[macOS] hive-mind spawn hangs: printed Swarm ID never shows up in hive-mind status, workers stay idle, tasks remain 0

[mgredenberg]

Summary

I cannot get claude-flow running properly on my system.

Following the README instructions, spawning a new Hive Mind reports that the swarm is initialized and workers are starting, but then everything stalls. The Swarm ID printed by spawn never appears in hive-mind status. Instead, status shows a different swarm with 0 tasks and all workers idle. Waiting >3,000 seconds shows no progress.

There also appears to be an ID mismatch/normalization issue: the spawn output uses an ID with underscores, while status shows a different ID with hyphens and different numeric segments. An MCP memory_usage store event writes to yet another swarm namespace, suggesting the components aren’t referencing the same swarm.

Steps to Reproduce

1. In a clean project directory, initialize:

   npx claude-flow@alpha init --force

2. Spawn a hive mind (per docs):

   npx claude-flow@alpha hive-mind spawn "gameplay" --namespace gameplay --claude

3. Observe the spawn output and then check status:

   npx claude-flow@alpha hive-mind status

Expected Behavior

• The Swarm ID printed by spawn appears in hive-mind status.
• Initial tasks are created (>0), and workers progress from idle to in progress.
• Ongoing activity/updates are visible within a reasonable time.

Actual Behavior

• spawn prints an initialized swarm and “deploying” message, then stalls indefinitely. Example:

  🐝 HIVE MIND INITIALIZED - Swarm ID: swarm_1755248730668_xjfiuzvrf
  
    Active Workers:
    - 🔬 researcher-alpha: Game mechanics, player psychology, market analysis
    - 💻 coder-beta: Game engine, physics, rendering, optimization
    - 📊 analyst-gamma: Performance metrics, UX, balancing
    - 🧪 tester-delta: QA testing, bug tracking, playtesting
  
    Now deploying the swarm to begin gameplay development. Let me coordinate the workers on initial tasks:
  

• MCP events show mixed IDs/namespaces:

  ⏺ claude-flow - swarm_status (MCP)(swarmId: "swarm_1755248730668_xjfiuzvrf")
  
  ⏺ claude-flow - memory_usage (MCP)(action: "store", key: "hive/phase1/complete", value: "Research and architecture phase
                                    complete. Core mechanics defined: 3-layer loop system, hybrid-casual balance, ECS
                                    architecture, performance targets set.", namespace: "swarm-1755248699199")
  

  Note: swarm_status references swarm_1755248730668_xjfiuzvrf (underscores),
  while memory_usage writes under swarm-1755248699199 (hyphens, different numbers).

• hive-mind status shows another swarm entirely, with no tasks and all workers idle:

  🐝 Active Hive Mind Swarms
  
  ════════════════════════════════════════════════════════════
  Swarm: hive-1755248699193
  ID: swarm-1755248699199-e0m7d9vlr
  Objective: gameplay
  Queen Type: strategic
  Status: active
  Created: 15.8.2025, 09:04:59
  
  Agents:
    👑 Queen: Queen Coordinator (active)
    🐝 Workers:
      - Researcher Worker 1 (researcher) idle
      - Coder Worker 2 (coder) idle
      - Analyst Worker 3 (analyst) idle
      - Tester Worker 4 (tester) idle
  
  Tasks:
    Total: 0
    Completed: null
    In Progress: null
    Pending: null
  
  Collective Memory:
    Entries: 4
  
  Consensus Decisions:
    Total: 0
  ════════════════════════════════════════════════════════════
  

• After >3,000 seconds, nothing progresses.

Observations / Hypothesis

• Swarm ID normalization mismatch (underscores vs hyphens) and differing numeric suffixes suggest the CLI, MCP event handlers, and persistence layer are not referring to the same swarm record.
• The “Now deploying … initial tasks” message doesn’t result in any tasks; the task seeding/enqueue step may be failing or writing to the wrong namespace.
• MCP memory_usage storing under a different swarm namespace further indicates cross-wiring between IDs.

Workarounds Tried

• Re-initializing with init --force.
• Re-running spawn and polling status for >3,000 seconds.
  No effective workaround found.

Environment

• OS: macOS 15.6 (Sequoia)
• Hardware: Apple Silicon (M-series)
• Editor: Claude Code (latest version)
• CLI invocation: npx claude-flow@alpha
• Network: No special proxy/VPN involved (standard home connection)

If helpful, I can re-run with verbose/debug (e.g., DEBUG=claude-flow*) and share full logs.

[roman-rr]

This isn't a configuration issue — it's the expected behavior of the current implementation.

We audited the codebase and found exactly why this happens:

1. hive-mind spawn does start a real Claude process — child_process.spawn('claude', ['--dangerously-skip-permissions', '<system prompt>']). That part works.

2. But the swarm ID doesn't show up in hive-mind status because the spawned process and the status command read different state. The spawn writes to one state path, status reads from another.

3. Workers stay 'idle' because agent_spawn (the MCP tool) only creates a Map entry {status: 'idle', taskCount: 0}. No actual worker process starts from MCP calls.

4. Tasks remain 0 because task_assign fires an EventEmitter event that no worker process listens to. The coordinator polls task.status every 100ms waiting for changes that never come.

This is the fundamental architectural gap: the MCP tools create state records but have no execution backend. The CLI spawn creates real processes but they're not connected to the MCP tool layer.

Workaround: Use Claude Code's native Agent tool for multi-agent work (it spawns real subprocesses). Use ruflo only for memory_store/search (HNSW vector search — this actually works).

Full audit: Independent Audit