Repository navigation
Per-model npm override in custom provider config is ignored #31919
Description
Description
When configuring a custom provider with a per-model npm override (as documented in the Providers page under "Troubleshooting"), the override is silently ignored. The provider-level npm package is always used regardless of model-level overrides.
Use case
We run a LiteLLM proxy that serves both Claude (via Bedrock) and GPT (via Azure). LiteLLM exposes both /v1/messages (Anthropic-native) and /v1/chat/completions (OpenAI-compatible) on the same endpoint.
We need Claude models to use @ai-sdk/anthropic (→ /v1/messages) and GPT models to use @ai-sdk/openai-compatible (→ /v1/chat/completions) under a single provider.
Configuration
{
"provider": {
"litellm": {
"npm": "@ai-sdk/anthropic",
"options": {
"baseURL": "https://litellm.example.com/v1",
"apiKey": "{env:LITELLM_VIRTUAL_KEY}"
},
"models": {
"claude-sonnet-4-6": {
"name": "Claude Sonnet 4.6"
},
"azure/gpt-4-1": {
"npm": "@ai-sdk/openai-compatible",
"name": "GPT-4.1"
}
}
}
}
}Expected behavior
litellm/claude-sonnet-4-6→ uses@ai-sdk/anthropic(provider default)litellm/azure/gpt-4-1→ uses@ai-sdk/openai-compatible(per-model override)
Actual behavior
Both models use the provider-level @ai-sdk/anthropic. The per-model npm field is ignored.
Also tested the inverse (provider default @ai-sdk/openai-compatible, Claude models with "npm": "@ai-sdk/anthropic" override) — same result: override ignored.
Documentation reference
The Providers docs state under Troubleshooting:
"For mixed setups under one provider, you can override per model via
provider.npm."
Workaround
Define two separate providers with different npm packages:
litellmwith@ai-sdk/anthropicfor Claude modelslitellm-gptwith@ai-sdk/openai-compatiblefor GPT models
OpenCode version
1.17.1
Operating System
macOS
Activity
OpeOginni commentedon Jun 13, 2026
Hey @christianherweg0807 You have an issue with your config, the docs says its overridden by
provider.npm, which means like<model_id>.provider.npmSo what you should have is
{ "provider": { "litellm": { "npm": "@ai-sdk/anthropic", "options": { "baseURL": "https://litellm.example.com/v1", "apiKey": "{env:LITELLM_VIRTUAL_KEY}" }, "models": { "claude-sonnet-4-6": { "name": "Claude Sonnet 4.6" }, "azure/gpt-4-1": { "provider": { "npm": "@ai-sdk/openai-compatible", }, "name": "GPT-4.1" } } } } }Try this and should work. If it does please go ahead and close the issue!
christianherweg0807 commentedon Jun 15, 2026
We tested the per-model
npmoverride workaround from the docs and confirmed the bug reproduces with the following stacktrace:The root cause is visible in the trace: LiteLLM's Anthropic pass-through streaming adapter (
experimental_pass_through/adapters/streaming_iterator.py) is being used to process responses from the Azure OpenAI backend (azure/gpt-4-1-nano). Because the per-modelnpmoverride is ignored, the request goes to Azure via the correct endpoint but the response is then deserialized by the Anthropic adapter, which crashes when it encounters an OpenAI-style chunk with an emptychoiceslist.This confirms the bug: the per-model
npmfield has no effect on which streaming adapter is used. The provider-level@ai-sdk/anthropicis applied unconditionally.We are currently using the two-provider workaround as described in the issue.
OpeOginni commentedon Jun 15, 2026
@christianherweg0807 Did you make use of the new config shape I provioded in my reply?
github-actions commentedon Aug 15, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.