Architecture

Neural Inverse OSS is built on VS Code. All Neural Inverse features live under src/vs/workbench/contrib/.

Modules

Module	What it is	Window
`void/`	Core LLM integration — chat, inline edit, autocomplete, settings, threads	Sidebar (Ctrl+L, Ctrl+K)
`powerMode/`	Autonomous coding agent with tool calling	Cmd+Alt+P
`neuralInverse/`	Agent manager, model management, deployments	Cmd+Alt+A
`neuralInverseFirmware/`	Embedded development — MCU database, SVD, serial, debug	Cmd+Alt+F
`neuralInverseModernisation/`	Legacy migration — 5-stage pipeline with KB	Cmd+Alt+M

Process Model

Process	Folder pattern	Has access to
Browser (renderer)	`browser/`	DOM, window, React components
Main (Node.js)	`electron-main/`	node_modules, filesystem, network
Common (shared)	`common/`	Types and utilities — no process-specific APIs

LLM requests route from browser to main process via IPC. This avoids CSP restrictions and lets us use provider SDKs (Anthropic, OpenAI, Ollama, etc.) that require Node.js.

Service Registration

Every module has a .contribution.ts file that imports services as side-effects:

// void.contribution.ts
import './editCodeService.js'
import './sidebarPane.js'
import './autocompleteService.js'
// ...

Each imported file calls registerSingleton() which registers the service with VS Code's DI container.

Services inject dependencies via decorators:

constructor(
  @ILLMMessageService private readonly llm: ILLMMessageService,
  @IVoidSettingsService private readonly settings: IVoidSettingsService,
) {}

How Features Connect

User types in Sidebar Chat (Ctrl+L)
  -> ChatThreadService stores message
  -> LLMMessageService routes to main process
  -> Main process calls provider SDK (streaming)
  -> Response streams back via IPC events
  -> If tool calls: IToolsService dispatches to registered tools
  -> If code edits: EditCodeService creates DiffZones
  -> Inline diffs render in the editor (red/green)
  -> Checkpoints created for undo

User opens firmware workspace
  -> ProjectDetector scans for firmware markers
  -> MCUDatabase loads matching MCU profile
  -> SVDParser loads register maps
  -> HardwareContextProvider injects MCU context into LLM prompts
  -> fw_* tools become available in chat + Power Mode

User starts modernisation session
  -> ModernisationSessionService tracks state
  -> Discovery/Planning/Translation engines run per stage
  -> KnowledgeBase persists all decisions
  -> Tools auto-register for Power Mode access

Shared Infrastructure

Service	What it does	Used by
`ILLMMessageService`	Routes LLM requests to providers	Everything
`IVoidSettingsService`	Provider configs, model selections, preferences	Everything
`IVoidInternalToolService`	Tool registry for agent tool calling	Power Mode, Sidebar, Modernisation
`IEditCodeService`	Code modification (fast/slow apply, diffs)	Chat, Ctrl+K, Power Mode
`IChatThreadService`	Conversation threads + checkpoints	Chat, Power Mode

Key Rules for Contributors

ASCII only in TS/JS string literals — non-ASCII breaks the esbuild release build
No any casts — find the correct type
No changes outside contrib/ without discussion — upstream VS Code code is maintained separately
BYOLLM attribution — AI-assisted commits need dual Co-authored-by trailers (see CONTRIBUTING.md)
Semicolons — follow existing convention per file
Service registration — always through .contribution.ts side-effect imports, never direct instantiation

Was this page helpful?