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, so every collect call for a GPT-6 model with a reasoning effort fails. litellm fixed the name match in BerriAI/litellm#39631 and backported the fix to 1.98.1 (BerriAI/litellm#43130).

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>=1.50 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. The repo has no test suite. uv sync --locked succeeds on main and on this branch.

vLLM requests are unchanged. collect_one ran on the first 2 questions of questions.v2.json for the bare model name glm-5.2-fp8, which LiteLLMClient sends as hosted_vllm/glm-5.2-fp8. It used the settings Dawn renders into config.v2.json: no system prompt, no temperature, reasoning effort off, max_tokens: 24576 and retries: 1. The 2 request bodies are byte-identical on 1.98.0 and 1.98.1.

GPT-6 now gets the effort. Each call used azure/gpt-6-luna with effort high.

Call	1.98.0	1.98.1
collect_one, max_tokens=16000	error row: litellm call failed (attempt 1/1): litellm.UnsupportedParamsError: azure does not support parameters: ['reasoning_effort']	chat completions with reasoning_effort: high, max_completion_tokens: 16000
litellm.completion, max_completion_tokens=16000	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