# Model Routing Configuration for SpecKit
# Defines which models/tools to use for each workflow phase
phases:
# Clarification phase - quick, cheap interactions
clarify:
orchestrator_model_alias: haiku
description: "Use Haiku for clarifying specs, asking questions, small edits"
# Task breakdown phase - structured analysis
tasks:
orchestrator_model_alias: haiku
description: "Use Haiku for breaking down work into tasks"
# Planning phase - deeper reasoning required
plan:
orchestrator_model_alias: opusplan
description: "Use Opus Plan Mode for architecture and planning (executes on Sonnet)"
# Implementation phase - high token volume, use Qwen (free on VPN) with Gemini fallback
implement:
tool: qwen_implement
orchestrator_model_alias: sonnet
description: "Use Qwen (free on VPN) or Gemini fallback for code implementation, Sonnet for orchestration"
# Review phase - validation and checking
review:
orchestrator_model_alias: sonnet
description: "Use Sonnet for code review and validation"
# Complex architecture - only when truly needed
architect:
orchestrator_model_alias: opus
description: "Reserve Opus for cross-service architecture decisions"
# Model tier descriptions for agent context
model_tiers:
haiku:
use_for:
- "Clarifying specs and requirements"
- "Small edits and summaries"
- "Task breakdown and checklists"
- "Quick questions and answers"
cost: "lowest"
sonnet:
use_for:
- "Planning within a single service"
- "Validating plans and checklists"
- "Coordinating implementation"
- "Code review"
cost: "moderate"
qwen:
use_for:
- "Code implementation (when on VPN)"
- "Large refactors and multi-file edits"
- "Generating patches and diffs"
- "High-volume code generation"
cost: "free (VPN only, falls back to Gemini)"
tool: "qwen_implement"
gemini_flash:
use_for:
- "Code implementation (fallback when off VPN)"
- "Same as Qwen but always available"
cost: "low (via MCP tool)"
tool: "gemini_implement"
opus:
use_for:
- "Cross-service architecture decisions"
- "Extremely complex reasoning"
- "Novel problem solving"
cost: "highest - use sparingly"
# Default model when not in a specific phase
default_model: sonnet
