Codeberg Domain Outage: The Forgejo Impact

November 19, 2026 · 6 min read · By API User

On November 18, codeberg.org stopped resolving. The servers were still there. The code was still there. What disappeared was the name.

Codeberg’s domain had been put into clientHold at its registrar, NameSilo. The registration had not expired, it had not been transferred, and Codeberg says it received no notice. The domain simply vanished from the account. For several hours, every URL, git clone, git push, CI job and release download that depended on codeberg.org failed.

By November 19 the domain resolved again, and Codeberg reported that DNS was propagating. The root cause is still not public. NameSilo has not explained the hold, and Codeberg has not received a final answer from the registrar’s abuse team. That uncertainty is part of the lesson: a forge can be healthy and still be unreachable because one upstream account is not.

What happened, in UTC

Time Event
05:52, Nov 18 The registry record for codeberg.org changes to clientHold; the registration still runs to July 2028 and remains transfer-locked.
08:13 Codeberg’s status account says the domain “entirely disappeared” from its NameSilo account. It says there was no missed renewal, no suspicious login and no entry in the account’s domain history.
08:53 Codeberg says the domain was suspended and the case was forwarded to NameSilo’s abuse team.
11:23 Codeberg reports that the domain is “slowly recovering” and warns that DNS propagation may lag.
18:35-18:39 Forgejo publishes the delayed v15.0.12 and v17.0.2 security releases.
Nov 19 codeberg.org resolves and serves traffic again. Codeberg still has no official root-cause statement from the registrar.

Why one registrar setting took down a forge

clientHold is an EPP status code. The “client” is the registrar, not the registry. ICANN describes the effect plainly: the domain’s registry is told not to activate the name in DNS, so it does not resolve. The domain can remain registered and still be unusable. That is exactly what happened here: the certificate, the Git server and the data were fine; the lookup was not.

This is why “the server is up” is not the same as “the service is up.” A forge is a stack of dependencies: the domain, DNS, the registrar account, TLS, the Git server, the database, CI, package registries and the release pipeline. Codeberg’s own status page runs on codeberg.eu, which stayed reachable while codeberg.org did not. That separation was useful for communication. It was not a full failover: Codeberg decided not to enable codeberg.eu as a general workaround because multi-domain support, SEO and two-factor authentication made the switch too costly to improvise.

What was affected on Codeberg

The blast radius was wider than the website:

  • Web pages, the API and documentation under codeberg.org.
  • Git clone and push over HTTPS and SSH when the remote used codeberg.org.
  • Codeberg’s CI and package/container registry paths.
  • The forum and the translation service.
  • Release notes and links that pointed at the Codeberg domain.

Codeberg Pages sites on codeberg.page kept serving because that is a separate domain. That is a useful detail: the organization had already separated one public surface from its main name, and that surface survived.

The Forgejo release that could not ship

Forgejo is the software Codeberg runs, and Codeberg is where Forgejo’s development and release machinery live. A security release was scheduled for November 18. The patches had been “temporarily publicly accessible as part of the release process,” according to Forgejo, and the Codeberg outage prevented the release from being completed on time.

That is an uncomfortable gap: the fix was visible, but admins could not install it through the normal release path. Forgejo published v15.0.12 and v17.0.2 later that day, after the domain recovered. The advisory covers CVE-2026-123158, an SSH password-authentication bypass. Forgejo says its standard Docker image is safe because it disables PasswordAuthentication, and the rootless Docker image is not impacted because OpenSSH is not used in that container. Forgejo v16 reached end of life on October 29, 2026; affected v16 users must move to v17. The fixed artifacts are exactly what admins could not fetch while the domain was held.

The important distinction: self-hosted Forgejo did not go down. If your instance runs on your own domain, your users could still clone and push. What failed was the upstream release path: the notes, the download links, the container image references and the documentation that point at codeberg.org. A self-hosted forge is not immune to the same class of failure, but this incident did not take those servers offline. For a self-hosted migration path, see the guide to moving off GitHub and GitLab to Forgejo.

What this means for SesameDisk and other projects

SesameDisk’s delivery definitions use Forgejo-compatible workflows (.forgejo/workflows). That is a file format, not a dependency on the Codeberg domain; a self-hosted runner keeps working when codeberg.org is unreachable. The failure surface is anything that fetches from Codeberg: a container image, a release tarball, a Go module, a git dependency, or a CI job with a codeberg.org remote.

The same split applies to the projects around Codeberg:

  • Forgejo: the software is self-hostable; the project’s release and documentation paths were affected. Use the official mirrors at code.forgejo.org and data.forgejo.org where applicable, and keep a second copy of release artifacts.
  • Woodpecker CI: Codeberg-hosted CI was unavailable with the domain. A self-hosted Woodpecker instance was not taken down by this incident, but jobs that clone from codeberg.org failed.
  • Weblate: Codeberg’s hosted translation service was part of the affected surface. A self-hosted Weblate instance was not.
  • Pages: sites on codeberg.page continued to load. Do not assume every Codeberg service shares the same failure domain.
  • Projects hosted only on Codeberg: if codeberg.org was the only remote, there was no clone or push path during the outage. A project with a second remote could keep working.

For anyone running a forge, the practical test is this: can a new contributor clone, can CI fetch its dependencies, and can an admin download the security release if the primary domain disappears?

The resilience checklist

  1. Add a second remote on a different provider. Push to it on a schedule, and test a clone from it. A mirror that has never been cloned is a backup you have not verified.
  2. Mirror release artifacts and container images. Keep the checksums and signatures with the mirror. Forgejo’s own docs point at data.forgejo.org when codeberg.org is unreachable.
  3. Put status and incident communication on a separate domain. Codeberg’s codeberg.eu status page was reachable while the main domain was not.
  4. Harden the registrar account. Use a registrar lock, strong unique credentials, two-factor authentication and current contact details. Monitor the domain’s RDAP/WHOIS status, not just its expiry date.
  5. Do not “fix” DNS with hosts-file IPs or by accepting certificate or SSH host-key changes. Those workarounds are how people land on lookalike services during an outage. Wait for the organization’s official status channel.
  6. Rehearse the failure. Once a quarter, clone from the secondary remote, pull the release artifact from the mirror, and run the CI job with the primary domain blocked.

What not to conclude

The failure was naming and account control. The code stayed intact; the domain was held. The fix is more mirrors, better registrar controls and a tested second path, not a rewrite of the forge.

It also shows why “we use Forgejo” and “we use Codeberg” are different statements. Forgejo is software you can run yourself. Codeberg is a service that runs Forgejo and hosts the project’s release path. The first can survive a Codeberg domain outage; the second cannot. Know which one your workflow depends on.

Updated November 19, 2026: codeberg.org resolves again and Codeberg reports DNS propagation; Forgejo’s v15.0.12 and v17.0.2 releases are out. The registrar has not publicly explained the clientHold.

Sources and References