Models and Budget

Plan and verify with a strong model while a cheaper or local model writes the code:

/goal --planner-model anthropic/claude-opus-5-5 --planner-reasoning high --worker-model local/qwen-code Add a CSV export for the filtered report

When a cheap worker keeps stalling, retry with a stronger one:

/goal:retry --worker-model anthropic/claude-opus-5-5

--model sets all three roles, --verifier-model picks the model for both the review and the judge, and a --<role>-reasoning value without a model applies to the current model. Every model is resolved before planning starts, so a typo fails at once. The choice is saved with the goal, so answers, approval and /goal:resume keep it after a restart. /goal:approve and /goal:retry take the same flags.

Give Long Goals a Bigger Budget

A goal stops as over_budget after 10 million tokens (completion plus uncached prompt tokens across all roles) or 120 minutes by default. Raise the limits, or set role defaults, in a profile; 0 disables the token limit:

{
  "profiles": {
    "general": {
      "goal": {
        "maxTokens": 25000000,
        "maxDuration": "4h",
        "planner": {"model": "anthropic/claude-opus-5-5", "reasoning": "high"},
        "worker": {"model": "local/qwen-code"},
        "verifier": {"model": "anthropic/claude-opus-5-5", "reasoning": "high"}
      }
    }
  }
}

A role you leave out uses the session’s current model. /goal:status and the final message show what a goal spent, for example Spent: 1.2M/25.0M tokens, 38/240 min.