chore(deps): litellm 1.98.0 -> 1.98.1, so Azure GPT-6 models take reasoning_effort

[RubenBranco]

Why

litellm 1.98.0 recognises only GPT-5 names as Azure reasoning models. For azure/gpt-6-luna, gpt-6-sol and gpt-6-astra it rejects reasoning_effort with UnsupportedParamsError. tau2 sets litellm.drop_params = True, so here litellm drops the effort without an error, and GPT-6 runs at its default effort. litellm fixed the name match in BerriAI/litellm#39631 and backported the fix to 1.98.1 (BerriAI/litellm#43130). On 1.98.1, tau2's agent calls, which always carry tools, go to the Azure Responses API with reasoning.effort.

1.98.1 is a patch on the 1.98 line. It stays below 1.100, where litellm's hosted_vllm path drops reasoning_content from tool-loop history.

Changes

• uv.lock moves litellm from 1.98.0 to 1.98.1. No other package moves.
• pyproject.toml is unchanged, because litellm>=1.98.0 already allows 1.98.1.

Evidence

No model was called. A local stub HTTP server recorded every request litellm sent and returned canned replies.

Tests. 190 passed, 17 failed and 1 xfailed, on main and on this branch. The same 17 tests fail on both with AuthenticationError: they call a real LLM API, and the run had no provider credentials. The command was uv run --locked --with pytest --with pytest-asyncio pytest tests. As in #4, it excluded tests/test_domains/test_banking_knowledge, tests/test_gym, tests/test_streaming and tests/test_voice, which need optional dependencies.

vLLM requests are unchanged. LLMAgent on hosted_vllm/glm-5.2-fp8 ran a 3-request airline tool loop. The requests were a user turn, the result of a get_user_details call from the airline environment, and a second user turn. Every stub reply carried reasoning_content. The 3 request bodies are byte-identical on 1.98.0 and 1.98.1. The second and third requests replay 1 and 2 assistant messages, each with reasoning and reasoning_content.

GPT-6 now gets the effort. Each call used azure/gpt-6-luna with reasoning_effort="high" and max_completion_tokens=16000.

Call	1.98.0	1.98.1
LLMAgent, 14 tools	chat completions, no effort sent	/openai/responses with reasoning.effort: high, max_output_tokens: 16000, model gpt-6-luna
generate(), no tools	chat completions, no effort sent	chat completions with reasoning_effort: high, max_completion_tokens: 16000
generate(), no tools, litellm.drop_params = False	UnsupportedParamsError	chat completions with reasoning_effort: high, max_completion_tokens: 16000

Every result above is the same with litellm's live model cost map and with LITELLM_LOCAL_MODEL_COST_MAP=True.

🤖 Generated with Claude Code