CODORQ: ogrodje za orkestracijo razvoja programske opreme z različnimi agentskimi orodji

This paper presents CODORQ, a local-first framework for orchestrating software development with various agent tools. It is based on the observation that when multiple agents work simultaneously, the limitation is not only the capacity of the models, but also the lack of a persistent, verifiable coordination layer. CODORQ connects Codex CLI, Claude Code and Antigravity CLI sessions via a local MCP server to repositories, separate Git worktrees, terminals and a common model of tasks, roles, statuses, reviews, QA results and memory. The system supports the separation of duties between execution, review and QA, tracks the identity and liveness of sessions and treats a human as the execution manager. Persistent records allow for context recovery after a crash or restart without relying on terminal rewriting. The paper describes the architecture, multi-level persistent memory, task statuses, operator web interface and defense mechanisms for restarts, unresponsive sessions, excessive MCP responses and concurrent changes. A demonstration example of the development of a small UI component shows the path from a request through a proof commit and independent review to a QA result. The paper does not demonstrate productivity gains, but rather shows how multi-agent work can be made traceable, reproducible, and operator-manageable.