Skill migrator

# Install the Skill Migrator

Create the files below exactly as shown, relative to the repository root.
After writing them, tell the user which migrator skill path was installed and that future migrations must still start with a plan for approval.

## Files

### File 1 / 1: `.agents/skills/skill-migrator/SKILL.md`

```markdown
---
name: skill-migrator
description: Migrate Claude Code skills, commands, and plugins to OpenAI Codex with a plan-first approval gate.
---

# Skill Migrator

Use this skill when the user asks to migrate AI coding-agent skills from Claude Code into another coding-agent format.

# Skill migration request

Source: Claude Code skills, commands, and plugins.
Target: OpenAI Codex.

You are migrating AI coding-agent skills in this repository. Work repo-locally. Do not ask the user to upload files elsewhere.

## Approval gate

Do not create, edit, move, or delete files before the user approves your migration plan. Your first response must be a plan only, except when the target is custom and the target contract is missing. In that case, ask only for the missing target contract first, then produce the plan after the user answers.

## Discover source files

Inspect the repository for these Claude Code inputs:

- CLAUDE.md
- CLAUDE.local.md
- **/CLAUDE.md
- .claude/skills/**/SKILL.md
- .claude/skills/**/references/**
- .claude/skills/**/scripts/**
- .claude/skills/**/assets/**
- .claude/commands/**/*.md
- .claude/settings.json
- .claude/settings.local.json
- .claude-plugin/plugin.json
- .claude-plugin/marketplace.json
- plugins/**/plugin.json
- plugins/**/skills/**/SKILL.md
- plugins/**/commands/**/*.md
- plugins/**/docs/**/*.md
- plugins/**/skills/**/references/**
- plugins/**/skills/**/scripts/**
- plugins/**/skills/**/assets/**

Classify each discovered item as one of:

- portable skill
- Claude-specific skill
- Claude command workflow
- Claude plugin package
- Claude project instruction
- Claude settings/context file
- support file
- unsupported or ambiguous item

## Plan format

Return a migration plan with:

- discovered source files
- proposed target files
- how every Claude skill, command, plugin skill, plugin command, project instruction, settings file, doc, reference, script, and asset will be handled
- content changes needed for the target agent
- risks or assumptions
- validation commands you will run after writing files
- explicit approval question at the end

## Target output

Use these target paths:

- .agents/skills/<slug>/SKILL.md

Codex target rule: use the Codex `skill-creator` skill/workflow before writing target skills when it is available. Use it to shape each migrated skill into a valid Codex Agent Skill, then write the final files under `.agents/skills/<slug>/`.
If `skill-creator` is not available in the current agent environment, say so in the plan and follow the Codex Agent Skills structure directly.

## Conversion rules

- Preserve the user's source files unless the user explicitly asks you to remove them.
- Preserve support files losslessly when the target format supports skill directories.
- Convert Claude commands into target-native skill workflows or AGENTS.md sections; do not pretend slash commands exist in agents that do not support them.
- Treat CLAUDE.md files and Claude settings as migration context. Convert durable project guidance into the target's repository-instruction format when appropriate, and do not copy local-only or machine-specific settings blindly.
- For Claude plugins, inspect `plugin.json` and migrate contained skills, commands, docs, references, scripts, and assets. Do not recreate Claude marketplace metadata unless the target is Claude Code.
- Keep frontmatter minimal and target-native. For Codex and Claude Code skills, use `name` and `description` frontmatter and put extra provenance in markdown body text.
- Use safe relative paths only. Reject paths with absolute roots, drive letters, null bytes, or `..` segments.
- If two source skills map to the same slug, stop and ask the user how to resolve the collision.

## After approval

After the user approves the plan, write the files, run the proposed validation, then summarize:

- files created or changed
- source files left untouched
- warnings that remain
- commands/tests that passed or could not run
```