feat: implement OpenAI compatible API provider support

[hkdeman]

This PR implements OpenAI compatible API provider support to resolve the issue where environment variables for OpenAI compatible endpoints were not working.

Changes

• Add WHODB_OPENAI_COMPATIBLE_LABEL environment variable for custom labeling
• Create OpenAI compatible provider when API key and endpoint are configured
• Add OpenAI-Compatible LLM type with request/response handling
• Support custom models from WHODB_CUSTOM_MODELS environment variable
• Reuse existing OpenAI-compatible API structure without requiring new API code

Fixes #496

Generated with Claude Code

[jazzberry-ai[bot]]

Bug Report

Name	Severity	Example test case	Description
Incorrect SSE Handling in OpenAI-Compatible Streaming	Medium	Use an OpenAI-compatible API that returns Server-Sent Events (SSE). Send a prompt and observe that the response parsing fails.	The parseOpenAICompatibleResponse function doesn't handle the 'data: ' prefix in SSE responses, leading to JSON unmarshaling errors.

_{Comments? Email us. Your free trial ends in 7 days.}

[modelorona]

@claude could you take a look at the issue by the Jazzberry bot?

[claude[bot]]

Claude finished @modelorona's task —— View job

---

Fixed SSE Handling Bug ✅

I investigated and fixed the SSE handling issue reported by Jazzberry bot.

Todo List:

• [x] Read the OpenAI compatible client implementation
• [x] Analyze the parseOpenAICompatibleResponse function
• [x] Identify the SSE parsing issue with 'data: ' prefix
• [x] Implement fix for proper SSE handling
• [x] Test and commit the changes

Bug Analysis:

Found the issue in core/src/llm/openai_compatible_client.go:74. The function was trying to unmarshal raw SSE lines directly, but SSE format includes "data: " prefix that must be stripped before JSON parsing.

Fix Applied:

• Strip "data: " prefix from SSE lines before JSON parsing
• Handle "[DONE]" control message to properly terminate streams
• Prevents JSON unmarshaling errors in streaming responses

Fixed in commit: dba2272 - The OpenAI-compatible streaming should now work correctly with proper SSE handling.

---

[jazzberry-ai[bot]]

Bug Report

Name	Severity	Example test case	Description
Truncated Streaming Response	Medium	Configure the application to use an OpenAI-compatible API that sends the "[DONE]" message immediately after the last content chunk, without a "stop" FinishReason. Send a prompt that generates a multi-chunk response. Observe that the last part of the response is missing.	The streaming response parsing logic in parseOpenAICompatibleResponse terminates prematurely when the "[DONE]" message is received, potentially truncating the last part of the response. The code should send the accumulated response to the receiverChan before breaking the loop upon receiving the "[DONE]" message to ensure the complete response is delivered. Also, the code should check for the FinishReason being "stop" before sending the final response.

_{Comments? Email us. Your free trial ends in 7 days.}

[modelorona]

@claude can you look at the latest Jazzberry comment?

[claude[bot]]

Claude finished @modelorona's task —— View job

---

Fixed Truncated Streaming Response Bug ✅

I've analyzed and fixed the "Truncated Streaming Response" bug reported by Jazzberry bot.

Todo List:

• [x] Read the OpenAI compatible client implementation
• [x] Analyze the parseOpenAICompatibleResponse function for truncation issue
• [x] Identify the problem with "[DONE]" message handling
• [x] Implement fix to ensure complete response delivery
• [x] Test and commit the changes

Bug Analysis:

Found the exact issue in core/src/llm/openai_compatible_client.go:72-74. The current logic:

// Skip SSE control messages
if line == "[DONE]" {
    break  // ❌ This breaks immediately without sending final response
}

Problem: When "[DONE]" is received, the function broke immediately, potentially losing the last accumulated response.

Fix Applied:

1. Send accumulated response before breaking on "[DONE]": Modified the "[DONE]" handler to check if there's accumulated content in responseBuilder and return it before terminating the stream
2. Handle normal stream termination: Added fallback logic to return accumulated response when the stream ends normally (EOF)

Changes in commit d28ec08:

• Lines 72-78: Send final response before breaking on "[DONE]" message
• Lines 127-131: Return accumulated response when stream ends normally

This ensures complete responses are delivered regardless of whether the stream ends with "[DONE]" or EOF, preventing truncation issues.

---

[jazzberry-ai[bot]]

Bug Report

Name	Severity	Example test case	Description
OpenAI-Compatible SSE Injection	High	Set up a malicious OpenAI-compatible endpoint that injects JavaScript code into the SSE stream.	The application blindly trusts the SSE stream from the OpenAI-compatible endpoint, leading to potential injection attacks and client-side code execution.

_{Comments? Email us. Your free trial ends in 6 days.}

[nicopace]

Any update on this? Thanks!!

[modelorona]

@nicopace we're just doing some local testing on our end!

[modelorona]

@nicopace we've merged this to main and are creating a new release! make sure to pull again from docker :)

let us know how you get on. I was playing around with the same API you tried and didn't run into issues.

you might need to clear your browser cache/storage due to how the providers are stored/managed in the frontend