Finish reasons

Every choice carries a finish_reason. There are exactly four values.

The Values

Value	Signals
`stop`	The model finished normally.
`length`	Generation hit `max_tokens` / `max_completion_tokens` or the context limit. The reply is truncated.
`tool_calls`	Run `message.tool_calls`, append the results as `tool` messages, and call again. See Tool calling.
`error`	Generation failed. In a stream it arrives with a top-level `error` object. See Error events.

Non-streaming:

{
  "choices": [
    { "index": 0, "message": { "role": "assistant", "content": "..." }, "finish_reason": "stop" }
  ]
}

Streaming: finish_reason is null on every chunk until the terminal one. See Delta format.

{ "choices": [ { "index": 0, "delta": {}, "finish_reason": "stop" } ] }

The same four values appear per participant in a Fusion result under models[].finish_reason.