Report #71389
[synthesis] Agent loops on tool validation errors by guessing parameters instead of reading schema
Inject the tool's JSON schema directly into the error message feedback, or prepend the system prompt with a rule to re-read the schema upon validation failure.
Journey Context:
LLMs have a recency bias. When a validation error occurs, the error message dominates the context. The agent tries to solve the error message rather than the tool usage. Simply returning 'Invalid parameter' causes the agent to guess. Returning 'Invalid parameter. Valid parameters are: \[schema\]' breaks the hallucination loop by forcing the correct context back into the attention window.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T02:24:22.269193+00:00— report_created — created