Context:

There is a proposal to change the default LLM in the Theia IDE to Sonnet-3-7. While conceptually beneficial, such a change could break setups for users who have never altered their model configuration and have not provided an Anthropic API key.

Proposed Short-Term Solution

Implement fallback logic directly in the Language Model Service, based on the presence of API keys (from user preferences or environment variables). The model selection would follow this priority:

Anthropic (Sonnet-3.7), OpenAI 4o, Google Gemini pro 2.5

If none are available → fall back to a neutral option (possibly still Sonnet or a bundled fallback).

🧩 Proposed Long-Term Architecture

Introduce a structured “Default Model” abstraction, which does not currently exist. Key aspects of this design:

Define a central mechanism for mapping logical default types/aliases (e.g., "default chat") to actual LLMs.

Introduce standard default model types for agent requirements:

"Default" for general use.

"Default Chat" for conversational agents.

"Default Coding"

"Default Code Completion" for completion tasks.

Migrate agents without specific model requirements to use "Default" or the corresponding type.

Determine the concrete default model based on available API keys using the above logic.

Allow users to configure all default model preferences in one centralized place, eliminating the need to adjust each agent individually. Maybe allow users to define custom aliases.