{
  "title": "Reasoning pauses for a tool",
  "caption": "This request example has visible output, model work, tool waits and retained history. Until the answer, the user may see only an optional preamble. A tool wait is different from active model computation. Longer work can keep state for a longer time, but the exact hosted sequence and retention stay unknown. The author made this explanatory diagram with example playback, and animation time is not measured runtime.",
  "alt": "Request flow: the model continues reasoning after a tool wait, then generates an answer. Computation and retained history are in different lanes.",
  "sources": [
    {
      "path": "content/blog/2026-07-22_reasoning-effort-inference-systems/post.md",
      "description": "Article explanation and retained source citations"
    }
  ],
  "limitations": [
    "The execution is an example. Playback duration is not measured runtime.",
    "State-owner changes are examples, not measurements or a description of a proprietary KV/cache scheduler. Tool latency and animation time are independent."
  ]
}