Model Routing

Baton routes each task to the cheapest agent + model that can handle it: heavy reasoning to Opus/Claude, UI to Gemini, small fixes to Codex/Cursor, trivial fixes to a local model. The routing is deterministic keyword + severity scoring — no LLM call, instant, and every decision shows its reasoning.

Configure it

Routing lives in a committed, team-shared baton.config.json at the repo root:

{
  "routing": {
    "mode": "auto",
    "rules": [
      { "match": ["plan", "architecture"], "tier": "heavy" },
      { "match": ["ui", "css", "component"], "agent": "gemini" },
      { "match": ["bug", "crash", "fix"], "tier": "light" }
    ],
    "default": "standard",
    "tiers": {
      "heavy":    [{ "agent": "claude", "model": "opus" }],
      "standard": [{ "agent": "cursor" }, { "agent": "claude", "model": "sonnet" }],
      "light":    [{ "agent": "codex" }, { "agent": "gemini" }],
      "local":    [{ "agent": "aider", "model": "ollama/qwen2.5-coder" }, { "agent": "opencode" }]
    }
  }
}

Modes

Severity & tiers

When no rule matches, Baton scores the task 0–100 from deterministic hints:

• Heavier (+): architecture, refactor, migrate, security, concurrency, performance, research, plus a long, detailed description.
• Lighter (−): typo, rename, comment, bump, lint, minor, quick, plus a very short description.

The score maps onto a tier, and each tier is an ordered fallback chain — the first installed agent in the chain wins, so a missing CLI (e.g. Ollama down) falls through to the next entry automatically.

Preview a route

baton route "refactor the auth module for concurrency"
# → claude (opus)   severity 88/100 → heavy tier

Routing is also applied automatically by baton pass when you omit --to, and surfaced in the dashboard's Launch and Handoff dialogs with the matched keywords and confidence shown.