dowafu 0.1.0 → 0.3.0

package/README.md

@@ -1,6 +1,10 @@
 # dowafu
 
-Send a section of your design doc to external LLMs for review, from the terminal.
+**English** ｜ [繁體中文](https://github.com/eyesofkids/dowafu/blob/main/README_zh-tw.md)
+
+A read-only review harness: it sends a section of a document to external models, decides
+what each of them may read, records what they did, and audits the shape of what they
+returned.
 
 You write a ticket. `dowafu` calls each provider's API. Each reviewer — a *spoke* —
 reads only the files you whitelisted, and returns observations with the evidence it
@@ -8,16 +12,24 @@ read them from. Everything lands on disk for you to check.
 
 **Spokes produce observations, not verdicts.** What to do about them stays with you.
 
-> ### Tickets work in English or Chinese. The CLI's own output is Chinese.
+> ### English and Traditional Chinese are both fully supported.
+>
+> One flag decides the language for the whole run: `--lang en` or `--lang zh-tw`. Without
+> it, `DISPATCH_LANG` applies; without that either, the default is **English**. The flag
+> wins over the environment variable, and an unrecognized value in either is rejected
+> rather than guessed at.
 >
-> Write the ticket's section headings in either language and the rest follows: the
-> reviewer is prompted in that language, asked for its report in that language's
-> template, and the audit checks it against the matching one. The language is decided by
-> the headings themselves — there is no flag to forget — and the dry run prints which one
-> it resolved to, per reviewer.
+> The language reaches everything: the reviewer's prompt and report template, the audit
+> that checks the report against it, and the CLI's own output — `--help`, error messages,
+> the dry-run report, `summary.md`. The dry run prints the resolved language per reviewer,
+> so you can see it before anything is sent.
 >
-> What is still Traditional Chinese: `--help`, error messages, the dry-run report and
-> `summary.md`. That is the development language and it has not been translated.
+> A ticket's section headings may be written in either language regardless — they are
+> field names, not a language switch. See [Ticket format](#ticket-format).
+>
+> The skills and reviewer definitions come in both languages too, under `publish/en/` and
+> `publish/zh-tw/`. **Install one or the other, never a mix** — a reviewer's closing line
+> has to match the template the audit checks it against.
 
 ## Install
 
@@ -29,26 +41,27 @@ The command is `dowafu`.
 
 ## API keys
 
-Keys are read from `$DISPATCH_HOME/.env`, which defaults to `~/.config/dispatch/.env`
+Keys are read from `$DISPATCH_HOME/.env`, which defaults to `~/.config/dowafu/.env`
 (`DISPATCH_HOME` or `XDG_CONFIG_HOME` override it). Variables already present in the
 environment win over the file, so CI and one-off overrides need no file at all.
 
 ```bash
-mkdir -p ~/.config/dispatch
-cat > ~/.config/dispatch/.env <<'EOF'
+mkdir -p ~/.config/dowafu
+cat > ~/.config/dowafu/.env <<'EOF'
 DEEPSEEK_API_KEY=
 GEMINI_API_KEY=
 OPENAI_API_KEY=
 ANTHROPIC_API_KEY=
 EOF
-chmod 600 ~/.config/dispatch/.env
+chmod 600 ~/.config/dowafu/.env
 ```
 
 Only the providers you actually dispatch to need a key. The file is plain text — it is
 protected by nothing but its file permissions.
 
-**The current directory's `.env` is never read.** That directory is the project under
-review, and its secrets have no business in a process that talks to three external APIs.
+**The current directory's `.env` is never read.** That is where you ran the command, and
+usually the project under review; its secrets have no business in a process that is
+talking to external APIs.
 
 ## Usage
 
@@ -101,9 +114,10 @@ parser matches** — use one of the two sets below, exactly as written.
 - prisma/schema.prisma
 ```
 
-The set you use decides the reviewer's language: `# Questions` gets an English prompt and
-an English report template, `# 具體問題` gets the Chinese ones. Mixing the two sets inside
-one reviewer's file is not supported — the first heading that matches wins.
+Either set is accepted, and the choice does **not** decide the reviewer's language —
+that comes from `--lang` / `DISPATCH_LANG` (see above). The two sets are aliases for the
+same fields, so an English ticket can run in Chinese and vice versa. Mixing both sets
+inside one reviewer's file is not supported — the first heading that matches wins.
 
 Reviewer definitions live in `.claude/agents/<agent>.md` under the repo root — they are
 the source of each spoke's system prompt, and the CLI reads them directly. Results are
@@ -117,7 +131,7 @@ requests and responses.
   refused, and the refusal is recorded.
 - **`_docs/` is off limits**, whatever the whitelist says.
 - **Nothing is billed before you confirm.** The dry run prints the resolved repo root,
-  models, token estimates and output path, and calls no API.
+  each reviewer's model and language, token estimates and the output path, and calls no API.
 - **Secrets are masked** in `run.jsonl`, `raw/*.json` and stdout.
 - **Failures stop the run.** A missing key, an unknown model, a file that does not
   exist — each aborts with the path or name that caused it, before any spend.
@@ -141,17 +155,22 @@ an agent working there knows how to write a ticket, what to check before spendin
 how to read the results. Copy both directories — `.claude/` holds the reviewer
 definitions the CLI itself reads, so it is required no matter which agent you use.
 
+It ships in both languages, `publish/en/` and `publish/zh-tw/`. **Pick one.** A mix does
+not work: a reviewer's fixed closing line has to match the template the audit checks it
+against.
+
 ```bash
 TARGET=<your project>
+SRC=publish/en                  # or publish/zh-tw
 mkdir -p "$TARGET/.claude/skills" "$TARGET/.claude/agents" "$TARGET/.agents/skills"
-cp -R publish/.claude/skills/. "$TARGET/.claude/skills/"
-cp -R publish/.agents/skills/. "$TARGET/.agents/skills/"
-cp publish/.claude/agents/*.md "$TARGET/.claude/agents/"
-cp publish/workflow_spec.md "$TARGET/"
+cp -R "$SRC/.claude/skills/." "$TARGET/.claude/skills/"
+cp -R "$SRC/.agents/skills/." "$TARGET/.agents/skills/"
+cp "$SRC"/.claude/agents/*.md "$TARGET/.claude/agents/"
+cp "$SRC/workflow_spec.md" "$TARGET/"
 ```
 
-See `publish/README.md` for the details. Those documents are currently written in
-Traditional Chinese.
+See `publish/en/README.md` — or `publish/zh-tw/README.md` — for the details, each written
+in its own language.
 
 ## License
 

package/README_zh-tw.md

@@ -0,0 +1,197 @@
+# dowafu
+
+[English](https://github.com/eyesofkids/dowafu/blob/main/README.md) ｜ **繁體中文**
+
+**一支唯讀審查用的 spoke harness**：把文件段落派給外部模型，限定它讀什麼、記下它做了什麼、機械稽核它的回報格式。
+
+講具體一點——你(與AI agent協同)寫一份工單，`dowafu` 呼叫各家 provider 的 API；每個審查者（*spoke*）**只讀你列進白名單的檔案**，回傳帶依據的觀察，全部落檔供你查核。
+
+相容於各種支援標準SKILL的AI Agent程式與VS Code擴充套件包含以下，但還有更多:
+
+- Claude Code
+- VS Code Codepilot 
+
+**spoke 產出的是意見，不是裁決。** 那些意見要怎麼處理，仍然由你決定。
+
+> ### 英文與繁體中文都是完整支援的語言。
+>
+> 整次執行的語言由一個旗標決定：`--lang en` 或 `--lang zh-tw`。沒帶旗標時看 `DISPATCH_LANG`，兩者都沒有時**預設為英文**。旗標優先於環境變數；任一方填了無法辨識的值都會直接中止，不會猜。
+>
+> 語言涵蓋所有地方：審查者收到的 Prompt 與報告範本、稽核用來檢查報告的範本，以及 CLI 自己的輸出——`--help`、錯誤訊息、dry-run 報告、`summary.md`。Dry run 會針對每個審查者印出判定的語言，送出前就看得到。
+>
+> 工單的段落標題兩種語言都可以寫，與語言選擇無關——它們是欄位名稱，不是語言開關。見[工單格式](#工單格式)。
+>
+> Skill 與審查者定義同樣有兩種語言，分別放在 `publish/en/` 與 `publish/zh-tw/`。**擇一安裝，不可混裝**——審查者的固定收尾句必須與稽核檢查用的範本是同一種語言。
+
+## 安裝
+
+```bash
+npm install -g dowafu
+```
+
+執行指令為 `dowafu`。
+
+## API Keys
+
+API Key 會從 `$DISPATCH_HOME/.env` 讀取，預設位置為：
+
+`~/.config/dowafu/.env`
+
+可以透過 `DISPATCH_HOME` 或 `XDG_CONFIG_HOME` 覆寫這個位置。
+
+如果環境變數中已經存在相同的變數，環境變數會優先於檔案中的值。因此在 CI 或臨時覆寫設定時，完全不需要建立 `.env` 檔案。
+
+```bash
+mkdir -p ~/.config/dowafu
+cat > ~/.config/dowafu/.env <<'EOF'
+DEEPSEEK_API_KEY=
+GEMINI_API_KEY=
+OPENAI_API_KEY=
+ANTHROPIC_API_KEY=
+EOF
+chmod 600 ~/.config/dowafu/.env
+```
+
+只需要為實際要執行 Dispatch 的 Provider 設定 Key 即可。
+
+這個檔案是純文字檔——它唯一的保護機制就是檔案權限。
+
+**永遠不會讀取目前工作目錄的 `.env`。**
+
+那個目錄是你下指令的地方，通常就是被審查的專案；它的 secrets 不該被一個正對外部 API 發請求的行程讀進去。
+
+## 使用方式
+
+```bash
+dowafu <ticket-dir> --dry-run   # 解析、驗證、估算。不呼叫 API，也不產生成本。
+dowafu <ticket-dir> --yes       # 執行審查。這個操作會產生成本。
+dowafu --help                   # 查看所有旗標
+```
+
+沒有指定 `--yes` 時，指令會要求使用者確認。
+
+當 stdin 不是 TTY 時——例如由 Agent 代替你執行——沒有人可以回答確認，因此程式會在呼叫任何 API 之前停止。
+
+## 工單
+
+工單是一個目錄，其中包含三種檔案。
+
+其中的標題是**字面標記（literal markers）**，Parser 會直接比對這些標記——必須完全使用以下兩組標題之一。
+
+| English | 中文 |
+| --- | --- |
+| `# Questions` | `# 具體問題` |
+| `# Allowed reads` | `# 允許讀取` |
+| `# Under review` | `# 待審段落` |
+| `# Premises` | `# 前提（不受審）` |
+
+| 檔案 | 內容 |
+| --- | --- |
+| `_dispatch.md` | 要執行哪些審查者，以及各自使用哪個 Provider 與 Model |
+| `_shared.md` | 前提與正在審查的段落，以原始內容直接貼入 |
+| `<agent>.md` | 每個審查者一份：包含它的問題，以及允許讀取的檔案 |
+
+```markdown
+<!-- _dispatch.md -->
+<!-- format: v1 -->
+# dispatch auth-review
+
+| agent | provider | model | effort |
+| --- | --- | --- | --- |
+| hole-finder-safety | deepseek | deepseek-v4-flash | |
+| hole-finder-feasibility | openai | gpt-5.6-luna | |
+```
+
+```markdown
+<!-- hole-finder-safety.md -->
+# Questions
+1. Does the permission check described here hold under concurrent requests?
+
+# Allowed reads
+- lib/auth-guard.ts
+- prisma/schema.prisma
+```
+
+兩組標題都收，而且**與審查者使用的語言無關**——語言由 `--lang`／`DISPATCH_LANG` 決定（見上方）。兩組只是同一批欄位的別名，所以英文工單可以用中文跑，反之亦然。
+
+不支援在同一個審查者檔案中混用兩組標題——第一個符合的標題會決定用哪一組欄位名，不是決定語言。
+
+審查者定義放在 Repository 根目錄下的：
+
+```text
+.claude/agents/<agent>.md
+```
+
+它們是每個 spoke 的 System Prompt 來源，CLI 會直接讀取這些檔案。
+
+結果會寫入：
+
+```text
+tmp/spoke/<ticket-id>/
+```
+
+其中包含：
+
+- 每個 spoke 的審查報告
+- `summary.md`：包含稽核表格與預估成本
+- `run.jsonl`：每個事件一行
+- `raw/`：精確保存實際送出的 Request 與收到的 Response
+
+## 工具保證的事項
+
+- **每次呼叫的讀取範圍都有白名單限制。** 如果 spoke 要求讀取白名單以外的檔案，會被拒絕，而且拒絕事件會被記錄。
+- **`_docs/` 永遠禁止讀取。** 不論白名單如何設定。
+- **在你確認之前不會產生任何費用。** Dry run 會列出解析後的 repo 根目錄、每個 spoke 的模型與語言、token 預估值與輸出路徑，而且不會呼叫任何 API。
+- **Secrets 會被遮罩。** `run.jsonl`、`raw/*.json` 以及 stdout 中都會遮罩 Secrets。
+- **發生錯誤就停止整個執行流程。** 例如缺少 API Key、未知的 Model、指定的檔案不存在等，都會在產生任何費用之前中止，並指出造成問題的路徑或名稱。
+
+## Models
+
+| Provider | Model |
+| --- | --- |
+| `openai` | `gpt-5.6-luna`、`gpt-5.6-terra`、`gpt-5.6-sol` |
+| `deepseek` | `deepseek-v4-flash` |
+| `gemini` | `gemini-3.1-flash-lite`、`gemini-3.5-flash-lite`、`gemini-3.6-flash` |
+| `anthropic` | `claude-opus-5`、`claude-sonnet-5` |
+
+這份清單會隨套件以 `providers.json` 一起提供。
+
+如果想使用其他 Model，可以透過 `--providers` 指定自己的 Provider 設定檔。
+
+## 從 Agent 驅動
+
+`publish/` 包含可以複製到專案中的 skill 與審查者定義，讓在該專案中工作的 Agent 知道：
+
+- 如何寫工單
+- 在產生成本之前要檢查什麼
+- 如何讀取審查結果
+
+請將兩個目錄都複製過去。
+
+其中 `.claude/` 包含 CLI 本身會讀取的審查者定義，因此無論你使用哪一種 Agent，都必須複製它。
+
+`publish/` 提供兩種語言：`publish/en/` 與 `publish/zh-tw/`。**擇一複製，不可混裝**——審查者的固定收尾句必須與稽核檢查用的範本是同一種語言。
+
+```bash
+TARGET=<your project>
+SRC=publish/zh-tw               # 或 publish/en
+
+mkdir -p "$TARGET/.claude/skills" "$TARGET/.claude/agents" "$TARGET/.agents/skills"
+
+cp -R "$SRC/.claude/skills/." "$TARGET/.claude/skills/"
+cp -R "$SRC/.agents/skills/." "$TARGET/.agents/skills/"
+cp "$SRC"/.claude/agents/*.md "$TARGET/.claude/agents/"
+cp "$SRC/workflow_spec.md" "$TARGET/"
+```
+
+詳細內容請參考該語言目錄下的 README：
+
+```text
+publish/zh-tw/README.md   （或 publish/en/README.md）
+```
+
+兩份各自以該語言撰寫。
+
+## License
+
+MIT

package/dist/adapters/anthropic-messages.js

@@ -21,21 +21,24 @@
 import { normalizeAnthropicUsage, normalizeFinishReason } from "../usage.js";
 import { ProviderHttpError } from "../mask.js";
 import { checkRawObjectIntegrity } from "../raw-integrity.js";
+import { readFileToolDescription } from "./read-file-tool-description.js";
 const ANTHROPIC_VERSION = "2023-06-01";
 // §29 規格四：max_tokens 是 Anthropic 專屬必要參數，SendOptions 沒有這個概念（openai／gemini
 // 都不需要）。寫死在 adapter，不提升到 SendOptions——那會逼另外兩個 adapter 處理用不到的欄位。
 // 初始值推導見規格四：十次派工單支最高 40,970 output token，32k 有 20% 餘裕；本版預設 effort
 // 不會用到官方建議 64k 起跳的 xhigh／max。撞到 stop_reason:"max_tokens" 再調，不在此臆測。
 const ANTHROPIC_MAX_TOKENS = 32768;
-const READ_FILE_TOOL = {
-    name: "read_file",
-    description: "讀取指定路徑的檔案內容",
-    input_schema: {
-        type: "object",
-        properties: { path: { type: "string" } },
-        required: ["path"],
-    },
-};
+function buildReadFileTool(lang) {
+    return {
+        name: "read_file",
+        description: readFileToolDescription(lang),
+        input_schema: {
+            type: "object",
+            properties: { path: { type: "string" } },
+            required: ["path"],
+        },
+    };
+}
 // §29 規格二：thinking.type:"enabled" 在 Opus 5／Sonnet 5 會 400（實測確認，見
 // facts_dispatch.md）。正確路徑是 adaptive thinking ＋ output_config.effort，兩者一起送——
 // 即使 output_config.effort 的 JSON 片段與 deepseek 的 reasoning.style 完全相同，這是巧合
@@ -45,6 +48,10 @@ function buildReasoningParams(style, effort) {
         return {}; // §4：留白則不送任何 reasoning 參數（validate.ts 已解析為 default，實務上不會是 undefined）
     if (style === "anthropic")
         return { thinking: { type: "adaptive" }, output_config: { effort } };
+    // T7a〈D. 全域中文字串最終覆核〉：providers.json 的 schema 已限定 reasoning.style 只能是
+    // "openai"／"deepseek"／"gemini"／"anthropic"／null（見 providers.ts 的 parseReasoning），
+    // 此分支理論上不可達，同 raw-integrity.ts 的實作缺陷斷言。明確決定維持中文不動，不搬進
+    // messages.ts——多語系對一段不可達的防禦性文字沒有實質效益。
     throw new Error(`anthropic-messages adapter 不支援 reasoning.style=${style}`);
 }
 // §29 規格五：把 conv.turns 轉成 Anthropic 的 messages 陣列。tool turn 需要前瞻合併——
@@ -82,7 +89,7 @@ function turnsToMessages(turns) {
 // 不需打 API 就能測試——見 anthropic-messages.test.ts）。
 export function buildAnthropicRequest(conv, opts, config) {
     const messages = turnsToMessages(conv.turns);
-    checkRawObjectIntegrity(conv, messages);
+    checkRawObjectIntegrity(conv, messages, config.lang);
     return {
         model: opts.model,
         max_tokens: ANTHROPIC_MAX_TOKENS,
@@ -91,7 +98,7 @@ export function buildAnthropicRequest(conv, opts, config) {
         // §29 規格八：top-level 自動快取，由 API 自行管理斷點位置並隨對話推進。低於該模型最低
         // 可快取 token 數時 API 靜默不快取、不報錯，故此欄位可無條件加，不需前置估算。
         cache_control: { type: "ephemeral" },
-        ...(opts.enableTools === false ? {} : { tools: [READ_FILE_TOOL] }),
+        ...(opts.enableTools === false ? {} : { tools: [buildReadFileTool(config.lang)] }),
         ...buildReasoningParams(config.reasoning.style, opts.effort),
     };
 }

package/dist/adapters/gemini-native.js

@@ -8,19 +8,22 @@
 import { normalizeFinishReason, normalizeGeminiUsage } from "../usage.js";
 import { ProviderHttpError } from "../mask.js";
 import { checkRawObjectIntegrity } from "../raw-integrity.js";
-const GEMINI_TOOL = {
-    functionDeclarations: [
-        {
-            name: "read_file",
-            description: "讀取指定路徑的檔案內容",
-            parameters: {
-                type: "object",
-                properties: { path: { type: "string" } },
-                required: ["path"],
+import { readFileToolDescription } from "./read-file-tool-description.js";
+function buildGeminiTool(lang) {
+    return {
+        functionDeclarations: [
+            {
+                name: "read_file",
+                description: readFileToolDescription(lang),
+                parameters: {
+                    type: "object",
+                    properties: { path: { type: "string" } },
+                    required: ["path"],
+                },
             },
-        },
-    ],
-};
+        ],
+    };
+}
 function turnToContent(turn) {
     if (turn.role === "user") {
         return { role: "user", parts: [{ text: turn.text }] };
@@ -52,13 +55,13 @@ function buildReasoningParams(effort) {
 }
 // 純函式：組出實際會送出的 request body，並執行 §8 的 raw 完整性自我檢查（違反即拋
 // RawIntegrityError，不需打 API 就能測試——見 raw-integrity.test.ts）。
-export function buildGeminiRequest(conv, opts) {
+export function buildGeminiRequest(conv, opts, lang) {
     const contents = conv.turns.map(turnToContent);
-    checkRawObjectIntegrity(conv, contents);
+    checkRawObjectIntegrity(conv, contents, lang);
     const body = {
         contents,
         system_instruction: { parts: [{ text: conv.systemPrompt }] },
-        ...(opts.enableTools === false ? {} : { tools: [GEMINI_TOOL] }),
+        ...(opts.enableTools === false ? {} : { tools: [buildGeminiTool(lang)] }),
         ...buildReasoningParams(opts.effort),
     };
     return { contents, body };
@@ -66,7 +69,7 @@ export function buildGeminiRequest(conv, opts) {
 export function createGeminiAdapter(config) {
     return {
         async send(conv, opts) {
-            const { body } = buildGeminiRequest(conv, opts);
+            const { body } = buildGeminiRequest(conv, opts, config.lang);
             const res = await fetch(`${config.baseURL}/models/${opts.model}:generateContent`, {
                 method: "POST",
                 headers: { "x-goog-api-key": config.apiKey, "Content-Type": "application/json" },

package/dist/adapters/read-file-tool-description.js

@@ -0,0 +1,5 @@
+const READ_FILE_TOOL_DESCRIPTION = "讀取指定路徑的檔案內容";
+const READ_FILE_TOOL_DESCRIPTION_EN = "Read the contents of the file at the given path.";
+export function readFileToolDescription(lang) {
+    return lang === "en" ? READ_FILE_TOOL_DESCRIPTION_EN : READ_FILE_TOOL_DESCRIPTION;
+}

package/dist/adapters/responses.js

@@ -6,17 +6,21 @@ import OpenAI from "openai";
 import { normalizeFinishReason, normalizeResponsesUsage } from "../usage.js";
 import { ProviderHttpError } from "../mask.js";
 import { checkRawArrayIntegrity } from "../raw-integrity.js";
-const READ_FILE_TOOL = {
-    type: "function",
-    name: "read_file",
-    description: "讀取指定路徑的檔案內容",
-    parameters: {
-        type: "object",
-        properties: { path: { type: "string" } },
-        required: ["path"],
-    },
-    strict: false,
-};
+import { readFileToolDescription } from "./read-file-tool-description.js";
+import { m } from "../messages.js";
+function buildReadFileTool(lang) {
+    return {
+        type: "function",
+        name: "read_file",
+        description: readFileToolDescription(lang),
+        parameters: {
+            type: "object",
+            properties: { path: { type: "string" } },
+            required: ["path"],
+        },
+        strict: false,
+    };
+}
 // plan_dispatch_v1.8.md §5「reasoning.style 的三種轉換」（實測確認，v1.7 兩處填錯已修正）：
 // deepseek 正確路徑是 output_config.effort，不是 chat/completions 章節的
 // thinking/reasoning_effort——後者對 /v1/responses 端點回顯欄位但靜默不生效
@@ -28,6 +32,10 @@ function buildReasoningParams(style, effort) {
         return { reasoning: { effort } };
     if (style === "deepseek")
         return { output_config: { effort } };
+    // T7a〈D. 全域中文字串最終覆核〉：providers.json 的 schema 已限定 reasoning.style 只能是
+    // "openai"／"deepseek"／"gemini"／"anthropic"／null（見 providers.ts 的 parseReasoning），
+    // 此分支理論上不可達，同 raw-integrity.ts 的實作缺陷斷言。明確決定維持中文不動，不搬進
+    // messages.ts——多語系對一段不可達的防禦性文字沒有實質效益。
     throw new Error(`responses adapter 不支援 reasoning.style=${style}`);
 }
 function turnToInputItems(turn) {
@@ -44,12 +52,12 @@ function turnToInputItems(turn) {
 // RawIntegrityError，不需打 API 就能測試——見 raw-integrity.test.ts）。
 export function buildResponsesRequest(conv, opts, config) {
     const input = conv.turns.flatMap(turnToInputItems);
-    checkRawArrayIntegrity(conv, input);
+    checkRawArrayIntegrity(conv, input, config.lang);
     return {
         model: opts.model,
         instructions: conv.systemPrompt,
         input,
-        ...(opts.enableTools === false ? {} : { tools: [READ_FILE_TOOL] }),
+        ...(opts.enableTools === false ? {} : { tools: [buildReadFileTool(config.lang)] }),
         ...(config.store === false ? { store: false } : {}),
         include: ["reasoning.encrypted_content"],
         ...buildReasoningParams(config.reasoning.style, opts.effort),
@@ -67,7 +75,7 @@ export function createResponsesAdapter(config) {
             catch (err) {
                 if (err && typeof err === "object" && "status" in err) {
                     const anyErr = err;
-                    throw new ProviderHttpError(anyErr.message ?? "responses adapter 呼叫失敗", anyErr.status ?? 0, anyErr.headers ?? {}, anyErr.error, params);
+                    throw new ProviderHttpError(anyErr.message ?? m(config.lang, "responsesAdapterCallFailed"), anyErr.status ?? 0, anyErr.headers ?? {}, anyErr.error, params);
                 }
                 throw err;
             }

package/dist/audit.js

@@ -17,6 +17,11 @@ function getSection(sections, names) {
     return undefined;
 }
 const SUSPECT_PHRASES = ["應該改成", "建議採用", "嚴重度", "高風險", "應廢止"];
+// plan_i18n_v1.2.md §4.1／i18n_classification_t2.md §五 #8-12：英文回報若只比對中文詞，
+// 稽核會靜默失效——不報錯、不變紅，只是什麼都抓不到。兩套並存同時比對，不隨 lang 切換
+// （spoke 可能用另一種語言作答，稽核本來就不看工單語言）。summary.md 需要分開標示是
+// 哪一套命中，供日後調整這份清單時有資料可依據，故 auditSpoke 分開回傳兩個陣列。
+const SUSPECT_PHRASES_EN = ["should be changed to", "recommend adopting", "severity", "high risk", "should be deprecated"];
 // 抓看起來像相對路徑的引用：至少一層目錄＋副檔名，容許前後有反引號與 :行號。
 // plan_dispatch_v1.12.md §15：字元類須容許中括號，否則 Next.js 動態路由段（[id]、
 // [...slug]、[[...slug]]）會把路徑從中括號後截斷——截斷後的字串當然不在允許清單內，
@@ -34,6 +39,11 @@ const OBSERVATION_PATTERNS = [
     /^#{2,4}\s*觀察\s*\d+/, // 標題形式："### 觀察 1"
     /^(?:\*\*)?Observation\s*\d+(?:\.\d+)?(?:\*\*)?[:：]/i, // 英文模板的巢狀標記
     /^#{2,4}\s*Observation\s*\d+/i, // 英文模板的標題形式
+    // real-run-i18n-lang（2026-08-12）：`deepseek-v4-flash` 的中文格寫成 `## 1. 「比照 tags 路由」…`
+    // ——標題形式但編號後面直接是內容，沒有「觀察」二字，上面兩條標題樣式都認不出來，於是整份
+    // 判「無法計數」。**那是本次唯一一項「中文格看起來比英文格差」的來源，而它與語言無關。**
+    // 放在最後：前面任何一條命中就不會走到這裡，所以不會蓋掉既有判讀。
+    /^#{2,4}\s*\d+[.、]/,
 ];
 function countObservations(observationsBody) {
     if (observationsBody.trim().length === 0)
@@ -73,6 +83,8 @@ export function auditSpoke(finalText, allowedRelativePaths) {
             citedPathsOutsideAllowlistDetail: [],
             cannotVerifySectionPresent: false,
             suspectPhrases: [],
+            suspectPhrasesZh: [],
+            suspectPhrasesEn: [],
         };
     }
     const lines = finalText.split(/\r?\n/).map((l) => l.trim());
@@ -105,7 +117,9 @@ export function auditSpoke(finalText, allowedRelativePaths) {
     const pathsCitedElsewhere = new Set([...elsewhereText.matchAll(PATH_REGEX)].map((m) => m[1]));
     const pathsOnlyInCannotVerify = new Set([...pathsCitedInCannotVerifySection].filter((p) => !pathsCitedElsewhere.has(p)));
     const citedPathsOutsideAllowlist = citedPaths.filter((p) => !allowedSet.has(p) && !pathsOnlyInCannotVerify.has(p));
-    const suspectPhrases = SUSPECT_PHRASES.filter((phrase) => finalText.includes(phrase));
+    const suspectPhrasesZh = SUSPECT_PHRASES.filter((phrase) => finalText.includes(phrase));
+    const suspectPhrasesEn = SUSPECT_PHRASES_EN.filter((phrase) => finalText.includes(phrase));
+    const suspectPhrases = [...suspectPhrasesZh, ...suspectPhrasesEn];
     const citedPathsOutsideAllowlistDetail = citedPathsOutsideAllowlist.map((p) => ({
         path: p,
         section: findSection(sections, p),
@@ -119,5 +133,7 @@ export function auditSpoke(finalText, allowedRelativePaths) {
         citedPathsOutsideAllowlistDetail,
         cannotVerifySectionPresent,
         suspectPhrases,
+        suspectPhrasesZh,
+        suspectPhrasesEn,
     };
 }