Review Kedro PR

Review a Kedro pull request. Start with general review guidelines (any project), then apply Kedro-specific checks. Output findings to chat by default, or post to GitHub when asked.

Workflow

1. Identify the PR

Detect the PR from the current branch or accept a PR URL/number from the user:

gh pr view --json number,title,body,baseRefName,headRefName

If the user provides a PR URL or number, use that instead.

2. Gather the diff and existing comments

gh pr diff <number>
gh pr view <number> --comments
gh api repos/{owner}/{repo}/pulls/{number}/comments

Read existing review comments and discussion. Don't repeat what's already been said — skip issues that are already covered by existing comments. If you agree with an existing comment, you can reference it. If you disagree, explain why.

3. Analyze

Work through the general review guidelines first, then the Kedro-specific review areas. Focus only on changes introduced by this PR. Only flag issues not already raised in existing comments.

4. Verify findings

Before delivering, go through every finding one by one and check:

• Is this actually a problem, or a false positive?
• Is the file path and line number correct?
• Is the severity (Critical vs Suggestion) appropriate?
• Is the change in scope? gh pr diff includes cherry-picked or rebased-in commits from other branches. Cross-check with gh pr view <number> --json commits --jq '.commits[].messageHeadline' — if a commit clearly belongs to a different PR (e.g. references a different #NNNN), drop findings on those changes.

Drop anything you're not confident about. Fewer accurate findings are better than many noisy ones.

5. Format and deliver

Format findings using the "Output format" section at the end of this file. Two modes:

• Review only (default): Output all findings as chat messages. The user can read, edit, or discard before deciding whether to post.
• Review and post: When the user explicitly says "post", "submit", or "review and post", post findings directly to the PR on GitHub. See "Output format" for the exact commands.

---

General review guidelines

These apply to any project. Work through them in order.

1. Understand context and scope

• Understand what problem the PR is solving.
• Read the PR description thoroughly, including:
  • Dev notes
  • Linked issues
  • Any prior discussions or comments

2. Evaluate PR scope

• Check if the PR is focused on a single purpose.
• If there are too many changes or unrelated updates, suggest splitting into smaller PRs.
• Verify the changes actually align with the stated goal.

3. QA and validation

• Check if QA/testing steps are mentioned in the PR description.
• If not, ask the author to add clear steps to validate the changes.
• Mentally walk through the described QA steps and verify:
  • Expected functionality
  • Edge cases and failure scenarios
• Watch out for regressions or unintended side effects.

4. Code review and readability

• Go through the file changes and understand the logic.
• Suggest improvements where needed:
  • Better variable/function names
  • Cleaner structure or readability

5. Code design and modularity

• Look for opportunities to improve structure:
  • Can large functions be split?
  • Can logic be broken into smaller, reusable pieces?
• Suggest helper/utility functions if the same logic is repeated.
• Aim for simple, maintainable code.

6. Docstrings and inline comments

Match the style of existing docstrings in the same module — short, functional, and to the point.

Docstrings

• Check if docstrings are present on new/changed classes, functions, and methods.
• Docstrings must use Google style (Args:, Returns:, Raises:) to avoid rendering issues in the API docs.
• Keep docstrings concise. State what the function does, its arguments, return value, and any exceptions it raises. Do not restate what the signature already makes obvious.
• Do not include long explanations, design rationale, or implementation notes in docstrings. Those belong in PR descriptions or docs pages, not in code.
• For public APIs, ensure inputs and outputs are clearly documented. Add examples only when the behaviour is non-obvious and a short example genuinely helps — not as boilerplate.
• Suggest adding examples if helpful.

Flag as Suggestion when a docstring is verbose, repeats the function name, or documents design decisions instead of behaviour. Flag as Critical only when a public API is missing required documentation.

Inline comments

• Code should be self-explanatory through function names, docstrings, and clear structure. Readers should understand what it does without narrating comments.
• Do not add comments that restate the code (e.g. # increment counter above counter += 1).
• Inline comments are acceptable only for genuinely non-obvious logic — edge cases, platform quirks, or constraints that cannot be expressed clearly in code alone. See existing comments in files like kedro/utils.py for the expected bar.
• Flag unnecessary inline comments as Suggestion.

7. Testing and coverage

• Check if there are enough unit tests.
• Make sure tests cover:
  • Core logic
  • Edge cases
• Tests should be meaningful and easy to follow, not just for coverage.

---

Kedro-specific: PR checklist compliance

Check the PR against the Kedro PR template requirements.

RELEASE.md entry

Any change to code under kedro/ must have a corresponding bullet in RELEASE.md under # Upcoming Release, in the correct H2 section:

• ## Major features and improvements — new features, significant enhancements
• ## Bug fixes and other changes — bug fixes, minor changes, dependency updates
• ## Documentation changes — docs-only changes
• ## Community contributions — added by maintainers for external contributors

Format: * Description of the change.

If the PR modifies kedro/ files but RELEASE.md is not in the diff, flag as Critical.

Project documentation

If the PR changes public API behavior or adds new features, check whether the Kedro project docs (docs/) are updated. Docs use MkDocs with standard markdown — check for correct formatting to avoid rendering issues.

Backward compatibility

• Public API changes must not break existing callers without a migration path.
• Check for: removed/renamed parameters, changed return types, changed default values, removed public classes/functions.
• If breaking, there should be a deprecation warning or a migration note in RELEASE.md.

Kedro-Viz impact

If the change affects pipeline structure, metadata, catalog behavior, or dataset output, flag for Viz team coordination.

---

Kedro-specific: API usage

Check that the PR uses Kedro abstractions correctly.

• Datasets must subclass AbstractDataset and implement load, save, _describe. Versioned datasets subclass AbstractVersionedDataset.
• Runners must subclass AbstractRunner and implement _run and _get_executor.
• Hooks must use the @hook_impl marker from kedro.framework.hooks.
• Config loaders must subclass AbstractConfigLoader.
• Naming conventions should match existing patterns in the same module (method names, parameter names, class names).

For full API signatures and the complete list of base classes, read reference.md section "Kedro public API surface".

Non-obvious patterns: Some files use complex patterns on purpose (e.g. __getattr__, __init_subclass__ wrapping). If the PR touches kedro/io/core.py, kedro/framework/hooks/manager.py, or kedro/io/shared_memory_dataset.py, read reference.md section "Non-obvious patterns" before flagging anything — those patterns are intentional.

---

Out of scope

Do NOT flag:

• Pre-existing issues — only review changes in the diff.
• Security issues — handled by the separate kedro-security-review skill.
• Anything CI already checks mechanically — the following are handled by the kedro-babysit skill, which runs and fixes them automatically:
  • Linting and formatting (ruff)
  • Running unit tests (pytest)
  • Import-linter contract execution (architecture compliance)
  • DCO sign-off verification
  • Any other CI check with a pass/fail outcome

This skill is about reviewing code, not running checks.

---

Output format

Each finding gets an inline comment on the specific line, plus there's a summary of all findings. Delivery depends on the mode.

Review only (default) — output to chat

Present all findings as chat messages. For each finding, include:

• The file path and line number.
• Critical: or Suggestion: prefix.
• A concise explanation and suggested fix.

After all findings, output the summary using the template below.

Review and post — output to GitHub PR

Write a JSON file with the full review payload and post it via the script:

{
  "event": "COMMENT",
  "body": "## Kedro PR Review\n...(summary)...",
  "comments": [
    { "path": "file.py", "line": 42, "side": "RIGHT", "body": "**Critical:** ..." },
    { "path": "other.py", "line": 15, "side": "RIGHT", "body": "**Suggestion:** ..." }
  ]
}

Then post:

bash .agents/scripts/post_github_review.sh <review_json_file>

This creates one review with the summary as the main message and inline comments on specific lines. Delete the JSON file after posting.

Summary template

Used in both modes:

## Kedro PR Review
> Generated with `review-kedro-pr` skill.

### Overview
- **Critical:** <count> | **Suggestions:** <count>

### Critical (must fix before merge)
- [ ] `file/path.py:L42` — brief description

### Suggestions (consider improving)
- [ ] `file/path.py:L15` — brief description

### Notes
- General review: <summary or "no issues">
- PR checklist: <summary or "no issues">
- API usage: <summary or "no issues">

Classification

• Critical = must fix before merge (violations, missing requirements, bugs).
• Suggestion = worth improving but not blocking.
• If a section has no findings, write "No issues found."
• If a section has multiple points, use bullet points for clarity.

Communication guidelines

• Keep feedback clear and constructive.
• Ask questions instead of assuming.
• Be respectful of the author's approach while suggesting improvements.