A recorded run on the open reference app Google Online Boutique, with prepared defects. In a pilot we use your repository. Harborline's code is a copy of Online Boutique, Google's open-source microservices sample (Apache-2.0); we planted the bugs its tickets describe. The runs were recorded earlier — nothing runs while you look.
Harborline runs a storefront. This is its engineering cockpit: an AI pipeline that researches each ticket, writes the fix, proves it with the service's own tests and opens the pull request — stopping at a gate after every stage.
Request access → I have accessPersonal access, read-only: leave your e-mail and we send you a link that signs you straight in. Nothing to install. Harborline, its people and its tickets are fictional; the runs, pull requests and test results are real. The copy is not deployed anywhere, so the evidence ends at tests and merged pull requests — where a session says «no evidence it is deployed», that is the honest answer.
Who approved each gate. Autopilot was on for this run: it passed research and verification; a person approved the diff, the pull request and the close.
Three recorded runs from the demo seed, one click each once you have access. Display dates may be shifted by the nightly reset; they are not new live work, and nothing here calls a model.
The bug fix in the cart service, the unit test the agent added, and the full record of what it did.
HBL-101 · Implement the fix Open the fix → The test evidenceThe service built the way its Dockerfile builds it, its own tests run, the ticket's scenario reproduced before and after. Verdict: PASS.
HBL-101 · Verify the fix Open the evidence → The human controlThe ticket was too vague to fix safely, so the pipeline stopped and asked four questions. Autopilot was on and still did not guess: nothing moves until a person answers.
HBL-109 · Research the ticket Open the gate →The agent reads the services the ticket points at before it writes a line. Research names files, reproduces the bug, lists acceptance criteria.
Every fix is built the way its Dockerfile builds it and run against the service's own tests. A failing verdict sends the run back, not forward.
Approve, or send it back with a note — every decision is on record. Autopilot is a policy a team chooses: in these recorded runs it passed research and verification, a person approved every diff before it was pushed, every pull request and every close, and a question from the agent always waits for a person.
Harborline's platform is Go, Java, C#, Python and Node. The same pipeline works on GitLab or GitHub, Jira or another tracker.
Optional: 22 short steps through one ticket, from the runs list to the metrics. It starts only from the button below; close it at any time, and the button in the corner of the screen brings it back.
Take the guided tour → Or look around on your own
Tell us who you are and we send you a personal link — usually the same working day. The demo is read-only and recorded: you look around at your own pace, nothing you click changes anything.
We use your e-mail to send you the access and to follow up about the demo. Inside the demo we record which pages you open, what you click and for how long, so we know what to show you next — nothing else, and we delete it on request: hello@rubka.ai.