Feature Request: Model-Specific Instructions Independent of Provider

[KJ7LNW]

Feature Request: Model-Specific Instructions Independent of Provider

Description

Currently, instructions for AI models in Roo are typically managed at the provider level. This creates challenges when specific instructions need to be applied to particular models regardless of which provider is being used. Users often need to duplicate instructions across different providers or create workarounds when the same model is available through multiple providers.

Proposed Solution

Implement a system that allows model-specific instructions to be defined once and applied consistently across all providers that offer that model. This would create a layer of abstraction where instructions are associated with the model itself rather than the provider-model combination.

Implementation Details

1. Create a configuration system that maps instructions to specific model identifiers (e.g., gpt-4, claude-3-sonnet)
2. Implement a mechanism to merge model-specific instructions with provider-specific settings
3. Develop a priority system that determines which instructions take precedence when both model-specific and provider-specific instructions exist
4. Add UI elements to manage model-specific instructions separately from provider settings

Benefits

• Consistency across providers for the same model
• Reduced duplication of instructions when the same model is available through multiple providers
• Easier management of model-specific optimizations and prompting techniques
• More portable configurations when switching between providers
• Better separation of concerns between model capabilities and provider implementation

Implementation Suggestions

• Store model-specific instructions in a separate configuration file or database table
• Implement a middleware layer that applies model-specific instructions before provider-specific ones
• Create a model registry that maps variant names to canonical model identifiers (e.g., mapping both "gpt-4-turbo" and "gpt-4o" to a common identifier)
• Develop a template system for instructions that can be parameterized based on model capabilities

Discussion Points

• How should conflicts between model-specific and provider-specific instructions be resolved?
• Should there be a UI for managing these instructions or should it be configuration-file based?
• How can we handle model versioning when instructions may need to change between versions?
• Should we implement a community sharing mechanism for effective model-specific instructions?
• How can we handle models that are exclusive to one provider but might benefit from the same abstraction?

[kvokka]

Wanna the same idea but on a deeper level, making the changes MODE specific rather than MODEL specific, pls see #3773.

[KJ7LNW]

See also #3482 but it does not yet handle models, only modes.