fix(deepseek): map reasoning_effort=none to thinking disabled

[Jwrede]

Summary

• When reasoning_effort="none" is passed, DeepSeek V4 now receives thinking: {"type": "disabled"} instead of silently ignoring the value and leaving the model in its default thinking-enabled state.
• Also passes through thinking: {"type": "disabled"} when set directly, which was previously dropped by the enabled-only filter.

Test plan

• test_reasoning_effort_none_disables_thinking -- verifies reasoning_effort="none" maps to thinking: {"type": "disabled"}
• test_reasoning_effort_high_enables_thinking -- verifies non-none values still enable thinking
• test_thinking_disabled -- verifies direct thinking: {"type": "disabled"} is passed through
• test_thinking_strips_budget_tokens -- verifies budget_tokens is stripped (DeepSeek does not support it)
• test_no_thinking_params_leaves_thinking_unset -- verifies no thinking param when neither is set

Fixes #27453

[greptile-apps]

Greptile Summary

This PR fixes two gaps in DeepSeek's thinking-parameter mapping: reasoning_effort="none" now forwards thinking: {"type": "disabled"} instead of being silently dropped, and a directly-supplied thinking: {"type": "disabled"} is no longer filtered out by the enabled-only guard.

• transformation.py: Replaces the enabled-only if check with a tuple membership test covering both "enabled" and "disabled", and adds the "none" branch in the reasoning_effort handler.
• New tests: Five focused unit tests cover all code paths (disabled via reasoning_effort, disabled via direct thinking, enabled, budget_tokens stripping, and the no-op case); all use only mocked calls, consistent with the repo's test policy for this folder.

Confidence Score: 4/5

Safe to merge; the change correctly extends an if/elif guard to handle the disabled path and is fully covered by new unit tests.

The logic change is small and well-targeted: a tuple membership check replaces the single-value equality, and a new branch maps "none" to {"type": "disabled"}. Tests cover all five relevant cases. The only open item is a one-line docstring that still describes the old behaviour.

No files require special attention beyond the minor docstring fix in transformation.py.

Important Files Changed

Filename	Overview
litellm/llms/deepseek/chat/transformation.py	Fixes silent drop of reasoning_effort="none" and thinking: {type: "disabled"} by adding the disabled path; logic is correct and the inline docstring needs a one-line update.
tests/test_litellm/llms/deepseek/test_deepseek_chat_transformation.py	New unit tests covering all four new/changed code paths (disabled, enabled, budget_tokens stripping, no-op case); all mocked, no network calls.
tests/test_litellm/llms/deepseek/init.py	Empty __init__.py to make the new test directory a Python package.

Comments Outside Diff (1)

1. litellm/llms/deepseek/chat/transformation.py, line 35-36 (link)

   The inline docstring is now out of date — it still says DeepSeek only accepts {"type": "enabled"}, but the code now also accepts and forwards {"type": "disabled"}.

_{Reviews (1): Last reviewed commit: "fix(deepseek): map reasoning_effort=none..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

This pull request has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.