Inspiration
AI coding agents are getting better at writing code, but they still struggle with project-specific context. They may know how to edit a file, but they often do not know which functions are affected, what dependencies matter, what tests should run, or what hidden repository rules human reviewers already assume.
GitLab Orbit inspired this project because it gives AI agents something they usually lack: structured, queryable context about a software project.
Orbit Skill Forge builds on that idea by turning GitLab Orbit's codebase graph into practical merge-request review context.
What it does
Orbit Skill Forge uses GitLab Orbit Local to index a repository and query the codebase knowledge graph. It then generates a Context Pack for reviewers and AI agents.
The Context Pack includes:
- files and languages detected by Orbit
- function definitions with file paths and line numbers
- import dependencies
- blast radius assessment
- review checklist
- suggested test plan
- reusable AGENTS.md instructions for future AI agents
The workflow does not only generate a local Markdown file. It can also post the generated Context Pack directly to a GitLab merge request as a comment.
That turns GitLab Orbit from a code graph into a real developer workflow.
How I built it
I built Orbit Skill Forge using:
- GitLab Orbit Local for repository indexing
- GitLab CLI,
glab, for Orbit setup and local graph access - Orbit SQL queries over tables such as
gl_file,gl_definition,gl_imported_symbol, andgl_edge - Python for the Context Pack generator
- GitLab Merge Request API for posting the Context Pack directly to an MR
- Markdown for readable reviewer output
- a reusable
SKILL.mdartifact for future AI-agent workflows
The main script, forge.py, runs Orbit queries, formats the results into a Context Pack, saves the output, and posts it to a configured GitLab merge request when a GitLab token is available.
How it uses GitLab Orbit
Orbit Skill Forge meaningfully uses GitLab Orbit through the local CLI and SQL interface.
The workflow runs commands such as:
glab orbit local index .
glab orbit local sql "SELECT name, definition_type, file_path, start_line FROM gl_definition LIMIT 10"
The generated Context Pack is based on real Orbit graph data, including files, definitions, imports, and relationships.
What makes it different
Most AI coding tools ask the codebase a question once.
Orbit Skill Forge uses GitLab Orbit as a compiler for reusable operating knowledge. It turns graph context into durable instructions that reviewers and future AI agents can use before modifying code.
Instead of every AI agent starting from zero, every merge request can carry Orbit-powered context about what changed, what might break, and what should be tested.
Challenges I ran into
The biggest challenge was working quickly with a new platform and learning GitLab Orbit while building. I had to set up glab, install Orbit Local, index a repository, understand the local graph tables, and connect that output to a GitLab merge request workflow.
Another challenge was making the project more than a terminal demo. To fit the hackathon theme, the tool needed to perform a real workflow action, so I added merge-request comment posting through the GitLab API.
Accomplishments that I'm proud of
I am proud that the project uses real GitLab Orbit data instead of hardcoded examples. Orbit indexes the repository, returns actual definitions and file paths, and the tool turns that information into useful review context.
I am also proud that the final workflow posts directly into a GitLab merge request, making the output visible where developers already work.
What I learned
I learned how GitLab Orbit can act as a structured knowledge graph for software projects. I also learned how important context is for AI-native development. The model is not enough by itself; the agent needs reliable project context and reusable operating instructions.
What's next
Next, I would expand Orbit Skill Forge to automatically detect changed files from a merge request, query only the impacted parts of the Orbit graph, generate AGENTS.md patches, and run as a CI job on every MR.
I would also publish a richer version to the AI Catalog and add support for GitLab pipeline, issue, ownership, and security-finding context when available.
Log in or sign up for Devpost to join the conversation.