Native architecture

Process boundaries

React renderer (sandbox, context isolation, no Node)
        │ typed window.reges calls + correlated events
Narrow preload bridge
        │ validated IPC from the owned main frame
Electron main broker
        ├─ native file dialogs → bounded GGUF inspector
        ├─ hardware snapshot → memory admission
        ├─ local settings/conversation store
        └─ runtime controller
           ├─ private Lemonade adapter → qualified llama.cpp backend
           └─ experimental Strata adapter → private JSONL helper → native engine

The renderer draws state and accepts user intent. Main owns paths, native dialogs, network credentials and child processes. A model response is untrusted text; it cannot turn into shell execution or grant a tool permission.

Source map

Source	Responsibility
desktop/src/shared/contracts.ts	Typed snapshots, options, results and events
desktop/src/main/main.ts	IPC validation, serialization, lifecycle and exports
desktop/src/main/artifacts.ts	GGUF inspection and coarse fit planning
desktop/src/main/hardware.ts	Captured host/GPU availability
desktop/src/main/runtime.ts	Owned private Lemonade runtime and streaming adapter
desktop/src/main/runtime-controller.ts	Explicit engine dispatch and ownership across both adapters
desktop/src/main/strata-runtime.ts / strata-config.ts	Experimental private Strata protocol, exact artifact guard and memory admission
desktop/native/	Private tokenizer/template and token-stream helper
desktop/src/main/store.ts	Local atomic JSON persistence
desktop/src/renderer/	Functional native interface and optional presentation

The Strata adapter completed installed functional qualification for the exact reference profile; the controller dispatches to one explicit model engine and does not silently substitute one on failure. Lemonade remains the fresh/legacy default.

The public REST API proposed elsewhere is not this bridge. The current preload is an internal desktop boundary, not a supported general-purpose SDK for arbitrary web pages.

State and concurrency

Runtime phases distinguish stopped, starting, ready, installing, loading, loaded, generating and error. The host serializes mutating actions and tracks one active generation. Request IDs correlate chunks and cancellation. Completion must settle owned work before a conflicting operation is accepted.

Settings and conversations use atomic replacement in a per-user profile. Models remain in their selected locations. A future durable agent system needs a richer run store, recovery model and permissions; it should not be represented by adding imaginary tool events to the current chat stream.

Packaging and presentation

The Windows shell packages compiled code in ASAR and verified runtime resources beside it. ASAR is packaging, not encryption or a security boundary. Native security depends on the sandbox, isolated preload, validation and ownership rules.

The startup and About nebula are disposable presentation components. Chat must not retain their WebGL contexts or animation loops. This keeps decorative work separate from inference; it does not by itself make the inference engine faster.