fix: prevent duplicate BOS tokens with DeepSeek V3.1 in llama.cpp

[roomote]

This PR addresses Issue #7500 by adding a configuration option to prevent duplicate BOS tokens when using DeepSeek V3.1 with llama.cpp.

Problem

When using DeepSeek V3.1 through OpenAI Compatible provider with llama.cpp (with --jinja flag enabled), users were getting a warning about duplicate BOS tokens. This happens because llama.cpp automatically adds BOS tokens, but Roo Code was also sending messages in a format that triggered another BOS token addition.

Solution

Added a new configuration option openAiSkipSystemMessage that, when enabled for DeepSeek models:

• Merges the system prompt into the first user message instead of sending it as a separate system message
• Prevents the duplicate BOS token issue
• Maintains backward compatibility as an opt-in feature

Changes

• Added openAiSkipSystemMessage boolean option to OpenAI provider settings schema
• Updated OpenAI handler to detect DeepSeek models and apply the skip logic when configured
• Handles both streaming and non-streaming modes consistently
• Added comprehensive test coverage for the new functionality

Testing

• ✅ All new tests pass (10 test cases covering various scenarios)
• ✅ All existing OpenAI provider tests pass (no regression)
• ✅ All existing DeepSeek provider tests pass
• ✅ Linting and type checking pass

Usage

Users experiencing the duplicate BOS token issue with DeepSeek V3.1 and llama.cpp can enable the openAiSkipSystemMessage option in their OpenAI Compatible provider configuration.

Fixes #7500

---

Important

Adds openAiSkipSystemMessage option to prevent duplicate BOS tokens with DeepSeek V3.1 in openai.ts, with comprehensive tests.

• Behavior:
  • Adds openAiSkipSystemMessage option to prevent duplicate BOS tokens with DeepSeek V3.1 in openai.ts.
  • Merges system prompt into the first user message when enabled, avoiding separate system messages.
  • Applies to both streaming and non-streaming modes.
• Configuration:
  • Updates openAiSchema in provider-settings.ts to include openAiSkipSystemMessage.
• Testing:
  • Adds openai-deepseek-bos.spec.ts with 10 test cases for various scenarios, ensuring correct behavior with and without the new option.
• Misc:
  • Updates OpenAiHandler in openai.ts to detect DeepSeek models and apply skip logic when configured.

^{This description was created by for bd283b7. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Reviewing my own code is like debugging in a mirror - everything looks backward but the bugs are still mine.

[roomote]

I notice there's duplicate logic here between streaming (lines 108-129) and non-streaming (lines 248-268) modes. Could we extract this into a helper method like prepareMessagesWithSkipSystemMessage() to reduce duplication and improve maintainability?

[roomote]

The model detection using includes("deepseek") might be too broad and could match unintended models. Would it be more robust to use a specific list of model IDs or a regex pattern?

[roomote]

This comment is helpful, but could we expand it to explain when users should enable this option? For example: 'Enable this if you see duplicate BOS token warnings with DeepSeek V3.1 and llama.cpp'

[roomote]

Great test coverage! Consider adding a few edge cases:

• What happens when the system prompt is empty?
• Behavior with complex message content (arrays with multiple text/image parts)?
• Interaction with R1 format when both openAiR1FormatEnabled and openAiSkipSystemMessage are true?

[roomote]

Instead of hardcoding this for DeepSeek, could this feature be useful for other llama.cpp deployments? Consider renaming the option to something more generic like mergeSystemIntoFirstUser to indicate the behavior rather than the specific use case.

[daniel-lxs]

Closing this PR as the approach has fundamental limitations. The core issue is that we cannot reliably detect if llama.cpp is being used at runtime - we can only guess based on model names, which is not a sustainable solution. Merging system messages into user messages also changes the semantic structure of the conversation in ways that could affect model behavior.