Next-edit prediction
How Arshas predicts your next edit — and how to steer it.
Classic autocomplete predicts the next characters at your cursor. Arshas predicts your next edit — including edits on other lines and in other files. Rename a function and the model anticipates updating its call sites; change a type and it queues the follow-on fixes.
How it works
Every keystroke updates a lightweight model of your editing session: recent edits, the diff so far, diagnostics, and the semantic neighborhood of your cursor from the codebase index. From that context, Arshas proposes the most likely next change and renders it as a ghost diff you can accept with Tab.
Predictions are ranked by confidence. Low-confidence guesses are suppressed rather than shown — an interruption that is wrong costs more than no suggestion at all.
Steering predictions
The model follows your lead. Accept a couple of predictions in a refactor and it learns the pattern, offering to complete the remaining instances in one jump. Ignore a suggestion and it backs off that pattern for the session.
You can disable predictions per language or per workspace in Settings → AI → Predictions, and adjust how aggressively multi-line edits are offered.