Team & Swarm

Beyond simple sub-agent spawning and the coordinator pattern, Claude Code implements a full team/swarm system — persistent multi-agent teams where members run concurrently with different backend strategies. This is the most complex multi-agent pattern in the codebase.

Team Architecture

A team is a persistent configuration stored as a JSON file, with members that can be spawned across different execution backends:

graph TD
    L[Leader Agent] --> TC[TeamContext in AppState]
    TC --> M1[Member: researcher<br/>InProcess Backend]
    TC --> M2[Member: implementer<br/>Tmux Backend]
    TC --> M3[Member: reviewer<br/>iTerm2 Backend]

    M1 ---|AsyncLocalStorage| P1[Same Node.js process]
    M2 ---|tmux session| P2[Separate terminal pane]
    M3 ---|iTerm2 tab| P3[Separate iTerm2 tab]

Team File Structure

Teams are persisted as JSON files managed by src/utils/swarm/teamHelpers.ts:

// TeamFile structure (from teamHelpers.ts)
type TeamFile = {
  name: string
  description: string
  members: TeamMember[]
}

type TeamMember = {
  agentId: string        // Format: "name@team" (e.g., "researcher@project-team")
  name: string
  agentType: string
  model?: string
  prompt: string
  color?: string
  planModeRequired: boolean
  worktreePath?: string  // Git worktree for isolation
  sessionId?: string
  subscriptions?: string[]
  backendType: 'tmux' | 'iterm2' | 'in_process'
  isActive: boolean
  mode?: string
}

The agentId follows a deterministic format name@team, generated by formatAgentId():

src/utils/agentId.ts

export function formatAgentId(name: string, teamName: string): string {
  return `${name}@${teamName}`
}

Three Backend Strategies

1. InProcess Backend

The in-process backend runs teammates within the same Node.js process using AsyncLocalStorage for context isolation:

src/utils/swarm/spawnInProcess.ts

export async function spawnInProcessTeammate(
  config: InProcessSpawnConfig,
  context: SpawnContext,
): Promise<InProcessSpawnOutput> {
  const agentId = formatAgentId(name, teamName)
  const taskId = generateTaskId('in_process_teammate')

  // Independent AbortController — not linked to leader's query
  const abortController = createAbortController()

  // Create teammate context for AsyncLocalStorage
  const teammateContext = createTeammateContext({
    agentId, agentName: name, teamName, color,
    planModeRequired, parentSessionId, abortController,
  })

  // Create and register task state in AppState
  const taskState: InProcessTeammateTaskState = {
    ...createTaskStateBase(taskId, 'in_process_teammate', description),
    type: 'in_process_teammate',
    status: 'running',
    identity,
    prompt,
    model,
    abortController,
    awaitingPlanApproval: false,
    permissionMode: planModeRequired ? 'plan' : 'default',
    isIdle: false,
    shutdownRequested: false,
    messages: [],
    // ...
  }

  registerTask(taskState, setAppState)
  return { success: true, agentId, taskId, abortController, teammateContext }
}

提示

Each in-process teammate gets its own AbortController that is independent from the leader’s query abort. Teammates should continue running even when the leader’s query is interrupted — they’re persistent team members, not ephemeral sub-agents.

2. Tmux Backend

The Tmux backend spawns teammates as separate terminal sessions in tmux panes. Each teammate runs as an independent Claude Code process in its own pane.

3. iTerm2 Backend

Similar to Tmux but uses iTerm2’s native tab/pane management on macOS, providing a more integrated experience for iTerm2 users.

Spawning Configuration

src/utils/swarm/spawnInProcess.ts

export type InProcessSpawnConfig = {
  name: string           // Display name, e.g., "researcher"
  teamName: string       // Team this member belongs to
  prompt: string         // Initial task/instructions
  color?: string         // UI color for the teammate
  planModeRequired: boolean  // Must enter plan mode before implementing
  model?: string         // Optional model override
}

Teammate Identity

Each teammate carries a TeammateIdentity that persists across its lifecycle:

// Stored as plain data in AppState
const identity: TeammateIdentity = {
  agentId,           // "researcher@project-team"
  agentName: name,   // "researcher"
  teamName,          // "project-team"
  color,
  planModeRequired,
  parentSessionId,   // Links back to the leader's session
}

Killing a Teammate

The killInProcessTeammate() function handles graceful shutdown:

src/utils/swarm/spawnInProcess.ts

export function killInProcessTeammate(
  taskId: string,
  setAppState: SetAppStateFn,
): boolean {
  setAppState((prev: AppState) => {
    const task = prev.tasks[taskId] as InProcessTeammateTaskState
    if (!task || task.status !== 'running') return prev

    // 1. Abort the controller to stop execution
    task.abortController?.abort()

    // 2. Call cleanup handler
    task.unregisterCleanup?.()

    // 3. Call pending idle callbacks to unblock waiters
    task.onIdleCallbacks?.forEach(cb => cb())

    // 4. Remove from teamContext.teammates
    const { [agentId]: _, ...remainingTeammates } = prev.teamContext.teammates

    return {
      ...prev,
      teamContext: { ...prev.teamContext, teammates: remainingTeammates },
      tasks: {
        ...prev.tasks,
        [taskId]: {
          ...task,
          status: 'killed',
          endTime: Date.now(),
          messages: task.messages?.length
            ? [task.messages[task.messages.length - 1]!]  // Keep only last message
            : undefined,
          abortController: undefined,       // Release references
          unregisterCleanup: undefined,
          currentWorkAbortController: undefined,
        },
      },
    }
  })

  // Post-state cleanup
  if (teamName && agentId) {
    removeMemberByAgentId(teamName, agentId)  // Remove from team file
  }
  evictTaskOutput(taskId)                      // Clean up disk output
  emitTaskTerminatedSdk(taskId, 'stopped')     // SDK notification
  unregisterPerfettoAgent(agentId)             // Release trace entry

  return killed
}

Team Cleanup

When teams are disbanded or sessions end, comprehensive cleanup runs:

flowchart TD
    A[cleanupTeamDirectories] --> B[Read team config.json]
    B --> C{For each member}
    C --> D[destroyWorktree]
    C --> E[Remove task directories]
    D --> F{Has .git file?}
    F -->|yes| G[git worktree remove --force]
    F -->|no| H[rm -rf worktree path]
    C --> I[Delete team config file]

Worktree destruction (destroyWorktree in teamHelpers.ts) is particularly careful — it reads the .git file to find the main repository, then uses git worktree remove --force. If that fails, it falls back to rm -rf.

Perfetto Trace Integration

Team members are registered in Perfetto traces for hierarchy visualization:

// During spawn
if (isPerfettoTracingEnabled()) {
  registerPerfettoAgent(agentId, name, parentSessionId)
}

// During kill
unregisterPerfettoAgent(agentId)

This enables viewing the entire team hierarchy — leader and all teammates — in performance trace tooling.

Team State in AppState

The team system integrates deeply with Claude Code’s reactive state:

• teamContext on AppState holds the live team configuration and teammate map
• Each teammate’s InProcessTeammateTaskState is stored in AppState.tasks
• The InProcessTeammateTask React component (Task #14) drives the actual agent execution loop
• State changes trigger UI updates showing teammate status, spinner verbs, and tool counts

The state machine for a teammate:

spawned → running → idle ↔ running → completed/killed
                     ↑         |
                     └─ pending user messages

Teammates alternate between running (actively executing tools) and idle (waiting for instructions), with the leader able to send messages to idle teammates via SendMessage.