Portal run desk for supplier, carrier, and insurer logins

Build me a portal run desk app for our ops team. We deal with supplier, carrier, and insurer portals that have no API, so today someone logs in by hand every month to pull invoices, statements, and certificates. I want those chores saved as reusable tasks that anyone on the team can run and watch.

The main screen is a library of saved portal tasks. Each task is a card or row with a plain-English instruction like "log into the carrier portal and download last month's invoices", the portal it belongs to, a Run button, a status pill for its most recent run, and when it last ran. I can add a new task by naming the portal and typing the instruction, and I can edit the instruction later, which matters because rewording is how we fix a task that keeps failing.

Pressing Run starts the task with Browserbase Run Agent, passing the saved instruction as the natural-language goal and attaching the task's pinned Browserbase context so the portal login is already in place. Agent runs are asynchronous, so the handler should start the run, store the returned run id and session id against a new run record, and return straight away rather than blocking. Never hold a request open waiting for a portal to finish.

Below the library is a run history table listing every run across all tasks, with the task name, the portal, status, start time, and duration. Statuses come from Get Agent Run and should render as PENDING, RUNNING, COMPLETED, FAILED, STOPPED, or TIMED_OUT. While any run is not yet in a terminal state, the app polls Get Agent Run on a short interval to refresh the pill and the duration in place.

Clicking a run opens a detail view. It shows the live progress transcript built from List Agent Run Messages, paging with the since and nextSince cursor so new assistant and tool messages append to what is already on screen instead of refetching the whole transcript. It also lists the documents the run retrieved, using List Downloads filtered to that run's session id, with Get Download to pull each file's bytes when someone clicks through. While the run is still going, a Cancel button calls Stop Agent Run, and the run then settles as STOPPED in the history with whatever it had already done.

When a run fails, the detail view should offer a video replay instead of leaving people to guess. Use Get Session Replays for that run's session to list the available replay pages and play them back inline, right next to the transcript and the task instruction that produced them. The whole value here is that a non-engineer sees the recording in context rather than hunting for a session id in the Browserbase dashboard.

Each saved task pins its own Browserbase context, created once with Create Context, so cookies, local storage, and user data persist between runs and the portal stays logged in. When a portal session expires and a run fails on a login wall, the task shows a Re-authenticate button. That flow needs a live session behind it: open a fresh session with Create Session using the task's context with persistence enabled, then call Get Session Debug URLs and open the interactive debugger URL so a person can sign in by hand in that live window. The interactive URL only exists while the session is RUNNING, so only fetch it after the session is live, and release the session once the person confirms they are done. Because the session saves back into the pinned context, the task keeps working on later runs without anyone re-entering credentials.

When a run finishes successfully, its documents get filed automatically. Each portal has a Google Drive destination folder, created with Create Folder the first time and stored on the task, and each retrieved file is uploaded there with Upload File (Multipart). That operation covers files up to 5MB, which fits ordinary invoices and statements, so fall back to Upload File (Resumable) for anything larger. Then post a one-line result to a chosen Slack channel with Slack Bot Send a Message, naming the task, the portal, how many documents landed, and a link to the run detail view. The channel is configurable per task, and I want the option to post on every run or only on failures and re-authentication requests.

The run history also needs a reliability view that surfaces which portals fail repeatedly. Group runs by portal and task, show the failure rate and the last few failure causes from Get Agent Run, and sort the worst offenders to the top. The point is to tell someone that a specific instruction is badly worded or that a portal changed its layout, so they go fix that task rather than rerunning it forever.

Persist saved tasks (name, portal, instruction text, pinned context id, Drive folder id, Slack channel, notification preference) and run records (run id, session id, task id, status, start and end time, duration, failure cause, and the list of retrieved documents). The library and the run history are shared across the whole ops team, so anyone can see and rerun anyone else's task, and each run record should note who started it.