Response schema for agent outputs that may include tool calls.

This schema allows the LLM to either provide a final text response or request tool executions.

:thinking holds the response's thinking and redacted_thinking blocks in their wire shape (string keys, with signature or data). They are sent back unchanged, first, when the turn is replayed: a thinking model needs them on the assistant message that carries the tool calls.

:prefix_key (see prefix_key/3) records the system prompt, tools and model the response was produced under. The API accepts a thinking block only under that same prefix and model, so Normandy.Components.AgentMemory.drop_stale_thinking/2 drops thinking whose key no longer matches. It is never sent to the API.

:content_blocks holds the assistant turn as the API returned it: string-keyed, API-shaped blocks in their order (thinking, text with its citations, tool_use, server_tool_use, mcp_tool_use and their results), as Claudio.Messages.Response.to_assistant_content/1 makes them. When it is not empty, the turn is replayed from it unchanged, so the next request carries exactly what the model produced; :content, :tool_calls and :thinking are what Normandy reads. A response built without it (by hand, or stored before it existed) is rebuilt from those three fields.

:stop_reason is why the model stopped (Normandy.LLM.Response.stop_reason/0), or nil when the client did not say. The turn continues a :pause_turn response. It is never sent to the API.