Run the ticket handoff
On Linux with Node.js 22.22.0 or newer and npm, run the command below. It installs OrgSDK 0.2.0 from its GitHub release into npm’s execution cache, starts a temporary workspace and creates a sample support desk on loopback. The lab uses synthetic tickets and needs no model key or support account.
Look for the weekend result. All three routes should return T-32, T-35 and T-37, in that order. The command prints a JSON report and closes the temporary workspace. Your existing installation stays separate.
npm exec --yes --package=https://orgsdk.ai/labs/orgsdk-ticket-handoff-1.0.0.tgz -- orgsdk-ticket-handoffWhat each client received
We froze four queues and their expected answers before running the comparison. The direct client follows each page and filters locally. The purpose-built MCP report does that same work inside the support desk. OrgSDK starts with the generic page tool and saves the filtering code as a tested action in the project.
Each route read the same six pages. With direct paging, the client received twelve requester addresses it did not need. The OrgSDK action and the purpose-built report returned only the matching ticket IDs as the business result. OrgSDK also returned run metadata and source with each status poll.
| Route | Exact answers | Requester addresses in replies | Client tool calls |
|---|---|---|---|
| Direct paging | 4 / 4 | 12 | 6 |
| Purpose-built MCP report | 4 / 4 | 0 | 4 |
| OrgSDK saved action | 4 / 4 | 0 | 111 |
What happened when a model used the tools
We gave local Qwen3 8B the same four handoff questions. Each trial started with a fresh conversation and one available tool. To pass, the model had to call the tool on the requested queue and return the exact JSON array. The report tool and the saved action had the same model-facing description and input. The OrgSDK adapter waited for completion before returning the IDs to the model.
With a generic page description, Qwen issued no tool calls. We then repeated direct paging with a description that named the fields and explained the filter. Qwen got three queues right. On the weekend queue it read two of three pages, missed T-37 and returned an object instead of the requested array.
To repeat the model trial, first make qwen3:8b available in Ollama on this machine, then run the command below. Both report routes answered all four questions in one model tool call per queue. In this trial, putting the paging rule into a reusable tool helped. OrgSDK supplied a way to build, test and delegate that tool using the provider’s existing list operation.
| Tool offered to the model | Grounded exact answers | Model tool calls |
|---|---|---|
| Direct pages | 0 / 4 | 0 |
| Direct pages with filtering instructions | 3 / 4 | 5 |
| Purpose-built report | 4 / 4 | 4 |
| Saved action with a waiting adapter | 4 / 4 | 4 |
npm exec --yes --package=https://orgsdk.ai/labs/orgsdk-ticket-handoff-1.0.0.tgz -- orgsdk-ticket-handoff --with-modelWhen the saved action helps
Choose the existing report when the provider already offers exactly the handoff you need. It matched every answer in one call per queue. With a generic list tool, OrgSDK let the project author define that report once without modifying the support desk.
The author validated the code and ran a mocked test. A separate client had permission to replay only version 1. It ran the approved handoff after the author saved version 2. Requests to execute arbitrary code and to run version 2 were both denied.
This installation took 28 administrator requests and 44 author tool calls, including polls and creating a second version for the permission check. Across the four replays, OrgSDK returned 225,177 bytes of JSON-RPC responses. Direct paging returned 1,733 bytes and the dedicated report returned 335. The lab counts every status poll at a 250 ms interval. Counts and bytes vary with runner speed. The address count covers ordinary replay responses. The experiment measures a synthetic queue and leaves customer time, model cost and security of other data paths unmeasured.
Try the same desk with your agent
Start the CLI in its own data directory using the first command below. Open http://127.0.0.1:18081, create the workspace and account, and wait for the project runner to connect. Use a free loopback port if 18081 is already occupied. In a second terminal, run the second command to keep the sample desk available. Copy its printed MCP URL.
In the project’s Plugins page, choose Connect MCP server. Name the connection desk, use the printed URL and select No authentication. Choose Check connection / refresh tools. Review and enable list_tickets. The project SDK exposes it as desk.listTickets.
In Connect MCP, create an Authoring client token for this project and copy the MCP endpoint. Configure a client that supports HTTP MCP protocol 2025-03-26 with that URL and the token as a bearer credential. Keep the token in the client’s credential settings. Give the client the prompt below.
npm exec --yes --package=https://github.com/OrgSDK/OrgSDK/releases/download/v0.2.0/orgsdk-cli-0.2.0.tgz -- orgsdk start --host 127.0.0.1 --port 18081 --data "$PWD/.runtime/ticket-handoff"npm exec --yes --package=https://orgsdk.ai/labs/orgsdk-ticket-handoff-1.0.0.tgz -- orgsdk-ticket-handoff --serve-fixtureFind the open P1 tickets in the weekend queue. Use list_projects and get_project_sdk to find the desk tool. Follow every page. Save the working code as a parameterized action, test it with a closed P1 and an open P1, then run version 1. Return the ticket IDs and run ID.Hand the tested version to a second client
In Connect MCP, create an Action runner token. Grant actions:run and runs:read and select the exact tested action version. Give your second client this token, the workspace MCP URL, project ID, action ID and version number. Keep the support desk terminal running.
Ask it to call run_action with the mixed queue and a fresh idempotency key, then poll get_run. The answer should contain only T-21. Open Runs in the project to inspect the recorded version, status and output. Stop the sample desk with Ctrl+C when finished.
