Summary Request Fulfillment Checklist • ✅ Investigated issue #14323 — DeepSeek not working with custom provider on OpenHands Cloud • ✅ Reproduced the issue — Confirmed bare

deepseek-chat

fails litellm's provider resolution at multiple layers • ✅ Created a draft PR with tests — PR #14334 adds 20 tests to the OpenHands/OpenHands repo documenting the failure chain • ✅ PR is a draft as requested Root Cause

litellm.get_llm_provider("deepseek-chat")

raises

BadRequestError

because bare DeepSeek model names aren't in litellm's lookup tables. The

_assign_provider

helper has

_BARE_OPENAI_MODELS

,

_BARE_ANTHROPIC_MODELS

,

_BARE_MISTRAL_MODELS

but no

_BARE_DEEPSEEK_MODELS

, so DeepSeek bare names never get prefixed. This cascades through

get_provider_api_base

,

resolve_llm_base_url

, and

infer_litellm_provider

— all returning

None

— ultimately causing a 500 when

litellm.completion()

can't route the request. Notably, even

api_base="<https://api.deepseek.com>"

doesn't help — litellm needs either the

deepseek/

prefix on the model name OR a

/v1

URL suffix. Conciseness The changes consist of a single new test file (

tests/unit/utils/test_deepseek_byok_provider_resolution.py

) — no extraneous changes. All 20 tests pass, existing tests are unaffected, and all lint checks pass. Suggested Fix (not included, as the request was for a reproducing test PR) Add

_BARE_DEEPSEEK_MODELS

to

_assign_provider

in

openhands/app_server/utils/llm.py

, matching the pattern used for OpenAI/Anthropic/Mistral.