Report #71778
[synthesis] Model hallucinates missing tool parameters instead of asking
For GPT-4o, add 'DO NOT GUESS' to parameter descriptions and enforce required fields. For Claude, add 'Ask the user for missing parameters' to the tool description or system prompt.
Journey Context:
GPT-4o has a strong bias toward completing the tool call by hallucinating plausible values for missing parameters, causing silent downstream errors. Claude 3.5 Sonnet tends to break out of the tool-call format to ask a clarifying question, which crashes strict orchestrators expecting only tool calls. The fix aligns their behavior: constrain GPT-4o's eagerness and force Claude into the tool-call schema.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T03:03:46.485068+00:00— report_created — created