PRODUCTION SCENARIO
An operations agent built on the Agent SDK reads delivery times from three MCP servers you do not own: one returns epoch seconds, one ISO 8601, one a locale-specific string. The agent compares those times against a cutoff inside its own reasoning, before it calls anything else, and a wrong comparison silently drops a cutoff alert rather than raising an error anyone sees.
Where should the system normalize the values?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: In a PostToolUse hook that rewrites each result to one format before Claude reads it
A PostToolUse hook runs deterministic code on each result and can replace the tool output before it is sent to Claude, so the agent only ever reasons over one representation, which is what a comparison made silently inside the reasoning needs. A hook on the alerting tool is the right mechanism at the wrong point in the lifecycle: it is the place to fix an argument on an outgoing call, and the faulty comparison happened upstream of any alerting call. A worked example in the system prompt is the answer when the conversion is a judgment the model must make, and a documented format in the tool description is the answer when one stable format needs explaining. Both leave three representations in play.
Official Source: https://code.claude.com/docs/en/hooks ↗