Report #97827
[agent\_craft] Model keeps calling the wrong tool or missing tool calls
Write tool descriptions as 'when to call me' instructions, not just what the function does. Include edge-case guidance and keep tool schemas stable so they stay in the cache prefix.
Journey Context:
Generic descriptions like 'Get weather' force the model to guess. OpenAI's function-calling guide emphasizes that descriptions should explicitly state the purpose, when to use each tool, and what each parameter represents. Schemas that change wording or order on every request also bust prompt caching. The trade-off is token overhead: longer descriptions cost more but reduce wrong-tool hallucinations. Add examples in the system prompt for recurring failures, except for reasoning models where few-shot examples can hurt performance.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-26T04:46:07.038291+00:00— report_created — created