Model Router

Decision Algorithm

Step 1: Task Classification

task_signals[20]{signal,complexity,weight}:
  # Trivial signals (+1 each, trivial if sum ≥ 3)
  typo|spelling,trivial,+3
  rename variable,trivial,+3
  fix lint error,trivial,+3
  update comment,trivial,+2
  format code,trivial,+2
  single file mention,trivial,+1

  # Simple signals (+1 each, simple if sum ≥ 2)
  fix bug,simple,+2
  add validation,simple,+2
  refactor function,simple,+2
  write test for,simple,+2
  update config,simple,+1

  # Complex signals (+1 each, complex if sum ≥ 2)
  new feature,complex,+2
  multiple files,complex,+2
  API endpoint,complex,+2
  database migration,complex,+2
  integration,complex,+1

  # Architectural signals (+2 each, opus if sum ≥ 2)
  system design,architectural,+3
  architecture,architectural,+3
  security audit,architectural,+3
  major refactor,architectural,+2
  performance optimization,architectural,+2
  breaking change,architectural,+2

Step 2: Override Rules

overrides[6]{condition,force_model,reason}:
  user says "use opus",opus,Explicit request
  user says "use haiku",haiku,Explicit request
  user says "cheap/fast",haiku,Cost preference
  user says "thorough/careful",opus,Quality preference
  security/vulnerability task,opus,Safety critical
  production deployment,opus,Risk management

Step 3: Context Modifiers

modifiers[4]{context,adjustment,reason}:
  unfamiliar codebase,-1 complexity,Need more exploration
  well-documented project,+1 complexity,Can work faster
  critical path code,-1 complexity,Need more care
  test/mock code,+1 complexity,Lower risk

Output Format

When model routing applies, include in response:

🎯 Model: [model] | Complexity: [level] | Reason: [why]

Examples:

🎯 Model: haiku | Complexity: trivial | Reason: Single typo fix
🎯 Model: sonnet | Complexity: simple | Reason: Bug fix in auth module
🎯 Model: opus | Complexity: architectural | Reason: Security audit requested

Cost Savings Examples

*Opus costs more but prevents costly mistakes and rework

Integration with Agent Detector

The model-router works WITH agent-detector:

1. agent-detector runs → selects primary/secondary agents
2. model-router runs → selects optimal model for task
3. Both results shown in banner

Updated Banner Format:

⚡ 🐸 AURA FROG v1.20.0 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
┃ Agent: ui-expert │ Phase: 4 - Implement                ┃
┃ Model: haiku │ Teams: ✓ enabled                         ┃
┃ 🎯 Trivial task: typo fix │ 🔥 Quick fix incoming!      ┃
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

User Controls

Users can override model selection:

# Force specific model
"Fix this typo, use opus"  → Opus (overridden)

# Request cost optimization
"Quick fix, minimize cost" → Haiku (preference)

# Request thoroughness
"Audit this carefully"     → Opus (preference)

Complexity Escalation

If task becomes more complex during execution:

1. Detect escalation: Task grew from "fix typo" to "refactor module"
2. Notify user: "Task complexity increased. Consider using Sonnet for better results."
3. Continue or restart: User decides

Metrics Tracking

Track model usage for optimization:

metrics[4]{metric,purpose}:
  tasks_by_model,Distribution of task complexity
  cost_per_task,Average cost by complexity
  rework_rate,Tasks needing follow-up by model
  user_overrides,How often users override routing

Skip Conditions

Don't route when:

• User explicitly specifies model
• In middle of multi-turn conversation (maintain consistency)
• Task is ambiguous (ask for clarification first)

Related Files

• skills/agent-detector/SKILL.md - Agent selection (runs first)
• rules/context-management.md - Token optimization rules
• docs/REFACTOR_ANALYSIS.md - Cost optimization analysis

Version: 1.0.0 | Last Updated: 2026-01-21