build-go and build-gui now build straight to their final release filename
(e.g. dist/Completed-api_linux_amd64, dist/Completed-gui_windows_amd64.exe)
and upload-artifact uploads that directly - no separate Archive step, no
zip/tar.gz wrapping a single or paired executable.
build-web stays archived: it's not an executable, it's the built Angular
site (a directory of HTML/JS/CSS/asset files), so bundling it into one
tar.gz is still the right call there.
Trade-off worth knowing: a raw upload doesn't preserve the Unix executable
bit the way tar.gz did, so linux/darwin downloads need `chmod +x` before
running; .exe downloads are unaffected. Documented in release.yml's
comments.
The release job's download/upload loop needed no changes - it already
walks whatever's under artifacts/ with `find -type f`, regardless of how
many raw files land in a given artifact's subdirectory.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found by actually running release.yml against a real v0.0.1 tag: every
build-* job failed at actions/upload-artifact@v4 with "GHESNotSupportedError:
@actions/artifact v2.0.0+, upload-artifact@v4+ and download-artifact@v4+
are not currently supported on GHES." - exactly the risk flagged (but not
yet hit) when this workflow was first written. This Gitea Actions instance
doesn't support the v4 artifact protocol; v3 does.
v3 has no merge-multiple option: downloading all artifacts with no `name`
given lands each one in its own artifacts/<artifact-name>/ subdirectory
instead of flattened into artifacts/. Changed the generic-package-registry
upload loop from a flat `artifacts/*` glob to `find artifacts -type f` so
it finds files regardless of that nesting.
Not yet confirmed: whether build-docker (needs secrets.TOKEN, still unset)
and the rest of the release job succeed once the artifact step itself
works - that's the next real run to check.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found by actually running ci.yml on the real Gitea instance: the go job
failed at go vet with "wayland-client-core.h: No such file or directory"
building go-gl/glfw (a Fyne dependency), which compiles in both its X11
and Wayland backends by default. libgl1-mesa-dev + xorg-dev alone aren't
enough; added libwayland-dev and libxkbcommon-dev to both ci.yml and
release.yml's linux packages list.
I'd only verified cmd/gui's build on macOS and via the Windows mingw-w64
cross-compile - never on an actual Linux machine, so this gap wasn't
visible until the real CI run hit it. sqlc, web and docker all passed on
that same run (confirmed from the job logs, not just the status column).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
It set container: gitea/gitea:latest, which has no node - actions/checkout
fails immediately with "node: executable file not found" (confirmed via
tea actions runs logs 102 on the live instance: conclusion failure, 15s).
ci.yml replaces it - same go test coverage plus sqlc/web/docker, and no
container override, so it uses the runner's normal image instead of the
Gitea server's own image.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
cmd/api/main.go: registerWebUI serves a built web/ as a single-page app -
Echo's static middleware in HTML5 mode, with a Skipper that exempts
/api/*, /hello, /docs, /schemas/* and openapi.json/.yaml so it can never
shadow a real API route. It reads WEB_DIST_DIR (default /frontend, where
the Dockerfile puts it) and no-ops if that directory doesn't exist, so
`go run ./cmd/api` and CI stay Node-free and API-only; only the Docker
image actually turns the web UI on.
Dockerfile: new node:22-alpine stage builds web/, copied into the final
image at /frontend alongside the existing Go-built binary. Added
.dockerignore (web/node_modules, web/dist, web/.angular, .git) now that a
second language's tooling is in the build context.
Looked at how your other Gitea projects (MusicServer et al.) structure
this, since I can't exercise Gitea Actions itself from here: MusicServer's
own Dockerfile already does "copy a prebuilt frontend into the Go image,
serve it from Go" (there, from a separately-published frontend image) -
registerWebUI does the same thing in spirit, just within one image rather
than two, since this UI isn't meant to be published separately.
release.yml:
- build-gui now also cross-compiles the Fyne GUI for windows/amd64 via
mingw-w64, alongside the existing linux/amd64. Verified for real, not
just written: installed mingw-w64 locally and cross-built cmd/gui,
first plain (got a working PE32+ console exe), then with
-ldflags "-H=windowsgui" (got the GUI-subsystem exe actually wanted, no
console window). macOS isn't attempted - cgo has no equivalent
cross-toolchain for Darwin and there's no registered macOS runner here.
- new build-docker job pushes the API image (now carrying the web build)
to gitea.sanplex.xyz/sansan/completed, tagged with the version and
latest.
- replaced the hand-rolled curl+jq Gitea-release creation with
akkuman/gitea-release-action, and artifact upload with the Gitea generic
package registry - both are what your other projects (HanjaKanaMun,
MusicPlayer) actually use, found by reading their workflows rather than
guessing. Same for the registry host, owner, and secrets.TOKEN name
(previously secrets.GITHUB_TOKEN, which nothing else here uses).
ci.yml: new docker job builds the same image on every push to main/develop
(push: false) as a health check, matching MusicServer's build-vs-release
split.
Verified for real wherever I could: a full `docker build` of the new
Dockerfile, then running that image against a real Postgres container and
hitting it with curl - / serves the actual built Angular index.html,
/games (a client-side route) correctly falls back to it, /api/games and
/hello are untouched (still return real API responses, not shadowed by
the static handler), and register+list-games works end to end through the
container. Also ran it the documented way, `docker compose up --build`,
with the same results. Separately confirmed the Windows Fyne cross-build
produces a real, correctly-subsystemed .exe. What I could not run from
here is Gitea Actions itself - the workflow YAML is written to match your
other projects' established, working conventions, but its first real
execution on your Gitea instance is still the first real test of it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
internal/client (new): a typed Go HTTP client for the API - login,
register, logout, me, and full CRUD for games/platforms - shared by the
TUI and the Fyne GUI so request-building, auth and error handling (an
APIError with the HTTP status) live in one tested place instead of being
duplicated. 8 tests against an httptest server covering success, 401, 403
and that the token is stored/cleared at the right times.
cmd/tui (rewrite): drops the old unauthenticated read-only viewer. Now a
login screen (bubbles/textinput) -> a scrollable games table
(bubbles/table) with 'r' to refresh, built as an explicit state machine
(screenLogin/screenLoading/screenGames) so Update's transitions are
testable without a real terminal. 9 tests driving Update() directly.
cmd/gui (rewrite): drops raylib for Fyne. Same shape as the TUI - a login
form, then a games table, with networking on a goroutine and fyne.Do for
the UI updates. The table-cell and login-validation logic are plain
functions with no Fyne/display dependency, so they're tested directly (7
tests) rather than needing Fyne's headless test driver.
internal/api: tagged every operation (Auth/Platforms/Games/Misc) so the
generated OpenAPI spec groups into named services instead of one bucket;
api/openapi.json regenerated. Added CORS middleware (wildcard origin, no
credentials needed since clients use a bearer token, not the cookie) -
needed for the new web client's browser requests to the API's a different
origin.
web/ (new): an Angular 22 + Material app, API client generated from
api/openapi.json via ng-openapi-gen (services: true for per-resource
classes - AuthService/GamesService/PlatformsService - instead of one
generic Api facade). A Session service holds the token/user as signals,
persists the token to localStorage, and is the only thing that calls the
generated AuthService; an HttpInterceptorFn attaches the bearer token to
every request; a CanActivateFn guard protects /games. Login and Games are
standalone, lazy-loaded route components. 21 tests (Vitest, Angular's new
default - no browser/Karma needed) covering the service, interceptor,
guard and both components, run via `npm test` exactly as CI will call it.
Justfile: fixed run-tui/run-gui, which used `go run .../main.go` - broken
now that both packages span multiple files, since that only compiles the
one named file. Added web-install/web-dev/web-build/web-test.
.gitea/workflows/ci.yml (moved from .github, which Gitea doesn't read):
on push to main/develop - gofmt, go vet, go build, go test, go mod tidy
diff (with the Fyne build's system deps installed), sqlc diff, and the
web app's build+test.
.gitea/workflows/release.yml (new): on a vX.Y.Z tag, cross-compiles
api+tui for linux/darwin/windows (amd64+arm64, pure Go, no cgo), builds
the Fyne GUI for linux/amd64 only (cgo + its OpenGL/X11 link means it
can't cross-compile from one Linux runner the way api/tui do, and this
repo has no registered macOS/Windows runner to build it on natively -
documented in the workflow's comments), builds the web app, and creates a
Gitea release with all of it attached via the Gitea API. Written to
Gitea's documented GitHub-Actions-compatible syntax but not executable
from here - there's no Gitea server in this environment to run it
against, unlike everything else in this commit, which was.
Verified: every Go package builds/vets/tests clean; sqlc diff and go mod
tidy -diff are clean; the Angular app builds for production and its test
suite passes via `npm test`/`npm run build` (the exact commands CI runs),
including after a clean `npm ci` from the committed lockfile; the new CORS
config was checked with real preflight and GET requests against a running
server, not just read.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Anyone authenticated can still read the shared platform list, but only an
admin may add, rename or delete from it.
- users.is_admin boolean, defaulting to false. The very first account ever
registered becomes an admin automatically (CreateUser's INSERT uses
NOT EXISTS (SELECT 1 FROM users) for is_admin), so there's always at
least one without manual setup; promoting anyone beyond that is a manual
UPDATE for now (documented in the migration's comment on the column).
- GetSessionAndUser, GetUserByEmail and GetUserByID now also select
is_admin, and it flows through to model.User (and so to
/api/auth/{register,login,me}'s response) and to the auth middleware's
authedUser.
- New requireAdminMiddleware + addAdminAuth, layered on top of the existing
auth middleware. Wired onto create/update/delete-platform; list/get stay
on plain addAuth, since reading the list only requires being logged in.
Verified end to end: the first registered user comes back is_admin:true,
the second false; the non-admin gets 403 on create/update/delete-platform
but 200 on list; the admin can do all of it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Platforms no longer have a user_id: there's one global list, unique by
name, that every authenticated user reads from and writes to. Games are
unchanged - still private per user, still referencing platform_id.
- Migration: dropped platform.user_id and its FK, UNIQUE(user_id, name) ->
UNIQUE(name), index on change_seq alone instead of (user_id, change_seq).
Amended the one init migration in place rather than stacking a second one,
since nothing is deployed yet.
- db/query.sql: platform queries no longer take a user_id parameter.
DeletePlatform's NOT EXISTS guard already checked games from any user (it
never filtered the subquery by user_id), so deleting a shared platform is
correctly blocked by any user's games on it, not just the deleter's.
- routes.go: platform handlers still require auth (the list isn't public to
anonymous requests) but no longer scope by user.
- Tests: updated the mocks for the new signatures, added
TestListPlatformsIsSharedAcrossUsers.
Verified against a real Postgres: user A creates "PC", user B sees it
immediately, both can add games against it, each only sees their own
games, deleting the platform is blocked while either user has a game on
it, and the name stays globally unique.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Steps 1-4 of the review plan (docs/plans not checked in; see conversation):
Step 1 - make it boot from a clean checkout:
- Replace the old schema.sql + single ALTER-TABLE migration (which never
created the tables) with one init migration: users, sessions, platform,
game, all with UUIDv7-ready uuid ids, soft deletes (deleted_at), and a
shared change_seq sequence + trigger for a future sync feature.
- sqlc now reads the schema from db/migrations directly.
- Migrations are embedded in the binary (db/migrations/embed.go) instead of
being copied around and found via path fallbacks.
- ensureDatabase/runMigrations failures are now fatal, and main pings the
pool after connecting, instead of starting an API with no schema.
- Switched both the app and golang-migrate to pgx/v5 (dropped lib/pq).
- Fixed the Dockerfile's Go version (1.23 -> 1.25) and build target.
- Added a postgres healthcheck to docker-compose so the app waits for a
ready database instead of racing it.
- go.mod tidied, bin/ added to .gitignore, Justfile's docker-compose calls
switched to the v2 `docker compose` plugin.
Step 2 - auth and per-user data:
- internal/auth: argon2id password hashing, random session tokens stored
only as a hash.
- POST /api/auth/{register,login,logout}, GET /api/auth/me. Sessions work
via either an Authorization: Bearer header or an httpOnly cookie.
- Every platform/game query is scoped to the authenticated user.
Step 3 - API correctness:
- internal/model holds the DTOs, with no server-side dependencies, so
clients can import it later without pulling in Huma/Echo.
- dbError() maps ErrNoRows -> 404, unique violation -> 409, FK violation ->
409, instead of leaking Postgres error text with a 500.
- Soft-delete meant the FK on game.platform_id no longer blocks deleting a
referenced platform on its own; added a NOT EXISTS guard plus an
application-level check so that case now correctly returns 409 instead of
silently succeeding (caught by testing the live server, not by a mock).
- Added GET /games/{id} and GET /platforms/{id}. All JSON routes moved
under /api, freeing / for a future web client.
- ListGames/GetGame join platform_name in.
- cmd/openapi writes the generated spec to api/openapi.json.
Step 4 - tests and CI:
- Rewrote the mocks and handler tests for the new Querier shape; added
coverage for 401 (no session), 404 (not found / another user's row),
409 (duplicate email, platform referenced by games), 422 (score out of
range), and that queries are actually scoped by the authenticated user's
id, not just that the right status code comes back.
- .github/workflows/ci.yml: go vet, go test, go mod tidy -diff, sqlc diff.
Everything above was exercised against a real local Postgres (migrations
up/down, the trigger, constraints) and through the running server via curl
(register, login, scoping, every status-code case), not just unit tests.
Known regression, intentionally deferred to the next step: cmd/tui and
cmd/gui still call the old unauthenticated GET /games, which no longer
exists. They need internal/client and a login flow before they work again.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
SansanandClaude Sonnet 5
Your Name