Running a translation
Home is the language dashboard: one row per target language with a progress bar, a status chip, and the last run time. Click a row to open it in Review.
Start a run
Section titled “Start a run”Tick the languages you want and press Translate. Languages run one after another, each as a job with scope pending — every key that is missing in that language plus every key the checks flagged. The run strip shows live progress; the AI-tasks button in the title bar hosts the same view from anywhere. Reload the window or open a second one mid-run and it picks the run up again — its progress and its Cancel.
Rules a run always follows:
- One job at a time. Starting a second is refused, not queued.
- A run that can’t work doesn’t start. No engine set up for the feature, or a prompt that can’t work (AI features): the run is refused with the reason and the fix. A setting that breaks mid-run — the prompt edited, the preset deleted, the built-in engine unable to load its model — ends the run with the reason instead of failing every key; what was staged stays staged.
- Cancel keeps staged work. Everything already translated stays staged for review; nothing is thrown away.
- Exhausted keys are NAMED. If a key fails every retry (the engine kept breaking a placeholder, say), the run finishes with that key listed — never silently skipped.
The Runs page offers the same start controls plus scope choices
(pending / flagged / unsure / all) and the full run history. A running run
shows there as the same strip as on Home, with its progress and Cancel; its card shows
the counts, or the last run — how it ended and the keys it couldn’t deliver — after a
reload too. One honest
note: unsure selects keys the probe marked as disagreements —
and today the probe only runs from the CLI, so on an app-only workflow that scope
finds nothing.
The staged pile — read this once
Section titled “The staged pile — read this once”A run writes no locale files. It stages proposals. So after your very first
run the language row may say 0 done and 120 staged — that is a successful
run waiting for you, not a failure. Open Review, look at the pile, and press
Apply all when you’re ready: that click is the first time anything touches
<lang>.json.