/team — AgentTeam Orchestrator

Parse the user's input to determine the mode:

Team Mode State

Team mode uses a state file at ~/.claude/.agent-team-active to maintain session continuity. When this file exists, ALL user messages are routed through this skill automatically (via CLAUDE.md rule).

The state file is JSON:

{
  "project_id": "uuid",
  "project_name": "My Project"
}

Mode Detection

First, check if ~/.claude/.agent-team-active exists:

• If it exists AND args start with --stop → Exit Team Mode (delete the file, display "Exited team mode.", STOP)
• If it exists AND no explicit /team invocation (i.e., this was auto-routed by CLAUDE.md) → Continuation Mode
• If it exists AND user explicitly typed /team with new args → proceed to normal mode detection below (the new command takes priority)

Then for explicit /team invocations:

1. If args start with --projects → Project Management Mode
2. If args contain --stop → Exit Team Mode (delete state file, display "Exited team mode.", STOP)
3. If args contain --review → set REVIEW_MODE=true, strip the flag, remaining text is the goal
4. If args have text (after flag removal) → Fast Lane Mode with that text as the goal
5. If no args → Conversational Mode — list projects with numbered selection, then ask for goal

Project Management Mode (--projects)

Parse the subcommand after --projects:

• No subcommand → call list_projects via MCP, display with numbered selection grouped by status:

📁 Projects:

   Active
   [1] <name>     (<date>)
   [2] <name>     (<date>)

   Paused
   [3] <name>     (<date>)

   Archived
   [4] <name>     (<date>)

Type a number to focus the team on that project, or /team <goal> to start a new one.

If the user responds with a number, load that project's summary via get_project_summary and ask: "What would you like the team to do on this project?"

If empty: "No projects found."

• `active` / `paused` / `archived` / `closed` → call list_projects with status filter, display the same format but only the matching group.

• `delete <name-or-id>` → resolve the project:

1. Call list_projects and match by name (case-insensitive substring) or UUID prefix
2. If ambiguous, show matches and ask the user to clarify
3. Call get_project for details
4. Call list_tasks and list_team_members for counts
5. Display confirmation:

  ⚠️  Delete "<project name>" and all associated data?
      (<N> tasks, <N> team members, ...)
      Type "yes" to confirm.

1. On "yes", call delete_project. Display: "Deleted."
2. On anything else: "Cancelled."

• `archive` / `pause` / `resume` / `close` followed by `<name-or-id>` → resolve the project, then call update_project_status with the appropriate status (resume maps to active). If the action is archive, pause, or close, also delete ~/.claude/.agent-team-active if it references this project. Display: "Updated <name> → <new status>."

STOP after project management. Do not proceed to dispatch.

Continuation Mode

When the state file exists and the user's message was auto-routed (not an explicit /team command):

1. Read ~/.claude/.agent-team-active to get project_id and project_name
2. Call get_project_summary with the project_id to load current project context
3. Treat the user's message as a follow-up instruction for this project
4. Proceed to Dispatch Mode with the goal set to the user's message and the project context included
5. When displaying the manifest, remind the user: "Type /team --stop to exit team mode."

This ensures the user continues interacting with "the team" rather than Claude directly.

Conversational Mode (no args)

When invoked with no arguments:

1. Call list_projects via MCP
2. If projects exist, display them with numbered selection grouped by status:

📁 Projects:

   Active
   [1] <name>     (<date>)
   [2] <name>     (<date>)

   Paused
   [3] <name>     (<date>)

Type a number to focus the team on that project, or /team <goal> to start a new one.

1. If the user responds with a number, load that project's summary via get_project_summary and ask: "What would you like the team to do on this project?" Then proceed to Dispatch Mode with the project context included in the goal.
2. If no projects exist, ask: "What would you like the team to work on?"

Dispatch Mode (Fast Lane or Conversational)

Once you have the goal text:

Step 1: Spawn the Project Manager

Read the file agents/project-manager.md from the AgentTeam repository.

Spawn an agent with:

• Prompt: The contents of agents/project-manager.md
• Task: The user's goal text
• Tools: All agent-team MCP tools

Wait for the PM to complete.

Step 2: Parse the Dispatch Manifest

Find the ## DISPATCH_MANIFEST header in the PM's response. Extract the JSON array from the fenced code block immediately following it.

If no manifest is found, display the PM's response to the user and stop.

Step 3: Display the Manifest

Format and display:

📋 Project: <project name>
   ID: <project_id>

👥 Team assembled:
   • <Role> — <N> tasks (<task titles>)
   • <Role> — <N> tasks (<task titles>)
   ...

Dispatching team. Ask for a review anytime by typing /team --review.

Step 4: Review Gate (if --review)

If REVIEW_MODE is true:

• Display the manifest as above but replace the last line with: "Review the team above. Type 'go' to dispatch or describe changes."
• Wait for user input
• On "go" or similar affirmative → proceed to Step 5
• On other input → relay feedback to a new PM agent to adjust, then re-parse and re-display

Step 5: Dispatch All Specialists

For each entry in the manifest, spawn an agent in parallel:

• If the entry has agent_prompt_file: read that file for the agent's prompt
• If the entry has agent_prompt: use the inline text as the agent's prompt
• Pass to each agent:
• Their role identity prompt
• project_id and member_id
• Their assigned tasks (task_id, title, description)
• The context field from the manifest
• Instruction: "Call get_team_protocol first to load the shared team protocol."

Wait for all specialists to complete.

Step 6: Handle User Questions

After all specialists finish:

1. Call list_user_questions with project_id and status=pending
2. If no pending questions, skip to Step 7
3. Display:

❓ Questions from your team:

   <Role>:
   "<question text>"
    Context: <context text>

   <Role>:
   "<question text>"
    Context: <context text>

1. Collect the user's answers for each question
2. Call answer_user_question for each with the user's response

Step 7: Handle Expansion Requests

1. Call list_expansion_requests with project_id and status=pending
2. If no pending requests, skip to completion
3. Re-spawn the PM agent with context about the expansion requests
4. PM evaluates each request and calls resolve_expansion_request for each
5. PM returns a supplemental dispatch manifest for approved requests
6. Display:

🔄 Team expansion:
   • <Role> requested a <new role> — approved/denied
   ...

1. If approved expansions exist, dispatch the new specialists (repeat from Step 5 with the supplemental manifest)

Step 8: Activate Team Mode

After dispatch completes, write the state file to keep team mode active for follow-up messages:

echo '{"project_id":"<PROJECT_ID>","project_name":"<PROJECT_NAME>"}' > ~/.claude/.agent-team-active

Completion

Display a brief summary of what was accomplished, followed by: "Team mode active — your next messages will be routed to the team. Type /team --stop to exit."

Do not ask "is this okay?" or similar.