yet-another-agent-harness: Skill for Claude Code

Design Principles Audit

Audit the codebase against well-known design principles. Produce a scored DESIGN_AUDIT.md with prioritized remediation actions.

Core principle: Investigate, do not pattern-match. Read the actual code to determine whether a violation is genuine, intentional, or a false positive. Only flag findings with specific file:line evidence.

Prerequisite check

git --version 2>/dev/null && git rev-parse --is-inside-work-tree 2>/dev/null

If not in a git repository, warn but proceed — git history will be unavailable but file analysis still works.

Step 1: Detect context

Determine language, framework, and paradigm — this governs which principles apply and how violations manifest:

ls go.mod package.json pyproject.toml Cargo.toml *.csproj 2>/dev/null

Read README.md and CONTEXT.md (if present) for architectural intent.

Determine audit scope from input:

• Full audit (default) — entire codebase, excluding vendor/, node_modules/, generated files
• Module audit — specific directory or package
• Delegated mode — called by tech-debt; focus on changed files and surface findings for the scorecard

For language-specific violation patterns per principle, see references/principles-guide.md.

Step 2: Scan for violations

For each applicable principle, read actual code and identify violations. Do not grep for keywords — trace call paths and responsibilities.

SOLID

• S — Single Responsibility: Find files/classes/functions with multiple distinct reasons to change. Signals: files > 300 lines mixing concerns, functions doing I/O + business logic + formatting.
• O — Open/Closed: Find code that requires modification (not extension) to add new behavior. Signals: long switch/if-else chains on type tags, no interface/abstraction at extension points.
• L — Liskov Substitution: Find subtypes that weaken preconditions, strengthen postconditions, or throw unexpected errors. Signals: overrides that do nothing, panic on valid input, type assertions before use.
• I — Interface Segregation: Find broad interfaces forcing implementors to satisfy methods they do not use. Signals: interface with > 5 methods, stub/no-op implementations.
• D — Dependency Inversion: Find high-level modules importing concrete low-level types. Signals: direct instantiation of dependencies, no interfaces at package boundaries, untestable code.