[BUG] Sambanova endpoint returning API Error 400 and API Error to OpenAI Compatible API

[richtong]

Problem (one or two sentences)

OK this is a strange one, Sambanova used to work a few months ago, but getting back to it, both the SambaNova API Provider returns an API Error 400. (I know the API key is valid and I've tested the OpenAI Compatible API point with both curl and Open-WebUI and they work properly).

The SambaNova provider is returning the correct /models but a chat call does not work.

More confusingly, if I directly use the OpenAI Compatible API point with https://api.sambanova.ai/v1 and a working API key, I also get the correct /models output in the Model section, but when I try to use it I get an "API Error"

Context (who is affected and when)

If you are trying to use SambaNova fast inferencing with Roo Code it fails with both the dedicated provider and the OpenAI Compatible API Endpoint

Reproduction steps

1. Mac OS Tahoe 26.1
2. VS Code, Roo Code 3.36.2
3. Create a Provider
4. Select SambaNova
5. Enter a working API key
6. Press save and run "hello"

Expected result

Should show a result

Actual result

With the SambaNova provider I get a 400 error and with the OpenAI Compatible Provider I get invalid API

Variations tried (optional)

I tried this with Open-WebUI with the same api endpoint and token. Also tried this with a curl command. Both worked

App Version

3.36.2

API Provider (optional)

SambaNova

Model Used (optional)

gpt-oss-120b and DeepSeek-R1

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

is there a way to get a log to see what is exactly being sent. I could find it.

I'm exploring this issue with the SambaNova API integration. I'll investigate the provider implementation and propose a fix.

I've analyzed this issue and identified the likely cause. Here's my proposed implementation plan:

Root Cause

The SambaNova API appears to be rejecting the stream_options: { include_usage: true } parameter that's being sent by the base OpenAI compatible provider. This parameter is not universally supported by all OpenAI-compatible APIs.

Proposed Solution

1. Override the createStream method in SambaNovaHandler to exclude the stream_options parameter, similar to how the OpenAI provider handles Grok xAI
2. Update the SambaNova provider to handle streaming without usage metrics in the stream (usage can still be tracked separately if needed)
3. Add tests to ensure the fix works correctly

Implementation Details

• Modify src/api/providers/sambanova.ts to override the createStream method
• Remove stream_options: { include_usage: true } from the request parameters for SambaNova
• Ensure the provider still works correctly for both streaming and non-streaming modes
• Update tests in src/api/providers/__tests__/sambanova.spec.ts to verify the fix

This approach follows the existing pattern used for other providers with similar limitations (like Grok xAI) and should resolve both the 400 error with the dedicated SambaNova provider and the API error with the OpenAI Compatible endpoint when pointed at SambaNova.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[richtong]

@roomote brilliant, that makes sense. Approve this plan, please proceed

I've seen your comment and approval. I'll proceed with implementing the fix to remove the stream_options parameter from SambaNova API calls. I'll open a PR to share progress.

Opened a PR with my attempt at addressing this: #10009

The fix removes the stream_options parameter from SambaNova API calls, which was causing the 400 error. All tests pass. Feedback welcome!

[hannesrudolph]

Apologies for the delayed response.

Native tool calling has been implemented since this was reported, and this should no longer reproduce on the latest Roo Code extension.

If you still see the SambaNova 400 or generic “API Error” on the latest version, please open a new issue (linking back to this one) with:

• Roo Code extension version, VS Code version, OS
• Provider type used (SambaNova provider vs OpenAI-compatible), base URL, model name
• Concrete repro steps and the exact error text
• Relevant Roo Code output logs (with secrets redacted)