nyrvo-runners

A daily record of what GitHub-hosted runner images actually contain, captured with Nyrvo.

Runner images change without announcing it. Node moves a minor version, a toolchain disappears, Docker's compose plugin jumps — and you find out when a build that passed on Tuesday fails on Wednesday for no reason you can see in your own diff. This repository exists to make that a dated event instead of a mystery.

Git is the database

There is no server here, and no database. Each runner label has exactly one file:

data/ubuntu-latest/current.json
data/macos-latest/current.json
…

The file always holds the latest observation. The history is the git history, which means git log -p is the feed:

git log -p --follow data/ubuntu-latest/current.json

One file per label rather than one per day is a deliberate choice. Committing a dated snapshot every morning would grow this repository forever and bury the handful of days that mattered under a year of identical ones. Storing only the current state means a commit exists only when something changed, so the log carries events rather than heartbeats, and "when did this change?" becomes a question git answers natively.

How a day works

A scheduled workflow captures every label in parallel, uploads each snapshot as an artifact, and a single final job normalises them and makes at most one commit. Six jobs pushing to one branch would race; one commit a day also reads better, because the diff shows everything that moved together.

Three details that are load-bearing rather than incidental:

• The capture does not run inside a checkout. A Nyrvo snapshot describes the directory it runs in as well as the machine, so inside this repository it would record this repository's own commit and read its files as declared requirements. Every snapshot would then churn on data about itself. An empty directory yields the runner and nothing else.
• created_at is dropped before storing. It changes on every run, so keeping it would produce a commit a day whether or not a runner drifted. The commit's own date already records when we looked.
• The Nyrvo version is pinned. If this tracked the latest release, a change in the tool would look exactly like a change in a runner and the dataset would stop meaning what it claims. Bumping it is its own commit, so that commit is the one to blame when every label moves at once.

What a snapshot contains

Operating system, architecture and kernel; the versions and install paths of Go, npm, Node, Python, Ruby, PHP, Rust and Java; Docker's client, server and compose versions and whether the daemon answers; the images of any running containers; and the names of environment variables.

Never the values of environment variables. Nyrvo does not record them, which is why publishing a runner's environment here is safe at all.

What would prove this worth keeping

After roughly a month, three things should be true: no job broke unattended, at least one real drift was captured, and this repository answers "when did node change on ubuntu-latest?" faster than reading the runner image changelog.

If nothing drifts on any runner in a month, that is worth knowing too, and cheaper to learn here than by building anything larger on the assumption.

Licence

MIT for the tooling. The captured data describes public GitHub-hosted runner images and is published as fact rather than claimed as property.