OxCode: an AI software engineer for VS Code (82% SWE-bench Verified). Looking for feedback

OxCode, an AI software engineer built by the team at Oxlo AI for VS Code and Open VSX editors, scored 82% on SWE-bench Verified (410/500) on a single attempt and 73/89 on Terminal-Bench 2.1, with full results and eval setup published at oxcode.ai/benchmarks. The agent plans, writes, tests and debugs code, is free to start via the Visual Studio Marketplace, and its builders are soliciting feedback on where it fails on real repositories, which tasks to add to the evals, and what would make developers trust an agent's patch.

I’ve been building OxCode with a very determined team: an AI software engineer that runs inside VS Code and editors that use Open VSX. It plans, writes, tests and debugs code. On a single attempt it scores 82% on SWE-bench Verified 410/500 and 73/89 on Terminal-Bench 2.1. The full results and the eval setup are public at Benchmarks | OxCode https://www.oxcode.ai/benchmarks . I’d love feedback from people who work on coding agents: where does it fail on your real repos, which tasks should we add to the evals, and what would make you trust an agent’s patch? It’s free to start: OxCode - AI Software Engineer - Visual Studio Marketplace https://marketplace.visualstudio.com/items?itemName=oxloai.oxcode-ai . Replies here are welcome.