本文へ移動
cccskills
無料GitHub で公開

i18n-parity

MANDATORY for every coding agent and contributor touching localized content — keep all five localization surfaces (dashboard UI keys, wiki page, full mirrored guides, locale-aware formatting, language switchers) in parity across every supported language. Use automatically (without being asked) whenever you add or change user-visible UI copy, add an i18n key, edit README-EN.md or a localized guide, edit wiki/index.html, or add a language surface.

インストール方法を見る

含まれるファイル(5)

  • SKILL.md13.3 KB
  • references/new-language-checklist.md10.3 KB
  • references/translation-style.md6.3 KB
  • scripts/i18n-audit.sh15.0 KB
  • scripts/sync-agent-mirrors.sh5.4 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

i18n Parity

This repository is localized on five independent surfaces, each with its own mechanism. A change that lands on one and not the others leaves the product half-translated: the English falls through, and the gap is invisible to anyone working in English. This skill states the invariant, maps every surface to the exact files, and gives one command that proves parity.

The invariant: English is the source of truth on every surface, and no change is done until every supported language carries it in the same PR. A fallback to English is a safety net, never a completed translation.

Supported languages are declared in one place — supportedLngs in client/src/i18n/index.ts. Today: en, zh, vi, ko, es. Everything below derives from that list; the audit script reads it rather than hard-coding.

The five surfaces

#SurfaceEnglish source of truthTranslations live inAutomated gate
1Dashboard UIclient/src/i18n/locales/en/*.jsonclient/src/i18n/locales/<xx>/*.json (same 15 namespaces)client/src/i18n/__tests__/i18n.test.ts — key, type, and interpolation-token parity
2Wiki pageEnglish text in the wiki/index.html DOMwiki/script.js (T, ATTRIBUTE_TRANSLATIONS, META, languageLabels, the two language ladders) + wiki/i18n-content.js (window.__WIKI_CONTENT_I18N, both the body bundles and plain)client/tests/wiki-i18n.test.ts — live-DOM prose coverage, inline-tag preservation, block-length budgets, asset-version sync
3Full mirrored guidesREADME-EN.mdREADME-CN.md (zh), README-VN.md (vi), README-KO.md (ko), README-ES.md (es)partial — scripts/i18n-audit.sh (existence, heading count, cross-links) and server/__tests__/plugins-marketplace.test.js (documented counts); prose parity is review-only
4Language switchers—client/src/components/Sidebar.tsx, client/src/lib/paletteCommands.ts, the two .lang-select-menu blocks in wiki/index.html, nav.json languageNames / languageShortscripts/i18n-audit.sh
5Locale-aware formatting—client/src/lib/format.ts (SupportedLanguage union, getCurrentLanguage() whitelist, getCurrentLocale() BCP-47 map)client/src/lib/__tests__/format.test.ts

What is deliberately NOT localized

Verified against the tree — do not go looking for translation hooks in these, and do not add them without being asked:

  • The root landing page index.html has no i18n layer. Do not add data-lang markup; put localized long-form content in the wiki instead. Its one language-aware element is the Languages (en/zh/…) stat label, which just enumerates the codes.
  • The repository landing page README.md is a concise English-only entry point. It links every complete language guide but is not itself mirrored. Full technical content belongs in README-EN.md and its four translations.
  • client/index.html is an English shell: <html lang="en">, og:locale=en_US, and English <title>/meta. The React app never reassigns document.documentElement.lang when the user switches language — a known gap, not something a translation PR is expected to fix.
  • The CLI (bin/ccam.js + cli/), the MCP server (mcp/), the Express server (server/), the desktop shell (desktop/), the VS Code extension, and the statusline contain no i18n wiring at all. Their output is English.
  • client/src/lib/event-summary.ts and event-grouping.ts build tool-event headlines and bullets from English template literals (Last message: …, 3 lines stdout, 2 matches). Their MODULE_GUIDE boilerplate claims strings belong in i18n JSON, but neither file imports i18next. They sit outside the key system today; do not "fix" them as part of a localization change.
  • Number and date formatting is only locale-aware where the format.ts helpers are used. Most components call toLocaleString() / toLocaleString(undefined, …) directly, which follows the browser locale rather than the chosen UI language. In new code prefer getCurrentLocale() from format.ts; leave existing call sites alone unless asked.

Workflow A — you changed content (the common case)

Find what you touched in the left column and ship everything in the right column in the same PR.

You changedYou must also do
Added/renamed a UI string or i18n keyAdd the key to en and every other locale in the same namespace file. Same key path, same value type, same {{interpolation}} tokens.
Added a new namespace (new *.json)Create it for every locale, then register the imports, the resources entry per language, and the ns array in client/src/i18n/index.ts.
Added user-visible wiki text in wiki/index.htmlFollow .claude/rules/wiki-i18n.md: scannable layer (the PLAIN selector set — .logo-sub, .section-label, .nav-section, .nav-empty, .stat-label, .t-label, h2/h3/h4, th, .hero-desc, plus .nav-link / .hero-badge trailing text nodes) → T in wiki/script.js; body prose (the HTML_SEL set — p, li, td, th, captions, .callout-body > strong, .route-desc, footer) → wiki/i18n-content.js keyed by whitespace-normalized innerHTML; new alt/aria-label/title/placeholder → ATTRIBUTE_TRANSLATIONS. Then bump CACHE_NAME in wiki/sw.js and the matching ?v= query strings.
Edited the English-only landing README.mdKeep it concise and ensure it still links README-EN.md and every localized full guide. Do not copy landing-page prose into the full guides.
Edited a section of README-EN.mdMirror the same edit at the corresponding section of README-CN.md, README-VN.md, README-KO.md, and README-ES.md. All four, every time.
Changed behavior that the README/wiki document (env var, event type, route, CLI command, feature)Run the update-project-docs skill — it owns the change→docs mapping — then come back here for the translation propagation it triggers.
Changed a documented count (plugins, skills, namespaces, languages)The count can appear in the landing README, all five full guides, ARCHITECTURE.md, docs/*.md, index.html, wiki/index.html, wiki/i18n-content.js, and server/__tests__/plugins-marketplace.test.js. Grep the old number repo-wide; update every hit.

Workflow B — adding a new language

This is a large, exact, mechanical change. Read references/new-language-checklist.md and work through it top to bottom — it lists every file, in dependency order, with the exact edit for each.

The three things contributors most often ship incomplete, stated up front:

  1. The full-guide mirror must be complete. README-<XX>.md is a full mirror of README-EN.md — every section, every table row, every code block, every mermaid diagram, in the same order. Do not summarize, do not drop "less important" sections, do not stop halfway. Diff the heading list against README-EN.md before you open the PR — the audit script compares heading counts, which catches a truncated mirror but not a reordered or silently condensed one.
  2. Every app key must be translated. All 15 namespaces × every key. The parity test fails on a missing key, but it passes on a key you copied over in English — so read your diff for untranslated leftovers.
  3. The wiki must be translated completely, not just the headings. That means a full <xx> bundle in wiki/i18n-content.js (thousands of body strings), a full <xx> block in T, META, and every entry of ATTRIBUTE_TRANSLATIONS in wiki/script.js. A locale that only fills T renders a page with translated headings over English paragraphs, which is worse than English.

What stays in English (all surfaces)

Never translate: code inside <code>/backticks, commands, file and directory paths, URLs, env-var names, HTTP methods and status codes, CLI flags, code identifiers, numbers with units, brand and product names (Claude Code, MCP, Codex), Claude Code hook event names (PreToolUse, Stop, …), and Claude Code tool names (Bash, Agent, Read, Edit). Translate only the prose around them. A block that is entirely code or identifiers needs no wiki entry — it correctly falls back to English.

Two different things are spelled Agent. The Claude Code tool named Agent (alongside Bash, Read, Edit) is an identifier and stays literal in every locale, everywhere it appears — hook-event tables, tool lists, event names. The UI noun for an agent (common:agent / common:subagent) is product vocabulary with a per-locale contract, and only that noun has the Spanish exception below. Never carry the exception into the tool name.

The UI noun's contract, asserted by client/src/i18n/__tests__/i18n.test.ts: zh, vi, and ko keep Agent / Subagent literal; es renders them agente / subagente. A new locale must decide this explicitly and add its row to that test.

Terminology, per-locale conventions, and the shared glossary live in references/translation-style.md. Read it before translating; drifting terminology across pages is the most common review comment on localization PRs.

Two traps worth knowing before you edit the wiki

  1. client/tests/wiki-i18n.test.ts parses wiki/script.js by exact source markers. It slices the file between literal strings — " const T = " … "\n\n const PLAIN", " const ATTRIBUTE_TRANSLATIONS = " … "\n const ATTR", " const META = " … "\n const trH" — and evals what it finds. Renaming, reordering, or re-indenting those declarations breaks the test with a confusing error. Add locales inside the existing objects; leave the declaration lines alone.
  2. The wiki's chosen language lives in localStorage["wiki-lang"], not a URL parameter. There is no ?lang= support. First visit falls back to a navigator.language prefix ladder in wiki/script.js. To preview a locale, use the switcher, or run localStorage.setItem("wiki-lang", "<xx>") in the console and reload.

This skill is mirrored for every agent

The canonical copy — and both scripts — live at .claude/skills/i18n-parity/. It is mirrored, with links rewritten to repo-root-relative paths and an agents/openai.yaml interface added, to:

  • .agents/skills/i18n-parity/ (the shared/OpenAI skill tree)
  • .codex/skills/i18n-parity/ (Codex)

Editing this skill means regenerating the mirrors — edit the canonical copy, then run:

bash .claude/skills/i18n-parity/scripts/sync-agent-mirrors.sh

i18n-audit.sh runs that script in --check mode, so a stale mirror is a reported gap rather than silent drift. The mirrors carry no scripts: they point back at the canonical ones by repo-root path.

Verify (do not skip)

# 1. Cross-surface parity: locale sets, namespace files, key parity, switcher
#    entries, wiki bundles, full-guide mirrors and cross-links, agent-skill mirrors.
bash .claude/skills/i18n-parity/scripts/i18n-audit.sh

# 2. UI key/type/interpolation parity + locale formatting
npm run test:client

# 3. Wiki live-DOM coverage, inline tags, metadata, cache versions
cd client && npx vitest run tests/wiki-i18n.test.ts && cd ..

# 4. Documented counts asserted against the source tree
npm run test:server

# 5. The static wiki files and locale JSON are Prettier-managed
npm run format

i18n-audit.sh exits non-zero and names the exact file and locale for every gap. Every check is a structural one — a named thing is present or it is not — with a single exception: the wiki/i18n-content.js bundle-size check is a stub detector (it flags a locale holding under 60% of the largest bundle's entries). Exact per-string wiki coverage is asserted by client/tests/wiki-i18n.test.ts, which walks the live DOM.

State explicitly which surfaces you updated and which you intentionally skipped (with the reason), per the repo's verification policy in CLAUDE.md.

Tips

  • Write the full English guide first and get it right, on all surfaces, before translating anything. Every other locale is derived from it; re-translating because the English moved is the biggest waste in this workflow.
  • To find where a string already lives: grep -rn "<neighbouring English text>" client/src/i18n/locales/en wiki/i18n-content.js wiki/script.js.
  • When adding a language, one locale per subagent is fine for the wiki body bundle (it is large), but keep the full-guide mirror with a single author so the section order and terminology stay coherent.
  • Wiki edits are cache-first: forgetting the CACHE_NAME / ?v= bump means returning visitors never see the translation you just shipped.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Inspect fired CCAM alerts and manage alert rules for token thresholds, event patterns, inactivity, and status duration. Use when acknowledging alerts, creating or editing a rule, checking cooldowns, or connecting alert rules to webhook targets.

日本語の概要は準備中です。原文の説明を表示しています。

hoangsonww/Claude-Code-Agent-Monitor1,0692026年10月12日 更新

Identify anomalous sessions using Agent Monitor data — cost outliers from the pricing engine, token anomalies (cache miss spikes, compaction baseline surges), unusual event type ratios (PreToolUse/PostToolUse gaps, APIError clusters), behavioral deviations from workflow intelligence (complexity score outliers, error propagation anomalies), and sessions with abnormal metadata (extreme turn_count, high thinking_blocks, zero turn_duration).

日本語の概要は準備中です。原文の説明を表示しています。

hoangsonww/Claude-Code-Agent-Monitor1,0692026年10月12日 更新

Produce a detailed report on APIError events from Agent Monitor data — counts over time, which sessions and models are affected, and the likely root cause (rate limits, overload/529, or context-window pressure) inferred from each event's summary and data payload. Use when API errors spike or when you need to explain why requests are failing.

日本語の概要は準備中です。原文の説明を表示しています。

hoangsonww/Claude-Code-Agent-Monitor1,0692026年10月12日 更新

benchmark

無料

Benchmark one session (or a small recent set) against the rolling average using Agent Monitor data — cost, total tokens, tool count, and workflow complexity score — and report where each metric lands as a percentile of the population. Tells you whether a session was normal, cheap, or an outlier. Use when judging whether a session was typical or out of band.

日本語の概要は準備中です。原文の説明を表示しています。

hoangsonww/Claude-Code-Agent-Monitor1,0692026年10月12日 更新

Define a spend budget for Claude Code and, optionally, create a cost alert rule that fires when usage crosses the limit, via POST /api/alerts/rules on the Agent Monitor dashboard. Reads current spend from /api/pricing/cost to size the budget sensibly and explains every rule field before writing. Use when setting a spend limit or wiring up a budget guardrail.

日本語の概要は準備中です。原文の説明を表示しています。

hoangsonww/Claude-Code-Agent-Monitor1,0692026年10月12日 更新

Analyze prompt-cache effectiveness for Claude Code usage from the Agent Monitor dashboard — cache hit rate (total_cache_read / (total_cache_read + total_input)), cache_write vs cache_read reuse, cache-read vs cache-write spend, and the sessions with the poorest reuse. Pulls token totals from /api/analytics, per-session detail from /api/sessions, and dollar splits from /api/pricing/cost. Use when diagnosing cache spend or deciding whether prompt caching is paying off.

日本語の概要は準備中です。原文の説明を表示しています。

hoangsonww/Claude-Code-Agent-Monitor1,0692026年10月12日 更新

hoangsonww のスキルをすべて見る

このスキルの問題を報告する