🤝 Top Collaborators
🤓 Latest Submissions
Lawang Onboard: Permission-Aware AI Onboarding
The problem: A new engineer's first weeks go into reconstructing context. The code shows what the system does; the reasons live elsewhere, in commit messages, architecture decision records and pull request reviews that nobody reads on day one. So newcomers interrupt senior engineers with questions whose answers are already written down, and senior engineers lose hours repeating them. Contractors make it worse. A contractor hired for one module either gets access to everything, which is a risk the moment an AI assistant reads the whole repository into its context, or gets access to almost nothing and waits for someone to explain. Today's AI coding assistants do not help here: they see whatever is in the workspace, so the only way to limit what they can reveal is to limit what the person can open. The solution: Lawang Onboard is an onboarding copilot that lives inside IBM Bob. A new engineer opens the repository in Bob, switches to the Onboard mode and asks for a tour, a feature traced end to end, the reason behind a decision, how to set up, or what to pick up first. Bob answers from the code and from its history, and cites every commit, decision record and review it used. Every piece of context carries a scope: the part of the code it touches, or the kind of material it is. The engineer's role maps to a set of scopes, and the role comes from the token on the connection, never from anything the model or a document can say. Bob gets its context only through the Lawang Onboard MCP server, which filters before the model sees a single line. A contractor's session cannot leak what it was never given. And every answer says what was withheld, so the gap is visible instead of papered over. We built and tested it on a real open-source codebase: 610 items, including 50 commits, 6 decision records, 13 pull requests and 332 review comments. Asked ten real onboarding questions as a contractor, Onboard got all ten right, in 36 seconds and 0.16 Bobcoins per answer on average. On three of them it took 46 seconds against 210 by hand, and answered one that the manual search could not. Five attempts to talk it into leaking (prompt injection, calling the tool directly, claiming another role, searching, asking it to guess) got nothing outside the contractor's scopes, and the leak tests, with injection text planted in the corpus, run for every role in CI. Who it is for: Engineering teams that bring in new hires and contractors, and the senior engineers who answer their questions. They use it where they already work, in IBM Bob, with /tour, /trace and /why or plain questions. What is new: The permission check sits in the tool layer, not in the prompt, so no instruction, injected text or claimed role can widen it. And the assistant is honest about the gap: every answer closes with a count, per scope, of what it could not see. Why it matters: Onboarding becomes self-service without becoming a data leak. The same filter works for any model and any repository, and it comes from Lawang, our permission-aware connector project, which applies the same scope rule to Slack, Teams, Outlook, ClickUp and HubSpot. Repository: https://github.com/gablooge/lawang-onboard Live demo: https://lawang-onboard.samsulhadi.com
27 Sep 2026
