SessionStart hooks with mcp_tool type fire before MCP servers connect

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/36692
2. https://github.com/anthropics/claude-code/issues/35899
3. https://github.com/anthropics/claude-code/issues/41778

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

Update: PostCompact hooks also affected

Same MCP disconnection issue affects PostCompact hooks, not just SessionStart.

Reproduction

hooks.json with PostCompact event:

{
  "hooks": {
    "PostCompact": [
      {
        "matcher": "manual",
        "hooks": [
          {
            "type": "mcp_tool",
            "server": "concurai-ememory",
            "tool": "e_property",
            "input": { "action": "list" },
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Run /compact in a session with the plugin loaded.

Debug log evidence

2026-05-11T13:57:13.484Z [DEBUG] Forked agent [compact] finished: 1 messages
2026-05-11T13:57:13.489Z [WARN] Hooks: mcp_tool hook skipped — MCP server 'concurai-ememory' not connected

Hook fires 5ms after compaction completes, MCP still disconnected.

User-visible error

PostCompact [concurai-ememory/e_property] failed: MCP server 'concurai-ememory' not connected

Impact

This makes mcp_tool hooks unusable for all lifecycle events — SessionStart, PostCompact, and likely PreCompact and SessionEnd too. MCP disconnects during lifecycle transitions and doesn't reconnect before hooks fire.

The primary use case for PostCompact + mcp_tool is session rehydration: restoring context from a knowledge store after compaction destroys it. Without this, plugins that provide long-term memory have no automated recovery path after compaction.

Environment

• Claude Code v2.1.138
• macOS Darwin 25.3.0
• MCP transport: Streamable HTTP

Resolution: Not a race condition — plugin-scoped MCP limitation

After further testing, this is not a timing race. The root cause:

• Plugin-scoped MCP servers (defined in plugin .mcp.json) are torn down during all lifecycle transitions (SessionStart, PostCompact, PreCompact). Hooks firing during these events find the server disconnected.
• Connector-level MCP servers (account-level, claudeai-proxy transport) stay connected through all lifecycle transitions. Hooks targeting connectors work perfectly — PostCompact hook via connector executed successfully.

The fix (for plugin authors)

Don't put MCP server definitions in the plugin's .mcp.json. Configure the MCP connection as a connector at the account level. Plugin provides skills, CLAUDE.md, and hooks — hooks reference the connector server name. This works today with no changes to Claude Code.

Still potentially worth fixing

Plugin-scoped MCP servers could survive lifecycle transitions like connectors do, but it's not a blocker — the connector approach is architecturally cleaner anyway (plugin = logic, connector = infrastructure).

Closing as resolved via workaround.

Your workaround doesn't fix SessionStart, right? Feels like this should be reopened.

Correct. we just are avoiding the problem by setting up MCP outside of the plugin.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.