Real runs, start to finish. An agent gets a help-desk task in plain words, decides how to use the openjev tools, and acts on the typed answers. Each transcript is copied from the agent's session log: the task, every tool call and result, and the final answer.
Both transcripts replayed by demo/examples.mjs: the recorded
words and numbers, with the agent's thinking time cut and its answers excerpted. Watch it
in a terminal with node demo/examples.mjs; no key needed.
| Example | Tool | What it shows |
|---|---|---|
| One complaint, four judgments | jev_ask |
Category, urgency, safety hazard, and how upset the member is, from one call, routed by the co-op's own thresholds |
| This morning's queue | jev_batch |
Eight member messages triaged in one tool call and keyed by ticket id, with hazards dispatched first |
Both come from an electric cooperative's help desk, where a misfiled message can leave a live line down on a school route.
- Agent: DeepSeek Harness, the
reference client, in its one-shot
headlessprofile, runningdeepseek-flash. - Server: this repository's build (0.1.2 plus
jev_batch) with default settings, attached by a copy ofopenjev.patch.ymlpointed atdist/index.js. - Workspace: an empty directory, so the agent had nothing to work with but the tools.
- Edits: JSON is re-indented, the model's private reasoning is left out, and the malformed attempts the server rejected are described rather than reprinted. Nothing else is changed or removed.
The answers are live model output from 2026-10-12, so a new recording will differ in the details.
With OPENJEV_API_KEY in the harness environment (~/.dsh/.env), from the repository
root:
dsh --profile headless --patch examples/openjev.patch.yml "$(cat examples/coop-queue/task.md)"The final answer prints to stdout. The full session, tool calls and results included, is
kept in the harness's session history under ~/.dsh/sessions/.
Each task.md is a plain prompt, so it also works in Claude Code, Claude Desktop, Cursor,
or any other client with openjev-mcp configured.
