GitLab Transcend Hackathon: What developers built on GitLab Orbit (opens in new tab)
GitLab’s Transcend Hackathon showed how developers use GitLab Orbit’s live code graph to answer questions about dependencies, ownership, testing, deployments, and risk. Of 1,576 registered participants, teams submitted 265 eligible projects, while contributors also merged 61 improvements into Orbit itself. The strongest projects used graph-based context to improve change analysis, migrations, testing, security response, and agent governance. ## Problems Developers Targeted - Around 70 teams built tools to predict what a change might break before merging. - More than 30 focused on onboarding and codebase comprehension. - Other common use cases included: - Incident root-cause analysis - Architecture drift detection - Flaky-pipeline diagnosis - CVE tracing across repositories - The popularity of these projects reflected a shared problem: relevant information is scattered across Git, CI, deployment systems, and dashboards. - Orbit consolidates those relationships into a queryable graph that agents can access through MCP or engineers can query directly. ## Technological Implementation - **Winner: Sankofa** - Provides three agents triggered by different workflow events: - **Radar** analyzes merge-request blast radius, affected pipelines, and ownership. - **Guide** prepares briefs when issues are assigned. - **Shield** traces vulnerabilities through the dependency graph. - Shield can identify a vulnerability’s full reach in one graph traversal. - **Runner-up: Stayed Shipped** - Measures whether changes merged by AI agents remain in production. - Detects changes that were later silently fixed forward, a result standard dashboards often miss. ## Design and Usability - **Winner: Carver** - Estimates the cost and risk of legacy migrations using Orbit’s dependency graph. - Breaks work into units, estimates effort and generation cost, and highlights untested, high-risk services. - Refuses to invent estimates when the relevant service cannot be grounded in Orbit. - **Runner-up: Marshal** - Takes an organization-wide migration goal, identifies affected repositories, sequences the work, and creates merge requests in waves. ## Potential Impact - **Winner: CrossCut** - Selects only tests that a change can actually affect. - Traverses Orbit’s call graph from changed symbols to determine transitive test impact. - Can reduce CI workloads by 90% or more on large or cross-repository suites. - **Runner-up: OrbitWeaver** - Performs autonomous refactoring using exact graph-based blast radius rather than vector similarity. - Updates affected files in dependency order, reducing the risk of incomplete refactors and broken pipelines. ## Quality of the Idea - **Winner: Transcend** - Extends Orbit with OWL, SPARQL, RDF, and semantic-web reasoning. - Supports queries involving transitive closure, external knowledge bases, and complex joins beyond the native API. - Its demonstration connected code implementing knowledge-graph embedding methods with related papers, authors, and publication years. - **Runner-up: Universal Agent OS** - Focuses on agent governance rather than agent capabilities. - Requires agents to interview users, plan before coding, preserve evidence, and validate their work. - Emphasizes accountability as AI-generated code becomes more common. ## Contributions to Orbit - The Contribute Track produced 61 merged merge requests from 26 contributors. - Improvements included: - Support for C++20 concepts, Go package declarations, Kotlin coroutines, and Ruby lambdas - Ontology corrections - A CI SIGPIPE fix - The first Orbit query tutorial - Documentation fixes, including clarification of `max_depth` versus `max_hops` - Nineteen contributors received cash prizes, and all participants earned swag credits. GitLab Orbit’s strongest value is not simply agent orchestration but the structured context behind it. Teams building reliable developer automation should prioritize a live dependency and ownership graph, using precise traversal where correctness matters more than probabilistic retrieval.