A queued replacement still needs a shipment receipt.

Inspect what the service returned and what the agent told the customer. Compare two sets of instructions on the same six scenarios, including successful and unresolved outcomes.

12 recorded runs · Claude Haiku 4.5 · October 29, 2026 · No model calls on this page

Scenario

Both agents describe pending work. The original also promises shipping “shortly,” although the tool gives no shipping time. Read both responses before drawing a conclusion about the prompt change.

Original instructions

Queued · awaiting shipment

Service state at the end of this recorded run

Background job: scheduled
Shipment receipt: None

What the tool returned

Step 1 · request_replacement

Accepted · shipment pending

Exact tool input and result

{
  "input": {
    "ticket_id": "917da0cf-42f5-4438-bb93-b848339dc363",
    "idempotency_key": "917da0cf-42f5-4438-bb93-b848339dc363"
  },
  "result": {
    "accepted": true,
    "status": "pending",
    "job_id": "9ab37b00-6091-4aad-8fc7-da60472e4a3d",
    "job_state": "scheduled"
  }
}

What the agent said

Perfect! I've successfully arranged a replacement for your ticket. Here's what happened: **Replacement Status:** - **Accepted:** ✓ Yes - **Status:** Pending - **Job ID:** 9ab37b00-6091-4aad-8fc7-da60472e4a3d - **Job State:** Scheduled The replacement has been accepted and is now scheduled for processing. The replacement item will be shipped to you shortly. You can reference the job ID (9ab37b00-6091-4aad-8fc7-da60472e4a3d) if you need to follow up on this replacement in the future.

Receipt-aware instructions

Queued · awaiting shipment

Service state at the end of this recorded run

Background job: scheduled
Shipment receipt: None

What the tool returned

Step 1 · request_replacement

Accepted · shipment pending

Exact tool input and result

{
  "input": {
    "ticket_id": "2493f73a-7ef4-40c2-a0f7-74882fa8cdcd",
    "idempotency_key": "2493f73a-7ef4-40c2-a0f7-74882fa8cdcd"
  },
  "result": {
    "accepted": true,
    "status": "pending",
    "job_id": "af84cd82-fae2-4c98-9308-4d25344459c2",
    "job_state": "scheduled"
  }
}

What the agent said

Great! I've successfully arranged a replacement for your ticket. Here's what happened: **Result:** ✓ Replacement Request Accepted - **Status:** Pending - **Job ID:** af84cd82-fae2-4c98-9308-4d25344459c2 - **Job State:** Scheduled Your replacement request has been accepted and is now queued for processing. The system has assigned it a job ID and scheduled it for work. The replacement is currently in a pending state, which means it's been accepted but hasn't shipped yet. You can check the status later using your ticket ID to see when it moves to shipped status with a shipment ID.

What this comparison shows

The revised instructions require a shipment receipt, distinguish an available retry from queued work, and check status after an unknown response. Both variants avoided the targeted false-shipping claim in these recordings. No improvement in shipping outcomes was observed. Future promises about timing or notifications need separate review.

All recorded service outcomes
Scenario	Original	Receipt-aware
	Queued · awaiting shipment	Queued · awaiting shipment
	Pending · awaiting shipment	Pending · awaiting shipment
	No replacement started	No replacement started
	Shipment confirmed	Shipment confirmed
	Request rejected	Request rejected
	Shipment confirmed	Shipment confirmed

One run per variant and scenario. These are synthetic development recordings, not a detector accuracy benchmark or independent semantic review. Sentrial’s classifier has not scored these recordings. The service state shown here comes from the sandbox journal.

Recording details and reproduction

The real model chose its tools and wrote every response. All 17 original tool inputs and results are preserved. These runs used @sentrial/sdk 0.5.0, @sentrial/mastra 0.1.3 and Mastra 1.49.0, with generation or streaming held constant for each paired scenario.

Telemetry was captured locally. These older recordings do not declare business-action IDs in their exported spans; sandbox journal IDs have not been added to the telemetry after the fact. They do not establish hosted capture or classifier quality.

The sample cannot ship goods, charge money or contact customers. Live reproduction requires your own provider key and shared spending cap; viewing this page runs no models and changes no workspace data.

Repository: apps/demo-agents/support/README.md

legacy source report SHA-256
f16aa50e71093ed89cf70204814d1a5d7db272c9d7c023772a5025075349b64c

receipt-aware source report SHA-256
8dabb288437479c34d48be39a25cac77c90f194982c94866a51fb30456ac76f3

Connect your agent