openai-agents-python

Eight Archify diagrams. Each node carries SRC badges that link to the exact lines they were drawn from, all pinned to openai/openai-agents-python@71c2da4.

01architecture Architecture How the Runner loop, model boundary, tools and sandbox fit together Runner 循环、模型边界、工具与沙箱如何组合 Pinned to openai/openai-agents-python
02sequence Execution flow One Runner.run with a session, two turns, end to end 一次带 Session 的 Runner.run：两个回合，端到端 Pinned to openai/openai-agents-python
03architecture Core abstractions Which object owns what (Runner, Agent, Tool, Handoff, Model, RunState, SandboxAgent, Capability) 每个对象分别负责什么（Runner、Agent、Tool、Handoff、Model、RunState、SandboxAgent、Capability） Pinned to openai/openai-agents-python
04dataflow Context flow What reaches the model on each turn, and which hooks can change it 每个回合什么进入模型，哪些钩子可以改变它 Pinned to openai/openai-agents-python
05dataflow Memory flow Two memories: Session history and sandbox Memory files 两种记忆：Session 历史与沙箱 Memory 文件 Pinned to openai/openai-agents-python
06workflow Tool runtime Tool call path: classification, approval, invocation, error text 工具调用路径：分类、审批、执行、错误文本 Pinned to openai/openai-agents-python
07lifecycle Agent loop The Runner loop as a NextStep state machine 作为 NextStep 状态机的 Runner 循环 Pinned to openai/openai-agents-python
08sequence Sub-agent flow Handoff (control moves) vs Agent.as_tool (control stays) Handoff（控制权转移）与 Agent.as_tool（控制权保留） Pinned to openai/openai-agents-python

The diagram list (titles, types, questions, pins) lives in _data/studies.yml.

All eight were finalized with archify finalize --quality showcase --repo-root <clone at the pinned commit> (schema validation, verified delivery, strict provenance check, headless-browser check) and reviewed from visual-check captures. The JSON specifications (Archify candidates), the gate results and a regeneration recipe are in assets/openai-agents-python/archify/.

← Back to the research article / 返回研究文章：English · 中文