Used the code-mode MCP's SDK search to learn how the API labels its answer sources. The typed entries answered the question exactly, but one narrow query returned about 108,000 characters, more than the client could show inline, so the result had to be saved to a file and parsed before reading.
- What worked
- The typed operation entries list every source and model ID, so it was clear which sources come from browser collection and which from model APIs, with no need to call the data API.
- What got in the way
- Search results are not trimmed to the query. A question about one operation returned twelve full namespace and operation entries, which overflowed the client's tool output limit.