v3.8.0 — Lazy MCP tools (big token savings)

Lazy MCP tools — large MCP servers no longer cost you tokens every turn. Inspired by mcp2cli and Anthropic's Tool Search.

Added

• Lazy MCP tool loading (on by default). Instead of injecting every connected MCP tool's schema into every request, Kodiqa exposes 3 fixed meta-tools and discovers tools on demand:

  • mcp_search — find MCP tools by keyword (names + descriptions, ranked by how often you've used them)
  • mcp_tool_schema — fetch one tool's full input schema only when needed
  • mcp_call — execute a tool by name

  For a 50-tool MCP server this cuts the per-turn tool-schema cost from ~5,300 to ~310 tokens (~94%), and the cost stays flat no matter how many MCP tools are connected.

• Usage-aware ranking — mcp_search orders results by call frequency (persisted in ~/.kodiqa/mcp_usage.json).

• /mcp lazy [on|off] — toggle it; /mcp list shows the current mode and token trade-off. Set mcp_lazy: false in settings to keep the old always-inject behavior.

Tests

• test_mcp_lazy.py — manager primitives, lazy vs. non-lazy tool exposure, meta-tool handlers, usage ranking, and routing. 383 total; ruff + CI green.