Deferred Tool Loading

PROJECT CONTEXT:
My agent's tool definitions eat a huge slice of every context window, and most tasks use five of the forty tools. Harness analyses show the fix: a small always-on core plus discovery and on-demand loading for the long tail.

GOAL:
Implement deferred tool loading: an eager core set, a tool-discovery mechanism, dynamic registration mid-session, and measured context savings.

REQUIREMENTS:
- Tool usage analysis identifying the true core set
- A compact catalog of deferred tools (name + one-line purpose) discoverable by the model
- A load_tools mechanism that registers deferred tools mid-session
- Session-scoped loading: tools loaded stay loaded for the session
- Before/after measurement of context spent on tool definitions

CONSTRAINTS:
- Deferring must not break workflows - discovery has to be reliable enough that the model finds what it needs
- The catalog must cost a small fraction of what the full definitions cost

DEFINITION OF DONE:
- Context spent on tool definitions drops substantially on typical tasks
- The model successfully discovers and loads a deferred tool when a task needs it
- Loaded tools function identically to eagerly-loaded ones
- A task-completion regression suite shows no capability loss

COMMON FAILURES TO AVOID:
- Deferring a tool the model needs but never discovers, silently degrading capability
- A catalog so verbose it costs as much as the definitions it replaced
- Cache-hostile loading that invalidates the prompt prefix every registration
- No measurement, so the 'optimization' is faith-based