Mend agent skills

Mend Agent Skills

Purpose

This skill iterates through existing agent skills, verifies that their embedded commands (e.g., make lint, npm test) are still valid in the current codebase, and updates them if they are broken or outdated.

When to Use

• When the Makefile or task runner configuration (e.g., package.json, pyproject.toml) has changed.
• When an agent reports that a command in a skill failed to execute.
• As a periodic health check to prevent "bit rot" in the agentic layer.

Instructions

1. Identify Target Platform

• Check for existing configuration directories:
  • .claude/ -> Claude Code
  • .cursor/ -> Cursor
  • .gemini/ -> Gemini CLI
  • .codex/ -> Codex
• If multiple or none are found, or if you are unsure, use AskQuestion to ask the user which platform to prioritize.

2. Audit Skills

• Iterate through each SKILL.md file in the platform's skills directory (e.g., .claude/skills/).
• Identify commands defined in tables or instruction blocks. Look for keywords like Command, Step, or code blocks containing shell commands.

3. Verify and Mend Commands

• For each extracted command:
  • Determine its intended working directory.
  • Verify: Check if the command/target is still valid.
    • If it's a make command, check the Makefile.
    • If it's an npm/pnpm/yarn command, check package.json.
    • If it's a language-specific tool (e.g., pytest, ruff), verify the tool is configured.
  • Mend:
    • If the command is invalid, search for a replacement (e.g., if make test became make unit-test).
    • If a replacement is found, update the SKILL.md file using StrReplace.
    • If no replacement is found, mark the skill as "Broken" and report it.

4. Final Report

• Summarize:
  • Total skills audited.
  • Commands successfully mended.
  • Skills requiring manual intervention (no replacement found).