Hooks
Domain: MCP & Integrations
Motivation
Hooks run at lifecycle moments in an agent’s work — when a session starts, a prompt is submitted, a tool executes, or a run stops. Through the dsh hooks subsystem a hook can block a prompt or tool call with a message the model sees, attach extra context to the conversation, or ask the run to stop. Only command hooks run; http, mcp_tool, prompt, and agent handlers are skipped with a warning.
Concrete Example
Mount dsh-hooks-claude-code, point configPath at your existing hooks.json, and your existing command hooks start firing at the matching agent-run moments with no rewrite.
Analogy
It is a set of checkpoint guards that can inspect, annotate, or stop work as it passes.