Reviewing — flags, accepting, notes

Open Review. The language picker shows outstanding counts and lands you on the busiest language — where the work actually is.

The workspace

Three panes. On the left, the queue rail: buckets with live counts — Needs review (anything a hard check flagged), Unsure, Terminology, Missing, Came back identical, Proposed, All flagged — plus Accepted (every verdict you’ve recorded, with an Un-accept on each: a decision can always be revisited). Under the buckets, a per-check breakdown — ten spurious questions are ten instances of the same decision, so working one code at a time beats context-switching — and a search box over keys and text.

In the middle, the list: one terse row per key. Its job is only to let you move; judging happens in the detail pane. A reviewed key shows struck-through. In the Came back identical bucket the list grows a bulk bar: tick rows (or press “tick the N the engine calls correct” — a suggested selection, nothing more) and Approve them in one click, one undo.

On the right, the detail pane — everything a reviewer once left the page to find out: why it’s flagged in plain English, the source with placeholders marked, the engine’s proposal, a Google Translate second opinion (an independent reading — neither source is reliably better; copy it across only if you agree), a back-translation, the namespace siblings, and the note box.

The keyboard

j/k (or arrows) move · a accept · u undo · e edit the target · g toggle the Google panel · b back-translate · / search · Ctrl/⌘Z undo · Escape leaves a text box. Keys never fire while you type.

The staged pile

If the last run staged translations, a banner offers Apply all (with a confirmation naming the file it writes) and Discard them. Applying writes <lang>.json and immediately re-runs the checks on what was written — whatever they flag lands in the queue below.

What the flags mean

Each queued key shows its source, an editable target, and one or more flags:

• missing / blank — no translation, or an empty one. A string that is empty in the source on purpose comes back empty from a run: there is nothing to translate.
• placeholder-changed — a {name}-style token was lost or altered. The #1 ship-blocker; the app would crash or print garbage. A translation that adds plural forms its source doesn’t need (“{n} encontrado | {n} encontrados” for “{n} found”) is checked form by form, and so are numbers and brackets.
• plural-halves-lost / plural-halves-identical — a one|other plural lost a half, or both halves came back identical.
• glossary-translated — a do-not-translate term got translated.
• untranslated — the target equals the source (often correct — “OK” is “OK” in Spanish; that’s what accepting is for).
• startpunc / endpunc / spurious-interrogative — punctuation drift at the edges (¿…? handling in Spanish, trailing colons, …).
• numbers / brackets / doublewords / whitespace — numbers changed, brackets unbalanced, a word doubled, or stray whitespace.
• disagreement — advisory: a second engine pass translated this key differently; worth a look, never a failure.
• terminology — advisory: this key translates a term inconsistently with the rest of your catalogue.
• source-changed — the English changed since this was translated; the flag shows the English it was made from. Fix it and Save, or accept it if the translation still says the same. A run of pending keys picks these up. (Keys translated before docgen started noting this are taken as up to date the first time it sees them.)

The actions

• Save — edit the target and save; the checks re-run instantly and the row re-flags or clears on the spot.
• Accept as correct — “I looked; this is right.” The flag stops appearing.
• Unaccept — take it back.
• Apply proposal — take the engine’s staged suggestion for this one key.
• What does it say? — a back-translation of the current target into your language, from the same engine that translated it. A sanity mirror, not proof.
• Note for the next run — see below.
• Siblings — nearby keys (same prefix), for consistency checks.
• Undo — one click undoes the last action, even when that action was a bulk apply of two thousand keys.
• Accept N pre-ticked — bulk-accepts rows whose only flag is untranslated and which the confirmation pass verified as genuinely-the-same. The pass runs after CLI translate runs AND after app-started runs (since 2026-08-04): when a run finishes, its byte-identical proposals get the second opinion, and the rows arrive pre-ticked the moment you apply them. The engine still never signs off — pre-ticks are annotations; the acceptance is always your click.

Accepting — why acceptances expire

An acceptance is a statement about one exact pair of strings: it’s recorded against the key, the flag code, the source text, and the target text. Edit either side — even one character — and the finding comes back for fresh eyes. That’s deliberate: “I approved this sentence” must never silently cover a different sentence.

Acceptances live in <lang>.accepted.json, committed to git, each stamped with the reviewer name from Settings → Reviewer. Set that name — otherwise your review record says “unknown” forever.

Notes — teaching the next run

A note rides with its key into the next translation prompt. The canonical example: the engine translated a button labeled “Why?” literally, when the app wanted the interrobang tone of “¿Por qué?” — a one-line note fixed it on the next run and every run after. Notes live in <lang>.notes.json, committed.