agent-sdk-sidekick

Agent SDK Sidekick

Your deployment co-pilot. This skill bundles the current Claude Code and Agent SDK documentation as a local memory graph so Claude never guesses at SDK patterns from stale training data. When you want to ship a Claude Code prototype as a deployed agent, this skill reads the actual current docs and writes correct SDK code.

Why this exists: The Claude Agent SDK (formerly "Claude Code SDK") was renamed in late 2025. Managed Agents, plugins-in-SDK, subagents-in-SDK, checkpointing, and new hosting patterns all shipped after most models' training cutoff. Without current docs, Claude confidently generates broken SDK code from six months ago.

When to Use

• Converting a Claude Code prototype into a deployed Agent SDK project
• Writing Agent SDK code (Python or TypeScript) and need current API patterns
• Setting up hosting (Modal, AWS, GCP, self-hosted) for an agent
• Implementing subagents, hooks, plugins, or skills in the SDK
• Working with managed agents, channels, or agent teams
• Any question about the current Agent SDK API surface
• When the user says "deploy", "ship", "host", "convert to SDK", or "Agent SDK"

Bundled Documentation

This skill contains the complete Claude Code and Agent SDK documentation crawled and graph-linked for efficient traversal. Always read relevant docs before generating code.

Agent SDK Reference (read these first for SDK work)

• docs/agent-sdk/overview.md -- Agent SDK overview and architecture
• docs/agent-sdk/quickstart.md -- Getting started with the SDK
• docs/agent-sdk/python.md -- Python SDK reference
• docs/agent-sdk/typescript.md -- TypeScript SDK reference
• docs/agent-sdk/agent-loop.md -- How the agent loop works
• docs/agent-sdk/hosting.md -- Deployment and hosting patterns
• docs/agent-sdk/secure-deployment.md -- Security for deployed agents
• docs/agent-sdk/subagents.md -- Multi-agent orchestration
• docs/agent-sdk/hooks.md -- Hook system in the SDK
• docs/agent-sdk/plugins.md -- Plugin system in the SDK
• docs/agent-sdk/skills.md -- Skills in the SDK
• docs/agent-sdk/sessions.md -- Session management
• docs/agent-sdk/mcp.md -- MCP server integration
• docs/agent-sdk/custom-tools.md -- Building custom tools
• docs/agent-sdk/streaming-output.md -- Streaming patterns
• docs/agent-sdk/structured-outputs.md -- Structured output extraction
• docs/agent-sdk/permissions.md -- Permission system
• docs/agent-sdk/migration-guide.md -- Migrating from older SDK versions
• docs/agent-sdk/cost-tracking.md -- Usage and cost monitoring
• docs/agent-sdk/observability.md -- Logging and observability
• docs/agent-sdk/file-checkpointing.md -- File-based checkpointing
• docs/agent-sdk/user-input.md -- Handling user input
• docs/agent-sdk/modifying-system-prompts.md -- System prompt customization
• docs/agent-sdk/claude-code-features.md -- Claude Code features in SDK
• docs/agent-sdk/slash-commands.md -- Slash commands
• docs/agent-sdk/todo-tracking.md -- Todo/task tracking
• docs/agent-sdk/tool-search.md -- Tool search functionality
• docs/agent-sdk/streaming-vs-single-mode.md -- Streaming vs single turn
• docs/agent-sdk/typescript-v2-preview.md -- TypeScript v2 preview

Claude Code Reference (for understanding the prototype side)

• docs/agent-teams.md -- Agent teams / multi-agent coordination
• docs/headless.md -- Headless mode (CI/CD, scripting)
• docs/hooks.md -- Claude Code hooks
• docs/plugins.md -- Claude Code plugins
• docs/skills.md -- Claude Code skills
• docs/sub-agents.md -- Claude Code subagents
• docs/github-actions.md -- GitHub Actions integration
• docs/checkpointing.md -- Checkpointing in Claude Code
• docs/channels.md -- Channels
• docs/channels-reference.md -- Channels API reference
• docs/llms.txt -- Full page index

Each doc page has links_to and links_from metadata in its YAML frontmatter for graph traversal.

Workflow

When the user wants to deploy a prototype:

1. Read the prototype -- understand what the Claude Code orchestration does (tools, subagents, hooks, skills it uses)
2. Read the relevant SDK docs -- start with overview and quickstart, then follow links to specific features (subagents, hooks, plugins, etc.)
3. Read the hosting docs -- hosting.md and secure-deployment.md for the target platform
4. Write the SDK project -- translate the prototype into Agent SDK code using the patterns from the docs
5. Include deployment config -- Dockerfile, requirements.txt, hosting platform config

When the user asks an SDK question:

1. Find the relevant doc page -- use the links in the index above, or grep the docs/ directory
2. Read the page -- get the current API patterns
3. Answer with citations -- reference which doc page the information came from

Critical: Never guess at SDK patterns

The Agent SDK has changed significantly. Always read the bundled docs before:

• Writing import statements
• Constructing agent configurations
• Setting up tool definitions
• Implementing streaming patterns
• Configuring permissions or hooks
• Deploying to any hosting platform

If you're unsure about an API pattern, read the doc page. The docs are right here.