Model selection & effort

Each driver runs a model. You can pin one per workspace, and tune reasoning effort where the CLI supports it.

Pin a model

Set agentDriver.model to an id the driver’s CLI accepts:

version: 1
agentDriver:
  kind: claude-code
  model: claude-opus-4-8

Omit model and the driver uses its own default. An id the CLI rejects surfaces as an error from that CLI — quote the failing id and correct it.

OpenRouter models

With OpenRouter (Codex driver → OpenRouter gateway), the model is an OpenRouter model id, which opens up a wide catalog without installing a per-vendor binary.

Reasoning effort

Effort trades latency and tokens for depth. It is per chat, surfaced as low / medium / high, and maps to each provider’s native control:

• Claude Code passes an effort flag.
• Codex sets the model’s reasoning-effort config.
• Drivers without a native effort control leave it unset.

Higher effort spends more tokens for harder problems; lower effort is cheaper and faster for routine turns. Set it on the chat, not in the committed config.

A cheaper model for inline help

Inline Terminal AI calls default to a cheap, fast model independent of the main driver, overridable per workspace under terminalAi.model. See agentissimo.yaml.