Why Gitea

Gitea is the lab’s git origin. It is also its issue tracker and its container registry, and neither of those was planned. Both came from the same mistake being caught twice: a careful comparison of dedicated products that never asked what the fleet already ran.

Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by human, chat 2026-09-21. How this site is written →

The tracker nobody chose

The task queue had selected Redmine over OpenProject and Taiga, and moved it to the front so everything after it could be tracked there. Its deciding reason was that Redmine could take a login from the reverse proxy’s header, a pattern already proven twice in the lab. The agent read the application’s own source while building it. The header login is not in Redmine’s core; it lives in third-party plugins.

Human’s call

Given two imperfect options, the human declined the frame: “I need to weigh the options between setting up a tool with just OK auth, with another tool that doesn’t do exactly what I want, compared to just getting the job done which is tracking tasks and assignments.”

That reframing was the decision. The question stopped being which tracker and became whether one was already running.

One was. Reading Gitea’s own database showed the tables for issues, assignees, labels, milestones, kanban boards, dependencies, comments, time tracking and per-user API tokens, all empty and unused. Adopting it cost no new virtual machine, database, certificate, identity-provider configuration, backup job, monitor or credential store. It was already backed up nightly, copied offsite, and its dump had been opened in a restore drill.

What was given up was stated at the time. Status is open or closed, and anything finer is a label or a board column by convention, not an enforced workflow. There are no typed custom fields and no Gantt view. Reversibility is high, since leaving means moving a few dozen issues between two REST APIs.

It also removed a rule that would have needed policing. The design said tickets point at repository files and never duplicate them. In a separate tracker that is discipline. In the same forge, an issue that references an errata file is just a link inside one repository.

The registry that was already there

The container registry runbook said to deploy Zot beside Gitea. The agent tested the host first. Gitea already answers the OCI distribution API with the exact unauthenticated response a Docker client expects, and a nonsense path next to it redirects to a login page, so the 401 came from the registry handler and not from a generic catch-all.

The measured comparison went against the second registry on every row. It would have needed a new certificate, which meant a DNS change, a new identity-provider integration, a new backup chain and a new monitor. Its authentication would have been a shared password file, where Gitea offers per-user tokens that can be revoked one at a time. The argument for a separate registry, that losing Gitea should not lose the images, did not survive measurement: the second registry’s storage would have been local disk on the same machine, so the blast radius was identical with two services instead of one.

One thing Gitea does not do is pull-through caching, which the CI runner’s design wanted. That is a different product from an artifact registry and should be chosen as one if the runner needs it.

What was compared, and what was missing

A later review asked whether Gitea was the right home for packages once it held real artifacts. It compared Gitea against Zot, Devpi, Verdaccio, Harbor and Nexus. The Gitea half is measured on the running instance; the others are from general knowledge and were labeled as unverified.

Gitea has retention rules, deletion, and backup that checks the archive holds as many package bytes as the live store. It has no quotas, no pull-through mirror, no signature verification and no built-in vulnerability scanning. The largest practical risk is unbounded growth on the same disk that serves the git origin, and the answer is a cleanup rule plus an alert on the package directory.

The review’s own footnote matters more than its table. The stated goal was to be alerted if a compromised package ever landed or was deployed, and no registry delivers that. Even Harbor’s scanning is a check at push time, and the requirement was continuous. That is a separate layer, handled by generating a software bill of materials and scanning it continuously, and it was already specified elsewhere. A thorough comparison of registries could not have revealed that registries were the wrong question.

What generalizes

Before comparing options, ask what is already deployed that does part of the job. A well-executed comparison of the wrong candidate set is still the wrong answer, and nothing inside the comparison can tell you.