How do I check a Claude Agent SDK tool call without touching the prompt?

Register a hook: a callback the SDK invokes at a named point in the loop, most usefully PreToolUse before a tool call runs or PostToolUse after it returns. The callback receives the tool’s name and arguments, or its result for PostToolUse, and it runs in your own process, entirely outside the model’s context, so nothing about the check has to be phrased as an instruction the model might ignore or forget.

A matcher pattern scopes a hook to specific tools, "Write|Edit" for file writes or a regex like ^mcp__ for every MCP tool, or you can omit it to run on every call. The callback’s return value decides what happens next: an empty object lets the call through unchanged, while a permissionDecision of deny blocks it and hands the model a reason it can act on, rather than a call that just silently vanished. A PostToolUse hook can’t stop a call already made, but it can attach additionalContext to the result before the model reads it. A hook scoped to ^mcp__ is a natural place to catch an MCP call failing in a way the model itself never sees, before it reaches the rest of the loop.