agent-offload runs another program (OpenCode) that executes model-chosen shell commands inside your project. The kit limits it with explicit denies and checks the result, but it is not a sandbox. Read this before you use it on anything that matters.
- The executor's OpenCode agent profiles deny: leaving the project folder, reading
*.envand key files,git push,git reset --hard,rm -rf,pip installand globalnpm install, and editing the kit's own files (.offload/,.opencode/,opencode.json). These denies were verified against a live OpenCode: the engine rejects the calls, not the model's good will. - Work happens in a separate git worktree on a scratch branch; your tree is never checked out, reset or cleaned.
- After the run the kit discards everything if the executor touched its own limits, and drops out-of-scope and oversized files.
- Shell commands are not sandboxed beyond the denies. A command that is not denied runs with your user's rights. Run delegations on a machine and account where that is acceptable.
- Your code leaves your machine. Free models run on someone else's servers, and some providers train on prompts. Choose the privacy preset accordingly; unknown terms are denied by default.
Open a private security advisory on GitHub (Security → Report a vulnerability) rather than a public issue. A way for the executor to escape the denies, reach outside the project, or make the kit merge unchecked work is a security problem; please include the OpenCode version.