# Human review with an existing workflow conversation

Use Node 22.13+. Run `npm install`, copy `.env.example` to `.env`, and set your account and
an explicitly chosen workspace. These credentials stay on your machine; this is a Node example.
Optional `ARBI_DEPLOYMENT_DOMAIN` identifies the real deployment when its API is behind a proxy.

```sh
npm start -- start
npm start -- status con-ID
# Wait until draft is done and review is ready. Read the actual draft above.
npm start -- send-back con-ID "Make the tone warmer"
npm start -- status con-ID
# Review the revised draft, then approve explicitly:
npm start -- approve con-ID
npm start -- status con-ID
```

Replace `con-ID` with the ID printed by start. Each command can run in a new process: progress
and revisions live on the server. The same `con-` workflow conversation survives review and
send-back; each agent answer has its own `msg-` ID. Approval completes the review step using
the existing API. Sending back reopens the draft and review; it does not create a second run.
No email is sent and no document is changed by this example.

The owner is the reviewer. Existing access checks still apply. `run_now_temp` requests the existing
per-run authorization mode; it does not create standing agent membership. If the deployment reports
pending authorization, use its existing explicit authorization flow before proceeding. This example
never grants access automatically. Unattended execution still needs the platform's existing agent
identity and permission setup; a browser session alone is not a background worker.

For a UI, render `runProgress()` and collect an explicit decision before calling `completeStep()` or
`sendBackStep()`. Do not automatically approve based on an agent's text.