Demo 07 — DiffComplete (⌘.)

Note: This demo previously covered Inline Chat (⌘K), inline AI completions, Supercomplete, and next-edit prediction. All of those surfaces were removed from VibeCody on 2026-04-26 as part of the inline-completion patent-distance work. The only AI editing surface in VibeCody is now DiffComplete (⌘.), described below. (For background on the removal, see notes/PATENT_AUDIT_INLINE.md if you have a local working copy.)

Overview

DiffComplete is VibeCody’s explicit-chord, diff-mode AI editing surface. You select code (or none for a whole-file edit), describe the change, and the model returns a unified diff that you review hunk-by-hunk in a modal. You can edit the result in Monaco before applying, and you can refine the diff with a follow-up instruction without re-entering the original prompt.

This is deliberately not a ghost-text / inline-completion / FIM-style flow.

Prerequisites

• VibeCody installed (vibecli --version returns 0.5.5+)
• At least one AI provider configured in ~/.vibecli/config.toml:

[claude]
enabled = true
api_key = "sk-ant-..."
model = "claude-sonnet-4-6"

• VibeUI built and running:

cd vibeui && npm install && npm run tauri dev

• A project open in VibeUI (File > Open Folder)

Step-by-Step Walkthrough

1. Open a source file

Click a file in the Explorer sidebar, e.g., src/main.rs.

2. Trigger DiffComplete

Optionally select a region of code, then press Cmd+. (macOS) or Ctrl+. (Linux/Windows). With no selection the request applies to the whole file.

The DiffComplete modal opens.

3. Write your instruction

Type a natural-language instruction:

Convert this function to async and add proper error handling

4. (Optional) Attach extra files as context

Click + Add file… to open the OS file picker and select related files. The selected files are forwarded to the model as explicit context — there is no automatic embedding search or hidden retrieval.

The picker is the only path for cross-file context. Attached files appear as removable chips.

5. Submit (Cmd+Enter)

The backend calls diffcomplete_generate with your instruction, the selection (or whole file), the surrounding 200-line context window, and any attached files. The model returns a unified diff which is parsed by applyUnifiedDiff and handed to DiffReviewPanel.

6. Review hunk-by-hunk

DiffReviewPanel shows each hunk with Accept, Reject, Accept All, Reject All buttons. The currently-assembled result preview updates as you change accept/reject choices.

7. (Optional) Edit before applying

Click Edit ✎ to switch into a Monaco editor seeded with the currently-accepted hunks. Make any final tweaks, then press Apply (edited) to write the modified text. Use ← Hunks to discard your edits and return to the hunk view.

8. (Optional) Refine and regenerate

Below the diff view there’s a Refine input. Type a follow-up like “tighten the error path” or “use the helper from utils”, then press Regenerate (Cmd+Enter). The previous diff and your refinement are layered on top of the original instruction — the chain stays visible in the prompt.

9. Apply

When you’re satisfied, click Apply (or Apply (edited) if you went through the edit step). The buffer is updated via editor.executeEdits and a diffcomplete event is recorded in the Cascade Flow timeline.

What’s not in VibeCody anymore

The following Tauri commands and frontend modules were removed on 2026-04-26 and are intentionally not coming back:

• request_inline_completion / request_ai_completion — FIM-style code completion
• predict_next_edit — next-edit prediction
• inline_edit / generate_code — Cmd+K backend
• semantic_search_codebase / build_embedding_index — orphan after SupercompleteEngine was removed
• vibe-ai/src/completion.rs (CompletionEngine)
• vibeui/src/components/InlineChat.tsx
• vibeui/src/utils/SupercompleteEngine.ts
• vibeui-ai-inline-completion-enabled localStorage toggle
• Monaco.languages.registerInlineCompletionsProvider registration in App.tsx

If you need an alternative, use DiffComplete (⌘.) — it is the supported AI editing surface.