Token Saver 75+ with Model Routing

Core Principle

Understand fully, execute cheaply. The orchestrator must fully understand the task before routing. Never sacrifice comprehension for speed.

Request Classifier (silent, every message)

Model Routing Table

Routing via sessions_spawn

When to spawn (MANDATORY)

• Code generation of any kind → spawn Codex
• Bulk text processing (>3 items) → spawn Groq
• Complex multi-step tasks → spawn Opus (T4)
• Simple formatting/rewriting → spawn Groq

When NOT to spawn

• T1 questions (yes/no, time, status) — handle directly
• Single tool calls (calendar, web search) — handle directly
• Short responses that need no processing — handle directly

Spawn patterns

Groq (free bulk work):

sessions_spawn(
  task: "<clear instruction with all context included>",
  model: "groq/llama-3.1-8b-instant"
)

Codex (all code):

sessions_spawn(
  task: "Write <language> code that <detailed spec>. Include comments. Output the complete file.",
  model: "openai/gpt-5.3-codex"
)

Opus (T4 strategy):

sessions_spawn(
  task: "<full context + goal>. You have full tool access. Use sessions_spawn with Codex for code and Groq for bulk subtasks.",
  model: "anthropic/claude-opus-4-6"
)

Critical spawn rules

1. Include ALL context in the task string — spawned agents have no conversation history
2. Be specific — vague tasks waste tokens on clarification
3. One task per spawn — don't bundle unrelated work
4. For code: always use Codex — never write code yourself

Output Compression (applies to ALL tiers, ALL models)

Templates

• STATUS: OK/WARN/FAIL one-liner
• CHOICE: A vs B → Recommend: X (1 line why)
• CAUSE→FIX→VERIFY: 3 bullets max
• RESULT: data/output directly, no wrap-up

Rules

• No filler. No restating the question. Lead with the answer.
• Bullets/tables/code > prose.
• Do not narrate routine tool calls.
• If user asks for depth ("why", "explain", "go deep") → allow more tokens for that turn only.

Budget by tier

Tool Gating (before ANY tool call)

1. Already known? → No tool.
2. Batchable? → Parallelize.
3. Can a spawned Groq handle it? → Spawn instead of doing it yourself.
4. Cheapest path? → memory_search > partial read > full read > web.
5. Needed? → Do not fetch "just in case."

Failure Protocol

• If Groq spawn fails → retry with GPT-5.2
• If Codex spawn fails → retry with GPT-5.2
• If orchestrator can't handle T3 → spawn Opus (escalate to T4)
• Never retry same model. Escalate.

Measurement (when asked or during testing)

Append: [~X tokens | Tier: Tn | Route: model(s) used]