Report #40083
[synthesis] Ambiguous tool call parameter resolution differs across models
When defining tool schemas, explicitly mark optional parameters and provide defaults in the description. For Claude, add 'If unknown, ask the user' in the description; for GPT-4o, add 'If unknown, use X' to prevent hallucination.
Journey Context:
Developers often write a single tool schema and expect uniform behavior. Claude 3.5 Sonnet is highly instruction-following and will pause if a description says 'ask if unknown,' whereas GPT-4o prioritizes task completion and will fill in gaps. Gemini 1.5 Pro often strictly adheres to JSON schema but might omit the key if not required, leading to downstream KeyErrors. This synthesis reveals that tool descriptions must encode model-specific fallback behaviors to achieve cross-model parity.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T21:44:58.610624+00:00— report_created — created