Maximum quality

The brief: accuracy over everything. The strongest coding model works, an equally serious model from a different family judges, effort and thinking run high, and you pay list price for the privilege. Use this for gnarly multi-file changes, unfamiliar codebases, and anything you'd rather pay $2 for once than $0.20 for three times.

The config

{
  "agent_engine_config": {
    "pipeline_worker_model": "anthropic/claude-fable-5",
    "pipeline_judge_model": "openai/gpt-5.5",
    "pipeline_triage_model": "deepseek/deepseek-chat",
    "default_model": "anthropic/claude-fable-5",

    "allowed_models": [
      "anthropic/claude-fable-5",
      "openai/gpt-5.5",
      "gemini/gemini-3-pro",
      "deepseek/deepseek-chat"
    ],

    "auto_mode": "off",
    "effort_auto": "off",
    "reasoning_auto": "off",

    "agents": {
      "worker": {
        "effort": "xhigh",
        "reasoning": "on"
      },
      "judge": {
        "effort": "high",
        "reasoning": "on",
        "prompt": "You are a strict, evidence-first code judge. Claimed success without deterministic evidence is a FAIL.",
        "params": { "service_tier": "priority" }
      },
      "triage": { "effort": "low", "reasoning": "off" }
    }
  }
}

Why these choices:

• Fable 5 as worker. The strongest coding/agentic model in the catalog — worth its $15/Mtok output on work where a wrong change costs more than the tokens. xhigh effort + thinking on: let it deliberate.
• GPT-5.5 as judge — not another Claude. Cross-family judging is the point: a different family shares fewer blind spots with the worker. High effort, thinking on, and a stricter custom prompt. service_tier: "priority" buys faster verdicts on providers that support tiers.
• DeepSeek triage. Even a maximal setup shouldn't pay flagship rates for a one-token classification. This is the one dial that never needs to be expensive.
• Autos off. In a hand-tuned profile, explicit settings beat automatic ones.

Run it hard

# Arm the deterministic oracle and give it room to work.
stella --budget 8.00 run \
  "port the auth middleware to the new session store; all tests must pass" \
  --test-command "cargo test -p auth"

# Outcome-driven, judged rounds, capped.
stella --budget 15.00 goal "the payments integration test suite passes on CI"

What it costs

Estimates at list prices, medium task, caching healthy (see the cost anatomy):

The judge adds only cents per verification; the worker's output tokens are ~80–90% of the bill. If this profile stings, the balanced profile keeps a flagship judge while cutting the worker bill 5–7×.