The repository root is the canonical Agent Skill. Do not maintain separate instruction copies for each agent.
- Codex: user
~/.agents/skills/suffuseor project.agents/skills/suffuse. - Claude Code: user
~/.claude/skills/suffuseor project.claude/skills/suffuse. - OpenClaw: it reads the shared user and project
.agents/skillsroots. The managed installer deliberately shares that link with Codex.
The repository also includes a Claude skills-directory plugin manifest. Claude Code can load the root SKILL.md as its single skill and start the bundled .mcp.json server. Codex and OpenClaw use the standard Agent Skill directly.
Preview destinations without writing:
node scripts/manage.mjs plan --target all --scope userInstall canonical symbolic links:
node scripts/manage.mjs install --target all --scope userCheck the runtime without exposing the API key:
node scripts/manage.mjs doctorIf credentials live in a private .env, run node --env-file=/private/path/.env scripts/manage.mjs doctor. The manager reports only whether the key is configured, never its value.
Removal is conservative: uninstall removes only links that still point at this exact repository. It refuses real directories, files, or links to another source.
Use --scope project --project /absolute/project/path for repository-specific discovery. The default symlink workflow is for a trusted local checkout; registry or marketplace distribution should package real files rather than symlinks.
Run node scripts/manage.mjs mcp-config for a Codex configuration snippet and Claude/OpenClaw notes. Read mcp.md before enabling live tools.
Primary specifications and product documentation: