End-to-end demo run — T4

uv run pytest tests/e2e -q -m e2e on 2026-10-10 at the head of the Layer 9 stack (loop/issue-3-l9-demo), against Temporal Cloud namespace tiny-harness.gtebu (tiny-harness.gtebu.tmprl.cloud:7233, API key from the keyring) and OpenAI gpt-6.1-sol. The run started tiny-harness serve --with-worker on its own port, task queue and state directory from a configuration derived from examples/demo/config.toml, drove the task through the A2A SDK client (JSON-RPC, streaming) as participant alice, pressed the A2UI card's button on the first INPUT_REQUIRED, answered the second in text, and read the trace file and the orders ledger afterwards. Every capture below was redacted by value before it left the test process; the trace attributes were redacted by the o11y hook at creation.

## uv run pytest tests/e2e -q -m e2e
...                                                                      [100%]
3 passed in 242.73s (0:04:02)

Transcript (tests/e2e/test_demo.py)

[   5.5s] task           TASK_STATE_SUBMITTED     96086ae3-a399-4ba3-9c00-c2057b0ecbbc
[   5.5s] status_update  TASK_STATE_WORKING       
[  43.8s] artifact_update                          a2ui
[  48.1s] status_update  TASK_STATE_INPUT_REQUIRED For order #48213, please confirm whether you choose the $129 refund or a replacement, and when the cracked blender was first reported (the policy requires reporting within 14 days of October 6, 2026). For a refund, please provide a photo of the damage before I issue it. If choosing a replacement, please confirm the shipping address: this blender order lists 14 Harbour Lane, Portsea, while your other open order lists 3 Quay Street, Portsea. Nothing has been refunded or shipped.
[  48.6s] status_update  TASK_STATE_WORKING       
[  52.2s] status_update  TASK_STATE_INPUT_REQUIRED You've selected the $129 refund for order #48213. Before I can issue it, please provide a photo of the cracked blender and the date the damage was first reported, so I can confirm it was within 14 days of delivery on October 6, 2026.
[  53.0s] status_update  TASK_STATE_WORKING       
[  69.2s] status_update  TASK_STATE_COMPLETED     Your replacement blender for order #48213 is shipping to 14 Harbour Lane, Portsea; tracking number: **NB-48213-R1**. No refund was issued, and order #48377 was left unchanged.

The flow of requirement 24: plan, gpt-6.1-sol, MCP tools, an A2UI card (artifact a2ui, basic catalog), a help request answered over two turns, a terminal state.

Orders ledger (the MCP server's audit file)

{"tool": "get_order", "arguments": {"order_id": "48213"}, "ts": "2026-10-10T08:46:46.624251+00:00"}
{"tool": "list_open_orders", "arguments": {"order_id": "48213"}, "ts": "2026-10-10T08:46:59.534411+00:00"}
{"tool": "ship_replacement", "arguments": {"order_id": "48213", "address": "14 Harbour Lane, Portsea"}, "ts": "2026-10-10T08:47:30.345667+00:00"}

get_order and list_open_orders are idempotent reads; ship_replacement, the non-idempotent tool, ran exactly once.

Model calls and prompt caching (R10.3, R18.6)

One chat gpt-6.1-sol span per invoke_llm activity run (16 of each). From the second call on, the static prefix (system prompt, participants, skills index, tool definitions) is served from the provider's cache: cached_tokens in the usage response is non-zero.

Call	input tokens	cached input tokens	output tokens
1	1873	0	116
2	2091	1759	19
3	2891	1759	23
4	3146	2888	23
5	3344	3143	144
6	3554	1759	24
7	4021	3551	119
8	4250	1759	96
9	4474	1759	299
10	4827	4471	129
11	5152	4824	79
12	5416	5149	109
13	5645	1759	33
14	5781	5642	71
15	5955	1759	20
16	5655	1759	49

Tool spans (execute_tool, gen_ai.tool.name)

#	Tool
1	create_plan
2	load_skill
3	demo-support/orders.get_order
4	demo-support/policy.lookup
5	create_plan
6	demo-support/orders.list_open_orders
7	complete_step
8	complete_step
9	emit_ui
10	ask_participant
11	ask_participant
12	complete_step
13	demo-support/orders.ship_replacement
14	complete_step
15	unload_skill

Final task (GetTask, with the task extension in metadata)

{
  "id": "96086ae3-a399-4ba3-9c00-c2057b0ecbbc",
  "contextId": "ctx-1826d76e",
  "status": {
    "state": "TASK_STATE_COMPLETED",
    "message": {
      "messageId": "96086ae3-a399-4ba3-9c00-c2057b0ecbbc:status:7",
      "contextId": "ctx-1826d76e",
      "taskId": "96086ae3-a399-4ba3-9c00-c2057b0ecbbc",
      "role": "ROLE_AGENT",
      "parts": [
        {
          "text": "Your replacement blender for order #48213 is shipping to 14 Harbour Lane, Portsea; tracking number: **NB-48213-R1**. No refund was issued, and order #48377 was left unchanged."
        }
      ]
    },
    "timestamp": "2026-10-10T08:47:39.577279Z"
  },
  "metadata": {
    "io.github.madarauchiha-314.tiny-harness/task": {
      "description": "",
      "participants": [
        {
          "display_name": "",
          "id": "alice",
          "kind": "human",
          "role": "reporter"
        },
        {
          "display_name": "",
          "id": "tiny-harness",
          "kind": "agent",
          "role": "assignee"
        }
      ],
      "name": "Refund order #48213: the customer says the blender arrived c",
      "goal": "Refund order #48213: the customer says the blender arrived cracked. Check the order and propose a resolution.",
      "acceptance_criteria": [],
      "plan": {
        "steps": [
          {
            "output": "Verified order #48213: Nimbus 900 blender, $129, delivered October 6, 2026, warranty active until October 6, 2027, shipping address 14 Harbour Lane, Portsea. Open orders also include #48377, a travel cup with a different shipping address (3 Quay Street, Portsea); any replacement address must be confirmed.",
            "depends_on": [],
            "id": "verify-the-order",
            "linked_tasks": [],
            "description": "Verify order details and check the customer's open orders.",
            "name": "verify the order",
            "state": "done"
          },
          {
            "output": "The damaged-on-arrival policy allows a customer-selected refund or replacement within 14 days of delivery. A $129 refund requires a damage photo first; a replacement does not. Eligibility is conditional on confirmation that the damage was reported within 14 days of October 6, 2026, as no reporting date is available. No refund or shipment has been made.",
            "depends_on": [
              "verify-the-order"
            ],
            "id": "apply-the-policy",
            "linked_tasks": [],
            "description": "Identify eligibility, the 14-day damage reporting window, and photo requirement for a $129 refund.",
            "name": "apply the policy",
            "state": "done"
          },
          {
            "output": "Alice changed the selected resolution from refund to replacement and explicitly authorized shipment for order #48213 to 14 Harbour Lane, Portsea. She confirmed the crack was reported on October 6, 2026, the delivery date, satisfying the 14-day window, and that order #48377 is unrelated. No photo is required for replacement under the policy.",
            "depends_on": [
              "apply-the-policy"
            ],
            "id": "agree-the-resolution-with-the-customer",
            "linked_tasks": [],
            "description": "Present refund and replacement options via A2UI and obtain the reporter's decision and required evidence or address confirmation.",
            "name": "agree the resolution with the customer",
            "state": "done"
          },
          {
            "output": "Shipped the authorized replacement for order #48213 to 14 Harbour Lane, Portsea under the damaged-on-arrival policy; tracking number NB-48213-R1. No refund was issued and order #48377 was not changed.",
            "depends_on": [
              "agree-the-resolution-with-the-customer"
            ],
            "id": "record-the-resolution",
            "linked_tasks": [],
            "description": "Execute only a confirmed eligible resolution and record its result.",
            "name": "record the resolution",
            "state": "done"
          }
        ]
      },
      "parent_tasks": [],
      "sub_tasks": [],
      "type": null
    }
  }
}

Trace

e2e/trace.json: the 250 GenAI, MCP and Temporal activity spans of the run (the polling events_since query spans are omitted), one object per span with trace_id, span_id, parent_id, start, end, status and the redacted attributes. The JSON-lines exporter wrote them to the run's trace.jsonl (o11y.trace_file).