OpenAI Agent Runtime — Design skill for Claude Code

OpenAI Agent Runtime

[English](README.md) | [中文](README_zh.md)

**Agent Runtime** is an advanced execution engine built on top of the `openai-agents-python` SDK. It is designed to fill the gap in building long-running, stateful, and recoverable agent applications, providing capabilities similar to Anthropic's Claude Desktop or production-grade agent systems.

🌟 What We Built

While the native SDK (`Runner`) provides a basic execution loop, **Agent Runtime** extends it with critical engineering features required for real-world applications:

1. 🛡️ Robust Lifecycle Management

We introduced a structured **Phase** system to track the agent's exact state:

• INIT: Runtime configuration and initialization.
• PLANNING: Agent reasoning and LLM invocation.
• EXECUTING: Tool execution phase.
• OBSERVING: Processing tool results and environment feedback.
• TERMINATED / FAILED: Explicit terminal states.

2. 💾 Dual-Layer Persistence

To satisfy both machine replayability and human readability, we implemented a dual-write logging system:

• Machine Layer (events.jsonl): A structured event stream recording every state change. Used for debugging stats or precise replay.
• Human Layer (trace.md): A real-time Markdown rendering of the execution trace. Read the agent's thought process and tool outputs like an article.

3. ⏯️ Checkpoint & Resume

This is the core capability of the Runtime. The engine automatically saves a **Checkpoint (`state.json`)** after every key operation (like tool execution).

• Zero State Anxiety: Even if the process crashes or is interrupted, the Agent's memory (Conversation History) and current state (Turn Index) are preserved.
• Hot Resume: Simply use resume_from="run_id" to instantly restore the previous session state and continue execution.

4. 🔌 MCP & Local Model Support

• MCP Host: Native integration with the Model Context Protocol, allowing direct mounting of standard MCP Servers (e.g., Filesyst