[Feature Request] ZCode: Allow Subagents to Use Different Model/Effort Level Than Session Spine

🚀 Feature Request

Add the ability for ZCode subagents to be spawned with a different model and effort level than the main session's "spine" model.

Example Use Case:

• Session Spine: GLM-5.3 (full power — planning, architecture, review)
• Subagent Workers: GLM-5.3-Flash (fast/cheap — execution, file editing, searches)

---

📋 Current Behavior (Limitation)

❌ Issue

When using ZCode's multi-agent/subagent capabilities:

• All subagents inherit the same model as the parent session
• No way to specify "run this subagent on Flash model"
• No way to mix effort levels within a single session

User Experience

"In ZCode, I asked 5.3 to spawn subagents that run as 5.3-Flash and it told me that this is not possible so the subagents were 5.3 (non Flash) instead."

"It would be super cool if we can spawn subagents that have a different model and difficulty, so the session can act as a spine with 5.3 (non Flash) and workers can execute as 5.3-Flash."
— blockomoco [OC], Discord

---

✅ Proposed Solution

Architecture: "Spine + Workers" Pattern

ZCODE SESSION
|
+-- SESSION SPINE (Orchestrator)
|   Model: GLM-5.3 (Full Power)
|   Role: Planning, Review, Architecture
|   Command: "Spawn 3 subagents to refactor modules"
|
+-- SUBAGENT Worker 1  --> Model: GLM-5.3-Flash  Cost: $0.01
+-- SUBAGENT Worker 2  --> Model: GLM-5.3-Flash  Cost: $0.01
+-- SUBAGENT Worker 3  --> Model: GLM-5.3-Flash  Cost: $0.01
|
RESULT: 1x expensive + 3x cheap = HIGH QUALITY, LOWER COST

---

🎯 UI/UX Implementation

Option A: Session-Level Default Setting

Agent Configuration
---------------------
Main Session Model: [GLM-5.3        ▼]

Subagent Default:
( ) Same as session                    [current]
(*) Specify model:                       [NEW]
    Model: [GLM-5.3-Flash      ▼]
    Effort: [Low ▼]

Option B: Per-Invocation Override (Advanced)

User can prompt:

@spawn_subagent(
  task: "Refactor auth module",
  model: "glm-5.3-flash",
  effort: "low",
  max_tokens: 8000
)

Option C: Agent Pool Configuration

Agent Pool
---------
Orchestrator:  GLM-5.3      (planning)
Workers (1-5): GLM-5.3-Flash (execution)
Reviewer:      GLM-5.3      (quality gate)

[Add Role] [Remove] [Reorder]

---

💰 Cost-Benefit Analysis

Token/Cost Savings Scenario

Approach	# of Agents	Model	Est. Cost/Session
Current (all same)	4 agents	GLM-5.3	~$0.40
Proposed (spine+workers)	1× GLM-5.3 + 3× Flash	Mixed	~$0.16
Savings	—	—	~60% reduction ✅

Quality Retention

• Spine (GLM-5.3) handles: Planning, decomposition, review → High-quality decisions
• Workers (Flash) handle: Execution, file ops, simple tasks → Adequate for routine work
• Net result: Near-full-power quality at fraction of cost

---

🏆 Competitive Analysis

How Competitors Handle This:

Platform	Feature Name	Status
OpenAI Codex	Multi-model agent delegation	✅ Supported
Anthropic Claude	Subagent model override	✅ Supported
Cursor	Hierarchical agent models	✅ Supported
Windsurf	Cascade model selection	✅ Supported
ZCode	Subagent model selection	❌ Not supported ← WE ARE HERE

Quote from User:

"This is a capability that both Codex and Claude have and it allows to save considerable amounts of tokens/$ this way while retaining very high output quality close to what the stronger model would achieve by itself."
— blockomoco [OC]

---

🔧 Technical Implementation

Backend Changes Required

1. Agent Spawn API Extension

interface SpawnSubagentRequest {
  task: string;
  context: AgentContext;
  
  // NEW FIELDS:
  modelOverride?: string;       // e.g., "glm-5.3-flash", "glm-4.5"
  effortOverride?: 'low' | 'medium' | 'high';
  maxTokens?: number;
  temperature?: number;
}

2. Session Config Schema

interface SessionConfig {
  primaryModel: string;           // Existing: "glm-5.3"
  
  // NEW:
  subagentDefaults?: {
    model?: string;              // undefined = inherit from primary
    effort?: 'low' | 'medium' | 'high';
    maxConcurrent?: number;      // Limit parallel workers
  };
  
  agentRoles?: Array({
    name: string;                // "orchestrator", "worker", "reviewer"
    model: string;
    effort: string;
    responsibilities: string[];
  });
}

3. Quota/Billing Logic

def calculate_session_cost(session):
    base_cost = session.spine_model.cost_per_token * spine_tokens
    
    for subagent in session.subagents:
        model = subagent.model_override or session.primary_model
        base_cost += model.cost_per_token * subagent.tokens_used
    
    return base_cost

---

📊 Use Cases & Benefits

Primary Use Cases

Use Case	Spine Model	Worker Model	Benefit
Large refactoring	GLM-5.3	GLM-5.3-Flash	Plan smart, execute cheap
Code generation	GLM-5.3	GLM-4.5	Architect with best, generate with good
Testing/debugging	GLM-5.3-Flash	GLM-5.3-Flash	All fast for quick iteration
Documentation	GLM-5.3	GLM-5.3-Flash	Quality writing, cheap research
Multi-file edits	GLM-5.3	GLM-5.3-Flash	Coordinate carefully, edit rapidly

Who Benefits?

User Type	Impact
Power users	Optimize cost without sacrificing quality
Budget-conscious	Stretch quota further on free/lower tiers
Enterprise	Fine-grained cost control per project type
Quality-focused	Keep spine on best model, offload grunt work

---

🎯 Implementation Priority

Phase 1: Basic Model Override (MVP)

• Add model parameter to subagent spawn function
• Support GLM-5.3 ↔ GLM-5.3-Flash switching
• Show model assignment in agent panel UI
• Track per-agent token usage separately

Phase 2: Effort/Temperature Control

• Add effort parameter (low/medium/high)
• Map effort to temperature, top_p, max_tokens defaults
• Allow per-subagent overrides in prompts

Phase 3: Agent Roles/Pools

• Pre-configured role templates (orchestrator, worker, reviewer)
• Session-level default agent pool
• Save/load agent configurations
• Visual DAG of agent hierarchy with model labels

Phase 4: Smart Recommendations

• AI suggests optimal model per subtask
• Cost estimation before execution
• Auto-balance based on complexity detection

---

⚠️ Considerations & Edge Cases

Case	Recommended Handling
User doesn't have Flash access	Fall back to session model or show error
Context window mismatch	Truncate/prioritize context for smaller models
Subagent needs to spawn its own subagent	Inherit parent's model override or use session default
Billing display	Break down costs by model used
Error in Flash subagent	Retry with spine model (automatic escalation)

---

📈 Success Metrics

If implemented, track:

• Cost savings % for sessions using mixed models
• Task completion rate comparison (mixed vs uniform)
• User adoption rate of the feature
• Satisfaction scores from power users

---

📝 Additional Context

Full User Statement

"In ZCode, I asked 5.3 to spawn subagents that run as 5.3-Flash and it told me that this is not possible so the subagents were 5.3 (non Flash) instead."

"It would be super cool if we can spawn subagents that have a different model and difficulty, so the session can act as a spine with 5.3 (non Flash) and workers can execute as 5.3-Flash."

"This is a capability that both Codex and Claude have and it allows to save considerable amounts of tokens/$ this way while retaining very high output quality close to what the stronger model would achieve by itself."
— blockomoco [OC], Discord

---

🔗 Reference

Original Suggestion: Submitted by blockomoco [OC] via Discord community

---

Submitted by: Roman Galaxys10 (Roman) — Z.ai Volunteer Ambassador
Discord: bignavi_x
GitHub: romangalaxys10-spec
Source: Discord Community — User: blockomoco [OC]

No repository files or tests are named. Start by locating the SpawnSubagentRequest and SessionConfig entry points, then inspect the session cost calculation; done requires an agreed implementation scope, validation of model and effort overrides, and coverage for billing and fallback behavior.