Swap Control Center

Use this skill as the dynamic-provider preset for agent-control-center. Keep one orchestration engine and exactly one active controller.

Procedure

1. Read Agent Control Center and its four references. If the core skill is incomplete, stop and report the missing resource.
2. Freeze the objective, host, topology, scope, authorization, budget, and acceptance contract before route selection.
3. Build the live route inventory first. Classify every application, CLI, SDK, API, MCP bridge, and manual route independently. Never infer a ready route from a provider name or an application login.
4. Follow provider-selection.md. In the active parent conversation, use the host's exposed structured question tool, including AskUserQuestion or request_user_input when that exact tool is available. Otherwise ask one concise plain-text question at a time. Resolve controller and worker routes independently from current verified capabilities. Derive manual-handoff from a manual controller transport; otherwise select a compatible automatic or sequential relay mode. Never infer worker assignments from the selected controller.
5. If the chosen route needs setup, present the exact official source, version, install or configuration action, authentication step, expected billing or data boundary, rollback, and bounded smoke test. Obtain explicit user approval before each state-changing action or paid request. Rebuild the inventory after setup.
6. Bind the selected controller provider to controller=<provider-id>, bind the separately selected worker defaults or per-role routes, and run the Agent Control Center host-mediated loop. The host parent remains the user-facing relay and dispatch authority. The selected controller owns decomposition, work-order decisions, receipt review, and final acceptance. Topology skills execute controller-issued orders and do not create a second plan or controller. For provider=codex and transport=codex-cli, use the bundled stateful bridge in ../agent-control-center/references/codex-cli-bridge.md. Do not improvise a one-shot command. Other external providers require an equivalent verified preflight, start, reply, cancel, and close adapter before automatic use.
7. If the user requests a later role swap, use the one-transfer contract in controller-modes.md. Pause dispatch, settle outstanding work, obtain any required approval, invalidate unused old orders, and resume with one new controller. Never run co-controllers or recursive controller delegation.
8. Report the active host, controller provider, transport, requested and attested model settings, setup or smoke-test evidence, worker routes, transfer history, validation, fallback, and residual risk.

Read examples.md when a host needs a concrete Cursor, CLI, or unavailable-route interaction pattern.

Codex selection

Codex is a normal fixed-provider choice in this skill. When the user selects Codex, set controller=codex and follow the Codex route order and Cursor app-only safeguards in controller-modes.md. Require a verified Codex route and ask before any fallback to another controller. Do not require Cursor CLI or Cursor SDK merely because Cursor is the active host.

When Cursor hosts a Codex controller with Cursor workers, send the task envelope through the stateful bridge first. Codex returns bounded work orders, Cursor dispatches only the approved Cursor worker routes, Cursor returns proof receipts to the same Codex session, and Codex decides. Do not launch Codex analysis lanes unless the user independently selected a ready Codex worker route.

Before the first Codex controller request, use bridge preflight output to ask for the model and reasoning effort through the active parent's exact structured question tool, including AskUserTool when that tool is available. Preserve an explicit user choice. Do not offer values absent from the bridge preflight catalog.

Cursor-hosted Codex executable priority

When Cursor hosts provider=codex with transport=codex-cli, let the stateful bridge select and verify the executable in this order:

1. MVCK_CODEX_BIN, when the operator set one absolute executable path. This explicit route must pass preflight or stop without automatic fallback.
2. The openai.chatgpt extension path registered for the current codex_vscode Cursor host.
3. Other bounded, non-obsolete openai.chatgpt extension candidates that have matching local registry or manifest structure for the current platform.
4. The first executable codex entry on PATH.

Cursor environment, registry, and extension manifest values are untrusted local routing evidence, not cryptographic publisher or model attestation. Every automatic candidate must pass executable identity, CLI capability, login, and exact full CLI-to-cache version checks before selection. Never equate a prerelease version with its base release.

Bind preflight, catalog selection, start, reply, idle cancel, close, state, and public receipts to the same route-binding digest and executable SHA-256. Start a new bridge-owned child process from that executable. Never enumerate, attach to, signal, resume, or reuse the Cursor extension's app-server process or session.

Cursor Codex recovery

When this route fails preflight, start, or reply, read codex-extension-recovery.md. Run the non-mutating checks first and present the bridge's recoveryPlan as a proposal, not as authority. Before every fix, state the exact action, target, state change, quota or billing impact, risk, rollback, and verification, then wait for explicit approval in the parent conversation.

Keep approvals single-use and action-specific. A cache refresh does not approve login, install, update, state close, model selection, or a live controller start. Never manually edit models_cache.json, normalize a prerelease version, repair an open session in place, or reuse a route-binding receipt after the executable, cache, login, or host route changes.

Fail-closed rules

• Never choose a provider, model, effort, or route from documentation alone.
• Never install a CLI, start login, enable billing, edit user configuration, or send a paid test request without explicit approval.
• Never use a repository config file for personal provider credentials or a personal controller bridge.
• Never assign the controller provider as a worker merely because it is the controller. Require an independent verified worker selection.
• Never treat a command name as provider identity. Verify the binary's own version or authenticated metadata because aliases such as agent can collide.
• Never claim that an external controller directly controls the host UI or native agents. Plain MCP, CLI, SDK, and manual transports are request-response routes.
• Use requested-not-attested whenever the effective model or reasoning effort cannot be authenticated through the host boundary.

Report

Lead with the accepted task outcome or the exact missing capability. Then use the Agent Control Center output contract. Report relay mode, controller route, worker routes, and every user-approved setup action or controller transfer.