dsh-guardian

In a DSH environment with multiple plugins and agents, a single missing inject declaration can crash the entire plugin tree, a silent tool-call loop or runaway subagent recursion can stall a session indefinitely, and an ill-timed host restart can kill a task mid-execution. These failures share one trait: hard to diagnose after the fact, hard to notice before they happen.

dsh-guardian builds four lines of defense around four real-world lessons. A preflight scanner statically checks that every ctx usage in a new plugin matches its inject declarations, catching broken plugins before they can take down the host. A sliding-window loop detector flags repeated tool calls and steers the agent toward a different approach. A nesting-depth counter stops subagent recursion from running away. A safeToRestart status endpoint, backed by a safe-restart script, refuses to restart the host while any session is still running. Every action is detection, gentle steering, or event logging only. The plugin never terminates a task, and config changes take effect hot without a restart.

Ideal for users who maintain multiple DSH plugins, orchestrate multi-agent workflows, or need to restart the host regularly. It does not alter the existing workflow; it simply adds a quiet safety net that nudges the agent back on track when something goes sideways.