dsh-limit-timeout

When a DSH agent fills in timeoutMs for bash or timeout_ms for job_output, the value is entirely model-decided. Before this plugin, the deployment side had only a loose tool-level fallback; if the model requested an implausibly long wait, operators could only sit and wait or kill the process manually. dsh-limit-timeout inserts a check at tools/pre-execute: any call exceeding the current session ceiling is rejected outright, the rejection message spells out the active limit, and three alternatives are offered—lower the wait, switch to background execution, or request a session-scoped extension through the native approval channel.

The settings UI exposes two core fields—a global default wait cap and a hard ceiling—accepting natural notations like 120s, 2m, or 1h, with draft-style editing so accidental changes are one click away from reverting. With "allow model escalation" enabled, the model can request a longer wait via DSH's native approval flow; once approved, the extension applies only to the current session and resets on new sessions or process restarts. Additional toggles let you require explicit wait declarations on job_output calls and suppress repeat-tool-reminder injection, giving you finer control over resource scheduling.

This plugin suits admins who want to set a "wait budget" on the deployment side without either letting the model occupy tool slots indefinitely or flat-banning legitimately long queries. It works by intercepting rather than rewriting—tool parameters stay untouched—and composes orthogonally with existing permission and sandbox policies.