Routes coding tasks to optimal AI model tier by complexity: no LLM for mechanical edits, Haiku for simple refactors, Sonnet for multi-file bugs, Opus for architecture/security. Saves 50-65% API costs.
npx claudepluginhub gadaalabs/claude-code-on-steroidsThis skill uses the workspace's default tool permissions.
**VECTOR** — *A vector has both direction and magnitude — it points precisely at the right target.*
Creates isolated Git worktrees for feature branches with prioritized directory selection, gitignore safety checks, auto project setup for Node/Python/Rust/Go, and baseline verification.
Executes implementation plans in current session by dispatching fresh subagents per independent task, with two-stage reviews: spec compliance then code quality.
Dispatches parallel agents to independently tackle 2+ tasks like separate test failures or subsystems without shared state or dependencies.
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."
| 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 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)
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
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.
Single file edits, simple refactors, mechanical transformations, obvious test writing, documentation, format fixes.
Multi-file features, bug fixes requiring investigation, API integration, complex test design, module-level refactoring, code review.
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.
Escalate when:
De-escalate when:
| 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
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 |
Never:
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