用 n8n 自动评审 GitLab Merge Request：Gemini AI + JIRA 上下文

输入：GitLab MR 评论触发 → 处理：提取 diff 与可选 JIRA 工单摘要 → 输出：行级 AI 评审意见或 ALL_CLEAR

给谁

有 n8n 实例、GitLab 项目管理员权限，并持有 Gemini 与可选 JIRA 凭证的开发团队

你会得到

在 MR 评论中触发后，自动获取变更差异与 JIRA 上下文，由 LLM 生成逻辑/安全/性能问题并回写行级评论；无问题时发布 ALL_CLEAR 备注

你需要

n8n · gitlab · gemini · jira · webhook · http-request

步骤

1. 导入模板并配置触发器

   从 n8n 模板库导入 workflow JSON。打开 Webhook 节点，将路径占位符 REPLACE_WITH_UNIQUE_PATH 替换为自定义路径。在 GitLab 项目设置中配置 Webhook，将 MR 评论事件发送至该 URL。在 Gemini（或其他 LLM）节点选择凭证，可选在 JIRA 节点配置凭证。激活工作流。

   搭建指导 · 本站建议版本。搭建助手启动词；用于确认前置条件与权限

   复制提示词

   请协助我完成以下 n8n 模板搭建：
   1. 导入模板 JSON（来源：https://n8n.io/workflows/7924/）
   2. 配置 Webhook 节点：将 REPLACE_WITH_UNIQUE_PATH 替换为唯一路径
   3. 在 GitLab 项目 → Settings → Webhooks 中添加 Webhook URL，选择 Merge request events 或 Note events
   4. 配置凭证：Gemini 节点选择 googlePalmApi 凭证；JIRA 节点可选配置 jiraSoftwareCloudApi
   5. 设置环境变量 GITLAB_TOKEN
   6. 激活工作流后，在 MR 中评论触发词测试
   
   请确认以上每一步的输入和权限是否已就绪。

2. 触发监听与数据准备

   当 GitLab MR 收到包含触发词的评论时，Webhook 节点捕获事件。工作流提取项目 ID 与 MR IID，存入 Workflow Static Data 以便后续节点和错误处理复用。同时向 MR 发送「AI review started」状态消息。随后通过 GitLab API 获取 MR changes，解析 diff 为 original/new 代码块；若 MR 描述含 JIRA 工单号，则拉取工单摘要，如是子任务再拉取父工单摘要，合并为上下文。

3. LLM 代码评审

   将解析后的 diff 块与可选 JIRA 上下文送入 Basic LLM Chain。Gemini 2.5 Pro 按系统指令只关注逻辑正确性、安全与性能问题，忽略风格与文件名等。发现问题时以 🤖 **AI Review:** 为前缀输出可执行建议；无问题时返回 ALL_CLEAR。

   运行提示词 · 来源原文。Basic LLM Chain 节点的 messages.messageValues[0].message 字段（System Message）

   复制提示词

   You are an automated code review bot. Your primary goal is to identify high-value issues in code.
   
   **Instructions:**
   
   1. **Analysis Focus:** Concentrate only on business logic, correctness, security, and performance. Ignore style, naming, filenames, version bumps, and non-source files. Assume placeholders for yml files will be filled with correct values.
   2. **Output for Issues:** When you find an issue, begin your response with the prefix `🤖 **AI Review:** `. Follow it with a concise, actionable explanation. Prefer one-liners that are immediately actionable - what or where or why and how to fix.
   3. **Output for No Issues:** If your analysis finds **zero** issues worth reporting, your **ENTIRE** response must be the single keyword: `ALL_CLEAR`. Do not add any other text or explanation.
   4. ** Findings to ignore in yaml files:**
       - Ignore if a value for vaultSecretPath in a values file or a yaml contains a double slash (//)
       - Ignore errors where tag appears to be very or too specific
   
   **Important Rules:**
   
   * NEVER write praise, summaries, or conversational phrases like "I have no findings."
   * If you find issues, start every comment with the `🤖 **AI Review:** ` prefix.
   * If you find no issues, don't output anything.

   运行提示词 · 来源原文。Basic LLM Chain 节点的 text 字段（User Message / Prompt）

   复制提示词

   =First, consider the context from the associated JIRA ticket:
   ------------
   {{ $json.jiraParentContext || $json.jiraContext || 'No JIRA context was provided.' }}
   ------------
   
   File path：{{ $('Skip File Changes').item.json.new_path }}
   
   ```Original code
    {{ $json.originalCode }}
   ```
   change to
   ```New code
    {{ $json.newCode }}
   ```
   Please review the code changes in this section:

   运行提示词 · 来源原文。使用前确认粘贴位置、输入变量及输出格式。

   复制提示词

   ### ⭐️ Try It Out!
   
   Welcome! This template adds AI-powered code reviews to your GitLab Merge Requests. Simply comment **`ai-review`** on any MR and watch the workflow:
   
   - Captures the comment event.
   - Fetches diff & optional JIRA context.
   - Asks an LLM for insights.
   - Posts inline comments or an 'all clear' note.
   
   Perfect for catching logic, security and performance issues fast.

   运行提示词 · 来源原文。使用前确认粘贴位置、输入变量及输出格式。

   复制提示词

   ### 🟢 Step 1: Listen & Capture
   
   A Webhook node listens for merge request note events. When the comment text matches your trigger (default: `ai-review`), it:
   
   - Extracts the Project & MR IDs and stores them in a **Static Context** (see below) so later nodes always know where to post results.
   - Posts a short "AI review started" message back to the MR to let reviewers know the bot is working.

   运行提示词 · 来源原文。使用前确认粘贴位置、输入变量及输出格式。

   复制提示词

   ### 🟣 Step 2: Prepare & Review
   
   The workflow fetches the merge request changes via the GitLab API and parses the diff into original vs new code blocks.
   
   If the MR description contains a Jira key, the workflow fetches the issue summary and, if it's a subtask, also fetches the parent ticket for extra context. Both summaries are combined into the prompt.
   
   An LLM (Gemini by default) is prompted to find critical issues only—logic bugs, security flaws and performance problems—and to return comments with file and line numbers. If no issues exist, the model returns `ALL_CLEAR`.

   运行提示词 · 来源原文。使用前确认粘贴位置、输入变量及输出格式。

   复制提示词

   ### 🟠 Step 3: Post & Fallback
   
   Review comments are converted into GitLab discussion payloads and posted back to the MR at the correct file and line positions.
   
   If a comment’s position is not calculated properly, the workflow falls back to posting a comment as a thread at the MR level so nothing is lost.

   运行提示词 · 来源原文。使用前确认粘贴位置、输入变量及输出格式。

   复制提示词

   ### 🔵 Static Context & Error Handling
   
   Workflow Static Data is used to store persistent values—like the project ID and the MR ID so that they will be available across branches.
   
   This is critical for the Error Trigger: if something fails, the error path can still post to the correct MR using these stored IDs. Without static context, error comments wouldn’t know where to go.

   运行提示词 · 来源原文。使用前确认粘贴位置、输入变量及输出格式。

   复制提示词

   ### 🟡 Need Help & Customise
   
   - Change the trigger word in the **IF** node to match your own.
   - Swap in your preferred LLM or adjust the prompt to suit your guidelines.
   - Filter by file type or exclude certain directories in the parsing logic.

4. 结果回写与错误兜底

   LLM 输出经聚合后，由 Process Comments 节点过滤有效评论。若存在 issues，转换为 GitLab discussion payload 并尝试在行级位置发布；若位置计算失败，降级为 MR 级线程评论。若全部为 ALL_CLEAR，则发布「No significant issues were found. LGTM!」备注。Error Trigger 路径通过 Static Data 中存储的项目 ID 与 MR ID，向对应 MR 发送错误通知，最后清理 Static Data。

5. 自定义与调优

   可在 IF 节点（Need Review）修改触发词；在 Basic LLM Chain 调整 system message 以关注风格或文档；在 Skip File Changes 或 Prepare Code Changes 节点增加文件类型过滤；也可集成 Slack/Email 节点发送完成通知。

请协助我完成以下 n8n 模板搭建：
1. 导入模板 JSON（来源：https://n8n.io/workflows/7924/）
2. 配置 Webhook 节点：将 REPLACE_WITH_UNIQUE_PATH 替换为唯一路径
3. 在 GitLab 项目 → Settings → Webhooks 中添加 Webhook URL，选择 Merge request events 或 Note events
4. 配置凭证：Gemini 节点选择 googlePalmApi 凭证；JIRA 节点可选配置 jiraSoftwareCloudApi
5. 设置环境变量 GITLAB_TOKEN
6. 激活工作流后，在 MR 中评论触发词测试

请确认以上每一步的输入和权限是否已就绪。