Report #66773
[synthesis] Tool calls hallucinate missing required parameters
Set \`strict: true\` in OpenAI tool definitions to prevent parameter hallucination. For Claude, explicitly instruct in the system prompt: 'If required parameters are missing, do not invoke the tool; ask the user instead.'
Journey Context:
When faced with missing required parameters, GPT-4o exhibits a 'helpfulness bias' by guessing or fabricating parameters to complete the tool call, leading to silent failures. Claude 3.5 Sonnet exhibits an 'over-caution bias', often returning a text response asking for clarification, which breaks autonomous agent loops expecting a tool call. Setting OpenAI's \`strict: true\` forces validation, while Claude requires explicit system prompt instructions to align its behavior with autonomous execution.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T18:33:35.477767+00:00— report_created — created