Alibaba Cloud DevOps:Code review AI assistant

Configure AI review rules

1. As a code repository administrator, navigate to repository settings and select AI Assistant Settings.

2. Configure the Default AI Review Settings:

   1. If you select None, AI reviews are not automatically triggered for any merge request, regardless of the target branch.

   2. If you select All, AI reviews are automatically triggered for every merge request, regardless of the target branch.

   3. You can specify branches by name. When a merge request targets one of these branches, an AI review is automatically triggered. Branch names support wildcards using glob syntax:

      • ? matches a single character, excluding /.

      • * matches multiple characters within a single path segment.

      • ** matches multiple path segments.

      • For detailed rules, see glob.

3. Configure Custom AI Review Rules (Optional): You can define rules tailored to your team's preferences to improve review accuracy and reduce AI hallucinations.

reviews:
  # files that ignore to review, glob pattern
  ignores:
    - myfile.txt
    - myfolder/
    - "*.log"
    - "**/logs"
    - /debug.log
  # basic config
  language: "zh-CN" # Optional values: [zh-CN|en-US]
  problem_level: "CRITICAL" # Optional values: [BLOCKER|CRITICAL|MAJOR|MINOR] 
  review_mode: "DEFAULT" # Optional values: [DEFAULT|BY_COMMIT]
  enable_review_commit_message: true # Takes effect only when review_mode is BY_COMMIT. Optional values: [true|false]
  # instructions match specify file path
  path_instructions:
  - path: "**/*.js"
    instructions: 'Review the JavaScript code against the Alibaba JavaScript style
      guide and point out any mismatches'
  - path: "tests/**.*"
    instructions: |-
      Review the following unit test code written using the Mocha test library. Ensure that:
      - The code adheres to best practices associated with Mocha.
      - Descriptive test names are used to clearly convey the intent of each test.

1. Ignore specific files during review.

You can exclude certain files, such as configuration or test files, from AI reviews. Use the ignores keyword to define file path patterns. These patterns support glob syntax. For detailed rules, see glob.

reviews:
  # files that ignore to review, glob pattern
  ignores:
    - myfile.txt
    - myfolder/
    - "*.log"
    - "**/logs"
    - /debug.log

2. Set basic review configurations.

• language parameter: Specifies the language for review comments. Supported values are zh-CN (Chinese) and en-US (English). The default is zh-CN.

• problem_level parameter: Sets the minimum severity level for reported issues. The options, in decreasing order of severity, are BLOCKER (only blocker issues), CRITICAL (critical and above), MAJOR (major and above), and MINOR (all issues including minor suggestions). The default is CRITICAL. Setting this parameter to MINOR may generate a large number of comments.

• review_mode parameter: Sets the review mode. Supported values are DEFAULT (default review mode) and BY_COMMIT (single-commit review mode). The default is DEFAULT.

• enable_review_commit_message parameter: This parameter takes effect only when review_mode is set to BY_COMMIT. It determines whether to review the commit message in single-commit review mode. The default is true. When enabled, a Commit Message Suggestions section is added to the review summary, providing suggestions for each commit message.

reviews:
  # basic config
  language: "zh-CN" # Optional values: [zh-CN|en-US]
  problem_level: "CRITICAL" # Optional values: [BLOCKER|CRITICAL|MAJOR|MINOR]
  review_mode: "DEFAULT" # Optional values: [DEFAULT|BY_COMMIT] 
  enable_review_commit_message: true # Takes effect only when review_mode is BY_COMMIT. Optional values: [true|false]

3. Define custom review rules.

To align with your team's specific coding standards, you can use the path_instructions keyword to define custom review rules for specific file paths using natural language. These path patterns support glob syntax. For detailed rules, see glob.

reviews:
  # instructions match specify file path
  path_instructions:
  - path: "**/*.js"
    instructions: 'Review the JavaScript code against the Alibaba JavaScript style
      guide and point out any mismatches'
  - path: "tests/**.*"
    instructions: |-
      Review the following unit test code written using the Mocha test library. Ensure that:
      - The code adheres to best practices associated with Mocha.
      - Descriptive test names are used to clearly convey the intent of each test.

4. Sandbox configuration (Optional)

You can configure a sandbox to control the AI review runtime environment and enable advanced analysis capabilities. Enabling cross-file analysis may increase token consumption and processing time.

• enable parameter: Enables or disables running AI reviews in a sandbox environment. Supported values are true and false. The default is false. This option has no effect unless a specific capability switch is also enabled.

• enable_crossfile_analyze parameter: Enables cross-file change detection. Supported values are true and false. The default is false. This parameter is effective only when enable is true. When enabled, the AI review automatically detects breaking changes across files, such as changes in return value semantics, unhandled exceptions, or altered parameter order.

reviews:
  # Sandbox configuration (Optional)
  sandbox:
    enable: true
    enable_crossfile_analyze: true

More advanced sandbox capabilities will be available in the future.

Use AI-powered review

Limitations

The AI review feature has the following change size limitations:

1. Reviews are not supported for changes exceeding 1,000 modified lines.

2. Reviews are not supported for changes affecting more than 100 files.

Trigger a review

1. When you create a merge request, you can choose whether to immediately run an AI review. If you have configured branch rules in the Default AI Review Settings, the AI review option is automatically selected when the target branch matches a rule.

2. On the merge request details page, users with the Developer role or higher can manually trigger a review.

3. In a merge request comment, @mention Alibaba Cloud DevOps AI Assistant to trigger an AI review.

• If an AI review or commenting process is already in progress, you cannot trigger a new one until the current process is complete.

• An AI review cannot be re-triggered on the same version of the code. You must push new changes to start another review.

In the AI Review section of the right-side panel, click the Request Review button next to Alibaba Cloud DevOps AI Assistant to trigger a review.

Review results

1. A summary of code changes, including a list of modified files and their descriptions.

The review result also includes a sequence diagram that illustrates call relationships between the components involved in the changes.

2. Line-level issue detection and suggested fixes.

The AI Assistant adds comments directly to the code diff view, highlighting issues on specific lines and suggesting fixes. Each comment includes the severity level (for example, Critical | Bugs), a description, and a code diff with the recommended fix, such as adding a .refine() validation to the status field to restrict valid status IDs. The right-side panel shows the AI review status and the count of To Be Resolved issues. Inline review comments include Accept and False Positive buttons, which you can use to provide feedback on the AI review result.

3. Review summary

The review report summarizes findings with a breakdown by severity level: Blocker, Critical, Major, and Minor. Below the summary, each level lists its issues with descriptions, file paths, and line numbers.

4. Review modes: You can choose between the default review mode and the single-commit review mode. This can be specified in the review rule file using the review_mode keyword.

• Default review mode: This mode reviews all commits in a merge request as a single, consolidated code change. It is ideal for small, atomic changes.

• Single-commit review mode: This mode is ideal for complex merge requests with a large number of commits or extensive changes.

  • Provides a more granular review, making it easier to focus on small yet critical modifications.

  • Considers the commit message to better understand the developer's intent, leading to more relevant suggestions and fewer false positives. For example, the AI assistant might detect that the code FILE_MAX_SIZE = 1024 * 1024L (1 MB) contradicts the commit message, which specifies 1 GB. It would flag this as a Critical bug and suggest changing the code to FILE_MAX_SIZE = 1024L * 1024 * 1024 to match the intended requirement.

  • Reviews the commit message itself, providing suggestions to improve its quality and format. The tool generates a Commit Message Suggestions page that checks each commit for Format Issues (such as whether the subject line is in English and uses the imperative mood) and Content Quality (such as clarity and avoidance of redundant descriptions in the body). It then offers suggestions for improvement, like rewriting a Chinese subject line into an English imperative sentence, such as Add file meta info to file list API response.

Limitation of single-commit review mode: Compared to the default review mode, the single-commit review mode can significantly increase review time when there are many commits. In extreme cases, the processing time can be more than three times as long as that of the default review mode. Choose the mode that best balances review depth with performance for your project.

Enable the feature

As a repository administrator, go to Repository Settings > AI Assistant Settings and turn Use Project Convention Files on or off. The feature is enabled by default. If you disable it, newly initiated reviews no longer load project convention files. Other review capabilities remain available.

Supported project convention files

The AI Assistant automatically looks for the following files or rule directories in the repository root, the directories containing files changed in the current review, and each ancestor directory. No additional configuration in the review rules YAML file is required. Fixed file names and directory entry names support uppercase and lowercase variants.

Usage rules

• Files do not need to be included in the current changes: Even if a project convention file does not appear in the current diff, it can be discovered and used for review as long as it exists in the code version being reviewed.

• Apply conventions by directory scope: Root-level conventions apply to the entire repository. Conventions in a module directory apply only to that directory and its subdirectories. If applicable conventions conflict, those from a more specific directory take precedence. Path matching rules and activation conditions within the files are also used to determine their scope.

• Use the matching code version: The AI Assistant reads project convention files from the commit being reviewed. Newly added or modified files use their content from that commit. Deleted files are no longer used as active project conventions. In single-commit review mode, the project convention file version associated with each reviewed commit is used.

Enable the feature

Administrators of a repository can enable or disable review memory for the current repository by navigating to Repository Settings > AI Assistant Settings > AI Review > Review Memory.

Generate memories

In review comments, users can proactively inform the Alibaba Cloud DevOps AI Assistant about content they want remembered:

Please remember: When [situation] occurs, pay attention to [rule or requirement].

The Alibaba Cloud DevOps AI Assistant organizes natural language into a memory suitable for retention and displays the specific Add Memory content.

At this point, memories may have two states:

• Active: Memories take effect in the next review.

• Pending administrator approval: Requires administrator approval before taking effect in the next review.

Use memories

During reviews, the Alibaba Cloud DevOps AI Assistant automatically references all active memories and displays the specific Referenced Memory content.

Manage memories

Administrators of a repository can navigate to Repository Settings > AI Assistant Settings > AI Review > Review Memory to manage memory content.

Multiple operations are supported:

Memory content

To ensure memories are accurate and stable, prioritize long-term effective, clearly bounded, and reusable information.

• Recommended content to remember includes:

  • Development standards applicable long-term to the current repository or specific modules.

  • Business constraints related to historical compatibility, external systems, or online stability.

  • Technical solution boundaries explicitly chosen and repeatedly implemented by the team.

  • Risk points or verification items requiring continuous attention for specific change types.

  • Context repeatedly explained in reviews that will appear again in the future.

• Content not recommended for memory includes:

  • Temporary modification suggestions targeting only the current merge request.

  • Vague feedback without clear scope.

  • Personal emotional expressions or one-time discussion conclusions.

  • Issues already strongly constrained by tests, configurations, or static rules.

A high-quality memory should ideally contain four elements: