fix(sdk): prevent double provider-prefix strip for namespaced model IDs

[santhiprakash]

HUMAN:

---

AGENT:

Why

LLMProvider.from_model() parses a model string like openai/openai/example-model into model="openai/example-model" + provider="openai". When as_litellm_call_kwargs() forwards that as model="openai/example-model" + custom_llm_provider="openai", LiteLLM's Responses API path re-parses and strips the openai/ prefix again, sending example-model upstream instead of openai/example-model — which returns 400 Model not found.

This affects any namespaced model ID where the inner segment is also a LiteLLM provider name (e.g. openai/openai/o3-mini, mistral/mistral/large). The OpenRouter nested case (openrouter/anthropic/claude-sonnet-4) is unaffected because LiteLLM has a dedicated OpenRouter special-case that preserves the inner provider/model form.

Summary

• Re-prefix the model in as_litellm_call_kwargs() when the parsed model still starts with the provider prefix, so LiteLLM strips exactly one prefix and lands on the correct parsed model.
• Added regression test test_llm_provider_preserves_namespaced_model_after_litellm_reparse covering openai/openai/example-model.

Issue Number

Fixes OpenHands/OpenHands#16365

How to Test

uv run pytest tests/sdk/llm/test_litellm_provider.py -v

Result (12 passed):

tests/sdk/llm/test_litellm_provider.py::test_llm_provider_parses_nested_openrouter_model PASSED
tests/sdk/llm/test_litellm_provider.py::test_llm_provider_parses_bedrock_model PASSED
tests/sdk/llm/test_litellm_provider.py::test_llm_provider_strips_api_key_for_bedrock_calls PASSED
tests/sdk/llm/test_litellm_provider.py::test_llm_provider_handles_unknown_model_without_provider PASSED
tests/sdk/llm/test_litellm_provider.py::test_llm_provider_keeps_requested_api_base_verbatim PASSED
tests/sdk/llm/test_litellm_provider.py::test_llm_provider_preserves_namespaced_model_after_litellm_reparse PASSED
tests/sdk/llm/test_litellm_provider.py::test_split_kwargs_equivalent_to_full_model_string[...] PASSED (6 cases)
============================== 12 passed in 0.06s ==============================

Pre-commit verification:

uv run pre-commit run --files openhands-sdk/openhands/sdk/llm/utils/litellm_provider.py tests/sdk/llm/test_litellm_provider.py

All checks passed (Ruff format, Ruff lint, PEP8, pyright, import dependency rules, tool subclass registration).

Reproduction proof (before fix vs after fix):

import litellm
# Before: double-strip
litellm.get_llm_provider(model="openai/example-model", custom_llm_provider="openai", ...)
# -> ('example-model', 'openai')  wrong, 400 Model not found

# After: re-prefix prevents double-strip
litellm.get_llm_provider(model="openai/openai/example-model", custom_llm_provider="openai", ...)
# -> ('openai/example-model', 'openai')  correct

Video/Screenshots

N/A — backend correctness fix, no visual surface.

Type

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

Notes

The fix is minimal and only triggers when self.model.startswith(f"{self.name}/") — the exact condition that causes the double-strip. All existing test cases (openrouter nested, bedrock, mistral, litellm_proxy, openai local, unknown) are unaffected because their parsed models do not start with the provider prefix.