Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview

Source: Hacker News

Results

• Scored 65.2% vs Google’s official 47.8%, and the existing top closed‑source model Junie CLI’s 64.3%.

Clarifications

1. No {agents/skills}.md files were inserted at any point. No cheating mechanisms whatsoever.
2. The CLI agent was run in a leaderboard‑compliant way (no modification of resources or timeouts).
3. The full TerminalBench run was done using the fully open‑source version of the agent; there is no difference between what is on GitHub and what was run.

Context

I was originally going to wait for it to appear on the leaderboard, but after 8 days without a response from the maintainers (a large backlog of pull requests on their Hugging Face repo), I decided to post anyway.

References

• Hugging Face PR:
• Cheating reports on TerminalBench 2.0:
• Hacker News discussion:

Points: 101 Comments: 32