Add human approval to a side-effecting agent tool

Goal

Make tool calls that require human judgment wait before any side effect.

Steps

1. List tools that write, send, or delete data and include them in the approval policy. [openai-human] [openai-tools]
2. Set @function_tool(needs_approval=True) on the function tool—or a per-argument boolean policy—so Runner pauses before execution. [openai-human] [openai-tools]
3. Inspect result.interruptions, call state=result.to_state(), and show the reviewer the tool name, arguments, and target. [openai-human] [openai-tools]
4. Call state.approve(interruption) only for explicit approval; use state.reject(...) for denial, then await Runner.run(agent, state). [openai-human] [openai-tools]
5. Use a mock tool to verify approval executes, rejection does not, and the reviewed action matches actual arguments. [openai-human] [openai-tools]

How to verify

In a test environment, approve and reject once; only the approved path should call the side-effecting function.

Troubleshooting

If approval does not pause execution, check the SDK approval mechanism and ensure the tool has not run before the pause point.