Document Mode¶

A distraction-free plain-text editor that sits alongside Chat mode as a second way to use Orb. Chat mode is built around a character and a back-and-forth conversation; Document mode is just a blank page and a Generate button — good for freeform prose, one-shot stories, or raw prompt experimentation with no character card involved.

Switching to it¶

Click the 💬/📄 button next to "Orb" at the top of the sidebar. The sidebar swaps Worlds/Characters for a Documents list, and the main pane swaps the chat view for a full-page editor. Whatever document you had open reopens automatically the next time you switch back.

The mode-switch button, next to the Orb title in the sidebar header

Documents¶

+ New Document creates a blank one and opens it.
Click a document in the list to open it; ✏ renames it, ✕ deletes it.
Everything autosaves a second or two after you stop typing — the header shows Saved / Unsaved… / Saving… so you always know where you stand. Closing the tab mid-edit still flushes the save.
The token count next to the title is a rough estimate of the whole document's length, not just the last generation.

Generating¶

Put your cursor wherever you want the model to continue from and hit Generate (or Ctrl/⌘+Enter). Text streams in token by token and is tinted so AI-written text is visually distinct from your own — type anywhere, including inside a tinted run, and your keystrokes stay untinted. Stop (or Esc) cuts the stream off early. Each Generate adds up to your writer model's Max Tokens setting, shown live under "How to prompt."

Raw vs. Assisted¶

A toggle at the bottom of the editor picks how your document becomes a prompt:

Raw — sent to the model exactly as written, byte for byte. You type the chat template's own turn markers (e.g. <|im_start|>user) by hand and place the cursor right after the opening assistant tag. Full control, mikupad-style — best if you know your model's template.
Assisted — write plain notes instead of template syntax. Lines starting ### SYSTEM:, ### USER:, or ### ASSISTANT: are instructions; every other line is your story prose. Orb turns these into proper chat turns and lets the model's own template handle BOS/EOS and turn markers, so you never type them yourself.

### SYSTEM: You are a co-writer. Match the voice and style.
### USER: Write a story about a monkey. Ornate prose.
Once upon a time, beneath the gilded canopy…
### USER: Write tersely now. Short sentences.
The monkey woke. He▮

(Cursor at ▮.) Click How to prompt in the editor for this same cheat sheet in context. Generation stops when the model ends its turn — often just a sentence or two — hit Generate again to keep going.

Output Auditor¶

After each run the Output Auditor scans the new text with the chat editor's prose scanners (banned phrases, repetitive openers/templates, contrastive negation) and can fix findings with one forced JSON patch call that addresses each finding by its number — automatically when Auto-patch is on.

The patch call is KV-cache-friendly by construction: it byte-extends the generation prompt instead of building a fresh conversation, so llama.cpp reuses the document's warm prefix for the patch and the next Generate stays warm too (cache_prompt prefix matching). Per shape:

Text + Raw — patch prompt = the document verbatim + the run + the audit report appended as a raw continuation, JSON output forced by json_schema (decoding-only).
Text + Assisted — the client re-runs the exact /apply-template render the generation used (same parsed turns, same open prefill, same thinking kwargs) and extends that string the same way. Re-rendering rather than replaying the draft as a closed turn matters: templates like Qwen's inject <think></think> into the open generation prompt that a closed history turn wouldn't reproduce.
Chat (both) — the generation messages are replayed verbatim, the run is appended as the model's own turn, and the report as a trailing user turn; the patch schema stays out of the prompt (tools_in_prompt=False → forced via response_format). On templates that inject thinking bytes into the generation prompt (Qwen), reuse is bounded to the replayed context turns — text mode is the recommended path.

The frontend sends the full document text before the run as context (it is the generation prompt); truncating it would break prefix matching at character zero. The scanners themselves still see a server-side cap.

Undo/redo¶

Ctrl/⌘+Z and Ctrl/⌘+Shift+Z (or +Y) undo/redo both your own typing and anything the model generated, on one shared timeline — backing out a bad generation and backing out a typo work exactly the same way.

Document mode works on mobile too, with the same editor, generation, and shortcuts (Undo/Stop are also available as on-screen buttons).