codex-control-plane-mcp

Submit the task with codex_submit_task and wait_for_completion set to false to receive an operationId immediately. Then poll codex_get_operation_status until the state becomes completed or failed, handling waiting_for_approval states along the way.

Pass a unique client_request_id with every codex_submit_task call. If a retry occurs with the same ID, the control plane returns the existing operation instead of creating a duplicate, preventing repeated work after timeouts.

Start a workflow with codex_start_plan_workflow and poll codex_get_workflow_status. When the phase reaches review_plan, call codex_approve_plan with approved true to execute, or provide feedback to request a revised plan.

Yes, it runs as a standard stdio MCP server configured in your client's config file, such as claude_desktop_config.json. Set environment variables like CODEX_MCP_STATE_DB and CODEX_PROJECTS_ROOT in the server entry.

Stuck operations usually indicate an unresponsive app server or a turn awaiting interaction. Interrupt the operation with codex_interrupt_turn, or run codex_analyze_issue with the operation_timeout symptom to get severity-ranked checks.

The control plane returns a DUPLICATE_PROMPT error containing the existingOperationId in its details. Use that ID with codex_get_operation_status to track the already-active turn instead of starting a new one.