Known limitations

If a tool fails after changing files, the next one receives their list and continues from there. Nothing is undone: it can inherit something incomplete, and effects outside the folder, such as commands already run, are not detected.
The chat uses the CLIs' non-interactive modes. Native approval prompts (for example, "can I edit this file?") do not appear; in those cases, use the native CLI.
/parallel only analyzes: Codex in --sandbox read-only or OpenRouter, which does not access the files. The agents do not edit the project at the same time.
The simple/complex classification uses a model when there is a Groq, Cerebras, OpenRouter or Grok key. Without any of them, or if they all fail, it uses keywords and size. It can be wrong in both cases.
Switching models does not renew the account's quota.
On Windows the code uses the native equivalents (process groups and taskkill, msvcrt, folder junctions), but it has only been tested for real on macOS. The exception is running tools installed through npm, tested in CI on Windows.

This text comes from the project README. See the original on GitHub