Code Review

Code Review Skill

Perform structured, objective code reviews for Flutter/Dart projects following a repeatable checklist.

When to Use

Use this skill when:

• Asked to review a pull request, merge request, or branch.
• Evaluating changed, added, or deleted files for correctness and quality.
• Auditing a diff before merging.
• Checking whether new code meets project standards.

Review Workflow

Step 1 — Validate branch and merge target

1. Confirm the current branch is a feature, bugfix, or PR/MR branch — not the project's primary branch (e.g. main, master, develop).
2. Verify the branch is up-to-date with the target branch (no unresolved conflicts).
3. Identify the target branch for the merge.

Checkpoint: If the branch is behind the target, flag it before proceeding.

Step 2 — Discover changes

1. List all changed, added, and deleted files.
2. For each change, look up the commit title and review how connected components are implemented.
3. Analyze the change: is it clear why the change was made? If not, dig into the connected methods and files until it is. When you report, name which connected files/methods you analyzed and why — this shows the change was understood, not assumed.
4. Never assume a change is correct without investigating the implementation.
5. If a change remains unclear after investigation, note this explicitly in the report.

Step 3 — Review each file

Iterate through each changed file. For every file, verify the following:

For generated files (e.g., *.g.dart, *.freezed.dart): confirm they are up-to-date and not manually modified.

Scope discipline: Your job is not to comment on every change — it's to find errors and concrete improvement areas and comment on those. Don't manufacture comments where the code is fine.

Flutter-specific checks

(Note: The following is just an example using Bloc/Cubit; apply similar principles to Riverpod, Provider, or your chosen state management package.)

dartCopy code
// BAD — rebuilds entire tree on every state change
BlocBuilder<MyCubit, MyState>(
  builder: (context, state) => EntireScreen(state: state),
);

// GOOD — scope rebuilds to the widget that actually changes
BlocSelector<MyCubit, MyState, String>(
  selector: (state) => state.title,
  builder: (context, title) => Text(title),
);

• Verify Key usage on dynamically generated widgets.
• Check that dispose() is called for controllers, streams, and animation controllers.
• Confirm const constructors are used where possible.

Code Smells Reference

For each file, check for common code smells. Use refactoring.guru/refactoring/smells for definitions and suggested refactorings.

Step 4 — Evaluate the overall change set

1. Verify the change set is focused and scoped to its stated purpose — no unrelated changes.
2. Check that the PR/MR description accurately reflects the changes.

Test coverage

Verify test coverage explicitly — this is easy to skip and easy to fake, so be deliberate:

• For any new logic or significant change, search for the corresponding test file(s) and confirm tests actually exist.
• Check that tests cover the changed functionality including edge cases, not just the happy path.
• Evaluate whether tests could actually fail against real code, or only verify mocked behavior (a test that asserts a mock returns what the mock was told to return proves nothing).
• If tests are missing or insufficient, comment on the lack of coverage — don't let it pass silently.

Step 5 — Verify CI and tests

1. Ensure all tests pass in CI.
2. Check for new analyzer warnings or lint violations.
3. Fetch official documentation when unsure about best practices for a package.

Checkpoint: If CI is red or tests are missing for new logic, flag as a blocking issue.

Wrap-Up

After the per-file pass, decide the outcome:

• If everything looks good and no changes are needed: post an overall conclusion comment summarizing what the MR is about (what was done) plus any observations, and approve the MR.
• If the new changes are clean but you spotted smells in existing code: include those as optional refactor suggestions rather than blockers.
• If issues were found: summarize the key concerns clearly so the author knows what to address first.

Feedback Standards

• Be objective and reasonable — avoid automatic praise or flattery.
• Take a devil's advocate approach: give honest, thoughtful feedback.
• Provide clear, constructive suggestions for every issue found.
• Include requests for clarification for anything unclear.
• Classify each finding by severity: suggestion, minor, or major.

Output Format

By default, provide the review as a chat response — a structured response covering each file:

1. Summary — what changed and why.
2. Issues — each with severity (suggestion / minor / major) and a concrete fix suggestion.
3. Questions — specific clarification requests per file.
4. Verdict — one of: Approved, Approved with suggestions, or Changes requested.

Posting comments online (opt-in only). After presenting the chat review, ask the user whether they'd prefer you to also post these comments online on the PR/MR — so the team can see them, review them, and reply. Only post online if the user explicitly says yes. Never post to the platform on your own initiative.

When the user does opt in, post issues as inline comments anchored to the right file and line (use proper position fields), with the conclusion/key-concerns as a top-level review comment and an approval when warranted. This requires a review-bot access token for the platform (GitHub/GitLab); if one isn't configured, let the user know and ask them to set it up before posting.

Token safety. The token is a secret. You may check whether it exists and report its length to confirm it's configured, but never read, echo, log, print, or otherwise reveal the token value — not in chat, not in a file, not in a commit. Pass it to curl only by referencing the env var (e.g. $GITLAB_TOKEN), never by inlining the literal value, and avoid curl -v/--verbose (it prints the auth header). This is enforced by a PreToolUse hook (scripts/protect-token.sh) that blocks any Bash command which would expose the value. See the "Handling the token safely" section in each reference file for the safe existence/length check.

The hook fires in both the Claude Code CLI and the Agent SDK. (SDK apps that set settingSources/setting_sources explicitly must include "project" for skill hooks to load; it's included by default.)

For platform-specific API details, curl formats, and approval steps, follow:

• GitLab → references/gitlab-posting.md (uses the GITLAB_TOKEN env var)
• GitHub → references/github-posting.md (uses the GITHUB_TOKEN env var)