Not there yet

Said plainly, so you know what you are getting:

• No image input (pasting a screenshot).
• The OS sandbox is macOS only; on Linux and Windows you get the confined tools and shell path checks, not the kernel layer.
• /rewind restores files to before one of the last 20 file-changing requests, but it does not rewind the conversation to an earlier message. No IDE extension. Diagnostics come from tsserver for TypeScript and quick checkers elsewhere, not a general language server.
• Token counts are estimated (bytes / 4) rather than counted exactly. Outlines and signature checks for Python files use an indentation-based reader, not a real parser.
• The benchmark runner (bench/) did not come over from the Python version, so the numbers above are not re-measured on this binary.
• Several results above are tuned on one real repo; treat the defaults as measured there, not universal.

On "better than the others"Copy Anchor Link

The one head-to-head here is zerd vs Claude Code: 11 tasks, one run each, different models, 10/11 apiece. It shows both are in the same range, not that either is better. The design tables compare choices, not scores.