gitlab

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

GitLab Skill

Use the

glab

CLI to interact with GitLab. The primary workflow is MR Review — but this skill handles any GitLab operation the user needs.

使用

glab

CLI 与 GitLab 交互。核心工作流是 MR 评审，但本 skill 支持处理用户需要的任何 GitLab 操作。

URL Parsing

URL 解析

When given a GitLab MR URL like

https://gitlab.com/group/subgroup/project/-/merge_requests/42

repo_ref: strip

https://

and everything from

/-/

onward →

gitlab.com/group/subgroup/project

mr_id: extract the number after
```
merge_requests/
```
→
```
42
```

These two values power most glab commands:

glab mr <cmd> <mr_id> --repo <repo_ref>

当收到类似

https://gitlab.com/group/subgroup/project/-/merge_requests/42

的 GitLab MR URL 时：

repo_ref：去掉
```
https://
```
以及
```
/-/
```
之后的所有内容 →
```
gitlab.com/group/subgroup/project
```
mr_id：提取
```
merge_requests/
```
之后的数字 →
```
42
```

这两个值是大多数 glab 命令的核心参数：

glab mr <cmd> <mr_id> --repo <repo_ref>

MR Review Workflow

MR 评审工作流

When the user asks to review a MR, run this pipeline:

1. Fetch MR info, diff, and existing comments
2. Read & summarize existing comments (understand what's already discussed)
3. code-reviewer + security + qa → review in PARALLEL (with comment context)
4. Compose Thai comment from template
5. Post comment to the MR

当用户要求评审某个 MR 时，运行以下流程：

1. 拉取 MR 信息、diff 和已有评论
2. 阅读并总结已有评论（了解已经讨论过的内容）
3. code-reviewer + security + qa → 并行评审（传入评论上下文）
4. 基于模板撰写泰语评论
5. 将评论发布到对应 MR

Step 1: Fetch

步骤 1：拉取信息

bash

glab mr view <mr_id> --repo <repo_ref> --output json
glab mr diff <mr_id> --repo <repo_ref>
glab mr note list <mr_id> --repo <repo_ref>

Extract from the view output: MR title, source branch, target branch, author.

bash

glab mr view <mr_id> --repo <repo_ref> --output json
glab mr diff <mr_id> --repo <repo_ref>
glab mr note list <mr_id> --repo <repo_ref>

从 view 命令的输出中提取：MR 标题、源分支、目标分支、作者。

Step 2: Read Existing Comments

步骤 2：读取已有评论

Before diving into the review, read through all existing MR comments/notes fetched in Step 1. This gives crucial context — other reviewers may have already flagged issues, the author may have explained design decisions, or there may be ongoing discussions that affect how you should review. Summarize the key points:

Issues already raised by other reviewers
Author's explanations or decisions
Unresolved discussions that need attention
Resolved items (avoid duplicating feedback)

This prevents the review from repeating what's already been said and helps focus on gaps that haven't been addressed yet.

在开始评审前，通读步骤 1 拉取到的所有 MR 已有评论/备注。这会提供关键上下文——其他评审者可能已经标记了问题，作者可能已经解释了设计决策，或者存在仍在进行中的讨论，这些都会影响你的评审方式。总结核心要点：

其他评审者已经提出的问题
作者的解释或决策
需要关注的未解决讨论
已解决项（避免重复反馈）

这可以避免评审内容重复已有讨论，帮助聚焦于还没有处理的空缺部分。

Step 3: Parallel Review

步骤 3：并行评审

Read the project's

CLAUDE.md

if available (for conventions). Then spawn all three agents at the same time, including a summary of existing comments so reviewers have full context:

Agent(
  description: "Code review MR diff",
  subagent_type: "code-reviewer",
  model: "opus",
  prompt: """
[code-reviewer agent instructions — read from ~/.claude/agents/code-reviewer.agent.md]

---

如果项目存在

CLAUDE.md

则先读取（了解规范）。然后同时启动三个 Agent，传入已有评论的总结，让评审者获得完整上下文：

Agent(
  description: "Code review MR diff",
  subagent_type: "code-reviewer",
  model: "opus",
  prompt: """
[code-reviewer agent 指令 — 读取自 ~/.claude/agents/code-reviewer.agent.md]

---

Project Conventions

项目规范

[relevant sections from CLAUDE.md if available, else omit]

[如果存在 CLAUDE.md 则提取相关部分，否则省略]

Existing MR Comments

已有 MR 评论

[summary of existing comments from Step 2 — issues raised, author explanations, unresolved discussions] Do NOT repeat issues that other reviewers have already flagged unless you have additional insight to add.

[步骤 2 得到的已有评论总结——提出的问题、作者解释、未解决讨论] 除非你有额外的见解补充，否则不要重复其他评审者已经标记过的问题。

Task

任务

Review the following MR diff for convention compliance.

MR: !<mr_id> — <mr_title> Branch: <source> → <target>

评审以下 MR diff 是否符合规范。

MR: !<mr_id> — <mr_title> 分支: <source> → <target>

Diff

<full diff output> """ )

Agent( description: "Security review MR diff", subagent_type: "security", model: "sonnet", prompt: """ [security agent instructions — read from ~/.claude/agents/security.agent.md]

<完整 diff 输出> """ )

Agent( description: "Security review MR diff", subagent_type: "security", model: "sonnet", prompt: """ [security agent 指令 — 读取自 ~/.claude/agents/security.agent.md]

Existing MR Comments

已有 MR 评论

[summary of existing comments from Step 2 — issues raised, author explanations, unresolved discussions] Do NOT repeat security concerns that have already been raised unless you have additional findings.

[步骤 2 得到的已有评论总结——提出的问题、作者解释、未解决讨论] 除非你有额外的发现，否则不要重复已经提出过的安全隐患。

Task

任务

Security review the following MR diff.

MR: !<mr_id> — <mr_title> Branch: <source> → <target>

对以下 MR diff 进行安全评审。

MR: !<mr_id> — <mr_title> 分支: <source> → <target>

Diff

<full diff output> """ )

Agent( description: "QA review MR diff", subagent_type: "qa", model: "sonnet", prompt: """ [qa agent instructions — read from ~/.claude/agents/qa.agent.md]

<完整 diff 输出> """ )

Agent( description: "QA review MR diff", subagent_type: "qa", model: "sonnet", prompt: """ [qa agent 指令 — 读取自 ~/.claude/agents/qa.agent.md]

Project Conventions

项目规范

[relevant sections from CLAUDE.md if available, else omit]

[如果存在 CLAUDE.md 则提取相关部分，否则省略]

Existing MR Comments

已有 MR 评论

[summary of existing comments from Step 2 — issues raised, author explanations, unresolved discussions] Do NOT repeat QA concerns that have already been raised unless you have additional findings.

[步骤 2 得到的已有评论总结——提出的问题、作者解释、未解决讨论] 除非你有额外的发现，否则不要重复已经提出过的 QA 相关问题。

Task

任务

QA review the following MR diff. Focus on:

Test coverage gaps (are new code paths tested?)
Missing edge case tests
Regression risks
Acceptance criteria validation (if available)

MR: !<mr_id> — <mr_title> Branch: <source> → <target>

对以下 MR diff 进行 QA 评审。重点关注：

测试覆盖缺口（新代码路径是否有测试？）
缺失的边界 case 测试
回归风险
验收标准验证（如果有提供）

MR: !<mr_id> — <mr_title> 分支: <source> → <target>

Diff

<full diff output> """ ) ```

Read the agent files from

~/.claude/agents/

before spawning, so the full agent instructions are included in the prompt.

<完整 diff 输出> """ )


在启动 Agent 前先从 `~/.claude/agents/` 读取对应 Agent 文件，确保完整的 Agent 指令都包含在提示词中。

Step 4: Compose Thai comment

步骤 4：撰写泰语评论

Read references/mr-review-template.md and fill in findings from all three agents (code-reviewer, security, qa).

读取 references/mr-review-template.md，填入三个 Agent（code-reviewer、security、qa）的所有发现。

Step 5: Post

步骤 5：发布

bash

glab mr note <mr_id> --repo <repo_ref> -m "<thai_comment>"

glab

is not authenticated or fails, output the review in the conversation instead and tell the user.

bash

glab mr note <mr_id> --repo <repo_ref> -m "<thai_comment>"

如果

glab

未认证或运行失败，则直接在对话中输出评审结果并告知用户。

Common glab Operations

常用 glab 操作

Use these directly via

Bash

when the user asks for something other than a full review:

Task	Command
List open MRs	`glab mr list --repo <repo_ref>`
View MR details	`glab mr view <mr_id> --repo <repo_ref>`
Approve MR	`glab mr approve <mr_id> --repo <repo_ref>`
Check pipeline status	`glab ci status --repo <repo_ref>`
List pipelines	`glab ci list --repo <repo_ref>`
Retry a job	`glab ci retry <job_id> --repo <repo_ref>`
Add a note/comment	`glab mr note <mr_id> --repo <repo_ref> -m "<text>"`

For

--repo

, you can omit it if you're already inside the project directory (glab detects the remote automatically).

当用户要求的不是完整评审时，直接通过

Bash

运行以下命令：

任务	命令
列出打开的 MR	`glab mr list --repo <repo_ref>`
查看 MR 详情	`glab mr view <mr_id> --repo <repo_ref>`
批准 MR	`glab mr approve <mr_id> --repo <repo_ref>`
检查流水线状态	`glab ci status --repo <repo_ref>`
列出流水线	`glab ci list --repo <repo_ref>`
重试任务	`glab ci retry <job_id> --repo <repo_ref>`
添加备注/评论	`glab mr note <mr_id> --repo <repo_ref> -m "<text>"`

对于

--repo

参数，如果你已经处于项目目录下可以省略（glab 会自动检测远程仓库）。

Error Handling

错误处理

glab not authenticated: tell the user to run
```
glab auth login
```
glab command fails: output the review as conversation text instead of posting, explain what failed
Empty diff: note that the MR has no file changes and skip the review agents
Large diff (>500 lines): warn the user, proceed but note the review may miss details

glab 未认证：告知用户运行
```
glab auth login
```
glab 命令运行失败：不在 MR 中发布评论，直接在对话中输出评审结果，解释失败原因
空 diff：标注该 MR 没有文件变更，跳过评审 Agent
diff 过大（>500 行）：警告用户，继续评审但标注可能会遗漏细节