Snapp! · 2026 · AI tooling
Athena
Vercel AI SDK, eve, Vercel Sandbox, Workflow SDK, OKD, Grafana
Problem
Athena is a GitLab code-review agent for Snapp’s frontend group. It runs on every merge request, leaves inline comments and a summary, and answers in threads. v1 reposted its comments on every commit.
What I did
I was on the research team that chose its stack, and I made web-app its first consumer: I added the review job to the merge-request pipeline and wrote the guideline rules it reviews against, tightening them after its first runs.
Then I rebuilt v2 myself, as a self-hosted service on the eve durable-agent framework and the Vercel AI SDK, running on OKD. Comment state is rebuilt from GitLab rather than from model memory, so a re-run does not repeat itself, and a webhook router answers @athena review, resolve and stop.
A role gate and a project allowlist run before any token lookup, and host URLs are validated strictly so lookalike hosts are rejected. OKD offers no Docker or KVM, so the agent runs without a shell or filesystem: code is read through a GitLab API tool and the built-in tools are off. Large merge requests fan out to a sub-agent, behavior evals against the live model gate CI, and Grafana shows the metrics.
Result
About 50 merge requests a week across 15 projects. A review that took 15-20 minutes now takes about 5.
Agent surface
- Commands:
- @athena review, resolve, stop
- Skills:
- guideline rules it reviews against
- CI:
- behavior evals against the live model