bot -d "run-tests,write-docs"  --desc "Final review"

Step 3: View DAG Graph

$TM graph my-project

├─ ⬜ design [docs-agent]
│  ├─ ⬜ implement [code-agent]
│  │  ├─ ⬜ run-tests [test-agent]
│  │  │  └─ ⬜ review [monitor-bot]
│  │  └─ ⬜ write-docs [docs-agent]
│  └─ ⬜ write-tests [test-agent]
└─ ⬜ scaffold [code-agent]
   └─ ⬜ implement (↑ see above)

Step 4: Dispatch Ready Tasks

$TM ready my-project    # Shows all tasks whose deps are met

For each ready task, AGI follows this loop:

1. Get ready tasks:     task_manager.py ready <project> --json
2. For each ready task (can dispatch in parallel):
   a. Mark in-progress: task_manager.py update <project> <task> in-progress
   b. Dispatch agent:   sessions_send(sessionKey=..., message=<task + dep outputs>)
3. When agent replies:
   a. Save result:      task_manager.py result <project> <task> "<summary>"
   b. Mark done:        task_manager.py update <project> <task> done
   c. Check newly unblocked tasks (printed automatically)
4. Repeat until all done

Key DAG Features

• Parallel dispatch: ready returns ALL tasks whose deps are satisfied — dispatch them simultaneously
• Dep outputs forwarding: ready --json includes depOutputs — previous stage results to pass to agents
• Auto-unblock notification: When a task completes, shows which tasks are newly unblocked
• Cycle detection: add rejects tasks that would create circular dependencies
• Partial failure: If one task fails, unrelated branches continue; only downstream tasks block
• Graph visualization: graph shows tree view with status icons and dedup markers

Custom Pipelines

Linear (Mode A)

# Code + test only
python3 scripts/task_manager.py init quick-fix -g "Hotfix" -p "code-agent,test-agent"

# Docs first, then code
python3 scripts/task_manager.py init spec-driven -g "Spec-driven dev" -p "docs-agent,code-agent,test-agent"

DAG (Mode B)

# Diamond pattern: 2 parallel branches merge for review
$TM init diamond -m dag -g "Parallel dev"
$TM add diamond code -a code-agent --desc "Write code"
$TM add diamond test -a test-agent --desc "Write tests"
$TM add diamond integrate -a code-agent -d "code,test" --desc "Integration"
$TM add diamond review -a monitor-bot -d "integrate" --desc "Final review"

Choosing Between Modes

Data Location

Task files: /home/ubuntu/clawd/data/team-tasks/<project>.json

⚠️ Common Pitfalls

Mode A: Stage ID is agent name, NOT a number

In linear mode, the stage ID is the agent name (e.g., code-agent), not a numeric index like 1, 2, 3.

# ❌ WRONG — will error "stage '1' not found"
python3 scripts/task_manager.py assign my-project 1 "Build API"
python3 scripts/task_manager.py update my-project 1 done

# ✅ CORRECT — use agent name as stage ID
python3 scripts/task_manager.py assign my-project code-agent "Build API"
python3 scripts/task_manager.py update my-project code-agent done
python3 scripts/task_manager.py result my-project code-agent "Created main.py"

This applies to all stage-referencing commands: assign, update, result, log, reset.

The pipeline order is defined by -p at init time (e.g., -p "code-agent,test-agent,docs-agent"), and next automatically advances through them in order — but you always reference stages by agent name.

Tips

• One project per task — keep scope focused; create multiple projects for parallel work
• Meaningful project slugs — rest-api-v2, bug-fix-auth, refactor-db (not project1)
• Save results — always result before update done; this is the inter-agent context
• Log liberally — log is cheap; helps debug failed pipelines
• Reset and retry — reset --all for clean reruns; reset <stage> for targeted retry
• DAG fan-out — one root task can unblock many parallel tasks
• DAG fan-in — a task can depend on multiple predecessors (all must complete)