instructions text, not a new field or event. The reasoning behind them is in Prompting tool use.
Confirming before anything that can’t be undone
Name the tools that need a spoken yes first, don’t leave it to inference. A tool’s own description doesn’t carry a risk level, soinstructions has to.
instructions: confirm before a destructive tool call
Speaking about results, not tools
Talk about what was found, not what was called. Left unguided, a model narrating its own tool use describes the mechanism instead of the outcome.instructions: narrate results, not tool names
Handling a tool error
Afunction_call_output is just a string you wrote. Say what an error inside it looks like, or Dialog-RSN-1 can’t tell it apart from success. Match this to whatever shape your own tools actually return.
instructions: recognize and recover from a tool error
Searching, and getting citations back
Tell Dialog-RSN-1 when to search instead of answering from memory, and what to do when nothing relevant comes back. The search tool itself and thepoly_knowledge wiring are covered in the prompt guide; this is the instructions text on top of it.
instructions: when to search, and what to do with nothing
Next steps
Prompt guide
Why each of these works, and what the protocol already handles for you.
Reference
The function-calling wire format and the round trip example.

