Code Review AI Skills
Relevant source files
The following files were used as context for generating this wiki page:
Code Review AI Skills represent a specialized suite of autonomous capabilities within the Atmos platform designed to perform senior-engineer-level analysis of codebase changes. These skills automate the detection of security vulnerabilities, logic errors, architectural violations, and code smells across diverse technology stacks. By leveraging Git metadata and source-matched checklists, the AI provides structured feedback to developers through the Atmos interface or via generated Markdown reports.
The system is built on a "Review-First" philosophy, where the AI acts as a researcher and analyst before suggesting or implementing any fixes. It supports multiple review personas—ranging from general Atmos platform reviewers to "Fullstack Reviewers" and "Code Review Experts"—each tailored with specific checklists for frontend, backend, security, and SOLID principles.
Sources: skills/atmos-code-review/SKILL.md:1-12, skills/code_review_skills/fullstack-reviewer/SKILL.md:14-20, skills/code_review_skills/code-review-expert/SKILL.md:7-11
Review Architecture and Workflow
The code review process follows a structured multi-step workflow that ensures the AI operates with high context and accuracy. The system begins by scoping the review through Git commands, detecting the project's technology stack, and applying specialized checklists based on the categorized files.
Systematic Review Execution Flow
The following diagram illustrates the standard execution path for an AI code review skill:
This workflow ensures that the AI does not rely on assumptions. For every finding, the system is instructed to read the actual code, trace data flow, check related files (such as middleware or DTO layers), and verify that the identified fields or properties actually exist in the type definitions.
Sources: skills/code_review_skills/fullstack-reviewer/SKILL.md:32-75, skills/code_review_skills/fullstack-reviewer/SKILL.md:175-190, skills/atmos-code-review/SKILL.md:23-45
Categorization and Checklists
The AI skills utilize a directory of reference documents to perform deep analysis. Files are classified into categories to ensure the correct "senior engineer lens" is applied.
| Category | Typical Indicators | Targeted Analysis |
|---|---|---|
| Frontend | .tsx, .jsx, .vue, .ts, .css | Component design, state management, hooks, accessibility, and rendering performance. |
| Backend | .rs, .go, .py, .java, .rb | API design, concurrency, database queries, and error propagation. |
| Security | Auth modules, API keys, User Input | XSS, injection, CSRF, secrets leakage, and race conditions. |
| Architecture | Core abstractions, interfaces | SOLID principles, coupling, cohesion, and dependency management. |
| Infra/Config | Dockerfile, .yaml, .toml | Security and correctness of infrastructure definitions. |
Sources: skills/code_review_skills/fullstack-reviewer/SKILL.md:58-85, skills/atmos-code-review/SKILL.md:47-60
Severity Level Framework
Findings are prioritized using a standard P0-P3 severity scale to prevent "review fatigue" and highlight blocking issues effectively.
- P0 Critical: Security vulnerabilities or data loss risks. These must block a merge.
- P1 High: Logic errors or significant architectural violations.
- P2 Medium: Code smells and maintainability concerns.
- P3 Low: Style, naming, and minor nitpicks.
Sources: skills/code_review_skills/fullstack-reviewer/SKILL.md:22-30, skills/code_review_skills/code-review-expert/SKILL.md:13-21
Specialized Review Personas
The system supports multiple "skills" that act as different personas depending on the user's needs or the platform's triggers.
Fullstack Reviewer
A project-agnostic skill designed for universal review. It performs analysis across any language or framework by inspecting project configuration files (like package.json, Cargo.toml, or go.mod) to dynamically load appropriate checklists.
Code Review Expert
This skill focuses specifically on "Clean Code" metrics. It utilizes specialized references such as solid-checklist.md and removal-plan.md to identify:
- SRP Violations: Overloaded modules with unrelated responsibilities.
- OCP Issues: Frequent edits to add behavior instead of using extension points.
- Removal Candidates: Identifying unused or redundant code for deletion.
Sources: skills/code_review_skills/fullstack-reviewer/SKILL.md:14-18, skills/code_review_skills/code-review-expert/SKILL.md:32-45
Reporting and Integration
The results of an AI review are formatted into a structured Markdown report. When triggered by the Atmos UI, the report is written to a specific OUTPUT_FILE provided by the platform.
Report Structure
The generated report follows a strict template to ensure readability:
- Header: Meta-information including Date, Project Name, and Overall Assessment (APPROVE/REQUEST_CHANGES).
- Summary: Metrics showing the count of files reviewed and findings by severity.
- Findings by Severity: Detailed breakdown with file paths, categories, and suggested code fixes.
- Positive Highlights: Mandatory section acknowledging clean implementations or clever solutions.
Sources: skills/code_review_skills/fullstack-reviewer/SKILL.md:104-165, skills/atmos-code-review/SKILL.md:73-86
System Integration Sequence
The interaction between the user interface and the backend review engine is mediated through WebSocket actions and specific UI components like CodeReviewDialog.
Summary of AI Skill Capabilities
Code Review AI Skills transform raw Git diffs into actionable technical debt management plans. By using context-aware checklists and a tiered severity framework, the system ensures that AI-generated feedback remains relevant to the project's established patterns while maintaining high standards for security and architectural integrity. The distinction between various review personas allows developers to choose between general quality checks or deep architectural "expert" reviews.
Sources: skills/code_review_skills/fullstack-reviewer/SKILL.md:200-215, skills/code_review_skills/code-review-expert/SKILL.md:100-115