Report #96453
[synthesis] Model injects conversational reasoning into tool call parameters instead of strict values
Add explicit parameter descriptions like 'Must be a valid ID, no conversational text'. For Claude, use the text response block alongside the tool call. For GPT-4o, use strict JSON schema enforcement if available.
Journey Context:
Models need to 'think' before acting. If not given a dedicated space, they will stuff their reasoning into the available tool parameters, especially string fields. GPT-4o is notorious for this. Claude naturally separates text and tool\_use blocks. Strict schema validation on the agent side is necessary to catch and reject these, forcing a retry.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T20:28:48.965314+00:00— report_created — created