cookbook/90_models/ramp/TEST_LOG.md
Run against the live Router API on 2026-08-19 with .venvs/demo/bin/python.
Status: PASS
Description: Sync, sync streaming, async and async streaming runs on gpt-5.6-luna, half of
them through the RampRouter(...) constructor and half through the "ramp:gpt-5.6-luna" model
string.
Result: All four runs returned a two sentence horror story. The model string resolved to
RampRouter through the provider registry.
Status: PASS
Description: Four runs with WebSearchTools() on gpt-5.6-luna, covering sync, sync
streaming, async and async streaming.
Result: All four called search_news and answered from the results. Two of the four runs
logged No results found. from the search tool before retrying with a different query; no model
errors.
An earlier revision failed all four with Item 'fc_...' of type 'function_call' was provided without its required 'reasoning' item: 'rs_...'. Reasoning models bind the call to the reasoning
item that preceded it by item id, and Agno does not carry that item across turns, so RampRouter
now replays function calls without their id. call_id, which pairs the call with its output,
is untouched.
Status: PASS
Description: output_schema=MovieScript on gpt-5.6-luna, through the native structured
output path.
Result: Returned a fully populated MovieScript. The same schema also round-tripped on
claude-haiku-4-5, so structured output is not specific to the OpenAI backing.
Status: PASS
Description: models=[...] across three candidates instead of a single id, with
provider_timeout and metadata set.
Result: Streamed a haiku. Router selected a candidate and served it; model was omitted from
the request body, which is what models requires.