Report #94467
[synthesis] Catastrophic tool calls from parameter schema drift between LLM training data and live API
Enforce strict schema validation at the executor layer and dynamically inject current API defaults into the system prompt
Journey Context:
LLMs default to their training data for API parameters. If an API changed a default value \(e.g., overwrite from false to true\), the agent will construct a payload missing the parameter, assuming the old default, leading to destructive overwrites. Relying on the LLM to read API docs is unreliable. The executor must validate against the live schema, and the system prompt must explicitly state current defaults to override the model's internal weights.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T17:08:57.823825+00:00— report_created — created