Feature: Lazy-load skills and MCP servers (name-only stubs until invoked)

Problem

When starting a Claude Code session, all installed skills and MCP servers have their full descriptions/tool schemas loaded into context. This can consume 30-40% of the context window before the user even sends their first message.

For users with multiple plugins installed (Linear, ios-simulator, chrome automation, etc.), this overhead significantly reduces the available context for actual work.

Proposed Solution

Lazy-loading / stub mode for skills and MCPs:

1. At session start, only load skill/MCP names (minimal tokens)
2. When the user invokes a skill or the agent determines one is needed, then load the full description/schema
3. Optionally allow users to configure which skills/MCPs are "always loaded" vs "lazy loaded"

Per-session MCP enabling:

Allow users to start with minimal MCPs and add them mid-session:

• /mcp add linear - enables Linear MCP for this session only
• /mcp list - shows available but not loaded MCPs
• Config option: "defaultMcpMode": "minimal" vs "all"

Benefits

• Dramatically reduced context usage at session start
• Users can have many tools installed but only pay the context cost when actually using them
• More context available for the actual task
• Faster session startup

Current Workaround

Users must manually edit plugin .mcp.json files or disable entire plugins, losing access to useful tools entirely rather than having them available on-demand.