Smart Model Routing for Z.AI

Auto-route tasks to the cheapest z.ai (GLM) model that works correctly. Three-tier progression: Flash → Standard → Plus/32B. Classify before responding. FLASH (default): factual Q&A, greetings, reminders, status checks, lookups, simple file ops, heartbeats, casual chat, 1–2 sentence tasks, cron jobs. ESCALATE TO STANDARD: code >10 lines, analysis, comparisons, planning, reports, multi-step reasoning, tables, long writing >3 paragraphs, summarization, research synthesis, most user conversations. ESCALATE TO PLUS/32B: architecture decisions, complex debugging, multi-file refactoring, strategic planning, nuanced judgment, deep research, critical production decisions. Rule: If a human needs >30 seconds of focused thinking, escalate. If Standard struggles with complexity, go to Plus/32B. Save major API costs by starting cheap and escalating only when needed.

Security Scan

This is an instruction-only routing policy for selecting z.ai (GLM) models (Flash → Standard → Plus/32B); it requests no credentials, installs nothing, and its instructions are consistent with its stated purpose.

Purpose & Capability

The name and description (auto-route to the cheapest working GLM model) match the SKILL.md guidance. However, the skill is purely an instruction document — it contains rules and example session_spawn calls but no code, no integration, and no credentials. That means it cannot itself contact z.ai or enforce routing; it only tells an agent how to decide which model to pick. This is coherent but important to understand.

Instruction Scope

SKILL.md confines itself to classification and escalation rules for model selection and includes example usage (sessions_spawn). It does not instruct reading unrelated files, sending data to external endpoints, or harvesting secrets. The guidance assumes the agent/platform provides a sessions_spawn capability.

Install Mechanism

No install spec and no code files are present. This minimizes disk writes and third-party installs; lower risk and consistent with an instruction-only policy.

Credentials

No environment variables, credentials, or config paths are requested. That is proportionate to an instruction-only routing policy.

Persistence & Privilege

Skill does not request always:true, does not modify other skills, and keeps no persistent privileges. Autonomous invocation is allowed by platform default but the skill itself doesn't demand elevated presence or cross-skill access.

Assessment

This skill is a policy document (routing heuristics) rather than an integration: it won't call z.ai or incur costs on its own and it doesn't ask for any secrets. Before relying on it, confirm your agent platform actually supports and enforces model selection (sessions_spawn or equivalent) and respects these rules. Test the escalation rules with non-critical tasks to ensure they don't cause excessive upscaling (and cost). If you intend automatic spawning of z.ai models, you'll need appropriate platform wiring and credentials elsewhere — the skill itself does not provide or request them. Finally, monitor logs and set budget/timeout guards so repeated automatic escalation or retries can't run up unexpected API charges.

Smart Model Switching

Three-tier z.ai (GLM) routing: Flash → Standard → Plus / 32B

Start with the cheapest model. Escalate only when needed. Designed to minimize API cost without sacrificing correctness.

---

The Golden Rule

If a human would need more than 30 seconds of focused thinking, escalate from Flash to Standard.
If the task involves architecture, complex tradeoffs, or deep reasoning, escalate to Plus / 32B.

---

Model Reality (Relative)

Tier	Example Models	Purpose
Flash	GLM-4.5-Flash, GLM-4.7-Flash	Fastest & cheapest
Standard	GLM-4.6, GLM-4.7	Strong reasoning & code
Plus / 32B	GLM-4-Plus, GLM-4-32B-128K	Heavy reasoning & architecture

Bottom line: Wrong model selection wastes money OR time. Flash for simple, Standard for normal work, Plus/32B for complex decisions.

---

💚 FLASH — Default for Simple Tasks

Stay on Flash for:

• Factual Q&A — “what is X”, “who is Y”, “when did Z”
• Quick lookups — definitions, unit conversions, short translations
• Status checks — monitoring, file reads, session state
• Heartbeats — periodic checks, OK responses
• Memory & reminders
• Casual conversation — greetings, acknowledgments
• Simple file ops — read, list, basic writes
• One-liner tasks — anything answerable in 1–2 sentences
• Cron jobs (always Flash by default)

NEVER do these on Flash

• ❌ Write code longer than 10 lines
• ❌ Create comparison tables
• ❌ Write more than 3 paragraphs
• ❌ Do multi-step analysis
• ❌ Write reports or proposals

---

💛 STANDARD — Core Workhorse

Escalate to Standard for:

Code & Technical

• Code generation — functions, scripts, features
• Debugging — normal bug investigation
• Code review — PRs, refactors
• Documentation — README, comments, guides

Analysis & Planning

• Comparisons and evaluations
• Planning — roadmaps, task breakdowns
• Research synthesis
• Multi-step reasoning

Writing & Content

• Long-form writing (>3 paragraphs)
• Summaries of long documents
• Structured output — tables, outlines

Most real user conversations belong here.

---

❤️ PLUS / 32B — Complex Reasoning Only

Escalate to Plus / 32B for:

Architecture & Design

• System and service architecture
• Database schema design
• Distributed or multi-tenant systems
• Major refactors across multiple files

Deep Analysis

• Complex debugging (race conditions, subtle bugs)
• Security reviews
• Performance optimization strategy
• Root cause analysis

Strategic & Judgment-Based Work

• Strategic planning
• Nuanced judgment and ambiguity
• Deep or multi-source research
• Critical production decisions

---

🔄 Implementation

For Subagents

// Routine monitoring
sessions_spawn(task="Check backup status", model="GLM-4.5-Flash")

// Standard code work
sessions_spawn(task="Build the REST API endpoint", model="GLM-4.7")

// Architecture decisions
sessions_spawn(task="Design the database schema for multi-tenancy", model="GLM-4-Plus")
For Cron Jobs
json
Copy code
{
  "payload": {
    "kind": "agentTurn",
    "model": "GLM-4.5-Flash"
  }
}
Always use Flash for cron unless the task genuinely needs reasoning.

📊 Quick Decision Tree
pgsql
Copy code
Is it a greeting, lookup, status check, or 1–2 sentence answer?
  YES → FLASH
  NO ↓

Is it code, analysis, planning, writing, or multi-step?
  YES → STANDARD
  NO ↓

Is it architecture, deep reasoning, or a critical decision?
  YES → PLUS / 32B
  NO → Default to STANDARD, escalate if struggling
📋 Quick Reference Card
less
Copy code
┌─────────────────────────────────────────────────────────────┐
│                  SMART MODEL SWITCHING                      │
│              Flash → Standard → Plus / 32B                  │
├─────────────────────────────────────────────────────────────┤
│  💚 FLASH (cheapest)                                        │
│  • Greetings, status checks, quick lookups                  │
│  • Factual Q&A, reminders                                   │
│  • Simple file ops, 1–2 sentence answers                    │
├─────────────────────────────────────────────────────────────┤
│  💛 STANDARD (workhorse)                                    │
│  • Code > 10 lines, debugging                               │
│  • Analysis, comparisons, planning                          │
│  • Reports, long writing                                    │
├─────────────────────────────────────────────────────────────┤
│  ❤️ PLUS / 32B (complex)                                    │
│  • Architecture decisions                                   │
│  • Complex debugging, multi-file refactoring                │
│  • Strategic planning, deep research                        │
├─────────────────────────────────────────────────────────────┤
│  💡 RULE: >30 sec human thinking → escalate                 │
│  💰 START CHEAP → SCALE ONLY WHEN NEEDED                    │
└─────────────────────────────────────────────────────────────┘
Built for z.ai (GLM) setups.