This is the week the infrastructure layer stops being a pile of scripts and starts being something you can depend on.
shared-workflows cuts v1.0.0 on 05-16 and DockerContainers v1.0.0 on
05-21. Those two sit under everything else in the portfolio — every pipeline
pins the first, every image builds on the second — so versioning them was the
precondition for versioning anything at all. Both then take a handful of patch
bumps over the following days, which is what a real 1.0 looks like: not a
perfect tag, a tag you are allowed to move.
Coryp.Logging cuts v1.0.0 on 05-23. This is the first thing here published
as a library rather than a service — something another project takes a
dependency on. That is a different kind of release from deploying a container,
and worth marking separately.
On top of the infrastructure, the actual product surfaces move. The MCP
Server lands its tool surface in four PRs — file tools, a run_command shell
tool with workspace and environment confinement, git operations, and a tool
capability registry they all hang off. The registry is the interesting decision:
the tools are not exposed flat, they are gated. The IDE extension lands
activation, its config schema and its structured output channel in one unit.
Then there is the part nobody puts in a release note. On 05-24 the runner
labelling scheme — per-host, per-toolchain labels — is reverted in both
homelab-infra and homelab-stacks. It shipped days earlier as part of “Group B
Wave 1-3 — runner hardening”. The single-label scheme that replaced it is less
fine-grained and more correct. Two of the five patch bumps on DockerContainers
that same day exist because the first attempt at this was wrong.
Lessons learned
- Version the things everything else builds on first. Until shared-workflows and the base images had numbers, every other “release” in the portfolio was a deploy with no contract behind it.
- A revert three days after shipping is cheaper than a quarter of living with it. The runner labels were a reasonable idea that turned out to be operationally expensive. Reverting was the fix, not the failure.
- Gate a tool surface behind a registry rather than exposing it flat. It is the difference between adding tools and having to redesign when the fourth one needs a permission.