Orbit Sleuth: Graph-Driven Root-Cause Agent

The problem

When a bug lands, the slow part usually isn't writing the fix — it's the detective work: Where does this break? What does it depend on? What will my fix affect? Where do I even start? That context lives in senior engineers' heads, and AI assistants don't have it.

I feel this every day. I run AI-assisted 2nd-level support across a six-service production codebase, and my biggest time sink isn't the fix — it's tracing a symptom to the right code: which file, which dependency, which change. That discovery work is exactly what a code knowledge graph should automate.

What it does

Orbit Sleuth turns a bug symptom into an evidence-backed briefing before you write the fix. Give it a file, a symbol, a file:line, or a pasted stack trace, and it queries the GitLab Orbit knowledge graph to return:

Change-risk score (low → critical), derived only from graph structure
Upstream dependencies — where the defect may actually originate
Downstream blast radius — what a fix could break, direct and transitive, by depth
Ranked root-cause suspects + a minimal, ordered investigation pack
Receipts — the exact runnable Orbit query behind every single claim

The rule the agent enforces: no receipt, no claim. Every number is a graph fact you can reproduce, not a model guess.