code-audit-expert

name

code-audit-expert

description

Critically audit refactorings and design decisions against the real codebase, verify claims vs code, score only relevant quality axes, and write evidence-based tech reports. Use after refactors, when reviewing agent-reported work, or when asking for architecture / ISP / DRY / token-cost / reliability feedback. For builtin MCP tool audits, prefer critique-builtin-tool or lean-builtin-tool-auditor instead.

Code Audit Expert

Audit completed work against the codebase. Prefer skepticism over praise. Every claim needs a file/line or a measured number.

Scope routing

• Builtin MCP tools (schema, hints, Tool Design Manifesto): use critique-builtin-tool or lean-builtin-tool-auditor.
• This skill: architecture fit, claim verification, DRY/ISP, token/cost impact, side effects after refactors or feature work (TS/React, Rust, shared libs).

Workflow

1. Gather context — Changed files, PR notes, agent summary. Treat summaries as claims, not facts.
2. Verify against code — Open the actual implementations. Build a claim→evidence table (see template §2). Mark invents, mismatches, and unstated gaps.
3. Analyze only what applies — Interface fit (ISP), duplication/abstraction (DRY), cost (tokens/caching), reliability/side effects. Skip irrelevant axes.
4. Report — Follow reporting-template.md. Write under .libragent/work/ (e.g. .libragent/work/code_audit_report.md). Do not invent paths.

Hard rules

• No fabricated metrics. Do not invent token savings, coverage %, speedups, or “100% tests” unless you measured them in this session. Prefer qualitative statements or cite the command/output.
• Score only relevant axes. Irrelevant rows = N/A with one-line reason. Never force a 5/5 for flavor.
• Cite evidence. Prefer path:line or short quoted snippets. Star ratings without citations are invalid.
• Surface risks. Side effects, regressions, and remaining debt go in the report even when the change is otherwise good.
• Language. Skill instructions are English. Match the user’s language for the written report (Korean user → Korean report is fine).

Critique checklist (inline)

Use when scoring; omit items that do not apply.

• Claims in the agent/PR summary match the code
• Public contracts (schemas, APIs, tool responses) stay coherent for consumers
• Shared logic is extracted once (not copy-pasted) without over-abstraction
• Hot paths / LLM context avoid redundant payload echo
• Error handling and limits (size caps, LCS bounds, etc.) are explicit
• Tests cover the new contract fields and failure modes that matter
• Feature-flag / build-variant behavior is consistent where dual builds exist

Consumer / dependency check

When changing a core type or service (example: BaseAIService / src/lib/ai-service/base-service.ts), inspect call sites for correct type narrowing and broken assumptions—not only the edited file.