AI Chat & Inline Edit

The core AI features that every user interacts with daily.

---

Sidebar Chat (Ctrl+L)

The sidebar chat panel supports multiple modes:

Adding context

• Ctrl+L with code selected: sends that selection as context
• @file mentions: attach specific files to the conversation
• Automatic: the chat sees your current file, cursor position, and recent edits

Thread Management

Conversations are persisted as threads. Each thread tracks:

• Message history (user + assistant)
• File checkpoints (snapshots before each edit)
• Staging selections (code context attached to messages)

Threads survive IDE restarts and can be navigated via history.

---

Inline Edit (Ctrl+K)

Quick code transformation without leaving your editor:

1. Select code (or place cursor on a line)
2. Press Ctrl+K
3. Type what you want changed
4. LLM generates the edit as an inline diff (red/green)
5. Accept or reject

How it works internally

• Creates a "CtrlKZone" — a bounded region in the editor with an input textarea
• User instruction sent to LLM with the selected code as context
• Response streams back as search/replace blocks
• Diffs render inline in real-time (Fast Apply mode)
• Checkpoints created before and after for undo support

---

Autocomplete (Tab)

AI-powered completions as you type:

• Fill-in-middle: completes the current statement
• Multi-line: suggests entire blocks
• LRU cached: repeated patterns complete instantly
• Uses FIM (Fill-In-Middle) capable models for speed
• Toggle: enableAutocomplete in settings

Best with a fast local model (Ollama) — no round-trip latency.

---

Fast Apply vs Slow Apply

Neural Inverse uses two strategies for applying LLM-generated code:

Fast Apply (default):

• LLM outputs search/replace blocks with conflict markers
• Changes stream directly into the editor as inline diffs
• Accept/reject buttons appear inline
• Near-instant visual feedback

Slow Apply (fallback):

• LLM rewrites the entire file
• Traditional diff comparison
• Used when Fast Apply fails or for complete file rewrites

Controlled by enableFastApply in settings.

---

Commit Message Generation

When you stage changes in the Source Control panel, Neural Inverse can generate commit messages:

• Uses the model selected for the "SCM" feature
• Analyzes staged diffs
• Generates a conventional commit message
• Includes BYOLLM attribution trailers automatically

---

Contributing

Key files:

• void/browser/sidebarPane.ts — sidebar panel registration
• void/browser/quickEditActions.ts — Ctrl+K action + zone creation
• void/browser/editCodeService.ts — core apply logic (Fast/Slow)
• void/browser/autocompleteService.ts — tab completion engine
• void/browser/chatThreadService.ts — thread persistence and management
• void/common/voidSettingsTypes.ts — all settings definitions

---

---