A benchmark for LLMs on complicated tasks in the terminal
Terminal-Bench is a benchmark and execution harness for testing whether AI agents can complete difficult, end-to-end tasks in real terminal environments. It is used by developers of language-model agents and benchmarking systems, with tasks covering activities such as compiling code, training models, and setting up servers. Catalogue add-ons help agents work with the benchmark and its task suite.
2 files for GitHub Copilot and Claude Code: terminal-bench-1 copilot-instructions.md, terminal-bench-1 CLAUDE.md — 1,389 tokens loaded in every session.
.github/copilot-instructions.md A 599 tok CLAUDE.md A 790 tok These files are harbor-framework/terminal-bench-1's own configuration — they tell GitHub Copilot and Claude Code how to work on this repository, so they are not mods to install elsewhere. Copy one as a starting point and replace the parts that are about this project.