Load large tool catalogs only when needed

Goal

Reduce prompt overhead by starting with a tool-description lookup and loading only the selected tool definitions.

Steps

1. Start each request with a describe_tool lookup and a compact list of available tool names. [source04]
2. When selected, return a tool_reference and append that definition with defer_loading=True to the active tool list. [source04]
3. Keep a per-request loaded_tools set and test that a second lookup does not duplicate definitions or alter unrelated tools. [source04]

How to verify

Request the same deferred tool twice; confirm its definition appears once in the active request and unrelated tool schemas remain byte-for-byte unchanged.

Troubleshooting

If the prompt grows with duplicate definitions, deduplicate by stable tool name in the per-request loaded_tools set and reset that set for the next request.