Changelog

2026-09-17

“You can send up to ten a day” now means exactly that on every test alert, not “usually” — the counter behind it is shared rather than per-server. The Settings Test card and the setup checklist also tell you how many you have left, right after you send one.

The data processing agreement now states where each part of Bobbin holds your data, service by service: reports, transcripts and your tenant record in Firestore in europe-west1; connection credentials in Secret Manager under Google’s automatic replication; sign-in identities in Firebase Authentication, which Google operates from the United States; and every model request on Vertex AI’s global endpoint. It previously said everything was pinned to europe-west1, which was true of Firestore and Cloud Run and not of the rest. Corrected rather than reworded.
Three sentences about the roles you grant were narrower than the permissions they describe, and now say what the permission allows and what Bobby’s tool does with it: roles/run.viewer returns environment-variable values (Bobby keeps the names and discards the values); the Compute Engine configuration role’s instances.get returns instance metadata, startup script included (discarded by an allowlist); and the GKE configuration role’s clusters.get is the permission get-credentials uses, though it holds no permission on anything inside the cluster.
The public bobbin-onboarding repository’s Slack page now matches the app as it has been since follow-up questions shipped: six bot scopes, mentions delivered as events, and interactivity on for the card’s buttons.

When an investigation needs a service’s settings and not only its telemetry — was it configuration or load — you can give Bobby one more read-only role for that service, and only that service: Cloud SQL, Memorystore, GKE, Compute Engine or load balancing. Each is a custom role holding exactly the get and list permissions his tool calls, granted by the console, the script (--family) or the Terraform module (families), and deleted again by the revoke script. Nothing changes if you do not.
Bobby now investigates GKE alerts from what GKE already exports — Kubernetes events, the cluster autoscaler’s reasons, OOM kills, container restarts and memory against limits, and the Kubernetes audit log — on the roles you granted on day one. He reads the GKE API and Cloud Logging, and never connects to your cluster.

Bobby investigates the stack you actually run, not just Cloud Run. When an alert is about a database, a queue, a load balancer or a cluster, he now knows what that kind of thing fails on and asks accordingly.
For Cloud SQL he reads connections, replication lag, disk and transaction age together, in one step rather than six.
Before he concludes, he checks what else is firing in the project — the quickest way to tell one service breaking from the platform underneath it breaking.
None of this needs new access. It all runs on the read-only roles you granted on day one.

One bad month is free. We only add capacity to your subscription if you finish two of your last three billing periods over your monthly allowance — a single period over, however heavy, changes nothing, and we email you when one does so you know where you stand.
When it does happen it is sized to the middle of what you went over by rather than the worst of it, counted in investigations rather than in blocks of 25, and you get a week’s notice naming the quantity, the price and the date.
And it works the other way without being asked: if you are paying for extra investigations none of your last three periods needed, we lower the quantity, credit the difference and tell you.
Extra investigations now raise your daily cap as well as your monthly allowance, in the same proportion — buying capacity moves the one limit that ever declines an alert.
Test alerts are now ten a day rather than three.
At your project limit the console works out whether the next tier is cheaper than another project add-on, and leads with whichever it is.

Every investigation now ends with a card, whether Bobby found the culprit or ran out of road. An honest miss is a result too.
Daily investigation caps per plan, so a runaway alert storm cannot spend a month’s fair use in an afternoon.
Connect GitHub from the console, and Bobby can answer “what changed?” as well as “what broke?”.