openai-agents-runtime

OpenAI Agents runtime receipt example

This example targets the JavaScript @openai/agents local function-tool and handoff lifecycle boundaries. It creates real Agent instances, runs them through the SDK's run() loop, uses a real tool() definition, invokes a real handoff(), and signs hash-only atrib records from the SDK's agent_tool_end and agent_handoff lifecycle events.

Run It

pnpm --filter @atrib/integration openai-agents-runtime-smoke

The smoke is local, deterministic, and credential-free. It uses a scripted model that implements the SDK's Model interface so it can exercise real agent and tool execution without OPENAI_API_KEY.

What It Proves

• @openai/agents can run an Agent through the public run() API.
• A local tool() function executes through the SDK's normal function-tool loop.
• The SDK exposes the successful tool execution through agent_tool_end with the tool call id, function name, arguments, and result.
• A real SDK handoff() transfers control from the source agent to the receiver agent.
• The SDK exposes the transfer through agent_handoff after the handoff target is resolved.
• The atrib recorder signs one tool_call record and one https://atrib.dev/v1/types/handoff extension record with tool_name, args_hash, and result_hash.
• Public records stay hash-only and do not include raw tool arguments or tool results.
• Local sidecars keep the inspectable agent name, tool name, tool call id, arguments, result, handoff source, and handoff target.

What It Does Not Prove Yet

This is not a Python Agents SDK proof. It also does not call a hosted OpenAI model, the Responses API, computer-use tools, MCP transports, sessions, or OpenAI tracing export. The handoff receipt follows D073's extension-URI path; it is not a promoted handoff event byte and not a Pattern 3 verifier-gated handoff packet. Those need their own proof packets before public claims cite them.