platform/docs/localization.md

# Localization

**Status: DESIGN, NOT BUILT — with a working 5% skeleton already in the tree.**

This is an execution plan. It is written to be handed to an agent who has not had the conversation
that produced it, so it states the decisions and the reasons for them, not just the steps.

Every count here is a **measurement taken on 2026-08-23**, not a claim that stays true — the whole
point is for it to fall. Re-derive with `bun scripts/i18n/scan-hardcoded.ts` rather than trusting the
number beside it, and do not wire these into `docs:check`: a guard that fails when the count drops
would be a guard against progress.

---

## What this is

Every word the UI shows a person should come from a catalog keyed by locale, so the platform can be
read in a language other than English. Routes stay English (`/minesweeper`, `/chat`) — that is a URL
decision, deferred, and nothing here depends on it.

The scope is **the web UI only**. Not the server, not the plugins, not email or push. See D1 for what
the server does instead, which is the one decision that reaches outside this boundary.

---

## What exists today, and what it cannot do

**Updated 2026-08-23, after phases 0, 1 and 2.** This section described the pre-phase-0 world for long
enough that the next reader would have started from a false picture — bare `en.json`, 46 keys, three
components. All of that is now history and is kept below only as the "before" column, because the shape of
the gap is the argument for the plan and deleting it would leave the phases looking arbitrary.

|                            | before phase 0         | now                                           |
| -------------------------- | ---------------------- | --------------------------------------------- |
| catalogs                   | `en.json`, `pt.json`   | `en-GB.json`, `pt-PT.json` per workspace      |
| keys                       | 46                     | **581** — 538 `app.*`, 43 `ui.*`              |
| workspaces with a catalog  | 1 (`officer-web`)      | 3 (`officer-web`, `components`, `officerdev`) |
| components using it        | 3                      | all of `officer-web` and `components`         |
| plural rules               | none                   | `Intl.PluralRules`, per locale                |
| dates and numbers          | browser locale         | `useFormat()`, the account's locale           |
| a missing key              | rendered its dot-path  | falls back to English, warns in dev           |
| locales offered vs shipped | ~20 offered, 2 existed | only shipped locales are offered              |
| parity enforcement         | none                   | **a gap fails the build**                     |

Still standing from the original description, and still true: the package is hand-rolled, has no npm
dependency, and gives type-safe keys — a wrong key is a typecheck failure, which is worth more than it
sounds when one agent is editing hundreds of files.

**What remains:** `src/workspaces/officerdev` — the panels — and then translation. `panel.*` is scaffolded,
wired into the bridge, and empty.

Two things learned doing phases 1 and 2 that the plan did not anticipate:

- **`TranslationKeys` walked into a plural object**, so `t('key', { count })` — the entire point of D7 —
  did not compile while `t('key.one')` did. The documented feature was uncompilable and the workaround was
  not, which no test would ever have noticed because phase 1 contained no plurals. Fixed with type-level
  assertions.
- **The typed wrappers rebuilt `t` on every render**, which is the trap `CLAUDE.md` already records against
  `useClient()` and the Jellyfin player. It matters the moment `t` lands in a `useMemo` dependency list,
  which phase 2 needed immediately. Memoised.

**A shape this format cannot express:** a sentence with an inline link. Splitting around the anchor gives a
key whose Portuguese value is `"."`. The cheap fix is rephrasing so the link falls at a clause boundary;
there is no node-level interpolation helper and adding one is a design decision, not a patch. Worth
raising if a phase turns out to be full of them rather than absorbing it forty times.

### The size of the remaining work

Measured, not guessed — `bun scripts/i18n/scan-hardcoded.ts`:

| area                  | files   | occurrences |
| --------------------- | ------- | ----------- |
| `officer-web` (shell) | 132     | 556         |
| `officerdev` (panels) | 213     | 567         |
| shared `components`   | 91      | 47          |
| **total**             | **436** | **1170**    |

**884 distinct strings**, of which **88 need interpolation**. Against 46 keys today.

That is a floor, not a ceiling. The scanner cannot see multi-line JSX, ternaries inside JSX, strings
passed down as props, or anything a plugin renders. Budget for **1200–1500 keys**.

---

## Decisions

### D1. The web UI only — and the server returns codes, not prose

The server should **never emit user-facing prose**. It returns a stable machine-readable code; the UI
owns every word a person reads and localizes it like any other string.

```jsonc
// not this
{ "error": "The Minesweeper session is not running" }
// this
{ "code": "MINESWEEPER_NOT_RUNNING" }
```

This is better than translating the server, for reasons that have nothing to do with translation. A
code maps to exactly one failure site, so it is greppable and a support conversation has something
precise to quote. Prose drifts; a code does not. And the server stops needing to know who is reading.

- **Phase 1 establishes the convention** so new code stops adding debt: new error paths emit codes,
  and the catalog gets an `errors.*` namespace.
- **Phase 2 migrates the existing ~53** prose strings in `src/servers`. Not in this plan's scope.
- **Unmapped code** renders a generic localized line **with the code visible**. Never a blank, never
  a raw key.

**The trap, and it is a real one: a code must not leak more than the prose it replaces.** Where a
refusal is deliberately vague, the code has to stay equally vague. `officer-secret-store` returns a
byte-identical `403` for "no such app" and "no grant" precisely so a caller cannot enumerate what
exists; splitting that into `APP_UNKNOWN` and `GRANT_MISSING` would hand back the enumeration oracle
the prose was careful not to be. When in doubt, one code for one _response_, not one per branch.

### D2. Regional locales, BCP 47, and every catalog is complete

`pt-PT`, `pt-BR`, `en-GB`, `en-US` — not bare `pt` and `en`. Portuguese is the case that forces it, and
it is the owner's own language, so getting it wrong is felt immediately.

**Every locale ships a COMPLETE catalog. There is no regional inheritance and no overlay.** A file holds
every key the UI has, translated for that locale, or it does not ship.

This reverses an earlier version of this decision, which had `pt-BR` inherit from `pt` so a base catalog
carried the shared bulk and a regional file carried only the differences. That is the conventional design
and it was rejected for a product reason that outranks the engineering one:

**A missing European Portuguese key would have fallen through to Brazilian.** European and Brazilian
Portuguese differ most in exactly the vocabulary this UI is full of — technical terms — and the platform is
being sold into the European Portuguese market first. So the failure mode of an overlay is a Portuguese
reader getting Brazilian phrasing on whichever strings nobody remembered to override, silently, and with no
signal that it happened. Owner's call, and it is the right one: a reader must never be served a neighbouring
variant's words.

The engineering objection to full catalogs was drift — two near-identical files diverging within a month.
That is answered by making the **key-parity test load-bearing** rather than advisory (see Tooling): a locale
missing a key that `en` has fails the build, so an incomplete catalog cannot ship. Completeness is enforced
by the test rather than by a fallback quietly papering over a gap.

It is also less code. `resolveKey` looks in the active locale and then, only if the key is genuinely absent,
in `en` — which per D8 is a bug the parity test should already have caught. There is no chain to walk, no
base-versus-overlay merge order, and no question about which file a given string lives in.

`Intl` takes BCP 47 tags natively, so dates, numbers and plurals come out right with no extra work.

**Migration:** `settings.languages.default` stores bare codes today (`pt`). The picker must move to regional
tags, and stored values need a one-time widening (`pt` → `pt-PT`, `en` → `en-GB`). Prefer widening on read
over a database migration — a stale bare code should resolve, not fail.

**Consequence for translation work:** a new locale is a whole catalog, not a diff. Machine-translate the
full file and have it reviewed, rather than hand-writing overrides against a base. For `pt-PT` and `pt-BR`
that means two complete passes, reviewed separately, because the whole point is that neither borrows from
the other.

### D3. RTL — prepare, do not ship

No right-to-left locale ships in this phase. But since 436 files are being opened anyway, they get
opened once:

- **CSS logical properties** as components are touched — `margin-inline-start`, `padding-inline`,
  `text-align: start` — instead of `left`/`right`. Tailwind has `ms-`/`me-`/`ps-`/`pe-` for this.
- **`dir` on `<html>`**, driven by the locale, wired now and set to `ltr` for every shipped locale.

The cost during a pass we are already making is close to nothing. The cost of a second pass over 436
files later is not. Hebrew stays in the picker but is not selectable until a catalog exists (see the
"twenty offered, two that exist" defect above — fix it by filtering the picker to shipped locales).

### D4. Extend the existing package. Do not adopt a library

`react-i18next`, `lingui` and `formatjs` are all better than what is here, and all of them are
dependencies. This repository froze its lockfile and hand-rolled a secret store rather than take
one, for reasons written down in `CLAUDE.md` → "Installs are frozen".

The gap is smaller than it looks, because the hard parts are in the runtime already: `Intl.PluralRules`,
`Intl.DateTimeFormat`, `Intl.NumberFormat`, `Intl.RelativeTimeFormat`. What is missing is roughly 150
lines of glue, listed in "What it cannot do yet".

Revisit this only if message _formatting_ gets genuinely complex — nested selects, gendered forms.
ICU MessageFormat is a real specification and reimplementing it would be the wrong kind of ambition.

### D5. Keys, not English source text

`t('settings.profile.changePassword.title')`, never `t('Change Password')`.

Source-text keys (the gettext and lingui style) read nicely and mean an untranslated string still
renders. They also mean **editing an English word invalidates every translation of it**, and this
codebase already has typed keys — a wrong key is a compile error today, and that is not worth
trading away.

### D6. Catalogs live per workspace, merged at the bridge

Not one god file. Three:

```
src/apps/officer-web/locales/          →  namespace  app.*
src/workspaces/officerdev/src/locales/ →  namespace  panel.*
src/workspaces/components/locales/     →  namespace  ui.*
```

`I18nBridge` deep-merges them before handing the map to the provider. Each workspace exports its own
typed `useTranslation` the way `officer-web/lib/i18n.ts` does today.

Two reasons, and the second is the one that matters for execution. Strings belong next to the code
that renders them, so a panel stays self-contained. And **one 900-key `en.json` is a merge conflict
on every parallel edit** — with several agents or several days of work, a single file is a queue.

Namespaces also make an unqualified key impossible to place wrongly: `ui.*` is shared, so anything
under it must not mention a feature.

### D7. Plurals and formatting go through `Intl`

A plural value becomes an object; a string stays a string:

```jsonc
"panel.files.selected": { "one": "{{count}} file selected", "other": "{{count}} files selected" }
```

`t()` picks the branch with `new Intl.PluralRules(locale).select(count)` when `count` is passed.
Categories are per-locale — English and Portuguese use `one`/`other`, Arabic uses six — so the shape
must be an open record, never a two-field type.

A `useFormat()` hook returns `date`, `time`, `dateTime`, `number` and `relative`, each bound to the
active locale. **The ~23 `toLocale*` call sites move to it.** Left alone they follow the browser's
locale and quietly contradict the rest of the UI.

### D8. A missing key falls back to English and is loud in dev

Never render a dot-path to a person. Resolution walks the chain, ends at `en`, and only if English is
missing too does it render the key — which at that point is a bug the parity test below should have
caught.

In dev, a miss logs once per key with the key and the locale. In production it is silent.

---

## The key naming convention

Three segments, sometimes four. **Name the role, never the text.**

```
<namespace>.<area>.<component>.<element>

app.settings.profile.changePassword.title
panel.files.contextMenu.rename
panel.chat.composer.placeholder
ui.button.cancel
app.errors.MINESWEEPER_NOT_RUNNING
```

- `namespace` — `app` / `panel` / `ui`, per D6.
- `area` — the route or panel it belongs to; matches the directory.
- `component` — the file, lowercased.
- `element` — what it _is_: `title`, `description`, `placeholder`, `confirmLabel`, `emptyState`.

Rules that keep 1200 keys navigable:

- **Never key by the English words.** `ui.button.save`, not `ui.button.saveText` and never
  `ui.button.Save`.
- **Never build a key at runtime.** `t(\`panel.${kind}.title\`)` defeats both the type checker and
  the extractor. Use an explicit map from the dynamic value to literal keys.
- **`ui.*` may not name a feature.** If a shared component needs feature words, the caller passes
  them as a prop.
- **One key per meaning, not per occurrence.** "Cancel" in twelve dialogs is `ui.button.cancel` once.
  But a word that is the same in English and different elsewhere — "Post" the noun and the verb —
  gets two keys. When unsure, split: merging later is safe, splitting later is not.

---

## What must never be translated

Getting this wrong is worse than missing a string, because it corrupts data or breaks a contract.

- **Agent and chat output.** Whatever Claude or opencode wrote. It is content.
- **User content** — file names, paths, terminal output, code, commit messages, note bodies.
- **A label quoted from ANOTHER product's interface.** "Click **Enable**", "set **App name**", "choose
  **Internal**" in a setup guide. These read like prose and are identifiers: their whole job is to match
  what the reader will actually see on Google's screen, and Google localises by the reader's _Google_
  account language, not by ours. Translating them sends somebody hunting for "Ativar" in a console that
  says "Enable".

  So translate the connective prose around them and leave the quoted label in English. Do not key it with
  identical values in both catalogs either — that is a key whose two sides can never legitimately differ,
  which is a maintenance cost bought for nothing.

  Decided 2026-08-23 on `GoogleOAuthConfig.tsx`, where 55 of the file's 67 strings are of this kind.

- **Identifiers that cross a boundary** — route paths, permission keys (`registry.ts`), error _codes_
  (the code is the key; its message is what gets translated), app ids, sidecar names, PM2 process
  names, database values.
- **Log lines and thrown `Error` messages** that are not rendered to a user. They are for whoever is
  reading a terminal at 4am, and they should stay greppable in one language.
- **"Officer"**, and product names generally.

---

## Execution

Four phases. Each is a commit that leaves the tree green — the app must run in English at every point,
because a half-converted UI is unreviewable.

### Phase 0 — infrastructure, no visible change

Everything in "What it cannot do yet". Nothing in this phase touches a component.

1. `resolve.ts`: fallback chain (D2), plural selection (D7), dev-mode miss warning (D8).
2. `I18nProvider`: accept a merged map; expose `useFormat()`.
3. `I18nBridge`: deep-merge the three catalogs (D6); set `dir` on `<html>` (D3).
4. `Languages.tsx`: regional tags, and **filter the picker to locales that actually exist**.
5. Settings migration: widen stored bare codes on read.
6. Scaffold `locales/` in `officerdev` and `components`, each with its typed hook.
7. Land the tooling below.

Phase 0 is the whole plan's risk. It is also small, and it is the only part where a wrong decision is
expensive to reverse — review it before phase 1 starts.

### Phase 1 — shared components (91 files, ~47 strings)

Smallest, highest reuse, and it establishes `ui.*` before anything depends on it. Do this first so
phases 2 and 3 can reach for `ui.button.cancel` instead of inventing it twice.

### Phase 2 — the shell (132 files, ~556 strings)

`officer-web`: settings, dock, header, screens. Highest-traffic text and the part a reviewer will
look at first. Convert per directory, one commit per area.

### Phase 3 — the panels (213 files, ~567 strings)

`officerdev` apps. The largest surface and the most parallelizable, because panels barely share
strings. **One panel per commit** — this is where per-workspace catalogs earn their keep, but two
agents in the same panel will still conflict.

### Then: `pt-PT`

Only after English is fully keyed. Translating against a moving catalog is wasted work.

---

## Tooling and guardrails

Without these, 1200 keys rot. All three are small and belong in phase 0.

**`scripts/i18n/scan-hardcoded.ts`** — already written and committed. Reports every candidate
hardcoded string by area and file. It is how the numbers above were produced, and it is the progress
meter: the count goes down as work lands.

**A key-parity test — and per D2 this is now load-bearing, not a nicety.** For each shipped locale, every
key that `en` has must be present IN THAT LOCALE'S OWN FILE. Fails the build on a gap, the same way
`bun docs:check` fails on a stale number.

With full catalogs and no overlays, this test is the only thing standing between "complete translation" and
"silently English in patches". It replaces the fallback that used to hide a gap, so it has to be as
unavoidable as the typecheck: a missing key is not a warning, it is a red build.

**A no-new-hardcoded-strings check.** Once an area reaches zero, add it to an allow-list the scanner
enforces, so the next feature cannot reintroduce debt. Add this per area as it completes — enabling
it globally on day one would just be red forever.

**Also worth having:** an unused-key report. Keys outlive the components that used them, and nothing
else will ever notice.

---

## Verification

- `bunx tsgo` clean. A wrong key is a type error, so this is the primary check.
  **Note for this machine:** `earlyoom` (`-r 3600`) SIGTERMs `tsgo` under memory pressure — it dies
  around 15s with `Command failed`. That is not a type error. Free memory or scope the check.
- `bun test` — parity test green for every shipped locale.
- `bun scripts/i18n/scan-hardcoded.ts` — count strictly decreasing per phase.
- By hand: switch the locale to `pt-PT` and walk the converted area. Untranslated strings should read
  as **English, never as dot-paths**.
- Check a plural at 0, 1 and 2, and a date, in both locales. These are the two things that look fine
  in English and wrong everywhere else.

---

## Open questions

1. **Routes.** Deferred deliberately. Nothing here depends on it, and localized URLs bring
   canonicalization and redirect questions that are not worth answering yet.
2. **Who translates?** Machine translation for `pt-PT` and `pt-BR` reviewed by the owner is the
   obvious first pass. Any locale beyond Portuguese needs an answer to this before it is offered.
3. **Plugin frontends** are out of scope and will need a catalog-merging contract in the plugin
   manifest. Their strings are not counted above.

   **One constraint that contract has to answer, found while converting the shared components
   (2026-08-23):** those components now call `useTranslation`, so they resolve against whatever
   `I18nProvider` is above them — and with none above them they render their own dot-paths. Today that
   is safe by accident: there are exactly two React roots, and the one that is not `officer-web`
   (`apps/marketplace`) imports nothing from `components/`. A plugin frontend mounting its own root, or
   rendering a shared component outside `I18nBridge`, would put `ui.button.cancel` in front of a person.

   So the contract needs to say where a plugin's provider comes from — inherit the host's, or mount its
   own seeded with the host's catalogs — rather than leaving each plugin to remember. Note that a
   provider-absent fallback inside the `components` package was considered and rejected: it would have
   split resolution across two mechanisms and contradicted D6, which puts the merge at the bridge.

4. **The `translateTo` setting already in `settings.languages`** is about translating _content_, not
   the UI. It is a different feature that happens to sit in the same object — do not conflate them.
5. **Does a member see the owner's locale anywhere?** Any server-rendered or cached string shared
   between accounts needs a locale decision. D1 mostly removes this by removing server prose.