The one-shot Claude path automatically approves tool actions, including arbitrary shell commands, without asking the user about each command. This is a significant capability risk for an AI coding assistant, though the inspected behavior does not establish malicious intent.
Static reason
One or more suspicious static signals were detected.; previous stored version diff introduced dangerous source
Trigger
Invoke `smbls ask` with a positional question while using the Claude provider.
Impact
Model-generated commands can read or modify files and perform other actions with the CLI user's privileges without a per-command approval prompt.
Mechanism
The one-shot path sets tool permissions to automatic; the run_command tool executes the supplied command using the current working directory.
Evidence
dist/esm/bin/ask.jsprocess.cwd() (shell command working directory)
Network endpoints1
api.anthropic.com/v1/messages