workflows

Native Responses workflows

This sample hosts built Agent Framework workflows directly with ResponsesHostServer(workflow=..., parse_response=...); it does not wrap them with .as_agent().

• main.py parses a typed SloganRequest, builds fresh agents, clients, credentials, executors, and a graph for every hosted request, and keeps stable workflow/executor IDs so a later turn can restore its exact checkpoint.
• approval.py pauses on an approval request and resumes only after the caller answers the complete pending reply batch. Its "publish" operation is deliberately simulated and makes no external change.

The Foundry sandbox (agent_session_id) is the trusted user-isolation boundary. Responses response.id/previous_response_id select caller-visible turns inside that sandbox. The MAF checkpoint ID is private workflow state and is bound to the exact outer response; none of these IDs is a downstream model service_session_id.

Use a request-aware factory for multi-turn workflows or pauses. A built Workflow instance is single-use and is not cloned. Every factory invocation must create fresh mutable resources, while workflow names and executor IDs remain stable. store=False is one-shot and cannot expose a resumable pause.

Run locally from this directory with the current workspace:

uv run --no-sync python main.py
# or:
uv run --no-sync python approval.py

Example typed request:

curl -X POST http://localhost:8088/responses \
  -H "Content-Type: application/json" \
  -d '{"input":"{\"topic\":\"an affordable electric SUV\",\"style\":\"retro\"}","store":true}'

An approval response uses the id of each returned mcp_approval_request. Reusing an old, forged, duplicate, partial, cross-user, or cross-sandbox reply is rejected.

For Foundry deployment instructions, see the parent README.