Codex capability router

Codex Capability Router

Use this skill to choose capabilities before doing substantial work. Apply active system, developer, user, authorization, privacy, and safety constraints first. This is a routing layer, not a replacement for the selected capability, project skill, or repository evidence.

Workflow

1. Identify task signals:

   • user intent, domain, requested artifact, current workspace, file types, and explicit tools mentioned
   • installed skills in the current session and plugin skills already exposed by Codex
   • local plugin cache and registry data when the answer depends on installed or candidate plugin inventory
   • Treat an explicit installed capability name as a mandatory route. Treat a direct dominant-artifact or specialist-workflow match as mandatory when the capability materially improves the result. Incidental overlap remains optional.

2. Prefer installed capability:

   • If a bundled or installed skill/plugin clearly matches, use it directly and tell the user briefly which one is being used.
   • Apply references/installed-capability-execution-gate.md. Selection alone is not completion: use the exact entrypoint and relevant native tools, modules, templates, examples, or assets, then preserve concrete use evidence.
   • For Product Design work, start with the Product Design index or get-context skill, then route to ideate, image-to-code, audit, prototype, research, share, or URL-to-code as needed.
   • For design work, treat OpenAI image generation as a normal design option for visual directions, realistic product/interface references, mood boards, campaign visuals, and asset exploration; do not default only to wireframes or line-art sketches.
   • For files and artifacts, prefer the installed document, spreadsheet, presentation, PDF, GitHub, browser, computer-use, or sites plugin skill instead of hand-rolling the workflow.
   • When a local external-skill registry is configured, apply references/external-skill-registry.md so overlapping skills have one owner and heavy workflows are loaded only when needed.

3. Use the local registry when routing is unclear:

   • Build or refresh installed capability data:

python "$env:USERPROFILE\.codex\skills\codex-capability-router\scripts\build_capability_registry.py"

• Query a task:

python "$env:USERPROFILE\.codex\skills\codex-capability-router\scripts\query_capability_router.py" --task "product design prototype from a screenshot"

4. Discover candidate plugins only as suggestions:

   • Use scripts/refresh_plugin_candidates.py or public-source research to find official or community candidates.
   • Never auto-install, auto-enable, or auto-trust candidate plugins, MCP servers, apps, or hooks.
   • When suggesting a candidate, state the project function in plain language and ask the user whether to install or investigate it.

5. Continue with the selected workflow:

   • Load the matched skill/plugin instructions.
   • Follow every directly relevant required, must, always, or hard-gate reference. Prefer installed-version native resources over a lower-fidelity hand-written substitute.
   • If no installed capability fits, proceed with normal local evidence and general best practices.
   • If a selected capability cannot be used natively, state the concrete incompatibility or authorization/cost boundary before falling back. Do not claim it was used when it was only mentioned or read.

References

• references/routing-categories.md: task signals and preferred installed capabilities.
• references/installed-capability-execution-gate.md: native-use, source, fallback, and proof contract after capability selection.
• references/plugin-discovery-policy.md: discovery sources, trust levels, and install boundary.
• references/external-skill-registry.md: selected external skill routes, precedence, and side-effect boundaries.