Skip to content

Cooperation and safety

Starting one in a repository can execute arbitrary Python from .one/extensions/*.py and shell commands from MCP configuration in .one/settings.json. Use it only in repositories you trust. To inspect an untrusted repository, start with:

Terminal window
one --no-extensions --no-mcp

Extensions and MCP servers are not sandboxed; they run with one’s permissions. Cooperation mode is an approval gate, not a sandbox or security boundary.

--cooperation requests approval before mutating tools (bash, write, edit, plan, and apply_patch). In the TUI, toggle it with Ctrl+Z and use the visible Approve/Reject controls (or Alt+A/Alt+R); normal prompt input remains available while an approval is pending. Use Ctrl+A in the interactive fallback. Rejections include a reason returned to the agent. Interactive modes provide /steer, /follow, and abort controls; RPC and one run provide the channels described above.

When cooperation and ask_user are both enabled, the model is instructed to ask one focused question for necessary preferences or genuinely blocking ambiguity, then resume the same task rather than putting a question in its final summary.

apply_patch validates collisions and symlink paths and uses staging/rollback, but it cannot guarantee recovery from power loss, hostile concurrent filesystem changes, or rollback I/O failure. Review changes and retained backups when an error is reported.