Skills Router Bridge: OpenAI Codex CLI
Trigger: user text starting /skills-router or skills-router is a
registry/routing request.
Cheapest path:
- Prefer MCP
run_slash_commandwith the full user text. - For structured calls, use
refine_routesorroute_task; never paste route tables. - Fallback:
skills-router chat "<request>" --target codex --agent-id local-agent --json.
Examples: /skills-router install <pkg> for me, skills-router install <pkg> for all agents, /skills-router refine, skills-router route <task>.
Scope: default workspace:local-agent; global or all agents uses
global routes. Blank/named refine discovers while comparing visible scopes.
Safety: uninstall removes only Skills Router metadata. Keep
needs_selection inactive until the human confirms. Use --yes or
auto_approve only when the user explicitly accepts risk.
Reply: prefer human_summary; otherwise status + next action.
Do not paste raw JSON unless asked.
Setup: AGENTS.md. MCP: Prefer a local stdio MCP server: command "skills-router", args ["mcp"].
CLI: If MCP is unavailable, run skills-router chat "<slash request>" --target codex --json.
Notes: Keep the bridge prompt in AGENTS.md or the project instructions Codex reads.; Prefer JSON output and summarize only the human-facing result.