Swift Missing Translations

swift-missing-translations

End-to-end loop: audit → translate → wire up Swift-side leaks → verify. Use when the user wants to localize an iOS app or hunt down strings that haven't been translated yet.

Step 0 — Reach shared understanding

Before running anything, derive what you can from the repo, then ask only what's left in a single batched message. Don't assume the source language.

Read first:

• find . -name "*.xcstrings" → catalog paths.
• For each catalog: sourceLanguage field, set of locales used (union of localizations keys).
• UI source roots — typical: Features/, Views/, App/, Components/, Screens/.
• In-app language picker? grep -rn "selectedLanguage\|@AppStorage.*lang" and look for \.locale injected at the root view. Determines whether references/wire-up.md §d applies.
• Sample 3 translated keys with placeholders to learn existing tone and the project's placeholder convention.

Ask only when not derivable:

• Tone and audience (kid-friendly second-person, professional B2B, etc.). Critical, not in the repo.
• Adding a brand-new locale? Only if the user named one not in the catalog.
• Dev/TestFlight quality vs App Store quality (the latter wants a native-speaker review pass after).

If the user invoked the skill bare, summarise what you found and what you intend to do, then wait for go-ahead.

Step 1 — Audit

Run the bundled script:

python3 scripts/audit.py <project-root>

It outputs JSON with coverage (per-language counts) and literals (UI literals not in the catalog, with the LocalizedStringKey shape Swift would build and an auto_localizes flag).

Render coverage as one markdown table (don't reprint per phase). Group literals by file. Surface only the highlights to the user — full output is for your reasoning, not the chat log.

Step 2 — Translate the gaps

For each missing-language cell, generate a translation. Hard rules:

• Preserve %@, %lld, %.1f, %@%% placeholders in position, same count, same order. Validate before writing — reject any translation where the multiset of placeholder tokens differs from the source.
• Identity-pass for symbol-only / format-only sources (%@, =, →, ·, ×, ✓, brand names).
• Plurals: if the source has separate forms (1 X and %lld Xs), translate each separately.
• Tone: match what the user set at session start.

Write back to the xcstrings JSON: state: "translated", preserve Xcode's pretty-print (2-space indent, alphabetical key order, trailing newline).

For >100-cell fills, checkpoint per language: write en, build, test compiled .strings, then de, etc. So a placeholder bug doesn't pollute every language at once.

Show 5 sample rows × all target languages before bulk-writing, so the user can sanity-check tone.

Step 3 — Swift-side wire-up

The catalog can be 100% covered and the UI still ships in the source language. Read references/wire-up.md when the audit's literals list is non-empty or when the user reports a string still rendering in the source language despite a translated catalog.

Quick index of what's in there — load the file when one of these matches:

• String-typed view properties → §a
• String enum returning UI labels → §b (watch for Hashable + .uppercased() traps)
• runtime-built String passed to Text(verbatim:) / AttributedString → §c
• in-app language picker disagreeing with device locale → §d (LocaleBundle helper)
• AppShortcuts.xcstrings for Siri phrases → §e
• AppEnum for Int parameters in shortcut phrases → §f
• legacy {name} placeholders from a JS i18n import → §g

Step 4 — Verify

1. Re-run audit.py. Coverage 100%, literals list empty (or only intentional cases — flag those).
2. xcodebuild build for the simulator. Catalog edits may not show until a clean build rewrites <lang>.lproj/Localizable.strings.
3. Spot-check compiled output for one or two new keys per locale:

   plutil -p <DerivedData>/.../<App>.app/de.lproj/Localizable.strings | grep "key fragment"
   

4. Install on a sim, kill the running app first (xcrun simctl terminate <udid> <bundle-id>), relaunch, walk the main flows. Running apps don't pick up new bundle resources.
5. Run the test suite. Localization shouldn't change behaviour; if a test fails it's likely a hard-coded source-language assertion that needs the same localization treatment.

Output style

• One coverage table per request, not per phase.
• For Swift-side fixes, show a unified diff per file (apply via Edit), not the whole file.
• Commit only when the user says so. Default message: how many keys filled across how many languages, which Swift files were touched, build/test result.
• AI translations are dev/TestFlight quality. Tell the user upfront if they're targeting App Store submission — recommend a native-speaker review pass after.

When to skip

• Project uses legacy .strings files (pre-iOS 17). Flag and offer to migrate; don't auto-edit.
• Repo has CI that auto-extracts catalog keys (Xcode build phase running xcstringstool sync). Work within that flow, don't manually edit auto-extracted keys.