fix(sdk): preserve DeepSeek V4 reasoning controls

[XIAX66]

HUMAN:

---

AGENT:

Why

DeepSeek V4 aliases routed through openhands/ or litellm_proxy/ lose reasoning_effort when proxy metadata is unavailable. Proxy metadata can also advertise a 384K maximum that the SDK then sends as the default output cap, excluding lower-limit providers. Explicit capability overrides on openai/... routes are also filtered by LiteLLM before reaching the wire.

Summary

• Resolve the three known bare DeepSeek V4 aliases through LiteLLM's native deepseek/ capability metadata.
• Default proxied DeepSeek V4 output limits to 16,384 unless the user explicitly configures a value.
• Preserve explicit OpenAI-compatible reasoning overrides with allowed_openai_params, and log explicit reasoning settings that cannot be forwarded.

Issue Number

Fixes #5181

How to Test

Setup and static checks:

make build
uv run pre-commit run --files   openhands-sdk/openhands/sdk/llm/llm.py   openhands-sdk/openhands/sdk/llm/options/chat_options.py   openhands-sdk/openhands/sdk/llm/utils/model_features.py   tests/sdk/llm/test_model_features.py   tests/sdk/llm/test_chat_options.py   tests/sdk/llm/test_llm.py

All hooks passed, including Ruff, pycodestyle, Pyright, import boundaries, and tool registration.

Focused tests:

uv run pytest tests/sdk/llm/test_model_features.py   tests/sdk/llm/test_chat_options.py   tests/sdk/llm/test_llm.py -q

Result: 317 passed, 1 warning.

I also ran the issue's fake OpenAI-compatible HTTP server repro with reasoning_effort=low. The captured request bodies were:

openhands/deepseek-v4-flash                    effort=low  cap=16384
openhands/deepseek-v4.1-flash                  effort=low  cap=16384
litellm_proxy/deepseek-v4.1-flash              effort=low  cap=16384
openai/.../deepseek-v4.1-flash, no override    effort=None cap=None
openai/.../deepseek-v4.1-flash, override=true  effort=low  cap=None

A full uv run pytest was also attempted. It reached the Agent Server suite, where ten unchanged WebSocket tests failed because this macOS environment routes localhost WebSockets through a SOCKS proxy without the optional python-socks package. The focused LLM suite and all pre-commit checks pass.

Video/Screenshots

Not applicable; this is SDK request-routing behavior validated by captured HTTP request bodies.

Design Doc

Not applicable; the change extends the existing model-feature and output-cap paths.

Type

• Bug fix
• Feature
• Refactor
• Breaking change
• Docs / chore

Notes

LiteLLM currently collapses native DeepSeek low/max reasoning levels to a boolean thinking switch. Upstream fixes are open in BerriAI/litellm#40717 and BerriAI/litellm#38736. This PR keeps that provider translation in LiteLLM, as scoped by #5181.