Pick the Model for Scheduled Runs

A check that runs every ten minutes rarely needs your best model. Use a profile’s schedules block to run every scheduled prompt on a cheaper or local model, and to cap its context window so long runs compact early:

{
  "profiles": {
    "general": {
      "schedules": {
        "model": "local/qwen-code",
        "reasoning": "low",
        "limit": {"context": 64000, "reservedTokens": 8000}
      }
    }
  }
}

Each field is optional. Without model, scheduled runs use the session’s current model, and limit still applies to that model. Caretz reads the block when each run starts, so changing it and reopening the session also changes schedules you already registered. /schedule-list shows the model each job will use.

Rules

• model is a configured provider/model ref. Model rules apply: a blocked model makes the run fail rather than switching to another.
• reasoning needs model.
• limit takes the same keys as a provider model’s limits: context, reservedTokens, keepRecentTokens and summaryTokens. An unset key keeps the model’s value. context may shrink the model’s window but not exceed it, and reservedTokens must stay below context; a run that breaks either rule fails with an error in the conversation.
• A project or profile layer that sets schedules replaces the whole block from earlier layers.