MCP servers are per-Controller, causing N× process duplication when multiple tabs are open

[ttmouse]

Problem

When a user opens multiple conversation tabs (desktop) or switches between topics without closing previous ones, each tab gets its own control.Controller → own plugin.Host → own copy of every MCP server (codegraph, Nowledge Mem, user-configured plugins).

This means:

Tab 1 → Controller 1 → Host 1 → codegraph + Nowledge + other MCP
Tab 2 → Controller 2 → Host 2 → codegraph + Nowledge + other MCP
Tab 3 → Controller 3 → Host 3 → codegraph + Nowledge + other MCP

With 10 tabs and 3 MCP servers, that is 30 long-lived processes running simultaneously, each consuming memory, CPU, and potentially API connections.

Industry comparison

Tool	MCP / LSP process model
Cursor (VS Code fork)	Shared globally — single extension host, all Composer tabs share one set of MCP connections
Windsurf (VS Code fork)	Shared globally — all Cascade conversations share one MCP client
Copilot Chat (VS Code)	Shared globally — single extension host per window
Claude Code CLI	One per CLI process (each claude invocation is independent by nature)
Reasonix (current)	❌ One per Controller — N× duplication

The industry standard for tabbed/multi-conversation tools is a shared MCP host — per-tab process isolation is not done by any major tool in this space.

Scope

Most MCP servers are read-only (codegraph, Nowledge Mem, database readers, etc.) — they serve the same data regardless of which conversation is asking. There is no correctness reason to duplicate them. The exception would be stateful MCP servers, but those are rare.

Possible approaches

1. Elevate plugin.Host to application level — one shared host for all Controllers, each Controller gets a lightweight MCP client handle. Requires significant refactoring because boot.Build() currently owns the Host lifecycle.

2. Lazy per-Controller MCP — only connect MCP servers on the active tab; suspend/disconnect when the tab loses focus, reconnect on focus. Smaller scope, solves the resource problem without a full architecture change.

3. Configuration flag — let users decide: global_mcp = true (default) or per_session_mcp = false for cases where full isolation is needed.

Related

• PR fix(macOS): reap entire codegraph process tree on exit (Setpgid + negative-PID kill) #3735 fixed the exit-time orphan process problem (child processes surviving after Reasonix quits). This issue is about runtime accumulation while Reasonix is running.
• The current per-Controller design also means each tab rebuilds its MCP connections on model switches (/model), adding latency.

Performance impact

Each codegraph process uses ~100-300MB RSS (Node.js runtime). With 10 tabs that is 1-3GB just for code intelligence processes. Shared MCP host would bring this back to ~100-300MB total.