Unified Delegation Skill

⛔ CRITICAL: No Claude Subagents (MANUAL ENFORCEMENT)

NEVER spawn Claude models (Haiku, Sonnet, Opus) as subagents.

⛔ The deny rules in .claude/settings.local.json DO NOT enforce this. ⛔ Manual verification required before EVERY Task() call.

See CLAUDE.md for the pre-Task() checklist and fallback protocol.

Provider Selection Matrix

Task Type	Best Provider	Why	Fallback 1	Fallback 2	Last Resort
Complex reasoning	Kimi K2.5	Most intelligent, 256K context	GLM-4.7	MiniMax	Claude direct
Image/vision (batch)	Kimi K2.5	Built-in vision capability	GLM-4.6v	—	Claude direct
Creative/brainstorming	GLM-4.7	Strong creative problem-solving	Kimi	MiniMax	Claude direct
Web research	MiniMax	Fast, reliable, cheap	GLM	Kimi	Claude direct
Simple file exploration	MiniMax	Quick turnaround	GLM	Kimi	Claude direct
Batch operations	GLM	Good parallelism	MiniMax	Kimi	Claude direct
Code review	MiniMax	Fast blind-spot check	Kimi	GLM	Claude direct

Automatic Fallback Protocol

When a Task() call fails, follow this chain (NEVER fallback to Haiku/Sonnet/Opus):

Primary Failed	Try Next	Then Try	Last Resort
Kimi K2.5	GLM-4.7	MiniMax	Claude does work directly
GLM-4.7	MiniMax	Kimi K2.5	Claude does work directly
MiniMax	GLM-4.7	Kimi K2.5	Claude does work directly

Example:

python

# ❌ WRONG - Kimi failed, fell back to Haiku
Task(model="kimi", ...) → Error
Task(model="haiku", ...) → VIOLATION

# ✅ CORRECT - Kimi failed, fell back to GLM
Task(model="kimi", ...) → Error
Task(model="minimax", ...) → Success

If all cheap models fail:

•Do the work yourself (Claude)
•Document why all subagents failed
•Flag for manual investigation with user
•NEVER use Haiku/Sonnet/Opus as fallback

Quick Decision Flow

code

┌─ Is it reasoning/decisions? ──────────────────┐
│  YES → Claude does it directly                │
│  NO  → Delegate to subagent ↓                 │
├───────────────────────────────────────────────┤
│                                               │
│  ┌─ Does it need vision? ───────────────────┐ │
│  │  YES → Kimi K2.5 (or GLM-4.6v fallback)  │ │
│  │  NO  ↓                                   │ │
│  └──────────────────────────────────────────┘ │
│                                               │
│  ┌─ Is it complex/creative? ────────────────┐ │
│  │  Complex → Kimi K2.5                     │ │
│  │  Creative → GLM-4.7                      │ │
│  │  Simple → MiniMax                        │ │
│  └──────────────────────────────────────────┘ │
└───────────────────────────────────────────────┘

Provider Profiles

Kimi K2.5 (Most Capable)

Context: 256K tokens | Vision: Yes | Thinking mode: Yes

Best for:

•Complex multi-step reasoning
•Batch image analysis (10+ images)
•Tasks requiring deep understanding
•Fallback for failed GLM tasks

Launcher: .\scripts\start-kimi.ps1

API Config:

code

Base URL: https://api.moonshot.cn/anthropic/
Models: kimi-k2.5-thinking, kimi-k2-turbo-preview

GLM-4.7 (Creative)

Context: 128K tokens | Vision: GLM-4.6v variant | Thinking mode: Yes

Best for:

•Creative brainstorming
•Mathematical reasoning (95.7% AIME 2025)
•Parallel batch tasks
•Tool use orchestration

MCP: .cursor/mcp.json (GLM-4.6v configured)

MiniMax M2.1 (Fast & Cheap)

Context: 128K tokens | Vision: VLM API | Speed: Fastest

Best for:

•Quick web searches
•Simple file exploration
•Structured data extraction
•Code review for blind spots

Launcher: .\scripts\start-claude-minimax.ps1

MCP: .cursor/mcp.json (MiniMax configured)

Delegation Patterns

Pattern 1: Research → Claude Decides

code

1. Claude receives task requiring research
2. Claude spawns MiniMax: "Find all uses of X in codebase"
3. MiniMax returns findings
4. Claude reasons and implements

Pattern 2: Batch Vision Analysis

code

1. Claude needs to analyze 20 sprites
2. Claude spawns Kimi K2.5: "Analyze quality of each sprite"
3. Kimi returns analysis for all 20
4. Claude makes decisions based on report

Pattern 3: Creative Exploration

code

1. Claude needs alternative approaches
2. Claude spawns GLM-4.7: "Brainstorm 5 solutions for X"
3. GLM returns creative options
4. Claude selects and refines best approach

Pattern 4: Code Review

code

1. Claude writes code
2. Claude spawns MiniMax: "Check for bugs, edge cases, security issues"
3. MiniMax returns concerns
4. Claude addresses or dismisses with reasoning

Parallel Delegation

Launch multiple subagents in a single message:

code

Task(prompt="Research X", subagent_type="general-purpose")  ←─┐
Task(prompt="Research Y", subagent_type="general-purpose")  ←─┼─ Parallel
Task(prompt="Research Z", subagent_type="general-purpose")  ←─┘

Rules:

•Independent tasks → parallel
•Dependent tasks → sequential
•Never chain Claude subagents

Background Execution (Token Suspension)

Problem: Claude tokens burn while waiting for subagent results. Solution: Use run_in_background=true + end turn early.

Pattern: Fire-and-Retrieve

code

1. Claude receives task requiring research
2. Task(prompt="...", run_in_background=true) → returns output_file
3. Claude ends turn: "Research agent dispatched. Say 'continue' for results."
4. User says "continue"
5. TaskOutput(task_id="...", block=true) → retrieves results
6. Claude synthesizes and responds

When to Use Background Execution

Scenario	Background?	Why
Research >30 sec	✅ Yes	Saves expensive Claude wait time
Batch image analysis	✅ Yes	Long-running, user can wait
Quick file lookup	❌ No	Faster to wait inline
Claude needs result to continue	❌ No	Would block anyway

Token Savings Calculation

code

Blocking:     Claude waits 60s = 60s of Opus tokens burned
Background:   Claude ends turn = 0s of Opus tokens burned
              (subagent tokens are 50x cheaper)

Example Usage

code

# Fire (spawn and end turn immediately)
Task(
  prompt="Analyze all 20 sprites in assets/sprites/",
  subagent_type="general-purpose",
  run_in_background=true
)
→ Returns: {task_id: "abc123", output_file: "/path/to/output"}

# ... Claude ends turn, tells user to say "continue" ...

# Retrieve (on next turn)
TaskOutput(task_id="abc123", block=true)
→ Returns: Full subagent analysis

Token Economics

Provider	Relative Cost	When to Use
Claude Opus	50x	Final decisions, complex reasoning
Claude Sonnet	10x	Medium reasoning (avoid as subagent)
Kimi K2.5	1x	Complex tasks, vision
GLM-4.7	1x	Creative, batch
MiniMax	1x	Fast, simple

Key insight: 1 hour Claude exploration = 50 hours subagent exploration (cost).

Common Mistakes

Mistake	Impact	Fix
Claude spawning Haiku	Expensive	Use MiniMax instead
Sequential when parallel possible	Slow	Single message, multiple Tasks
Kimi for simple lookup	Overkill	Use MiniMax
MiniMax for complex reasoning	Poor quality	Use Kimi K2.5
Claude reading 10+ files	Context bloat	Delegate exploration

Integration with Other Skills

•/skill kimi-k2.5 - Detailed Kimi setup and patterns
•/skill minimax-mcp - MiniMax MCP integration details
•/skill token-efficient-delegation - Full token economics
•/skill subagent-best-practices - General subagent patterns

[Opus 4.5 - 2026-01-29]