youlianvr

code-critic

MUST USE when user asks to review code, find bugs, check for issues, audit code, code review, find bugs, check code, review, re-review, critique this, review this file, is this code good, what's wrong with, check for bugs, find problems. Sends code to proxy:4000 (deepseek-v4-flash) for merciless review. Returns structured report: score (0-10), critical count, each issue with severity/file/line/description/suggestion.

youlianvr 1 1 Updated 2d ago
GitHub

Install

npx skillscat add youlianvr/oper-share/code-critic

Install via the SkillsCat registry.

SKILL.md

Code Critic — merciless code reviewer

Sends files to proxy:4000 for review. Model: deepseek-v4-flash (free, local).

When to use

  • "review this code" / review this code

  • "find bugs in proxy.py" / find bugs in proxy.py

  • "check what's wrong with this file" / check what's wrong with this file

  • "code review before commit" / code review before commit

  • "security audit" / security audit

Commands

Review a single file


python _scripts/critic.py path/to/file.py

Review a directory (recursively)


python _scripts/critic.py src/

With a focus hint


python _scripts/critic.py proxy.py --context "look for security issues and resource leaks"

JSON output (for scripts)


python _scripts/critic.py src/ --json

Different model


python _scripts/critic.py file.py --model glm-5.2

What it returns


{

  "overall_score": 4.0,

  "critical_count": 3,

  "total_issues": 27,

  "passed": false,

  "summary": "Code has several critical issues...",

  "issues": [

    {

      "severity": "critical",

      "file": "proxy_probe.py",

      "line": 18,

      "description": "Hardcoded port defaults to 4001...",

      "suggestion": "Use the same default as proxy.py (4000)"

    }

  ]

}

Severity: critical > major > minor > nit.

passed = true means 0 critical issues. The score is always low by design

(harsh reviewer).

Requirements

  • Proxy on localhost:4000 must be running

  • Model deepseek-v4-flash available on the proxy (check: curl -s localhost:4000/v1/models)

  • httpx installed (pip install httpx)

Interpreting results

The critic is merciless by design. It ALWAYS finds issues, even in good code.

  • critical > 0 → real bugs that need fixing

  • major > 0 → serious architectural/design problems

  • minor > 0 → style, readability

  • nit > 0 → taste

Not every finding is the truth. ~15% are false positives (hallucinated

issues, missing context). Always verify before fixing.

git pre-commit integration


# .git/hooks/pre-commit — review all staged Python files

for f in $(git diff --cached --name-only -- '*.py'); do

    python _scripts/critic.py "$f" --json || exit 1

done