GitHub Docs

Diffcomplete (⌘.)

Diffcomplete (⌘.)

AI editing without ghost-text. Press ⌘. anywhere in the editor, describe the change you want, review the unified diff hunk-by-hunk, accept or reject what you like.

Diffcomplete is VibeCody’s only AI code-editing surface. It deliberately replaces keystroke-driven inline completions with an explicit-trigger, diff-output flow — better for review, better for safety, and patent-distant from FIM / next-edit / ghost-text systems.

Quick start

  1. Open any file in VibeUI (the desktop editor).
  2. (Optional) Select a region of code. If nothing is selected, the whole file is the editing target.
  3. Press ⌘. (Cmd-Period on macOS, Ctrl-Period on Linux/Windows). Or open the command palette and run AI edit (diff).
  4. Type the change you want — e.g. “extract the validation into a helper function”.
  5. Press ⌘⏎ to generate. The model returns a unified diff which is rendered in a side-by-side review.
  6. Per hunk: Accept keeps the change, Reject drops it, Edit lets you tweak the proposed text in place.
  7. Click Apply to write the accepted hunks back to the file.

That’s it. There is no other AI editing surface in VibeCody. If something looks like inline completion, it isn’t — it’s a static lint or formatting hint.

How it works

Diffcomplete sends a single deliberate request per ⌘. press:

[ ⌘. press ]
   │
   ▼
DiffCompleteModal  ── user types instruction ──┐
   │                                            │
   ▼                                            ▼
diffcomplete_generate (Tauri command)     [ Enter / ⌘⏎ ]
   │
   ▼
vibe_ai::diffcomplete::generate()
   │
   ├─► Builds a system prompt that demands unified-diff output
   ├─► Includes selection + 200-line context window each side
   ├─► Includes any user-attached "additional files" (explicit picker only)
   ├─► Includes project_memory from VIBECLI.md / AGENTS.md / CLAUDE.md
   │   (author-authored only; never auto-extracted state)
   └─► Calls the active provider's chat() surface — never FIM
   │
   ▼
Unified-diff text  →  applyUnifiedDiff()  →  modified file content
   │
   ▼
DiffReviewPanel  ── user accepts / rejects per hunk ──►  file written

Key design rules:

  • Trigger is explicit. Never on keystroke, never on idle, never on selection-change. Only ⌘. (or the equivalent palette command).
  • Output is a unified diff. Not a code suggestion, not a completion. The model is instructed to emit exactly one fenced ` ```diff ` block and the daemon extracts it.
  • Review is mandatory. The modified content is never silently applied — DiffReviewPanel always renders the hunks for review.
  • No hidden state. The only context shipped to the model is what’s visible in the modal: selection + 200-line bounded window, attached files, project memory, and the previous diff if you’re refining.

Refinement

After a diff is generated, the modal shows a Refine this diff field below the review.

  • Type a refinement instruction (e.g. “tighten the error path”) and press ⌘⏎ to regenerate.
  • The previous diff is included in the next request as a “Previous attempt” block, so the model refines instead of starting over.
  • The refinement is layered on top of your original instruction — you can iterate without losing intent.
  • You can repeat refinement as many times as needed; only the most recent diff is kept as context.

The refinement chain is visible to you and to the model — there is no auto-retry or shadow-context.

Additional files (manual context)

The + Add file… button under the instruction box opens a native file picker. Selected files are read by the daemon (sandboxed) and included in the prompt under an “Additional files (user-supplied context)” section.

This is human-in-the-loop retrieval by design:

  • Files are added by you, never by automatic embedding search or call-graph traversal.
  • No usage telemetry is collected from this picker.
  • You can remove an attached file via its × button before submitting.

If you want a file in scope, attach it. Diffcomplete will not guess.

Configuration

Default provider

Diffcomplete uses the AI provider that’s currently active in VibeUI’s Settings → API Keys. Any provider with a chat() surface works — Anthropic, OpenAI, Gemini, Groq, OpenRouter, Mistral, DeepSeek, Cerebras, Together, Fireworks, SambaNova, Perplexity, and the local mistralrs / Ollama paths are all supported.

To change the active provider mid-session: Settings → API Keys → click the provider you want.

Required configuration

At least one provider key. No provider configured = no diffcomplete.

Set a key one of two ways:

  1. In-app: Settings → API Keys → choose provider → paste key → Save.
  2. From the terminal: 
    vibecli set-key anthropic sk-ant-...
vibecli set-key openai    sk-...
vibecli set-key huggingface hf_...   # for gated mistralrs models

Keys are stored in the encrypted ProfileStore at ~/.vibecli/profile_settings.db — never in .env files, never in plaintext.

Verifying readiness

The daemon’s /health endpoint exposes the canonical readiness signal:

curl http://127.0.0.1:7878/health | jq '.features.diffcomplete'

{
  "available": true,
  "requires": "providers.configured_count > 0",
  "transport": "tauri-desktop"
}

available is true whenever /health.providers.configured_count > 0. Diffcomplete inherits its readiness from the canonical providers signal — there’s no separate per-feature probe.

Troubleshooting

“No active AI provider configured”

You don’t have any provider keys set. Open Settings → API Keys and add one, or run vibecli set-key <provider> <value> in your terminal.

“Model response did not contain a diff block”

The model didn’t follow the unified-diff format. This happens occasionally with smaller open-weights models. Workarounds:

  • Switch to a stronger provider (Claude, GPT, Gemini Pro tier).
  • Simplify the instruction — vague asks (“make this better”) confuse the format.
  • Try Regenerate with a refinement.

“Model returned a diff that could not be applied cleanly”

The model invented context lines that don’t match your file. This is typically a model-quality issue, not a Diffcomplete bug. Workarounds:

  • Click Regenerate with a refinement that points at the real code (e.g. “the variable is called userId not user_id“).
  • Close the modal, re-select a smaller region (200-line context window applies around the selection), and try again.

“Provider rate limit hit” / 429

Your provider has throttled you. Wait, or switch to a different provider via Settings.

“401 Unauthorized” / “Invalid API key”

The provider rejected the key. Re-check it in Settings → API Keys. Some providers (Anthropic, OpenAI) rotate trial keys quickly.

Network timeout

Slow connection or the provider’s region is unhealthy. Switch providers, or wait it out.

Scope

Diffcomplete is currently desktop-only — it ships as a Tauri command in VibeUI, not as a daemon HTTP route. Mobile / Watch / IDE plugins do not surface it today.

If you need diffcomplete from another client surface, file an issue describing the use case. The backend (vibe_ai::diffcomplete) is already a clean library and can be exposed via daemon HTTP without protocol changes.

Observability

Every diffcomplete generation emits structured tracing events on the daemon under the vibecody::diffcomplete target:

  • debug on request entry — provider, language, file_path, instruction_len, context_lines, has_selection, is_refinement, extra_files
  • warn when the provider is unavailable, the chat call fails, or no diff block was returned
  • info on success — provider, diff_len, had_explanation

Tail the daemon log with:

journalctl --user -u vibecli  # systemd
# or
RUST_LOG=vibecody::diffcomplete=info vibecli serve

User content (instruction text, file paths) is not logged at any level — only shapes (length, count) and stable enums (provider name, language). No telemetry leaves your machine without explicit opt-in.

Security

  • No autocomplete from history. The model never sees your shell history, edit history, or any state you didn’t explicitly attach.
  • Project memory is author-authored only. The project_memory field carries content from VIBECLI.md / AGENTS.md / CLAUDE.md only — never auto-extracted user state, scratchpads, or orchestration logs.
  • Sandboxed file reads. Files attached via + Add file… are read through read_file_sandbox, which respects the active workspace boundary.
  • No eval, no exec. The applier is a pure unified-diff parser — applyUnifiedDiff() in the modal source. It will refuse to apply hunks whose context doesn’t match the current file content.

What Diffcomplete is NOT

For clarity (and because we removed these on purpose):

  • Not keystroke-driven ghost text. There is no registerInlineCompletionsProvider path in VibeCody.
  • Not FIM (fill-in-middle). The model receives prefix + selection + suffix as discrete labeled regions, not a single FIM template.
  • Not next-edit prediction. The model only sees the current file state, never your past edits.
  • Not auto-retrieval. Files in context come from your explicit picker. There is no embedding search, no call-graph walker, no symbol-server probe.

Diffcomplete is a deliberate, claim-distant alternative to those patterns. If you want a different shape of AI editing, that is a feature request — but the answer will not be re-introducing ghost text.