dsh-safe-delete

When a DeepSeek agent executes bash on macOS, a single rm -rf can wipe away a half-day of work with no way to undo it. Developers who grant models free terminal access—whether in GUI chats, CI pipelines, headless bridges, or multi-agent orchestration—face that risk every day.

dsh-safe-delete takes a direct approach: it installs one global guard on the DSH tools registry that intercepts every bash call containing rm in command position and re-executes it as /usr/bin/trash. The model receives a clear receipt of exactly what was moved, so nothing vanishes silently. Its quote-preserving lexer catches sudo rm, /bin/rm, env- or nice-prefixed rm, xargs rm, and rm buried inside compound commands joined by &&, ||, ;, or |. Constructs that could hide an rm—$(...), backticks, heredocs, eval, nested sh -c—are denied outright with guidance rather than rewritten. One registration covers GUI sessions, scheduled automations, headless bridges, and subagents with zero per-session wiring.

If you run DeepSeek Harness on macOS and grant your agent bash access—whether for casual chat, automation, or multi-agent workflows—this plugin is a zero-intrusion safety net. It ships enabled by default, toggles with a single switch under Settings → General, and persists across restarts. It does not aim to be a sandbox; it solves exactly one problem: making the most destructive verb recoverable.