dsh-netshell

When you manage servers in DSH through both a human terminal and AI tool calls, a core pain point remains: once a command is submitted it is irreversible, and an AI model can still trigger destructive operations by mistake. DSH-NetShell embeds an allow / ask / deny three-tier command guard into every input path of both the local terminal and the remote SSH sessions, so a human pressing Enter and an AI invoking netshell_run share the same adjudication logic and a clear human confirmation point before anything actually runs. It also eliminates credential leakage concerns—SSH passwords are written only to DSH's encrypted credential store and never appear in list responses, tool results, terminal output, or logs, meaning the AI neither sees them nor can forge a confirmation.

On the interface side, the local terminal launches bash, sh, PowerShell, or cmd on the current device with zero SSH configuration. Remote servers support password, private-key, and ssh-agent authentication with multiple parallel sessions; switching DSH tabs does not sever the background connection. The bottom bar shows live counts of executed, blocked, and permitted commands, and the history panel lets you replay every adjudication in the current session. Two model tools—netshell_servers and netshell_run—expose host listings and single-command execution to the AI within the same guardrail, without ever touching credentials.

DSH-NetShell is built for developers, SREs, and platform teams who blend manual operations with AI assistance inside DSH. If you already rely on DSH's generic shell tooling but feel the lack of per-command gating, or you want server-bound AI actions audited on the same trail as human input, this plugin offers a lightweight, pluggable guard layer. It is a command-operation guardrail, not a full security sandbox: for sensitive environments set the permission level to locked and still review what interactive programs execute internally.