Thrifty Task Delegation

Use this skill to delegate tasks to opencode/glm-5, a cost-effective model ideal for:

• Simple research and information gathering
• Code exploration and file searches
• Document processing and summarization
• Routine automation tasks
• Parallel processing of independent subtasks

Usage Pattern

Invoke glm-5 via the task tool:

task({
  description: "Brief task summary",
  prompt: "Detailed instructions for the agent",
  subagent_type: "general"  // glm-5 is the default model
})

Cost Guidelines

Task Type	Recommended Model	Est. Cost
Simple search/explore	glm-5 (default)	Low
Code review	code-review-a	Low
Complex architecture	oracle	Medium
Deep debugging	code-review-c	Higher

Best Practices

1. Break large tasks into smaller chunks - glm-5 handles focused tasks well
2. Provide clear, specific instructions - reduces iteration and context usage
3. Use for parallel work - delegate multiple independent tasks simultaneously
4. Verify complex outputs - have a stronger model review critical results
5. Batch similar operations - group related file operations together

Example Delegations

File exploration:

task({
  description: "Find auth-related files",
  prompt: "Search the codebase for all files related to authentication. Return a list of file paths and brief descriptions of what each does.",
  subagent_type: "general"
})

Documentation summarization:

task({
  description: "Summarize API docs",
  prompt: "Read docs/api.md and provide a 3-bullet summary of the main endpoints and their purposes.",
  subagent_type: "general"
})

Parallel processing:

// Delegate multiple independent tasks concurrently
const tasks = [
  task({ description: "Analyze utils.ts", prompt: "...", subagent_type: "general" }),
  task({ description: "Analyze helpers.ts", prompt: "...", subagent_type: "general" }),
  task({ description: "Analyze constants.ts", prompt: "...", subagent_type: "general" })
];
await Promise.all(tasks);

When NOT to Use glm-5

• Complex multi-step architectural decisions
• Security-critical code review
• Novel problem-solving requiring creativity
• Tasks requiring deep contextual understanding

For these, use oracle, code-review-c, or other specialized subagents instead.