vector

Model Routing

Overview

VECTOR — A vector has both direction and magnitude — it points precisely at the right target. When invoked: evaluates task complexity and routes to the least-capable model that can succeed — Tier 0 (no LLM) through Tier 3 (Opus). Systematic routing cuts API costs 50–65% vs treating everything as Opus.

Core principle: Use the least capable model that can handle the task — save cost and latency without sacrificing quality.

Announce at start: "Running VECTOR to assign the correct model tier."

Single-Session vs Subagent Context

Context	How VECTOR applies
Main session	Tier selection is advisory — informs reasoning depth, not model invoked
Spawning subagents (COMMANDER or PHANTOM)	Tier selection is literal — specify model when dispatching
Multi-agent swarms (LEGION)	Assign each agent a tier based on its sub-task complexity

Full cost savings come from subagent dispatch: Haiku for mechanical tasks, Opus only for critical reasoning.

Tier Summary

TIER 3: Opus 4.6     — Architecture, security-critical, cross-system, novel problems, review of Tier 2 work
TIER 2: Sonnet 4.6   — Multi-file features, bug investigation, API integration, complex test design
TIER 1: Haiku 4.5    — Single file edits, simple refactors, obvious test writing, docs, format fixes
TIER 0: No LLM       — Mechanical single-pattern transforms (sed/awk/prettier/eslint)

Model Selection

Security-critical OR novel problem OR cross-system integration → Tier 3 Touches 3+ files OR requires investigation → Tier 2 Single file, mechanical, obvious → Tier 1 No judgment needed, pure pattern → Tier 0

Tier Details

Tier 0: Direct (Zero-LLM)

Latency: <1ms | Cost: $0 | Accuracy: 100% for exact patterns

Gate — ALL must be true: mechanical transform, unambiguous pattern, single file, reversible.

Examples: var → const (sed), sort imports (eslint --fix), format (prettier --write), rename single identifier (sed).

If Tier 0 fails → escalate to Tier 1.

Tier 1: Haiku — Fast/Cheap

Single file edits, simple refactors, mechanical transformations, obvious test writing, documentation, format fixes.

Tier 2: Sonnet — Standard

Multi-file features, bug fixes requiring investigation, API integration, complex test design, module-level refactoring, code review.

Tier 3: Opus — Most Capable

Architecture decisions, security-critical code, cross-system integration, novel problem-solving, elusive bugs (survived 3+ fix attempts), review of Tier 2/3 work before merge.

Dynamic Model Switching

Escalate when:

• Tier 1 gets stuck → Tier 2
• Tier 2 encounters novel problem → Tier 3
• Security implications discovered → Tier 3
• Task scope grows (1 file → 5 files) → Tier 2 or 3

De-escalate when:

• Tier 3 completes architecture → Tier 2 for implementation
• Tier 2 completes design → Tier 1 for mechanical edits

Model Capabilities Reference

Capability	Haiku	Sonnet	Opus
Simple edits	✓	✓	✓
Test writing (basic)	✓	✓	✓
Multi-file features	✗	✓	✓
Bug investigation	✗	✓	✓
Architecture design	✗	△	✓
Security analysis	✗	△	✓
Novel problem-solving	✗	△	✓
Code review	✗	△	✓

✓ = Strong | △ = Adequate | ✗ = Not recommended

Routing Feedback Loop

After every routed task, append to routing_log.md:

| <date> | <task-1-line> | Tier N | Success/Escalated | <why> |

And add/update pointer in MEMORY.md:

- [Routing: <task-type>](routing_log.md) — Tier N → [Success|Escalated|Failed]

Pattern	Adjustment
Tier 1 escalated to Tier 2 ≥ 3x	Start that task type at Tier 2
Tier 2 escalated to Tier 3 ≥ 2x	That pattern = Tier 3 default
Tier 3 completed trivially	Over-routed — note for next time

Red Flags

Never:

• Use Tier 3 for mechanical tasks (waste of money)
• Use Tier 1 for security-critical code (false economy)
• Use Tier 1 for novel problems (will get stuck)
• Stay in Tier 2 after 3 failed attempts (escalate)
• Skip Tier 3 review for architecture decisions

Final Rule

Tier 0: mechanical single-pattern transforms — $0
Tier 1: single file, obvious behavior — cheap/fast
Tier 2: multi-file, investigation — standard
Tier 3: architecture, security, novel — capable

Default to lowest tier that can handle it
Escalate when stuck, de-escalate after hard parts resolved