[Feature]: Support models don't allow system instructions

[zygwin15]

概述 | Summary

When using Gemma 3 models via Google API, I got error "Developer instruction is not enabled". I found out the reason is Gemma 3 doesn't support system instruction. The system instruction needs to be inserted at the beginning of user prompt. https://ai.google.dev/gemma/docs/core/prompt-structure#system-instructions

Can we add an option in Agent to enable this? Thanks!

[hariprasathys22]

Hi @zygwin15

I have implemented a method to merge the system with user prompt in the begining..

Added new field to OpenAIAgent
merge_system_prompt = models.BooleanField( _("Merge System Prompt to User Message"), default=False, help_text=_("Enable for models that don't support system instructions (e.g., Gemma 3)") )

**Updated message formatting logic to conditionally merge system prompts based on the new field:**
messages = [
    {"role": "user", "content": "You are a translator...\n\nHello, how are you?"}
]

Can I put a PR In that i will deeply elaborate it.

[versun]

Thanks @zygwin15 for the feature request!

This feature has been implemented and merged. You can now use the merge_system_prompt option in the OpenAI Agent settings to support models that don't support system instructions (like Gemma 3).

What was done:

• ✅ Added merge_system_prompt field to OpenAIAgent model
• ✅ Implemented logic to merge system prompt into user message when enabled
• ✅ Fixed token calculation logic for the merged mode
• ✅ Added the field to admin interface for easy configuration

How to use:

1. Go to Admin → Agents → Edit your OpenAI Agent
2. In the "Advanced" section, check the "Merge System Prompt to User Message" option
3. Save the agent

Related PRs:

• #138 the system instruction inserted in the begining of user prompt #139 - Initial implementation by @hariprasathys22, Thanks!
• fix: Correct token calculation logic for merge_system_prompt option #141 - Bug fixes for token calculation

The feature is now available in the main branch. Thanks again for the suggestion! 🎉

[hariprasathys22]

Thank you @versun