wtf is cursor-director?

Cursor Director is a set of instructions and scripts that turn the Cursor AI coding editor into a coordinator for several other command line coding assistants running at the same time. Instead of doing the actual coding itself, Cursor acts like a project manager: it breaks a task into pieces, decides which AI tool should handle each piece, and reviews the results when they come back. Each piece of work runs inside its own tmux session, which is a way of keeping a terminal program running in the background so you can check on it or reconnect to it later. Every Cursor conversation gets its own separate tmux setup, so several tasks running at once never interfere with each other. A lightweight helper watches each running task, starts the chosen coding tool, feeds it the task instructions, and checks in on its progress every so often, reporting back when it looks finished, without trying to do any of the actual coding or debugging itself. Different coding tools are chosen for different kinds of work. More complex reasoning, architecture, or debugging tasks go to Claude Code, ordinary implementation work goes to either Codex or Grok, and simple repetitive tasks are handled by a cheaper option. The system also tracks how much usage quota each AI provider has left and avoids sending more work to a tool that is close to its limit. Rather than trusting a coding tool's own claim that a task is finished, the system checks the actual file changes, test results, and any other evidence before marking work as done. For apps with a visible interface, such as an iOS app, a task is not considered complete until the feature has actually been built, installed, and tested on a simulator with screenshots to show the result. Finished task sessions are kept around rather than closed immediately, so a person can look back at what happened, and they are only removed when requested or when there is a good reason to clean up. The project also includes a newer approach where two different coding tools review each other's work: one tool's changes are sent to a different tool to check over and either approve or send back for rework, with a limited number of retries.