cowork-sessions（协同工作会议）

Cowork Sessions Knowledge

Domain knowledge for launching and managing cowork sessions that coordinate multiple plugin agents.

Use For

• Starting new cowork sessions from marketplace items
• Managing running sessions (pause, resume, cancel)
• Understanding session lifecycle and agent coordination
• Monitoring session progress and resource usage

Session Lifecycle

1. INITIALIZING
   - Validate item is installed
   - Check plugin dependencies
   - Parse task description

2. PLANNING
   - Break task into subtasks
   - Match subtasks to available agents
   - Determine execution order and parallelism

3. RUNNING
   - Dispatch agents with their subtasks
   - Monitor progress (polling every 3-5 seconds)
   - Collect intermediate outputs

4. COMPLETING
   - Merge outputs from all agents
   - Generate session summary
   - Record metrics (tokens, cost, duration)

5. COMPLETED / FAILED
   - Present final outputs
   - Archive session for history

Agent Coordination Model

Sessions use the Task tool to spawn sub-agents. Key patterns:

Sequential Execution

For dependent tasks (e.g., "generate code then write tests"):

Agent 1 completes → output feeds Agent 2 → Agent 2 completes

Parallel Execution

For independent tasks (e.g., "review security AND check performance"):

Agent 1 ─────────────→ output ─┐
Agent 2 ─────────────→ output ─┤─→ merge
Agent 3 ─────────────→ output ─┘

Fan-out/Fan-in

Common pattern for comprehensive analysis:

Coordinator decomposes task
    ├──→ Specialist Agent A
    ├──→ Specialist Agent B
    └──→ Specialist Agent C
         ↓
    Coordinator merges results

Session Configuration

Each marketplace item defines defaults:

• maxParallelAgents - How many agents run simultaneously (1-10)
• estimatedDuration - Expected session length
• avgSessionMinutes - Historical average
• completionRate - Success rate (0.0-1.0)

Resource Tracking

Sessions track:

• Tokens used - Total input + output tokens across all agents
• Estimated cost - Based on model pricing (opus/sonnet/haiku)
• Duration - Wall-clock time from start to completion
• Agent count - How many agents were activated

Error Handling

If an agent fails during a session:

1. The error is captured but other agents continue
2. Failed subtasks are marked in the session output
3. The session completes with partial results
4. User is informed of which subtasks succeeded/failed

Session Controls