Export Configuration

Convert the current project's dotforge configuration into formats compatible with other AI coding tools.

Input

$ARGUMENTS contains the target format: cursor, codex, windsurf, or openclaw.

If no argument provided, show available targets and ask.

Step 1: Read current configuration

Read these files from the current project:

• CLAUDE.md — project instructions
• .claude/rules/*.md — contextual rules (strip YAML frontmatter)
• .claude/settings.json — permissions and hooks

If none exist, error: "No dotforge configuration found. Run /forge bootstrap first."

Step 2: Transform based on target

cursor → .cursorrules

Generate a single .cursorrules file at project root:

1. Extract content from CLAUDE.md (skip forge markers)
2. Append all rules from .claude/rules/*.md (strip YAML frontmatter — globs:, paths:, alwaysApply:, etc. — keep content)
3. Convert deny list to text: "DO NOT: read/modify files matching: .env, *.key, *.pem, credentials"
4. Convert hooks to text instructions: "Before executing bash commands, check for destructive patterns: rm -rf, DROP TABLE, force push"
5. Wrap in a single markdown document

codex → AGENTS.md

Claude Code reads AGENTS.md natively (v2.1.277+), so this file is no longer Codex-only — whatever you write here Claude may load too. Before generating:

• If CLAUDE.md is a symlink to AGENTS.md, or already contains an @AGENTS.md line: skip — the project is already sharing one file; say so and stop.
• If AGENTS.md exists without an import: warn that Claude ignores it (a CLAUDE.md is present) and that regenerating creates a second drifting copy. Recommend adding @AGENTS.md to CLAUDE.md and moving the shared content into AGENTS.md instead of exporting.
• Never include the .claude/settings.json permission lists in AGENTS.md when Claude will load it — they are enforced natively and only add stale duplicates.

Generate AGENTS.md at project root:

1. Start with project context from CLAUDE.md (skip forge markers)
2. Append rules as "## Rules" section
3. Convert permissions to "## Permissions" section: list allowed and denied commands — Codex-only; omit when AGENTS.md will also be read by Claude (see above)
4. Add "## Workflow" section from agent orchestration rules if present
5. Format as flat markdown (Codex expects simple instructions)

windsurf → .windsurfrules

Generate .windsurfrules at project root:

1. Same content extraction as cursor
2. Windsurf format is similar to .cursorrules — single markdown file
3. Add Windsurf-specific header: "You are an AI assistant working on this project."
4. Append all rules and converted hooks/permissions

openclaw → ~/.openclaw/skills/{project}/SKILL.md + workspace agent config

Generate an OpenClaw workspace skill for the current project:

1. Create skill directory: ~/.openclaw/skills/{project-slug}/
2. Generate SKILL.md with frontmatter:

   ---
   name: {project-slug}
   description: "AI assistant for {project-name} — {stack description}"
   user-invocable: true
   metadata: {"openclaw":{"requires":{"bins":["claude"]}}}
   ---

3. Body content:
   • Project context from CLAUDE.md (skip forge markers, keep substance)
   • Build/test commands as executable instructions
   • Rules converted to behavioral instructions (strip globs: frontmatter)
   • Deny list as explicit "NEVER read or modify" instructions
   • Hook logic as text: "Before executing commands, check for destructive patterns: ..."
   • Agent roles summarized: "For architecture decisions, think like an architect. For security, scan for OWASP top 10."
4. Add execution section:

   ## Execution
   When asked to work on this project:
   1. cd to {project-path}
   2. Use `claude --print "<task>"` to execute via Claude Code
   3. Return the result formatted for the current channel

Additionally, if the project has specific commands (.claude/commands/*.md), list them as available actions:

## Available commands
- /audit — audit project configuration
- /health — health check
- /debug — assisted debugging
- /review — code review

What OpenClaw export preserves vs loses

Step 3: Handle conflicts

Before writing:

1. Check if target file already exists
2. If exists: show diff preview and ask user to confirm overwrite
3. If user declines: suggest alternative filename (e.g., .cursorrules.dotforge)

Step 4: Report

Show:

Export complete: {{target}}
  Output: {{filename}}
  Sources: CLAUDE.md, {{N}} rules, settings.json
  Note: hooks and deny list converted to text instructions (no enforcement outside Claude Code)

Warn: "Exported rules are advisory only. Hook enforcement (destructive command blocking) only works in Claude Code."

Limitations

• Hooks cannot be enforced outside Claude Code — converted to text instructions
• Agent orchestration is Claude Code-specific — simplified to workflow instructions
• Stack-specific rules are included but glob-based auto-loading is Claude Code-only