Every package inside a domain shares one version. Domains move independently. Helix goes to 1.4.0 as a set; Cortex is on whatever number Cortex is on. The obvious alternative is full-repo lockstep, one version for everything, which is simpler to explain and simpler to automate. It is wrong here for a specific reason: the docs packages carry the version of the thing they document. Publishing a Cortex fix under full-repo lockstep would bump @fusion.dev/helix-docs to a version in which nothing about Helix changed, and a reader landing on those docs would be reading a version number that describes an unrelated repair to an unrelated product. Per-domain keeps the property that actually matters, which is that @fusion.dev/helix-headless@x works with @fusion.dev/helix@x. That is the guarantee a consumer needs. Everything else about the numbering is bookkeeping. The version lives in the tag Every manifest in the repository holds 0.0.0-dev. The release stamps the real number in CI seconds before packing, and never commits it back. Committing it back would require a bot whose push access outranks branch protection, in order to record something git tag already holds. The tag is the record. helix-v1.4.0 means a release that promoted, and helix-attempt-v1.4.0 means a number that was spent and did not, so version computation reads both. A failed release burns a number rather than poisoning the next one. The split that has now happened The Rust workspace held only Cortex crates, so [workspace.package] version was the Cortex version by accident of there being nothing else in it, and the day Helix gained a Rust crate the two domains would have been silently back in full lockstep. Helix has two Rust crates now, and the field is gone. [workspace.package] keeps the edition, the license and the toolchain floor, which are facts about the workspace; each crate declares its own 0.0.0-dev, which is a fact about a domain. Three identical lines whose whole purpose is to be able to differ, and a check that fails both on a crate inheriting a version and on the workspace declaring one at all. The split was made ahead of the crate rather than with it, which is the order worth keeping: a constraint that holds because of a current absence is not a constraint, it is a coincidence, and coincidences do not survive the commit that ends them.