Model-Agnostic Architecture
Idea
Make Opus portable across LLM providers — not just Claude Code. Goal: switch to another model (cloud or local) with minimal rework.
Why
- Avoids vendor lock-in
- Enables local/offline usage
- Future-proofs the system as the model landscape evolves
What this would involve
- Thin LLM abstraction layer in
src/(provider interface, swap via config) - Audit
.claude/skills/agents for Claude-specific assumptions - Prompt templates stored as model-neutral Markdown, with a rendering step per provider
- Test with at least one alternative backend (e.g., Ollama, OpenRouter)
Status
Idea — no urgency. The CLAUDE.md design principle guides day-to-day decisions in the meantime.