agenthub

SUSPICIOUS: the core behavior matches a multi-agent orchestration skill, but it carries medium-high risk because it amplifies autonomous code changes, executes eval commands, and can merge results after processing untrusted content. Install trust is better than a raw downloader, yet the public registry model weakens provenance assurance.

The code primarily acts as a coordinator that discovers branches and worktrees and then executes a user-supplied evaluation command in a shell inside a worktree. The biggest risk is the shell execution of eval_cmd derived from inputs (config/env/args). While the script itself is not inherently malicious, it provides a high-risk sink that could be abused to run arbitrary commands, leading to potential data exposure or system compromise if inputs are untrusted. Recommendations: validate inputs strictly, avoid shell=True for untrusted data, sandbox or constrain eval_cmd, implement robust error handling, add auditing/logging, and consider parameterizing commands to minimize shell invocation surface.