Point the agent at exact lines

What it is

When you select whole lines in the editor — or drag a box over nodes in the flow pane, which is the same selection — a floating bar appears with three actions: Convert to process, AI edit and Copy to chat. Copy to chat opens the panel and pins a chip above the message box that reads, for example, “Onboarding · L4–L9”, with the first selected line as a preview. The placeholder changes to “Describe the change for the selected lines…”.

Region-safe by construction

A message sent with a chip goes to a dedicated edit route. The model receives the whole document with the selection marked between <<<SELECTION>>> and <<<END SELECTION>>>, and is instructed to return only the replacement lines for that region — no markers, no commentary, no lines from outside it. Runic then splices the reply into the exact line range on the client. A bad reply cannot touch text outside the selection, because the text outside the selection is never replaced. Your own message posts with a “[selection: ‘Onboarding’ L4–L9]” note so the range is on record.

The result arrives as a normal proposal: “Here’s the change for L4–L9 — nothing is applied until you say so:”, with the diff, Apply and Discard, the diagram preview, and a checkpoint once applied. If the reply equals what was there, the chat says “That part already does what you asked — no changes needed.”

Whole blocks travel together

The range is expanded to complete blocks before it is sent, the same rule the extraction feature uses. Select a .then line and its .if comes along; select an .if and its branches and .end come along. The floating bar shows how many lines were pulled in (“+2 to keep the block whole”), so the replacement always fits the structure around it.

Two ways to use a selection

The AI edit button on the same bar uses the same region-safe route but skips the chat: an instruction bar appears in the editor, the replacement shows as a diff with Apply and Discard, and after Apply an “AI edit applied” toast offers Undo for 15 seconds. Copy to chat is the version with a conversation around it — the proposal waits, previews in the diagram, and keeps a checkpoint.

Limits

Selection context works on process documents in a signed-in session. The chip is dropped when you switch documents, and if the document changed since you selected, the chat asks you to select again. One intent per message: attaching a file to import replaces the chip. The edit route allows 30 requests an hour, shared with the inline AI edit; it reads the first 1,000 characters of the instruction and the first 6,000 of the selection, plus 6,000 characters of context on either side, and runs on Claude Sonnet 5.

How to use it

1. 1
   Select the lines

   Drag across whole lines in the editor, select one full line, or drag a box over nodes in the diagram.
2. 2
   Press Copy to chat

   Click 💬 Copy to chat on the floating bar. The chat opens with a chip such as “Onboarding · L4–L9”.
3. 3
   Describe the change

   Type in the box (“Describe the change for the selected lines…”) and press Enter. Click ✕ on the chip to detach it first if you changed your mind.
4. 4
   Review the scoped diff

   The reply is a diff limited to that range. Apply to write it in, Discard to drop it; the applied change keeps a ↩ Restore checkpoint.

Commands & shortcuts

Tips

• Draw the selection on the diagram when that is where you noticed the problem — the bar offers the same actions.
• Keep the instruction about that region only. The model sees the document around the selection for context but can only return lines for the range.
• Ask for structure, not just wording: “turn this into an .if with a Yes and a No branch” — the block expansion keeps the result well-formed.
• If the chip vanished, you switched documents; select again in the document you meant.

Questions