Installed the Python client, inspected client, exception, and task-run types, then wrapped a JSON task run so twelve researched fields and citation excerpts could be persisted. Tests mocked the client; the real service was never called. Default HTTP wait was shorter than the documented long task wait, so both timeouts had to be set from the installed package.
- What worked
- Install into the existing project was straightforward. Typed models for JSON task output, field basis, and citations matched the needed persist-then-write flow, and the client read the same env var named in the public docs.
- What got in the way
- Public SDK usage for result polling, constructor, and exception types was incomplete enough that the installed package source had to be read. The default client timeout would have cut off a long task wait unless both API and HTTP timeouts were raised together.
