TEST_LOG

Tested 2026-08-31 against gpt-oss-120b, agno @ main (1b7800746), with a live CEREBRAS_API_KEY (and OPENAI_API_KEY for the embedder) and Postgres started via cookbook/scripts/run_pgvector.sh. oss_gpt.py was not run.

basic.py

Status: PASS

Description: Runs the same prompt through all four variants: sync, sync + streaming, async, and async + streaming.

Result: All four variants returned complete responses.

db.py

Status: PASS

Description: Two sequential questions with add_history_to_context=True and session history persisted through PostgresDb.

Result: Both questions answered; the second ("What is their national anthem called?") correctly resolved "their" to Canada from the persisted history.

knowledge.py

Status: PASS

Description: Inserts the Thai recipes PDF into PgVector (OpenAI embedder), then asks the agent a question answerable only from the PDF.

Result: 14 documents upserted; the agent retrieved 10 documents and answered the Thai curry question with the recipe content from the PDF, citing the cookbook page.

structured_output.py

Status: PASS

Description: Structured output via output_schema=MovieScript.

Result: Returned a valid MovieScript JSON object. No strict-mode complaints from the API.

tool_use.py

Status: PASS (transient rate limiting disclosed)

Description: Web-search tool use through all four variants (sync, sync + streaming, async, async + streaming).

Result: Across two full passes: the first pass completed all four variants cleanly; the second pass completed 2 of 4, with the other two failing on transient Cerebras 429 queue_exceeded ("high traffic") errors under back-to-back load. Not a model or code issue, but expect occasional 429s when running the variants in quick succession. Tool calls now run in parallel (steps with 2 and 5 calls at once were observed) — the previous model id forced parallel_tool_calls=False via a library special-case that no longer matches.