Batuta Routing

Batuta Routing Table

Batuta's core opinion: route every task to the cheapest executor that can handle it. Lanes use the complexity vocabulary that cy-create-tasks writes into task frontmatter (low, medium, high, critical) — the same vocabulary runtime_rules[].match.complexity matches on.

Lane semantics (the durable opinion)

How batuta derives the concrete table (never copy an example)

1. compozy__provider_models_list (with costs) is the ONLY source of concrete provider/model IDs — it reflects the CLIs actually installed and the models actually discovered on this machine. A provider absent from the catalog is not installed; never route to it.
2. Map each lane's selection rule onto the catalog using the cost fields (input_per_million / output_per_million) as evidence.
3. Model enablement is account-side and invisible to the daemon — present the derived table (with costs) to the operator for confirmation before storing; ask what their accounts enable when in doubt.

Example only — derived on one machine on 2026-08-11, DO NOT reuse

On that machine the derivation produced: low → codex/gpt-5.6-luna, medium → codex/gpt-5.6-terra@high, high → codex/gpt-5.6-sol, critical → claude/claude-opus-4-8. Your catalog will differ; derive, do not copy.

Canonical rule shape

This is the exact JSON SHAPE batuta writes with compozy__loop_configure (stored per-workspace override for implement-tasks) after deriving the values from the catalog — the model/provider strings below are the same dated example as above and MUST be replaced by the derived ones. The stored override is what run-loop children resolve at execution — batuta never sends per-run rules on dispatch, because per-run rules freeze into the run and are not inherited by run-loop children anyway. Rule matching precedence inside the stored layer: id > type > complexity.

[
  { "match": { "complexity": "low" }, "runtime": { "provider": "codex", "model": "gpt-5.6-luna" } },
  {
    "match": { "complexity": "medium" },
    "runtime": { "provider": "codex", "model": "gpt-5.6-terra", "reasoning": "high" }
  },
  { "match": { "complexity": "high" }, "runtime": { "provider": "codex", "model": "gpt-5.6-sol" } },
  {
    "match": { "complexity": "critical" },
    "runtime": { "provider": "claude", "model": "claude-opus-4-8" }
  }
]

Provider quirks

• Some providers multiplex upstreams and require the model field to carry a prefix — e.g. opencode only binds opencode/kimi-k2.5, never bare kimi-k2.5. The catalog's exact model_id is authoritative; copy it verbatim into the rule.
• A model can exist in the catalog and still be disabled for the operator's account at the provider (invisible to the daemon). When a lane fails its bind with zero tokens, ask the operator what their account enables.

Escalation and reclassification

• Repeated failure in a lane: write a surgical id rule one lane up into the STORED override (compozy__loop_configure on implement-tasks, e.g. {"match":{"id":"task_NN"},"runtime":{...}} prepended to the rules), then re-dispatch batuta-deliver. id beats complexity; remove the rule after the task lands.
• Operator reclassification in conversation ("use luna for this one") becomes the same stored id rule before the next dispatch.
• The daemon persists resolved_runtime with per-field provenance on every generation — routing decisions are auditable via compozy__loop_status, never narrated.