feat(skills): add diff-based code review skill

[svetkis]

---

name: Add Code Review Skill
about: Add diff-based code review skill with stack detection and ANTI-HALLUCINATION evidence requirement
labels: enhancement, skill

Problem

Fast-ASDLC framework currently lacks a standardized code review skill. When AI agents review code, they suffer from:

• Hallucination: Reporting issues in unchanged code or inventing problems
• Inconsistent severity: No uniform classification (BLOCKER/CRITICAL/MAJOR/MINOR/NIT)
• Missing stack context: Generic reviews that ignore language/framework specifics (.NET CancellationToken, EF Core NoTracking, async void, etc.)
• No evidence requirement: Findings without exact line numbers or quoted snippets

Proposed Solution

Add a code-review skill under /.agents/skills/code-review/ that provides:

1. Diff-scoped analysis: Review ONLY + lines and directly related context
2. Stack detection: Language/framework-specific checks for .NET/C#, TypeScript, Python, Go, SQL
3. Severity classification: BLOCKER / CRITICAL / MAJOR / MINOR / NIT with clear definitions
4. ANTI-HALLUCINATION protocol: Every finding requires exact file path, line number, quoted snippet, verification method
5. Self-correction loop: "Can I point to exact line? Did I read the file myself?"
6. Verdict: APPROVED / APPROVED_WITH_NITS / CHANGES_REQUESTED

Acceptance Criteria

• Skill file exists at /.agents/skills/code-review/code-review.md
• Skill defines scope rules (only + lines)
• Skill defines severity levels with examples
• Skill includes stack-specific checks for at least .NET/C#, TypeScript, Python, Go
• Skill includes ANTI-HALLUCINATION evidence requirement
• Skill includes self-correction loop prompts
• Skill defines output format: [SEVERITY] Title | File:line | Rule | Evidence: snippet | Fix: action
• Skill defines verdict criteria
• Clean markdown, no lint errors
• LLM-friendly: concise, structured, minimal context window usage

Files Added

.agents/
└── skills/
    └── code-review/
        └── code-review.md

Methodology Alignment

• Everything-as-Code: Review procedures encoded as markdown skill
• Spec-Driven: Skill itself is a specification for review agents
• LLM-Friendly: Structured format with clear constraints minimizes hallucination

References

• Fast-ASDLC METHODOLOGY.md: Agentic review workflows
• Existing skills pattern: /.agents/skills/*/SKILL.md