scalpel

Surgical AI for your codebase. Plugs into any project, performs a 12-dimension diagnostic scan (git forensics, architecture, tech debt, security, infrastructure), delivers a Codebase Vitals health sco

SKILL.md

GitHub

Scalpel — Surgical AI for Your Codebase

12-dimension diagnostic. Adaptive surgical teams. Zero side effects.

Installation

# Claude Code
git clone https://github.com/anupmaster/scalpel.git ~/.claude/skills/scalpel

# Codex CLI
git clone https://github.com/anupmaster/scalpel.git ~/.agents/skills/scalpel

# Any project (pendrive style)
cd your-project && git clone https://github.com/anupmaster/scalpel.git .scalpel && .scalpel/install.sh

Usage

> Hi Scalpel, start work
> scalpel scan              # Diagnostic only, no changes
> scalpel skip recon        # Jump to task execution

Standalone Scanner (Zero AI)

./src/scanner.sh                # Pretty terminal output
./src/scanner.sh --json         # Machine-readable
./src/scanner.sh --score-only   # Just the number: "83"

What It Does

  1. Diagnose — Scans 12 dimensions: stack, architecture, git forensics, database, auth, infrastructure, tests, security, integrations, agent ecosystem, code quality, performance
  2. Consult — Delivers Codebase Vitals health score, asks targeted questions
  3. Assemble — Designs a custom surgical team calibrated to YOUR project
  4. Operate — Spawns parallel agents in isolated worktrees with file jurisdiction
  5. Close — Delivery report, session memory saved, optional regression monitoring

Features

  • Standalone scanner — Pure bash, zero AI, zero tokens. Runs anywhere.
  • 12-dimension diagnostic — Deeper than any linter. Reads git history, deleted files, abandoned branches.
  • Adaptive teams — Not templates. Teams shaped by YOUR project's needs.
  • File jurisdiction — Each agent owns specific files. Zero merge conflicts.
  • Agent scoring — Quality enforcement. Score < 50 = agent terminated.
  • Session memory — Remembers previous sessions. Gets smarter over time.
  • Pendrive architecture — Plug in, work, unplug. Zero trace.
  • 7 agent adapters — Claude Code, Codex, Gemini, Cursor, Windsurf, Aider, OpenCode.
  • GitHub Action — Automated PR health checks.
  • Extensible config — Custom scan dimensions, team roles, scoring rules.

Permissions

Read-only during diagnostic. Write access only during surgical team operation. Never modifies your existing CLAUDE.md, agents, commands, or settings. Never commits itself to your git history. Uninstall leaves zero trace.