mcpField
R
review

review

github.com/keboolaTesting & quality17191 installsUpdated today
Repository

About review

review is a skill for reviewing Keboola component code, combining code quality and backward compatibility checks in one pass. It requires ruff checks first and sets strict rules for protecting identifying information in public repositories.

Key features

  1. 01Combined quality and compatibility review
  2. 02Run ruff checks first
  3. 03Analyze configuration and output changes
  4. 04Group findings by risk level
  5. 05Require anonymized telemetry data

Use cases

  1. 01Reviewing a component pull request
  2. 02Checking compatibility of configuration changes
  3. 03Assessing change risks without telemetry

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/keboola/ai-kit --skill review

After installing, type/reviewin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

references/
breaking-changes.md
code-quality.md
telemetry.md

Details

From the project README

Keboola Component Reviewer

Run both code quality and backward compatibility reviews in a single pass, unless the user asks for only one.

CRITICAL: Repos are PUBLIC. Never write client names, project names, stack URLs, or any identifying information anywhere — including PR comments. Anonymized aggregate numbers only.

1. Scope

Default: git diff $(git merge-base HEAD main)..HEAD for a PR, or git diff for unstaged changes. Specified: files or directories the user provides.

2. Automated Checks (run first)

Run ruff before doing any manual review — it catches formatting and linting mechanically:

uv run ruff format --check . 2>&1
uv run ruff check --target-version py313 . 2>&1

If either fails, report it as a single item: "run ruff format . && ruff check --fix .". Don't enumerate individual formatting or lint violations manually — that's ruff's job. Skip any review findings that ruff would catch.

3. Code Quality Review

Focus on what ruff doesn't catch. In priority order:

Blocking — architecture violations:

  • run() is not a clean orchestrator (should read like a table of contents, < 30 lines)
  • Clients or configuration initialized inside run() instead of __init__ — breaks sync_actions
  • Config not encapsulated in a typed object (Pydantic BaseModel or dataclass)

Important — quality improvements:

  • Missing type hints on public methods
  • Deprecated typing imports (typing.List, typing.Dict, typing.Optional) — use built-ins
  • Config values accessed as raw self.configuration.parameters.get(...) scattered through the code

Only report issues with confidence ≥ 60. See code-quality.md for detailed patterns and examples.

4. Backward Compatibility Review

Step 1: Find component IDs
grep -E 'KBC_DEVELOPERPORTAL_APP|APP_ID' .github/workflows/push*.yml

Handles single env var, multiple suffixed vars, and matrix strategy patterns.

Step 2: Analyze diff
git diff $(git merge-base HEAD main)..HEAD -- \
  'component_config/' 'src/configuration.py' 'src/component.py' 'Dockerfile' '.github/workflows/'

Check each changed file against the breaking change vectors in breaking-changes.md: config schema, Pydantic models, sync actions, output tables, state file, Dockerfile.

Step 3: Telemetry

If Keboola MCP is available, query telemetry for each component ID — active configs, job stats, and property usage for any changed/removed field. See telemetry.md for exact queries and anonymization rules.

When MCP is unavailable, state it explicitly and treat all changes conservatively (assume configs exist).

5. Output

Start with a 1-2 sentence overall assessment, then:

Code quality — TODOs grouped by Blocking / Important / Nice-to-Have:

### TODO N: [Title]
**Location:** `src/component.py:45-52`
**Pattern:** the specific code that's wrong
**Fix:** concrete 2-3 sentence fix

Backward compatibility — findings grouped by HIGH / MEDIUM / LOW / SAFE risk. Include telemetry table when available. End with verdict: APPROVE / REQUEST CHANGES / WARN.

Use a constructive tone — direct but kind, giving the author agency. "I'd personally..." over "you must...". "LGTM" when there's nothing to flag.

Comments

Markdown supported