agent-sdk-development（Agent SDK 开发）

Agent SDK Meta Skill

🚨 MANDATORY: Invoke docs-management First

STOP - Before providing ANY response about Claude Agent SDK: 1. INVOKE docs-management skill 2. QUERY using keywords: Agent SDK, TypeScript SDK, Python SDK, sessions, custom tools, permissions, or related topics 3. BASE all responses EXCLUSIVELY on official documentation loaded Skipping this step results in outdated or incorrect information.

Verification Checkpoint

Before responding, verify:

• Did I invoke docs-management skill?
• Did official documentation load?
• Is my response based EXCLUSIVELY on official docs?

If ANY checkbox is unchecked, STOP and invoke docs-management first.

Overview

Central authority for Claude Agent SDK (TypeScript and Python SDKs). This skill uses 100% delegation to docs-management - it contains NO duplicated official documentation.

Architecture: Pure delegation with keyword registry. All official documentation is accessed via docs-management skill queries.

Reference Loading Guide

This skill follows a pure delegation architecture with three loading layers:

Layer 1: Always in Context

• SKILL.md (this file) - Keyword registry, decision trees, delegation patterns

Layer 2: On-Demand via docs-management

All official Agent SDK documentation is accessed through docs-management queries:

Layer 3: External Resources (GitHub)

• TypeScript SDK: anthropics/claude-agent-sdk-typescript
• Python SDK: anthropics/claude-agent-sdk-python

Loading Strategy

This skill has no local reference files. All documentation is accessed via docs-management queries using the keyword registry below. This ensures:

• Zero duplication of official documentation
• Always-current information from canonical sources
• Zero maintenance overhead for documentation sync

When to Use This Skill

Keywords: Agent SDK, Claude Agent SDK, TypeScript SDK, Python SDK, npm install claude-agent-sdk, pip install claude-agent-sdk, sessions, session resumption, forking sessions, streaming mode, single mode, custom tools, allowedTools, disallowedTools, permissionMode, system prompts, appendSystemPrompt, outputStyle, CLAUDE.md, settingSources, cost tracking, todo tracking, structured outputs, hosting, plugins SDK, MCP SDK, agent branding

Use this skill when:

• Installing or configuring Claude Agent SDK
• Setting up authentication (API key, Bedrock, Vertex)
• Managing sessions and resumption
• Choosing between streaming and single input modes
• Creating custom tools
• Configuring permissions (allowedTools, disallowedTools)
• Modifying system prompts
• Tracking costs and usage
• Working with todo lists in agents
• Getting structured outputs
• Hosting agents in production
• Integrating MCP with SDK
• Understanding SDK branding requirements

Keyword Registry for docs-management Queries

Use these keywords when querying docs-management skill for official documentation:

SDK Fundamentals

Authentication

Sessions

Input Modes

Custom Tools

Permissions

System Prompts

Cost and Tracking

Structured Outputs

Hosting and Deployment

Plugins and MCP

Branding

SDK Feature Integration

API Reference

Quick Decision Tree

What do you want to do?

1. Install the SDK -> Query docs-management: "Agent SDK installation"
2. Set up authentication -> Query docs-management: "Agent SDK authentication"
3. Manage sessions -> Query docs-management: "session management SDK", "resume option"
4. Fork a session -> Query docs-management: "forking sessions", "forkSession"
5. Choose input mode -> Query docs-management: "streaming vs single mode"
6. Create custom tools -> Query docs-management: "custom tools SDK"
7. Configure permissions -> Query docs-management: "allowedTools", "permissionMode SDK"
8. Modify system prompts -> Query docs-management: "modifying system prompts"
9. Use CLAUDE.md files -> Query docs-management: "CLAUDE.md SDK", "settingSources"
10. Track costs -> Query docs-management: "cost tracking SDK"
11. Add todo tracking -> Query docs-management: "todo tracking SDK"
12. Get structured outputs -> Query docs-management: "structured outputs SDK"
13. Host in production -> Query docs-management: "hosting Agent SDK"
14. Use subagents -> Query docs-management: "subagents SDK"
15. Integrate skills -> Query docs-management: "skills SDK", "agent skills SDK"
16. Add skills (commands) -> Query docs-management: "skills SDK", "SDK commands"
17. Load plugins -> Query docs-management: "plugins SDK"

Topic Coverage

SDK Installation and Setup

• npm package: @anthropic-ai/claude-agent-sdk
• pip package: claude-agent-sdk
• TypeScript and Python SDK options
• Migration from Claude Code SDK

Authentication Methods

• ANTHROPIC_API_KEY environment variable
• Amazon Bedrock integration (CLAUDE_CODE_USE_BEDROCK)
• Google Vertex AI integration (CLAUDE_CODE_USE_VERTEX)
• Third-party provider configuration

Session Management Features

• Automatic session creation
• Session ID capture from init message
• Session resumption with resume option
• Forking sessions (forkSession/fork_session)
• Session branching for experimentation
• Context preservation across sessions

Input Mode Configuration

• Streaming input mode (recommended)
• Single input mode
• Mode selection criteria
• Performance considerations

Custom Tool Development

• Tool schema definition
• Tool handler implementation
• Tool registration with SDK
• Tool execution patterns

Permission Configuration

• allowedTools list
• disallowedTools list
• permissionMode settings
• Permission callbacks and handling

System Prompt Customization

• CLAUDE.md file loading (settingSources)
• appendSystemPrompt option
• outputStyle for agent persona
• Project-level vs user-level instructions

Cost and Usage Tracking

• Token usage monitoring
• Billing integration patterns
• Usage reporting
• Cost estimation

Todo List Integration

• Todo tracking API
• Task display and management
• Progress monitoring
• Todo state persistence

Structured Output Features

• Output schema definition
• JSON result validation
• Type-safe results
• TODO tracking agent example

Hosting Patterns

• Ephemeral sessions pattern
• Persistent sessions pattern
• Production deployment strategies
• Scaling considerations

Full Claude Code Features

• Subagents (./.claude/agents/)
• Agent Skills (./.claude/skills/)
• Hooks (./.claude/settings.json)
• Plugins (programmatic loading)
• Memory (CLAUDE.md files)

Branding Requirements

• Allowed naming (Claude Agent, Claude)
• Not permitted naming (Claude Code)
• Product branding guidelines

Delegation Patterns

Standard Query Pattern

User asks: "How do I create a custom tool in the Agent SDK?"

1. Invoke docs-management skill
2. Use keywords: "custom tools SDK", "creating custom tools"
3. Load official documentation
4. Provide guidance based EXCLUSIVELY on official docs

Multi-Topic Query Pattern

User asks: "I want to build an agent with sessions, custom tools, and cost tracking"

1. Invoke docs-management skill with multiple queries:
   - "session management SDK", "session resumption"
   - "custom tools SDK", "creating custom tools"
   - "cost tracking SDK", "token usage"
2. Synthesize guidance from official documentation

Troubleshooting Pattern

User reports: "My session resumption isn't working"

1. Invoke docs-management skill
2. Use keywords: "session management SDK", "resume option", "session ID"
3. Check official docs for session patterns
4. Guide user through proper session handling

Troubleshooting Quick Reference

Related Skills

Skills that work well with agent-sdk-development:

• docs-management - Primary delegation target; provides all official Agent SDK documentation
• mcp-integration - MCP server configuration; Agent SDK can integrate with MCP servers
• settings-management - Settings configuration; SDK respects Claude Code settings
• hook-management - Hook configuration; SDK supports hooks via programmatic loading
• current-date - Get current UTC date for version history and audit timestamps

Delegation example:

When working with Agent SDK + MCP integration:
1. Use agent-sdk-development for SDK-specific guidance
2. Use mcp-integration for MCP server configuration
3. Both skills delegate to docs-management for official documentation

Additional Resources

For test scenarios, multi-model testing notes, repository-specific guidance, and external references, see:

• references/testing-and-notes.md - Test scenarios, multi-model testing, repository notes, external links

Version History

• v1.1.0 (2025-11-27): Comprehensive audit and enhancement

- Added Reference Loading Guide with explicit doc_id table for all 17 official Agent SDK docs - Added Test Scenarios section with 5 common use case scenarios - Added Related Skills section documenting skill composition patterns - Added Multi-Model Testing Notes section - Enhanced Last Verified section with Type B audit metadata - Verified keyword registry covers all official SDK documentation

• v1.0.0 (2025-11-26): Initial release

- Pure delegation architecture - Comprehensive keyword registry - Quick decision tree - Topic coverage for all Agent SDK features - Troubleshooting quick reference

Last Updated

Date: 2025-11-28 Model: claude-opus-4-5-20251101