LLMService.Generate does not support Anthropic-compatible endpoints using v1/messages

[helei-ai]

Describe the bug

LLMService.Generate currently assumes an OpenAI-style upstream path and payload shape. As a result, Anthropic-compatible gateways may fail even when the gateway itself is valid and reachable.

To Reproduce

Set the following environment variables:

LLM_API_ENDPOINT=https://<gateway>/api/anthropic
LLM_API_KEY=...
LLM_MODEL=feature/gpt

Then call the server-side LLM generation path (for example through LLMService.Generate).

Expected behavior

When LLM_API_ENDPOINT points to an Anthropic-compatible gateway, the server-side LLM client should be able to:

• call v1/messages
• send Anthropic-compatible headers
• send an Anthropic-compatible request body
• parse the Anthropic-compatible response

Actual behavior

The current implementation normalizes requests as OpenAI-style requests, which causes Anthropic-compatible endpoints to fail.

Additional context

The project already exposes generic environment variables:

• LLM_API_ENDPOINT
• LLM_API_KEY
• LLM_MODEL

However, the current client behavior is not protocol-aware enough to support Anthropic-compatible gateways behind those generic settings.

A protocol-aware implementation would likely need:

• OpenAI-style handling via /v1/responses
• Anthropic-style handling via /v1/messages

with the corresponding headers, payload shape, and response parsing logic.