Orbit Onboard

New to a project? One question gives you a friendly codebase guide: how the code is structured, what to read first, and who to ask. All pulled live from the GitLab knowledge graph.

Comment

Inspiration

I was trying to contribute to a GitLab project I had never seen before. Hundreds of files. No idea where to start. The README was outdated. The one person who knew the codebase was offline.

I thought the information had to exist somewhere. Every function call, every dependncy, every contributor. Orbit already had it all indexed. Nobody had just read it back out in a way a new developer could actually use.

I work remotely as an intern building AI agents. I have contributed to GitLab myself and felt this pain firsthand. This hackathon was the perfect excuse to finally build it.

What it does

Ask one question in GitLab Duo: "I'm new to this project, give me a tour."

You get:

  • What this project is: what the codebase does and why it exists
  • Architecture: how modules depend on each other, as a Mermaid diagram
  • Where the code lives: the densest files and modules
  • Learn these first: the definitions the rest of the code calls most
  • Who to ask: recent contributors from merge history

How I built it

A Python script and a GitLab Duo agent skill. The architecture view rolls up CALLS edges to modules using first-two-path-segments. "Learn these first" counts distinct callers per definition. Four queries total, all fast because Orbit already holds the resolved graph.

Challenges I ran into

Getting the module rollup to work across both flat repos and deep monorepos. The agent also needed a different prompt design than the CLI since it has to reason about which queries to run and in what order.

Accomplishments that I'm proud of

On the Orbit codebase, "learn these first" surfaces exactly the right functions with zero manual tagging. Pure graph signal, no human curation needed.

What I learned

Orbit's aggregation model handles deep graph questions with surprisingly simple queries. The hardest part wasn't the queries, it was knowing which four to pick. This was my first AI hackathon and I genuinely enjoyed every part of it.

What's next for Orbit Onboard

  • Directory-scoped tours for large monorepos
  • Line-level links so an agent can open definitions directly
  • "Recent changes" section pulling the last few MRs as onboarding examples

Built With

  • gitlab-duo-agent-platform
  • gitlab-orbit-knowledge-graph
  • glab
  • mermaid
  • orbit-query-dsl
  • python
Share this project:

Updates