Hive-Mind Session Resume Does Not Restore MCP Server State

[chmc]

Bug Report: Hive-Mind Session Resume Does Not Restore MCP Server State

Environment

• claude-flow version: 2.7.42 (alpha)
• Node.js version: v25.2.1
• Platform: macOS Darwin 25.1.0 (arm64)
• Claude Code: Latest

Summary

When resuming a hive-mind session with --claude flag, the MCP server starts with empty in-memory state instead of loading the persisted session/swarm data from .hive-mind/sessions/. This causes all mcp__claude-flow__* tool calls to fail with "Swarm not found" errors, even though the CLI correctly shows the session as active.

Steps to Reproduce

1. Create a hive-mind session:

   claude-flow hive-mind spawn --objective "Test task" --workers 4 --claude

2. Work with the session, then pause (Ctrl+C)

3. Resume the session:

   claude-flow hive-mind resume session-1765101696583-uqbb9abjt --claude

4. In the spawned Claude Code instance, try to use MCP tools:

   mcp__claude-flow__swarm_status(swarmId: "swarm-1765101696582-fbc3nfj1p")
   

Expected Behavior

The MCP server should load the persisted swarm state from .hive-mind/sessions/ so that:

• mcp__claude-flow__swarm_status returns the active swarm info
• mcp__claude-flow__memory_usage retrieves stored collective memory
• All swarm coordination tools work with the resumed session

Actual Behavior

{
  "success": false,
  "error": "Swarm swarm-1765101696582-fbc3nfj1p not found",
  "timestamp": "2025-12-08T10:34:05.020Z"
}

The CLI shows the session correctly:

claude-flow hive-mind status
# Shows: hive-1765101696578 with 5 agents, status: active

But MCP tools cannot find it because:

1. CLI stores state in .hive-mind/sessions/ directory (JSON files)
2. MCP server maintains separate in-memory state
3. On resume, --claude only injects context into Claude's prompt, but doesn't initialize MCP server state

Additional Issue: SQLite Module Incompatibility

The MCP server also fails to initialize the memory store due to native module version mismatch:

[ERROR] [memory-store] Failed to initialize: Error: The module
'/Users/aleksi/.npm/_npx/.../better_sqlite3.node'
was compiled against a different Node.js version using
NODE_MODULE_VERSION 137. This version of Node.js requires
NODE_MODULE_VERSION 141.

This occurs because:

• npx claude-flow@alpha caches the package with native modules compiled for an older Node.js
• User updates Node.js but the cached native module is stale

Workaround: npm cache clean --force and reinstall

Proposed Solution

Option A: MCP Server Loads State from Disk

On startup, MCP server should:

1. Check for active sessions in .hive-mind/sessions/
2. Load swarm state from checkpoint JSON files
3. Initialize in-memory state to match persisted state

Option B: CLI Initializes MCP State on Resume

The hive-mind resume --claude command should:

1. After spawning Claude Code, send initialization commands to MCP
2. Use MCP tools to recreate the swarm with same ID/state
3. Populate memory namespace with persisted data

Option C: Unified State Store

Replace separate CLI/MCP state stores with a single source of truth:

• Both CLI and MCP read/write to same SQLite database
• Use file locking for concurrent access
• MCP server watches for changes

Files Involved

• .hive-mind/sessions/session-*-auto-save-*.json - Checkpoint data
• .hive-mind/sessions/hive-mind-resume-*.txt - Resume prompt templates
• src/mcp/mcp-server.js - MCP server (starts with empty state)
• src/cli/ - CLI commands (manage .hive-mind/sessions/)

Workaround

Currently, users must either:

1. Initialize a new swarm via MCP after resume (loses continuity)
2. Use only CLI commands, not MCP tools
3. Manually track progress outside of hive-mind (e.g., TODO.md file)

Related Issues

• May be related to the general architecture of having separate CLI and MCP systems
• Similar to issues with multiple MCP server instances causing state confusion