When the assistant encounters a recoverable error during a tool call — for example, a misunderstood query parameter — it now retries with corrected inputs instead of stopping mid-response. This produces more complete answers for complex queries that involve multiple tool calls, with less manual follow-up needed from users.
