Fixed
- 🧠 Reasoning models no longer lose chain of thought across parallel tool calls. Multiple tool calls in the same turn are now grouped into a single assistant message with shared reasoning items, instead of being split into separate messages that broke reasoning model round-tripping (e.g. o3, o4-mini).
- 🛡️ Orphaned tool calls no longer break chat history. If a crash or data corruption left a function call without a matching result, the Responses API conversion now skips the orphaned call instead of sending invalid input that caused permanent 400 errors.
- 🔄 Tool output ordering prevents corrupted history on crash. Tool call outputs are now appended to the output list before marking the call as "completed", so a crash between the two steps can no longer produce a "completed" call with no output — which would corrupt the message history on reload.
- 🧹 In-progress tool calls scrubbed on error or cancellation. When a chat task is cancelled or hits an error, any tool calls still marked "in_progress" are now set to "failed" before persisting, preventing stale in-progress items from lingering in the database.