Inspiration
Context loss is a real cost. Every developer who has stepped away from a codebase — after a vacation, a handoff, two sprints on another project — knows the feeling: nothing looks familiar, decisions have no visible reason, and reconstructing the current state can take hours before a single useful line gets written.
This problem is personal. ColdStart was inspired by three years of building AI companion systems for a developer with a traumatic brain injury. The external memory architecture we built to solve session-to-session context loss for a person turned out to solve the same problem for any developer returning to a codebase cold. When GitLab Orbit arrived — a unified context graph for AI agents covering the full SDLC — the connection was immediate: Orbit is exactly the substrate this solution was waiting for.
What It Does
ColdStart is a GitLab Duo Agent Platform agent that queries the Orbit Knowledge Graph and produces a structured Rebuild Manual for any project.
Given a repository path, ColdStart answers eight questions at the end of each significant change cycle:
- What exactly was made?
- Why was it made?
- How was it made?
- How does it pitch in — what does it contribute?
- Is it interconnected the way an ecosystem should be?
- Is this stable?
- Did anything conflict or stop working?
- Where does this take us — what is the next most important need?
These aren't arbitrary questions. They are a discipline: the difference between 100 iterations of drift and 100 iterations of architecture. ColdStart encodes that discipline into Orbit's graph, producing a brief any developer — or any AI agent — can act on immediately after a cold return.
How We Built It
ColdStart uses the Orbit Remote REST API to traverse the Knowledge Graph: MRs, issues, pipelines, deployments, and code relationships queried through Orbit's Cypher-like DSL. The agent is implemented in Python, deployed on the GitLab Duo Agent Platform, and published to the AI Catalog. Authorization mirrors GitLab permissions — the agent sees exactly what its user can see.
The Step Questions are encoded as the agent's output schema, ensuring every Rebuild Manual answers all eight questions in order, with explicit stability and conflict assessments before recommending the next action.
Challenges
The hardest challenge was scope discipline: Orbit exposes 28 node types covering the entire SDLC, and the temptation is to query everything. ColdStart deliberately constrains its traversal to what a returning developer actually needs — the signal, not the noise. Getting that filter right required understanding not just what Orbit can answer, but what question a cold developer is actually asking.
What We Learned
That the same architecture that solves context loss for a person with a memory condition solves it for any developer. The problem was never unique — it was just unaddressed at the tooling level. Orbit makes it addressable.
Log in or sign up for Devpost to join the conversation.