HybridCoder is an on-device iOS coding studio that is being refocused into a React Native and Expo builder.
- SwiftUI host app
- local workspace and project state
- chat-driven local coding workflow
- SpeziLLM/SpeziLLMLocal orchestration over llama.cpp GGUF runtimes
- semantic retrieval backed by local embedding models
- Qwen 2.5 Coder generation and explanation services
- patch planning and editing flows
- imported repository and prototype handling
HybridCoder runtime summary (as of 2026-04-17): the app routes local LLM orchestration through FoundationModelService and QwenCoderService, both backed by SpeziLLM/SpeziLLMLocal + llama.cpp, with models resolved from Files > On My iPhone > Hybrid Coder > Models/ via ModelRegistry.externalModelsRoot.
The repository already contains a real local LLM coding core:
- chat-driven route selection
- semantic retrieval over indexed code
- code generation and code explanation
- structured patch planning
- patch application to the active workspace
The repository does not yet have a finished bytecoding runtime.
That means there is still no fully implemented agent layer that can autonomously:
- interpret a user goal as a multi-step coding plan
- choose and sequence coding sub-tasks
- create, modify, rename, and delete files as first-class actions
- iterate until the requested feature or app milestone is actually brought to life
Start with docs/architecture.md.
That file is the canonical source for:
- the current implementation
- the target builder architecture
- the bytecoding / agent-runtime gap
- the non-drift rules that should keep the product aligned
Then use:
docs/product-vision.mdfor goals and non-goalsdocs/agent-runtime.mdfor the bytecoding strategydocs/roadmap.mdfor the phased refactor plan