Integration-Automation · JE-Chen · Aug 2, 2026 · Aug 2, 2026 · Aug 2, 2026 · Aug 2, 2026
diff --git a/.github/workflows/dev.yml b/.github/workflows/dev.yml
@@ -70,8 +70,16 @@ jobs:
       uses: actions/download-artifact@v4
       with:
         name: coverage-xml
+    - name: Read package version
+      # Without a version, SonarCloud's "previous version" new-code baseline has
+      # nothing to anchor to and counts the whole history as new -- which is how
+      # new_lines came to exceed the project's total line count.
+      id: version
+      run: echo "value=$(grep -m1 '^version = ' pyproject.toml | cut -d'"' -f2)" >> "$GITHUB_OUTPUT"
     - name: SonarQube Cloud scan
       uses: SonarSource/sonarqube-scan-action@v8.2.1
+      with:
+        args: -Dsonar.projectVersion=${{ steps.version.outputs.value }}
       env:
         SONAR_TOKEN: ${{ secrets.SONAR_TOKEN }}
         SONAR_HOST_URL: https://sonarcloud.io
diff --git a/.github/workflows/stable.yml b/.github/workflows/stable.yml
@@ -70,8 +70,16 @@ jobs:
       uses: actions/download-artifact@v4
       with:
         name: coverage-xml
+    - name: Read package version
+      # Without a version, SonarCloud's "previous version" new-code baseline has
+      # nothing to anchor to and counts the whole history as new -- which is how
+      # new_lines came to exceed the project's total line count.
+      id: version
+      run: echo "value=$(grep -m1 '^version = ' pyproject.toml | cut -d'"' -f2)" >> "$GITHUB_OUTPUT"
     - name: SonarQube Cloud scan
       uses: SonarSource/sonarqube-scan-action@v8.2.1
+      with:
+        args: -Dsonar.projectVersion=${{ steps.version.outputs.value }}
       env:
         SONAR_TOKEN: ${{ secrets.SONAR_TOKEN }}
         SONAR_HOST_URL: https://sonarcloud.io

diff --git a/architecture_explore.md b/architecture_explore.md
@@ -243,20 +243,33 @@ call_X_multi_file_and_send()   → run_dir_files_with_package(..., True)
 
 ```
 extend_ai_gui/
-├── ai_gui_global_variable.py     模板檔名清單 + 檔名→模板內容對照表
+├── ai_gui_global_variable.py     模板檔名清單 + 檔名→內建模板內容對照表
+├── prompt_store.py               編輯過的 prompt 檔案解析（~/.pybreeze/prompts/）
 ├── code_review/
-│   ├── code_review_thread.py     SenderThread(QThread)：CoT 多階段審查
+│   ├── cot_chain.py              接線表（純邏輯，無 Qt）：哪步引用哪步
+│   ├── code_review_thread.py     SenderThread(QThread)：跑八步審查鏈
 │   └── cot_code_review_gui.py    UI
 ├── prompt_edit_gui/
-│   ├── cot_prompt_editor_widget.py     編輯 5 個 CoT 模板（QFileSystemWatcher 熱更新）
-│   ├── skills_prompt_editor_widget.py  編輯 2 個 Skill 模板
+│   ├── prompt_editor_widget.py         共用編輯器（QFileSystemWatcher 熱更新）
+│   ├── cot_prompt_editor_widget.py     8 個 CoT 模板的檔案清單＋語言鍵
+│   ├── skills_prompt_editor_widget.py  2 個 Skill 模板的檔案清單＋語言鍵
 │   ├── prompt_file_io.py               共用存檔（失敗跳警告對話框）
-│   ├── cot_code_review_prompt_templates/   7 個模板常數
+│   ├── cot_code_review_prompt_templates/   8 個模板常數＋global_rule
 │   └── skills_prompt_templates/            2 個模板常數
 └── skills/skills_send_gui.py     單次 prompt 發送（RequestThread）
 ```
 
-**CoT 審查鏈**（`code_review_thread.py`）：`first_summary` → `first_code_review` → `linter` → `code_smell_detector` → `total_summary`。前四階段的結果被收集起來餵給最後的 total summary。每階段都套 `build_global_rule_template()` 包一層全域規則。
+**CoT 審查鏈**（`cot_chain.py` 定義接線，`code_review_thread.py` 執行）八個步驟：
+
+```
+first_summary → first_code_review → judge_single_review ┐（評分前一步的審查）
+              → linter → code_smell_detector → step_by_step_analysis ┐（走過每條發現）
+              → total_summary → judge（帶 linter/code smell 脈絡評分總結）
+```
+
+**編輯過的 prompt 會生效**（`prompt_store.py`）：每個模板都以程式碼常數出貨，`~/.pybreeze/prompts/<名稱>.md` 存在且非空時覆寫它。編輯器讀寫的就是這個位置，所以在編輯器裡改 prompt 會改變審查實際送出的內容 —— 這正是編輯器存在的理由。檔案缺失、空白、讀不到都退回內建版本；編輯過的 prompt 若含有鏈填不了的 placeholder，記 log 後退回內建，不讓整次審查倒在使用者無法從 UI 診斷的 `KeyError` 上。讀取不會建目錄，只有存檔才會。
+
+`cot_chain.py` 用兩張表描述接線：`STEP_RESULT_KEY`（每步答案存在哪個 key）與 `STEP_ARGUMENTS`（每步的 placeholder 由哪個 key 填）。**順序即相依順序** —— 每步只能引用它上面的步驟，`test_cot_chain.py` 有結構性測試守住這件事。步驟失敗時錯誤訊息只顯示給使用者、不會被存進 results，避免後續步驟把「傳送失敗」當成審查內容引用。每步都套 `build_global_rule_template()` 包一層全域規則。
 
 安全處理：送出前 `validate_url()`、`allow_redirects=False`、`stream=True` 搭配 `read_capped_text()` 限制回應大小、單一 `requests.Session` 重用 TCP/TLS 連線、`isInterruptionRequested()` 讓 widget 關閉時能中止。
 
@@ -387,6 +400,7 @@ extend_ai_gui/
 |---|---|
 | `ssh_known_hosts` | TOFU 確認過的 SSH host key |
 | `prthinker_setting.json` | prthinker 後端/平台/金鑰設定 |
+| `prompts/*.md` | 編輯過的 CoT / Skill prompt，覆寫內建模板 |
 | `response_stats.txt` | AI 審查接受/拒絕統計 |
 | `urls.txt` | AI 審查端點歷史 |
 
@@ -396,10 +410,10 @@ extend_ai_gui/
 
 ## 18. 測試與 CI
 
-- **單元測試** `test/test_utils/` — 60 個 `test_*.py`。純邏輯 + headless Qt widget 測試（`QT_QPA_PLATFORM=offscreen`）。涵蓋 curl/HAR 解析、SSRF 驗證、SSH 安全、process reader EOF、queue pump、語言對齊、mermaid parser、diagram 序列化、prthinker 設定等。有 hypothesis fuzz 測試（`test_fuzz_pure_logic.py`）。
+- **單元測試** `test/test_utils/` — 65 個 `test_*.py`、985 個測試。純邏輯 + headless Qt widget 測試（`QT_QPA_PLATFORM=offscreen`）。涵蓋 curl/HAR 解析、SSRF 驗證、SSH 安全、process reader EOF、queue pump、語言對齊、mermaid parser、diagram 序列化、prthinker 設定等。有 hypothesis fuzz 測試（`test_fuzz_pure_logic.py`）。
 - **整合測試** `test/unit_test/start_automation/` — 以 `debug_mode=True` 啟動 IDE，10 秒後自動關閉，驗證啟動流程與 extend tab
 - **CI** `.github/workflows/{dev,stable}.yml` — `unit-tests` job 跑 Windows runner、Python 3.10–3.14 矩陣，3.12 那一腳額外上傳 `coverage-xml` artifact；`sonarcloud` job 跑 ubuntu、`needs: unit-tests`。每日 02:00 排程 + push/PR 觸發。`stable.yml` 另有 `publish` job 負責版號遞增與 PyPI 發布
-- **覆蓋率** `.coveragerc` — `relative_files = True` 是必要的：報告在 Windows 產生、由 Linux 上的 scanner 讀取，路徑不能帶機器資訊。目前整體 57%（`utils/` 與 `tools_gui` 95–100%，UI 層 20–45% 拉低）
+- **覆蓋率** `.coveragerc` — `relative_files = True` 是必要的：報告在 Windows 產生、由 Linux 上的 scanner 讀取，路徑不能帶機器資訊。目前整體 60%（`utils/`、`tools_gui`、`dialog` 95–100%；`editor_main` 58%、`menu` 54%；仍低的是 `diagram_editor` 45%、`process_executor` 39%、`connect_gui` 28%）
 - **靜態分析** SonarCloud（`sonar-project.properties`，CI-based analysis；Automatic Analysis 已關閉且必須維持關閉，兩種模式互斥）+ Codacy（`.codacy.yml`）+ Bandit（`pyproject.toml` 中排除 test、skip B101/B404）
 - **SonarCloud 方案限制** 該組織的方案只開放 `main` 與 PR 的分析結果。非 main 分支的分析送得出去、CE 任務也會成功，但結果讀回來是 403（組織內每個專案都只有 `main` 一條分支）。因此 `dev.yml` 只在 PR 時掃描，`stable.yml` 另外掃 push to `main`
 

diff --git a/pybreeze/extend_multi_language/extend_english.py b/pybreeze/extend_multi_language/extend_english.py
@@ -134,6 +134,8 @@
     "prthinker_setting_extra_arguments_label": "Extra command-line arguments",
     "prthinker_setting_source_path_label": "prthinker source folder",
     "prthinker_setting_stored_at_label": "Stored at:",
+    # Prompt editors — where the edited prompts override the built-in ones from
+    "prompt_editor_stored_at_label": "Prompt files (these override the built-in prompts):",
     "prthinker_choose_source_path_label": "Choose the prthinker source folder",
     "prthinker_need_source_path_message":
         "prthinker is installed from source. Choose the folder holding its "

diff --git a/pybreeze/extend_multi_language/extend_traditional_chinese.py b/pybreeze/extend_multi_language/extend_traditional_chinese.py
@@ -134,6 +134,8 @@
     "prthinker_setting_extra_arguments_label": "額外的命令列參數",
     "prthinker_setting_source_path_label": "prthinker 原始碼資料夾",
     "prthinker_setting_stored_at_label": "設定檔位置：",
+    # Prompt 編輯器 —— 編輯過的 prompt 會覆寫內建版本
+    "prompt_editor_stored_at_label": "Prompt 檔案位置（會覆寫內建 prompt）：",
     "prthinker_choose_source_path_label": "選擇 prthinker 原始碼資料夾",
     "prthinker_need_source_path_message":
         "prthinker 是從原始碼安裝的。請選擇含有 pyproject.toml 的資料夾，"

diff --git a/pybreeze/pybreeze_ui/extend_ai_gui/ai_gui_global_variable.py b/pybreeze/pybreeze_ui/extend_ai_gui/ai_gui_global_variable.py
@@ -6,29 +6,45 @@
     FIRST_CODE_REVIEW_TEMPLATE
 from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.cot_code_review_prompt_templates.first_summary_prompt import \
     FIRST_SUMMARY_TEMPLATE
+from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.cot_code_review_prompt_templates.judge import \
+    JUDGE_TEMPLATE
+from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.cot_code_review_prompt_templates.judge_single_review import \
+    JUDGE_SINGLE_REVIEW_TEMPLATE
 from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.cot_code_review_prompt_templates.linter import \
     LINTER_TEMPLATE
+from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.cot_code_review_prompt_templates.step_by_step_analysis import \
+    STEP_BY_STEP_ANALYSIS_TEMPLATE
 from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.cot_code_review_prompt_templates.total_summary import \
     TOTAL_SUMMARY_TEMPLATE
 from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.skills_prompt_templates.code_explainer import \
     CODE_EXPLAINER_TEMPLATE
 from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.skills_prompt_templates.code_review import \
     CODE_REVIEW_SKILL_TEMPLATE
 
+# The order the chain runs in. Each step may only quote steps above it, so the
+# order is a dependency order, not a preference: judge_single_review scores the
+# review written just before it, step_by_step_analysis walks the linter and code
+# smell findings, and judge scores the finished summary.
 COT_TEMPLATE_FILES = [
     "first_summary_prompt.md",
     "first_code_review.md",
+    "judge_single_review.md",
     "linter.md",
     "code_smell_detector.md",
+    "step_by_step_analysis.md",
     "total_summary.md",
+    "judge.md",
 ]
 
 COT_TEMPLATE_RELATION = {
     "first_summary_prompt.md": FIRST_SUMMARY_TEMPLATE,
     "first_code_review.md": FIRST_CODE_REVIEW_TEMPLATE,
+    "judge_single_review.md": JUDGE_SINGLE_REVIEW_TEMPLATE,
     "linter.md": LINTER_TEMPLATE,
     "code_smell_detector.md": CODE_SMELL_DETECTOR_TEMPLATE,
+    "step_by_step_analysis.md": STEP_BY_STEP_ANALYSIS_TEMPLATE,
     "total_summary.md": TOTAL_SUMMARY_TEMPLATE,
+    "judge.md": JUDGE_TEMPLATE,
 }
 
 SKILLS_TEMPLATE_FILES = [

diff --git a/pybreeze/pybreeze_ui/extend_ai_gui/code_review/code_review_thread.py b/pybreeze/pybreeze_ui/extend_ai_gui/code_review/code_review_thread.py
@@ -5,9 +5,9 @@
 from PySide6.QtCore import QThread, Signal
 from je_editor import language_wrapper
 
-from pybreeze.pybreeze_ui.extend_ai_gui.ai_gui_global_variable import COT_TEMPLATE_RELATION
-from pybreeze.pybreeze_ui.extend_ai_gui.prompt_edit_gui.cot_code_review_prompt_templates.global_rule import \
-    build_global_rule_template
+from pybreeze.pybreeze_ui.extend_ai_gui.code_review.cot_chain import (
+    CODE_DIFF, STEP_RESULT_KEY, build_prompt
+)
 from pybreeze.utils.logging.logger import pybreeze_logger
 from pybreeze.utils.network.http_client import (
     ResponseTooLargeError, read_capped_text, CONNECT_TIMEOUT,
@@ -40,74 +40,36 @@ def run(self):
             session.close()
 
     def _run_templates(self, session: requests.Session, code: str) -> None:
-        first_code_review_result = None
-        first_summary_result = None
-        linter_result = None
-        code_smell_result = None
+        # Answers accumulate here as the chain runs; a later step quotes whichever
+        # of them its template asks for. See cot_chain for the wiring.
+        results: dict[str, str] = {CODE_DIFF: code}
         for file in self.files:
             # Stop promptly if the widget is closing instead of firing off the
             # remaining per-template POSTs.
             if self.isInterruptionRequested():
                 return
-            match file:
-                case "first_summary_prompt.md":
-                    first_summary_prompt = COT_TEMPLATE_RELATION["first_summary_prompt.md"]
-                    prompt = build_global_rule_template(
-                        prompt=first_summary_prompt.format(code_diff=code)
-                    )
-                case "first_code_review.md":
-                    first_code_review_prompt = COT_TEMPLATE_RELATION["first_code_review.md"]
-                    prompt = build_global_rule_template(
-                        prompt=first_code_review_prompt.format(code_diff=code)
-                    )
-                case "linter.md":
-                    linter_prompt = COT_TEMPLATE_RELATION["linter.md"]
-                    prompt = build_global_rule_template(
-                        prompt=linter_prompt.format(code_diff=code)
-                    )
-                case "code_smell_detector.md":
-                    code_smell_detector_prompt = COT_TEMPLATE_RELATION["code_smell_detector.md"]
-                    prompt = build_global_rule_template(
-                        prompt=code_smell_detector_prompt.format(code_diff=code)
-                    )
-                case "total_summary.md":
-                    total_summary_prompt = COT_TEMPLATE_RELATION["total_summary.md"]
-                    prompt = build_global_rule_template(
-                        prompt=total_summary_prompt.format(
-                            first_code_review=first_code_review_result,
-                            first_summary=first_summary_result,
-                            linter_result=linter_result,
-                            code_smell_result=code_smell_result,
-                            code_diff=code,
-                        )
-                    )
-                case _:
-                    continue
-
-            try:
-                # 傳送到指定 URL（重用 session 連線）
-                resp = session.post(
-                    self.url, json={"prompt": prompt},
-                    timeout=(CONNECT_TIMEOUT, 60), allow_redirects=False, stream=True,
-                )
-                reply_text = read_capped_text(resp)
-                match file:
-                    case "first_summary_prompt.md":
-                        first_summary_result = reply_text
-                    case "first_code_review.md":
-                        first_code_review_result = reply_text
-                    case "linter.md":
-                        linter_result = reply_text
-                    case "code_smell_detector.md":
-                        code_smell_result = reply_text
-                    case _:
-                        # total_summary.md has no intermediate result to store but
-                        # must still be emitted — `continue` here previously
-                        # dropped the final summary before it reached the UI.
-                        pass
-            except (requests.RequestException, ResponseTooLargeError) as e:
-                pybreeze_logger.error("CoT code review send failed for %s: %r", file, e)
-                reply_text = f"{language_wrapper.language_word_dict.get('cot_gui_error_sending')} {file} {e}"
-
+            prompt = build_prompt(file, results)
+            if prompt is None:
+                continue
+            reply_text, answered = self._ask(session, file, prompt)
+            result_key = STEP_RESULT_KEY.get(file)
+            # A failure message is shown but never stored: a later step must not
+            # quote "could not send" back to the model as if it were a review.
+            if answered and result_key is not None:
+                results[result_key] = reply_text
             # 發送訊號更新 UI
             self.update_response.emit(file, reply_text)
+
+    def _ask(self, session: requests.Session, file: str, prompt: str) -> tuple[str, bool]:
+        """Send one step's prompt; return its answer and whether it arrived."""
+        try:
+            # 傳送到指定 URL（重用 session 連線）
+            resp = session.post(
+                self.url, json={"prompt": prompt},
+                timeout=(CONNECT_TIMEOUT, 60), allow_redirects=False, stream=True,
+            )
+            return read_capped_text(resp), True
+        except (requests.RequestException, ResponseTooLargeError) as error:
+            pybreeze_logger.error("CoT code review send failed for %s: %r", file, error)
+            word = language_wrapper.language_word_dict
+            return f"{word.get('cot_gui_error_sending')} {file} {error}", False