name: model-router
description: Route work to the cheapest model tier that can reliably complete it.
Use when: (1) choosing a model for a task, (2) deciding model level for sub-agents,
(3) escalating after a failed attempt, (4) avoiding premium-model overuse on routine work.
category: principles
user-invocable: true
Model Router
Choose model tier by task complexity. Keep routine work cheap. Escalate only when required.
Core Policy
- Classify each task first:
ROUTINE, MODERATE, or COMPLEX.
- Start with the lowest tier that can succeed.
- Escalate one tier after a quality failure.
- Keep sub-agents on the cheapest viable tier by default.
Task Classification
ROUTINE
Use lowest-cost tier.
Signals:
- File reads/writes
- Status checks and health checks
- Deterministic formatting/transformations
- Bulk fetch/lookup tasks with clear success criteria
MODERATE
Use mid-tier.
Signals:
- Multi-step synthesis
- Standard code generation/refactoring
- Non-trivial summaries
- Typical implementation tasks with known patterns
COMPLEX
Use top-tier.
Signals:
- Ambiguous requirements
- Novel debugging/root-cause analysis
- Architecture tradeoffs
- Security-sensitive reasoning
- Repeated failure on lower tiers
Escalation And De-escalation
- Escalate when output quality is insufficient for the task goal.
- Escalate one tier at a time.
- De-escalate when work returns to deterministic or repetitive operations.
- Document escalation reason in one sentence before switching tiers.
Sub-Agent Rules
- Default sub-agents to the cheapest tier that matches task class.
- Reserve high-tier sub-agents for clearly complex reasoning only.
- If parent agent is high-tier, do not inherit that tier by default.
Anti-Patterns
- Premium tier for file I/O or status checks.
- Premium tier for repetitive extraction and formatting.
- Immediate jump from lowest tier to highest tier without justification.
- Leaving routine follow-up tasks on high tier after a complex step is complete.
Quick Decision Flow
- Classify task complexity.
- Pick the lowest viable tier.
- Run and evaluate output quality.
- Escalate one tier only if needed.
- De-escalate for routine follow-up.
