# Continuous integration and alpha releases (Gitea)

Single source of truth for how acdream builds, gates, and ships alpha builds.
Landed 2026-08-19. Companion to [`release-gate.md`](release-gate.md), which
owns the *local* bounded gate.

## What happens on a push to main

```
git push origin main
        │
        ├─ windows-gate    (RARE-win)        build + full lane-filtered suite
        ├─ linux-portable  (eriktestLinux)   portable closure, Linux lanes
        │
        └─ release   (needs BOTH green)      publish a Gitea Release
                                             + republish the `latest` pointer
```

Workflow: [`.gitea/workflows/ci.yml`](../.gitea/workflows/ci.yml). A red gate
cannot publish: `release` uses `needs:`, not a `workflow_run` trigger, whose
Forgejo support is unreliable.

## Why Gitea and not GitHub

GitHub Actions is **billing-blocked** on this account ("recent account payments
have failed"), and the repo is private, so hosted runners consume paid minutes.
Forgejo ships **no hosted runners at all**, so Actions there requires
self-hosted ones — which are free on both platforms. The same two machines can
serve GitHub later by registering a second agent; only the workflow's
`runs-on` labels change.

## The runners

| | Windows | Linux |
|---|---|---|
| Host | `RARE` (10.6.0.3) | `eriktestLinux` (10.0.0.202) |
| Agent | `act_runner` 0.2.13 | `forgejo-runner` 13.0.0 |
| Persistence | Scheduled task `ForgejoRunner`, at logon of `acbot` | systemd `forgejo-runner`, `Restart=always` |
| Labels | `windows`, `windows-latest`, `windows-x64` | `ubuntu-latest`, `ubuntu`, `linux`, `ubuntu-slim` |
| Execution | host mode (`:host`) — no Docker on either box | host mode |

Both **poll outbound** over HTTPS. Gitea never connects to them, so no inbound
ports, no port forwarding, and no static IP; they work behind NAT. The runner
does not have to live next to the Gitea container (which runs on `bluesnake`,
a host we have no shell on).

`forgejo-runner` publishes **no Windows binary in any release**, which is why
Windows uses Gitea's `act_runner`. Forgejo speaks the same Actions protocol.

### Prerequisites on a runner

- **.NET SDK in the `global.json` band** — currently `10.0.3xx`. `10.0.400` is a
  different feature band and `rollForward: latestPatch` rejects it.
- **Node.js** — `actions/checkout` and `actions/upload-artifact` are JavaScript
  actions. Docker images normally supply Node; in host mode the machine must.
- **Git**, and outbound HTTPS to `git.snakedesert.se`.
- **PowerShell 7** on Windows (`pwsh`); `tools/*.ps1` require it.

## Releases

Everything about distribution lives under **Releases** — nothing in git. A build
is ~120 MB, so payloads are release attachments; and the pointer the launcher
polls is itself a release asset, so there is no payload branch, no bot commit on
`main`, and no push that could retrigger the pipeline.

```
Release  0.1.0-build.<yyyyMMddHHmm>     <- the actual build
  client-win-x64.zip      AcDream.App.exe + acdream-headless.exe
  launcher-win-x64.zip    acdream-launcher.exe + acdream-bake.exe
  manifest.json

Release  latest                          <- pointer, replaced every publish
  manifest.json           names the version above and its asset URLs
```

The launcher polls the pointer at a URL that never changes
(`ReleaseManifestClient.ProductionManifestUri`):

```
https://git.snakedesert.se/erik/acdream/releases/download/latest/manifest.json
```

A pointer is needed because **Forgejo has no `/releases/latest/download/`
route** (verified: 404) — unlike GitHub, there is no built-in stable URL for
"the newest release". Publishing it recreates the `latest` tag each time, which
means deleting the old release *and* its tag; the tag outlives its release and
would otherwise block recreation.

Versioned releases are kept, so an older build stays downloadable and the
launcher's local rollback still has something to fall back to.

`tools/publish-bin.ps1 -BaseUrl <release asset base>` builds the payloads; CI
passes the tag's asset base. Running it locally is for inspection only —
publishing is CI's job.

### Verifying a release

```powershell
dotnet test tests/AcDream.Launcher.Core.Tests --filter Lane=Live
```

`LiveGiteaReleaseInstallTests` installs the advertised client from the real feed
through the production updater — real SHA-256/size verification, extraction, and
atomic activation — then asserts both hosts resolve out of the activated
directory and `current.json` names the installed version.

## Landmines

Each of these cost a red pipeline; none was a config typo. Two rows record a
fix that was tried and **disproved** — read those before repeating it.

| Symptom | Cause |
|---|---|
| `Cannot find: node in PATH` | JS actions need Node on the host in `:host` mode |
| `actions/setup-dotnet` never resolves | `data.forgejo.org` does not mirror it (404). `checkout` and `upload-artifact` **are** mirrored. Self-hosted runners carry the SDK anyway |
| Job "failed" while dotnet processes still run | `run-release-gate.ps1` redirects children to log files, so the step goes silent; Forgejo fails a non-reporting task as a zombie. CI runs `dotnet test` directly so output streams |
| ~40 tests fail on formatted numbers | Runner's `HKCU` locale was `en-SE` (comma decimal): expected `"update:0.25"`, got `"update:0,25"`. `Set-Culture` does **not** reach a scheduled task without a loaded profile — set the registry directly |
| `DOTNET_SYSTEM_GLOBALIZATION_INVARIANT=1` as the locale fix | Too blunt — it breaks tests that legitimately construct a culture. Fix the machine locale instead |
| `FileNotFoundException: client_cell_1.dat` | DAT-dependent tests missing `[Trait("Lane", "InstalledDat")]`. Build machines have no DATs |
| Timing-sensitive test fails only under load | `FakeAceTransportTests.PausedSelector_…` took 37-42 s under load vs ~350 ms isolated. Its harness drives a VIRTUAL clock but asserted on 2 s wall-clock windows; those are patience, not assertions, and now share a 60 s `HarnessPatience`. **Do not serialize the assembly to fix it** — that regressed Windows from 1000 passed in 7 s to 999/1000 in 17 s, breaking a loss-soak test that had never failed |
| Avalonia "calling thread cannot access this object" in cleanup | `MainWindowViewTests` needs a real desktop session and is `Lane=Manual`. Measured: PASSES on a dev desktop and on the CI Windows box over SSH; FAILS under `act_runner` and on Linux. Serializing the assembly does **not** fix it (tried via `xunit.runner.json` and a compiled-in `CollectionBehavior` attribute), and de-async-ing the test actively causes the failure. The stack shows a compositor being **constructed** during teardown — it is the headless session lifecycle, not parallelism |

## Culture note

The `en-SE` discovery is worth remembering beyond CI: config files, numeric
parsing, and the wire are all culture-safe (`System.Text.Json` is invariant by
spec, every `float/double.TryParse` passes `CultureInfo.InvariantCulture`, and
the protocol is binary). Only **diagnostic strings** format with the current
culture, so a European player sees `local=(8,00; 191,00)` in an F3 dump. The
client installs and runs correctly in both the US and Europe.