Two runs (447, 486) died with 'ld terminated with signal 9' linking the
solitaire_engine test binary — the runner OOMs on full dev debuginfo for
the Bevy dependency graph. line-tables-only preserves file:line panic
backtraces at a fraction of the link footprint.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Review feedback on #135: floating 'stable' + -D warnings lets every new
clippy release redden master with new lints — pin 1.95.0 like the
web-wasm-rebuild workflow. And ubuntu-latest lacks the ALSA/udev/X11/
Wayland dev packages the Bevy crates link against; install the standard
Bevy CI set (the project's own builder images have no desktop-Bevy
precedent — android-builder targets the NDK and web-e2e only builds the
server).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Until now no CI workflow ran the test suite or clippy at all — the
android-release, docker-build, web-e2e, web-wasm-rebuild, and
builder-image workflows cover packaging and e2e, but a direct push to
master (including the web-wasm-rebuild bot commit) never executed
cargo test or clippy. The gate discipline in CLAUDE.md §6 existed only
on developer machines.
test.yml runs the exact §6 commands (clippy --workspace --all-targets
-D warnings; cargo test --workspace) on master pushes and PRs, with
SQLX_OFFLINE against the checked-in .sqlx cache, path-filtered so
docs-only merges don't burn CI. Toolchain/cache mirror web-e2e.yml.
Found during the 2026-07-06 large-scale review (finding H1).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The rebuild-and-diff freshness gate (#93) could never pass: the Bevy wasm
artifacts aren't byte-reproducible across machines. Confirmed conclusively —
identical rustc 1.95.0 / LLVM 22.1.2, identical flags, Cargo.lock and remapped
source paths still yield host-dependent output (CI's build was 248 KB smaller
than a local one), while same-machine rebuilds are bit-identical. Path
remapping (kept in build_wasm.sh) is necessary but not sufficient.
Make CI the single source of truth instead: `web-wasm-rebuild` rebuilds pkg/ on
every master change to a wasm-feeding crate and commits it back. The deployed
artifacts can't silently rot and contributors no longer hand-run build_wasm.sh.
The pkg/ commit isn't in this workflow's trigger paths (no loop) but is in
docker-build's, so the refreshed wasm deploys.
Removes the false-failing web-wasm-freshness workflow.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The web-e2e job has been failing on every run with:
Error: Timed out waiting 120000ms from config.webServer.
Playwright's `webServer` runs `cargo run -p solitaire_server --quiet` and
waits 120s for /health. The workflow cached only npm — no Rust cache — so each
run cold-compiled the entire server graph (axum/sqlx/reqwest/aws-lc-sys) inside
that 120s window and never came up. The browser tests never executed; the
failure was infra, not a web regression.
Fix the harness so the tests actually run:
- add `Swatinem/rust-cache@v2` to warm the cargo cache across runs
- add a `Prebuild server` step (`SQLX_OFFLINE=true cargo build -p
solitaire_server`) before Playwright, so `webServer`'s `cargo run` reuses the
compiled binary and starts in seconds (.sqlx offline cache is committed)
- raise `webServer.timeout` 120s -> 300s as a safety margin for a cold cargo
cache (e.g. first run after a deps bump)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The web build shipped a ~3-week-stale pkg/ because the old "Check wasm pkg
drift" step in docker-build.yml only hard-failed on direct solitaire_web/
edits and treated solitaire_engine/_core changes as a non-blocking notice —
so the entire card_game migration (engine/core/data churn, v4->v5 save
schema) slipped through.
Replace it with a dedicated `web-wasm-freshness` workflow that rebuilds the
artifacts and diffs them against what's committed. A fresh build on a pinned
toolchain (rust 1.95.0 / wasm-bindgen 0.2.120 / wasm-pack 0.14.0 / binaryen
130) is byte-for-byte reproducible — verified locally — so this is precise:
it fails on *any* real drift (including gameplay-logic changes that don't
touch the JS API surface) and has no false positives on wasm-irrelevant edits.
Runs on pull_request as well as master push, so staleness is caught before
merge rather than only blocking the post-merge deploy.
Remove the superseded heuristic from docker-build.yml; master stays fresh via
the PR gate, so the deploy image is always built from current artifacts.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
solitaire_server/e2e/:
- smoke.spec.js: verifies /play-classic loads, exposes window.__FERROUS_DEBUG__
bridge, keyboard parity (Space=draw, U=undo), debug failure report, and
replay payload builder exports schema-v2 moves.
- gameplay_review.spec.js: HUD/controls render check, stock-click + undo
player flow, draw-mode toggle, autonomous play invariant batch, and
cycle-detection regression guard.
- cycle_metrics.js: headless cycle-rate analysis tool; run via
`npm run review:cycles` with configurable policy, game count, and
thresholds. Regression gate baked into package.json scripts.
- playwright.config.js: targets the local server at http://localhost:8080.
- package.json / package-lock.json: @playwright/test 1.60.0.
.gitea/workflows/web-e2e.yml:
- Runs on pushes to solitaire_server/, solitaire_wasm/, solitaire_core/,
or Cargo changes. Starts the server binary, waits for /health, runs
the full Playwright suite, uploads test-results/ on failure.
docs/testing-architecture.md: documents the three-tier test strategy
(unit → Playwright smoke → cycle regression) and the __FERROUS_DEBUG__
bridge contract.
scripts/update_quaternions_deps.sh: helper to bump the Quaternions
registry deps (klondike, card_game) by version and run the full
safety gate including deterministic replay checks.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Dockerfile:
- Drop --mount=type=secret,id=cargo_token: the Quaternions private
registry has been migrated to the public Cargo.io path so the build
secret is no longer needed. Removes the requirement for CI_TOKEN to
carry registry credentials.
CI workflow (docker-build.yml):
- Add solitaire_wasm/** and solitaire_web/** to the push-trigger paths
so changes to either WASM crate actually fire the build job.
- Add wasm drift check for solitaire_wasm artifacts (solitaire_wasm.js,
solitaire_wasm_bg.wasm) — exits 1 if solitaire_wasm/ or solitaire_core/
changed without updating the committed pkg files.
- Add hard canvas drift check: solitaire_web/ changes MUST update
canvas_bg.wasm or the deploy gets a stale Bevy binary.
- Add advisory notice for solitaire_engine/ / solitaire_core/ changes
that omit a canvas_bg.wasm rebuild (non-blocking; formatting commits
should not fail CI).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
cargo fetch --locked was failing with "failed to parse manifest" because
.cargo/config.toml (which registers the Quaternions sparse index) was
never copied into the build image, and the registry's auth token was
never supplied.
Changes:
- COPY .cargo/config.toml into the builder stage so Cargo knows the
Quaternions registry URL.
- Replace bare `cargo fetch` and `cargo build` with
`--mount=type=secret,id=cargo_token` variants that set
CARGO_REGISTRIES_QUATERNIONS_TOKEN from the mounted secret — token
never appears in image layers or docker history.
- Workflow: pass CI_TOKEN as the `cargo_token` build secret.
- Add solitaire_engine/** and solitaire_server/Dockerfile to trigger
paths so engine changes and Dockerfile edits kick off rebuilds.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
'git checkout deploy' is ambiguous because the repo contains both a
deploy/ directory and a deploy remote tracking branch. Switch to
'git switch' which is branch-only and unambiguous.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The CI bot was committing image-tag bumps back to master after every
Docker build, which forced a `git pull --rebase` before every developer
push. Moving the kustomization commit to a dedicated `deploy` branch
keeps master clean — the build bot no longer diverges it.
Argo CD / Flux should now watch the `deploy` branch (targetRevision:
deploy) instead of master.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Tag-push events are not reliably processed by the self-hosted Gitea
runner. workflow_dispatch with a tag input allows manual triggering
via the Gitea UI or API as a fallback.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the 5 per-run tool-install steps (~2m 30s) with a pre-built
container image (git.aleshym.co/funman300/android-builder) that ships
Ubuntu 22.04 + Java 17 + Android SDK/NDK + Rust stable + aarch64 target
+ cargo-ndk + sccache. android-release.yml now runs inside the container
and adds two cache steps instead: Cargo registry and sccache directory.
sccache (RUSTC_WRAPPER) caches at the translation-unit level so partial
hits survive Cargo.lock changes — far more resilient than caching the
full target/ directory.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
AndroidManifest.xml had hardcoded versionCode=1 / versionName=1.0, so
every shipped APK looked identical to Android and Obtainium could never
confirm the installed version matched the latest release tag — causing
a persistent false-update notification loop.
VERSION_NAME is now passed into the build script from the CI tag
(e.g. "v0.28.0" → versionCode=2800, versionName="0.28.0") and
forwarded to aapt2 link via --version-code / --version-name, overriding
the manifest without touching the file.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaced the curl|bash install_kustomize.sh approach (which makes an
unauthenticated GitHub API call to resolve the latest version) with a
direct pinned tarball download. Eliminates the tar glob failure that
broke run #226.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fires on any v* tag push. Steps:
- Installs Android SDK + NDK 30.0.14904198 (cached by SDK version key)
- Builds release APK for arm64-v8a via scripts/build_android_apk.sh
- Signs with the release keystore stored in RELEASE_KEYSTORE_B64 secret
- Creates (or reuses) the Gitea release for the tag
- Uploads solitaire-quest.apk as a release asset
Obtainium users can track releases by adding:
https://git.aleshym.co/funman300/Rusty_Solitare
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Switch docker-build.yml from paths-ignore to an explicit paths allowlist
so the workflow only fires on changes to solitaire_server/, solitaire_sync/,
solitaire_core/, Cargo.toml, Cargo.lock, or the workflow file itself.
Also harden the "commit and push updated kustomization" step:
- exit 0 early when the kustomization has no staged diff (nothing to push)
- retry the pull+push loop up to 3 times with a 5 s delay to handle
concurrent pushes that race the CI commit
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
All 10 warnings were caused by hotkey/keyboard UI code behind
#[cfg(not(target_os = "android"))] call sites whose definitions lacked
the matching gate. Fixes:
- help_plugin: gate keyboard-chip imports and font_kbd; #[allow(dead_code)]
on ControlRow (keys field is data, not dead)
- hud_plugin/ui_modal: replace cfg shadow pattern with cfg!() expression
so the hotkey parameter is read on every platform
- home_plugin: gate fn hotkey behind not(android)
- onboarding_plugin: gate HotkeyRow, HOTKEYS, spawn_slide_hotkeys and
their exclusive imports behind not(android)
- replay_overlay: gate keybind_footer_hint_text behind not(android)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Building locally or via a different pipeline; the self-hosted runner's
resource constraints made the 3-ABI release build unreliable.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
workflow_dispatch caused the entire android-release workflow to be silently
dropped; tag-triggered releases (v0.25.6) worked fine before this trigger
was added. Removing it restores tag-based release triggering.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Allows manually triggering the release build from the Gitea UI
(Actions tab → Android Release → Run workflow) without needing
to push a version tag.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
If KEYSTORE_BASE64 is unset, base64 -d writes an empty file silently,
cargo ndk then spends ~7 min compiling all ABIs, and only then does
apksigner fail. Add a size check after decode so the job fails in
seconds with a clear error message instead of wasting a full build.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The runner LXC was bumped from ~56 GB to 106 GB, giving ~70 GB of free
space — well above the ~40 GB a full 3-ABI release build needs. Revert
the disk-budget workarounds added in ab35fcf:
- Remove "Free disk space" step (no longer needed)
- Restore x86_64 target (arm64-v8a + armeabi-v7a + x86_64)
- Remove ABIS override so build script uses its full default set
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Run 181 (v0.25.0 tag) failed at "Build signed release APK" after ~7 min —
same disk-exhaustion pattern that hit the debug build. The debug workflow
was already fixed to arm64-v8a only; the release workflow still built all 3
ABIs and exceeded the runner's disk budget.
Changes:
- Add "Free disk space" step before system deps: removes /usr/local/lib/android,
/usr/share/dotnet, /opt/ghc, /usr/local/share/boost (~10 GB reclaimed).
- Limit ABIS to arm64-v8a + armeabi-v7a (drops x86_64, which is emulator-only).
- Remove x86_64 from rustup target add to match.
arm64-v8a covers all modern Android devices; armeabi-v7a covers legacy ARM.
x86_64 can be re-added later if a simulator-targeted test build is needed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The disk-budget fix worked — debug APK now builds, signs, and verifies
in ~6 minutes on a single ABI. But the upload step failed with:
GHESNotSupportedError: @actions/artifact v2.0.0+, upload-artifact@v4+
and download-artifact@v4+ are not currently supported on GHES.
upload-artifact@v4 rewrote the upload path to use a new artifact
service hosted on github.com; Gitea's GHES-compatibility layer doesn't
implement that endpoint. v3 still uses the older chunked HTTP upload
API that Gitea supports.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous run got all the way through compile + link + zipalign and
then died inside apksigner with `IOException: No space left on device`.
Cross-compiling all three Android ABIs (arm64-v8a, armeabi-v7a, x86_64)
in debug mode blows target/ past 25 GB, and by the time apksigner is
streaming the signed APK to disk the runner has nothing left.
Two changes:
1. build_android_apk.sh now reads `ABIS` from the environment (defaults
to all three for backwards compat) and uses it to assemble the
cargo-ndk `-t` flags.
2. android-build.yml passes ABIS=arm64-v8a, since the debug artifact
is consumed by adb-installing to a single arm64 device and the
other two ABIs were dead weight.
Also frees \$STAGING/app-unsigned.apk right after zipalign so it's not
sitting next to the aligned APK and the output APK during signing.
Release workflow is untouched — release APKs still ship all three ABIs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
cargo-apk 0.10 and its fork cargo-apk2 both failed to discover the
installed Android platform in this Gitea runner, despite ANDROID_HOME,
platforms;android-34, build-tools, and NDK all being present, readable,
and pointed at correctly. We never isolated whether the bug is in the
shared ndk-build crate's discovery logic or in the runner's env-var
propagation through cargo subcommand exec, so this commit stops fighting
either tool and assembles the APK from explicit toolchain steps instead:
cargo ndk -> per-ABI .so files
aapt2 compile/link -> manifest + resources -> base APK
zip -> bundle native libs into lib/<abi>/
zipalign -> 4-byte alignment
apksigner -> v2/v3 signing (debug keystore for CI, real for release)
The pipeline lives in scripts/build_android_apk.sh so it's reproducible
locally (same env vars, same commands). AndroidManifest.xml is now
checked in under solitaire_app/android/ and mirrors what cargo-apk would
have generated from [package.metadata.android] — keep them in sync if
either is changed. Local `cargo apk build` still works on developer
machines where cargo-apk is happy; CI just stops depending on it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
cargo-apk 0.10.0 has been unable to discover an installed Android
platform in this runner environment despite ANDROID_HOME, NDK,
build-tools, and platforms;android-34 all being present and readable.
cargo-apk2 is the maintained community fork on crates.io that reads
the same `[package.metadata.android]` block, so the solitaire_app
Cargo.toml needs no changes. Cache keys updated to apk2- so we don't
restore the broken cargo-apk binary from prior runs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Mirror the fix from android-build.yml: rename ANDROID_HOME -> ANDROID_SDK
in the env block to avoid the Docker-image-baked ANDROID_HOME overriding
the workflow value in run scripts. Use ${{ env.ANDROID_SDK }} template
expressions throughout, and explicitly export ANDROID_HOME/ANDROID_NDK_HOME
before cargo-apk build so it finds the SDK at the right path.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace shell variable $ANDROID_HOME references in run blocks with
${{ env.ANDROID_SDK }} template expressions. Gitea runner v1 may not
override Docker-image-baked ENV vars via docker exec; template expansion
happens at workflow compilation time, so the literal path is hardcoded
into the script before the shell runs it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Remove SDK detection logic and install directly to /opt/android-sdk,
matching the release workflow. The container Docker image has ANDROID_HOME
baked in at /usr/local/lib/android/sdk; installing there with sudo while
cargo-apk resolves ANDROID_HOME from the image ENV created a divergence.
Using a controlled path we own eliminates that class of conflict entirely.
Add SDK cache shared with the release workflow (same key prefix v2-).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ANDROID_SDK_ROOT was never set; zipalign and apksigner were resolving
to empty paths and would fail. All three occurrences replaced with
ANDROID_HOME which is defined in the workflow env block.
Also adds sudo to the cache-miss SDK install (mkdir/mv/sdkmanager) to
match the debug workflow pattern — /opt/android-sdk requires root on
a fresh runner.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
funman300