Phase Model Routing

Phase model routing lets you use a smaller, faster model for planning and a more capable model for execution — reducing cost and latency on the planning phase without sacrificing quality on the execution phase.

Configuration

[options.phase_model]
planning_model = "claude-haiku-4-20250514"
execution_model = "claude-sonnet-4-20250514"

How it works

1. Agent enters plan phase — BMO routes inference to planning_model
2. Plan is produced and (optionally) approved via /approve
3. Agent enters execute phase — BMO routes inference to execution_model

If planning_model is not set, the default agent model is used for both phases.

Architect / editor naming

This combination is sometimes called architect mode elsewhere: a reasoning or fast model proposes the plan, and a stronger model executes tool calls. BMO implements it with phase model + optional staged workflow (preset).

Related

• Staged Workflow — plan/execute gating
• Workflow map — where phase routing fits in the larger runtime loop
• Configuration — options.phase_model and adjacent runtime settings

Integration with Staged Workflow

Phase model routing works alongside Staged Workflow. When both are configured:

• Staged Workflow controls the plan → execute gate (requiring /approve)
• Phase model routing selects the model for each phase

If staged workflow is not enabled, BMO uses planning_model for the first turn and execution_model for subsequent turns automatically.

Choosing models

Phase	Recommended choice	Reason
Planning	Smaller/faster model	Plan phase is reasoning-heavy but output is text, not code
Execution	Larger/capable model	Execution requires precise code generation and tool use

Any model available in your provider config can be used for either phase. When the model ID exists in multiple providers, use provider/model to target the exact provider you want.