This file is a maintained navigation index. It is not a cached specification, a bibliography to reproduce, or permission to copy upstream prose. Open the relevant live pages before acting, record the URLs and verification date in the pull request, and keep only repository-owned decisions in this repository.
Read the smallest complete set that covers the affected surface.
| Surface | Live pointer |
|---|---|
| Plugin overview and builder routing | Build plugins |
| Skill and plugin distinction and explicit invocation | Skills and plugins |
Built-in image generation and $imagegen invocation |
Image generation |
| Plugin boundaries and component model | Plugin architecture |
| Skill format, triggers, resources, and tests | Build skills |
Skill loading, distribution, and the agents/openai.yaml keys interface, policy, and dependencies |
Build skills for Codex |
| MCP tools, authentication, and server behavior | Build an MCP server |
| Optional MCP-backed UI | Add optional UI |
| Manifests, paths, marketplaces, and packaging | Package plugins and marketplaces |
| Local connection and behavioral testing | Connect and test plugins |
| Durable repository instruction discovery | Custom instructions with AGENTS.md |
| Lifecycle extension and trust | Hooks |
| Installation and supported plugin surfaces | Use plugins |
Installed CLI inventory and /skills or /plugins behavior |
Codex CLI command reference |
| Scheduled task behavior and management | Scheduled tasks |
| Native import behavior from another agent | Import from another agent |
| Sandbox, approvals, and action boundaries | Agent approvals and security |
| Native project instruction and configuration discovery | Advanced configuration |
The marketplace is public and Git-backed, but it is not submitted to OpenAI's universal public directory. Keep submission requirements as a pointer for compatibility review only unless repository policy changes.
Use these when a component relies on the corresponding open standard. Agent Plugins is the packaging contract that compatible clients load; Agent Skills and the Model Context Protocol are the separate component contracts it packages, and reading one does not cover a claim that rests on another:
OpenAI's current product contract still controls how Codex packages, discovers, trusts, and runs those components.
Use public implementations to test assumptions after reading the product contract. Treat examples as evidence, not templates to copy blindly:
Read the current official source-host documentation for every migrated component and record the exact pages in the migration pull request:
If a source host moves a page, update this index in the same change that relies on the new location. Do not use third-party tutorials to establish a host contract when official documentation exists.
Use these only when the architecture decision they address is material. They do not override Codex documentation:
- Fetch live pages for every plugin action covered by
AGENTS.md; do not rely on this index alone. - Record URLs and the date checked, not quotations or copied examples.
- Verify CLI syntax with the installed release when commands are involved.
- When a pointer is unavailable, search only the upstream owner's official properties before using another source.
- When the current target contract remains uncertain, stop the contract change and record the unresolved question. Do not turn an assumption into policy.
- Review this index whenever a plugin surface changes and during each migration.