← 1.0.0 scope SAPE-4 In progress

Tag-driven release workflow: pushing a tag produces the artefacts, and somebody downloads one

Status
In progress
Component
release
Priority
High — the only repeatable way to run SAPE-3 twice
Blocked by
a tag — see below
Blocks
—

Where it stands, and why this is not Done

The workflow exists and has run green on GitHub's own runner, end to end: build for four platforms, checksums, every artefact proved to carry its version stamp rather than dev (the linux one by being executed), the container image built, started and connected to, with the welcome's product version matching the stamp. The three publishing steps skipped, because a manual run resolves publish: false — measured in the log, not inferred from the steps being grey.

This page briefly said Done. That was wrong, and reading the criteria below is what caught it. Four of the five need a tag that has actually been pushed:

So what is finished is the part that can be finished without cutting a tag, and cutting one is not this backlog's decision. The remaining work is not engineering.

Description

Add a workflow to packages/sapedb that fires on a version tag, builds the four targets and the container image from SAPE-3, stamps the tag into each binary, computes checksums, and publishes the lot. The pattern does not have to be invented: it already runs for real in this workspace and can be adapted rather than designed.

Why now

A release built by hand is a release that is correct once. The version stamp, the four targets and the checksums all depend on somebody remembering the same sequence, and the first time anyone finds out it was remembered wrong is the release after. This is also the "stranger cannot install it" category by a second route: artefacts that only appear when a human is available do not appear.

Measured 2026-09-20. packages/sapedb has no .github directory at all. Three sibling repositories in the same workspace do: packages/ecosy-next/.github/workflows/publish.yml, packages/ecosy-anchor/.github/workflows/publish.yml, and packages/ecosy-sapedb/.github/workflows/ with both publish.yml and test.yml. That last one is sapedb's own TypeScript client: the client already has a release workflow and the server it talks to does not.

Since superseded, twice. The server now has both a release workflow and a test workflow — the second because ISS-28 found that the release workflow's own go test would otherwise have run for the first time while cutting a release.

Acceptance criteria

  1. Pushing a test tag runs the workflow to completion with no manual step in the middle.
  2. The run produces every artefact SAPE-3 lists, attached to the release for that tag, with the checksum file.
  3. An artefact is downloaded onto a clean machine and executed there. It starts, and sapedb version prints the tag that was pushed. An artefact nobody has downloaded is an artefact that does not exist, so this criterion is not met by a green workflow run.
  4. The download in criterion 3 is fetched from the published URL, by a machine that has never had this repository checked out.
  5. Run twice, on two different tags. A release path that has run once has only proved it can be configured, not that it works.

Out of scope