Model Reasoning

Thinking output: Supported models can expose their internal reasoning before producing a response.

Supported models can expose their internal reasoning before producing a response. Enabling this gives you visibility into how the agent approached a problem, which is useful for debugging complex tasks or auditing model behavior. This is known as model reasoning or inline thinking.

Enable inline thinking by settings.json

Inline thinking is controlled by ui.inlineThinkingMode in your settings.json. The default is "off".

{
  "ui": {
    "inlineThinkingMode": "full"
  }
}

Value	Behavior
"off"	Thinking is not displayed (default)
"full"	The model's full reasoning is shown inline in the chat

Enable Inline Thinking by /settings

You can also use the /settings command, then scroll down to Inline Thinking and toggle it on by pressing Enter. Press esc to leave the menu.

This setting requires a restart to take effect.

Show thoughts in the terminal title

ui.showStatusInTitle streams the model's current reasoning to the terminal window title during the working phase. This provides ambient visibility without adding output to the chat.

{
  "ui": {
    "showStatusInTitle": true
  }
}

Scope thinking budgets to specific agents

Use the overrideScope match key to apply a different budget for a specific sub-agent without changing the global default. This example gives the codebaseInvestigator a higher budget than general chat:

{
  "modelConfigs": {
    "overrides": [
      {
        "match": {
          "overrideScope": "codebaseInvestigator"
        },
        "modelConfig": {
          "generateContentConfig": {
            "thinkingConfig": {
              "thinkingBudget": 8192
            }
          }
        }
      }
    ]
  }
}

The match object also accepts a model key to scope the override to a specific model name. Specificity determines precedence when multiple overrides match — a match on two keys wins over a match on one key.