Add GitHub Actions CI for the CPU-only test suites

.github/workflows/tests.yml

@@ -0,0 +1,100 @@
+name: Tests
+
+# CPU-only checks: everything here runs on a GitHub-hosted runner with no GPU
+# and no pycolmap/torch build, so it covers a subset of AGENTS.md's "Tests"
+# list. Skipped on purpose: queue/test_api.py (needs a running service),
+# scripts/test_fisheye.py (needs pycolmap), and the GPU half of
+# scripts/test_benchmark.py (skips itself when there is no CUDA device).
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+concurrency:
+  group: tests-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  queue:
+    name: Queue service tests
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+          cache-dependency-path: queue/requirements.txt
+
+      - name: Install queue dependencies
+        run: pip install -r queue/requirements.txt
+
+      - name: Install ffmpeg
+        run: sudo apt-get update -qq && sudo apt-get install -y --no-install-recommends ffmpeg
+
+      # Each suite sets its own SPLAT_ROOT/QUEUE_ROOT/QUEUE_GPUS before
+      # importing app.* (AGENTS.md, "Tests") -- nothing to configure here.
+      - name: test_stages.py
+        run: python3 queue/test_stages.py
+      - name: test_worker.py
+        run: python3 queue/test_worker.py
+      - name: test_regressions.py
+        run: python3 queue/test_regressions.py
+      - name: test_distance.py
+        run: python3 queue/test_distance.py
+      - name: test_telemetry.py
+        run: python3 queue/test_telemetry.py
+      - name: test_remote.py
+        run: python3 queue/test_remote.py
+      - name: test_benchmark.py
+        run: python3 queue/test_benchmark.py
+      - name: test_hoststats.py
+        run: python3 queue/test_hoststats.py
+
+  scripts:
+    name: Pipeline script tests (CPU)
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+
+      # numpy + OpenCV only: the same subset venv_gs/venv carry that these
+      # suites actually touch. No torch/gsplat/pycolmap here, so the GPU
+      # half of test_benchmark.py skips itself and test_fisheye.py is not run.
+      - name: Install script dependencies
+        run: pip install numpy opencv-python-headless
+
+      - name: test_avata_motion.py
+        run: python3 scripts/test_avata_motion.py
+      - name: test_imu_select.py
+        run: python3 scripts/test_imu_select.py
+      - name: test_upright.py
+        run: python3 scripts/test_upright.py
+      - name: test_benchmark.py (CPU part; GPU part skips without CUDA)
+        run: python3 scripts/test_benchmark.py
+
+  guardrails:
+    name: queue/app stays free of heavy deps
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - uses: actions/checkout@v4
+
+      # AGENTS.md: "The service's own venv has no numpy/OpenCV, so heavy
+      # work belongs in a script, not in queue/app/." Cheap to check for
+      # every push instead of finding out at deploy time.
+      - name: No numpy/cv2/torch imports in queue/app
+        run: |
+          if grep -rEn '^\s*(import|from)\s+(numpy|cv2|torch)\b' queue/app --include='*.py'; then
+            echo "::error::queue/app imported a heavy dependency; that work belongs in scripts/ (AGENTS.md)"
+            exit 1
+          fi

AGENTS.md

@@ -35,6 +35,13 @@ Every test must set `SPLAT_ROOT`/`QUEUE_ROOT` to a temporary directory and
 `QUEUE_GPUS=""` **before** importing `app.*` — `config.py` reads them at import
 time, and a test that forgot once wrote into a live queue database.
 
+**CI** (`.github/workflows/tests.yml`) runs everything above except
+`test_api.py` (needs a running service) and `test_fisheye.py` (needs
+pycolmap) on every push and pull request, on a GitHub-hosted runner with no
+GPU — `scripts/test_benchmark.py`'s GPU half skips itself there, same as
+locally. A separate job greps `queue/app/` for `numpy`/`cv2`/`torch` imports
+to keep the layout rule above honest.
+
 ## Rules
 
 - **Cache keys.** Each stage is cached by a hash of the options that affect

README.md

@@ -1,5 +1,7 @@
 # OSVplat
 
+[![Tests](https://github.com/pgodlews/OSVplat/actions/workflows/tests.yml/badge.svg)](https://github.com/pgodlews/OSVplat/actions/workflows/tests.yml)
+
 **Raw DJI `.OSV` dual-fisheye → Gaussian splat. No stitch.**
 
 OSVplat turns a DJI Osmo 360 or Avata 360 `.OSV` into a flyable 3D Gaussian

queue/README.md

@@ -341,7 +341,7 @@ POST   /api/cache/gc[?dry_run&budget_gb]
 
 ## Tests
 
-Five suites, installed by `deploy.sh`. Only the API suite needs a running
+Ten suites, installed by `deploy.sh`. Only the API suite needs a running
 service and reads `QUEUE_TOKEN` from the service's own env file:
 
 ```bash
@@ -352,10 +352,22 @@ set -a; . ~/splat/queue_app/.queue_env; set +a; python3 ~/splat/queue_app/test_a
 ~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_stages.py
 ~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_worker.py
 ~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_regressions.py
+~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_distance.py
 ~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_telemetry.py
+~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_remote.py
+~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_benchmark.py
+~/splat/queue_app/venv/bin/python ~/splat/queue_app/test_hoststats.py
 ~/splat/venv/bin/python ~/splat/scripts/test_fisheye.py
 ```
 
+None of the nine non-API suites need a GPU, and every one of them sets its
+own `SPLAT_ROOT`/`QUEUE_ROOT`/`QUEUE_GPUS` to a scratch directory before
+importing `app.*`, so they never touch a live queue. `.github/workflows/tests.yml`
+runs the CPU-only subset (everything above except `test_api.py` and
+`test_fisheye.py`, which need a running service and pycolmap respectively) on
+every push and pull request; see [AGENTS.md](../AGENTS.md#tests) for the full
+list, including the `scripts/` suites that need `venv_gs`.
+
 `test_worker.py` covers the lifecycle paths that used to be reasoned about and
 never exercised, each one a way a job could end up holding a GPU or a cache
 directory could end up with two writers: a cancel racing the dispatcher, a
@@ -377,8 +389,16 @@ key while each variant gets a distinct `train` key.
 
 `test_regressions.py` uses ffmpeg and an isolated database to check dual-lens
 trimming, concurrent thumbnails, scheduling estimates, pagination and fitted
-history. `scripts/test_fisheye.py` uses numpy and pycolmap on CPU to check stale
-model rejection and the distortion refit; neither suite starts GPU work.
+history. `test_distance.py` covers the Avata distance-selection option's cache
+keys and API surface (see below). `test_remote.py` covers CPU/GPU discovery and
+the output upload for rented-GPU runs (docs/cloud.md), against a local HTTP
+server standing in for S3. `test_benchmark.py` covers the host-benchmark
+service side (the machine hold, the run record); the workload itself
+(`scripts/benchmark.py`) is tested by `scripts/test_benchmark.py`, which needs
+`venv_gs`. `test_hoststats.py` covers the machine-load counters and the
+per-job time series against a faked `/proc` and cgroup tree. `scripts/test_fisheye.py`
+uses numpy and pycolmap on CPU to check stale model rejection and the
+distortion refit; none of these suites starts GPU work.
 
 ## Validated end to end
 

queue/deploy.sh

@@ -51,7 +51,9 @@ rsync -az "$REPO/queue/requirements.txt" "$DEST$APP/"
 # have to be there; test_api.py is stdlib-only and runs from anywhere.
 rsync -az "$REPO/queue/test_stages.py" "$REPO/queue/test_worker.py" \
   "$REPO/queue/test_api.py" "$REPO/queue/test_regressions.py" \
-  "$REPO/queue/test_distance.py" "$DEST$APP/"
+  "$REPO/queue/test_distance.py" "$REPO/queue/test_telemetry.py" \
+  "$REPO/queue/test_remote.py" "$REPO/queue/test_benchmark.py" \
+  "$REPO/queue/test_hoststats.py" "$DEST$APP/"
 # summarize_sweep.py reads the queue over HTTP and is meant to be run on the
 # target, where it can pick the token out of .queue_env; it was never shipped there.
 rsync -az "$REPO/queue/summarize_sweep.py" "$DEST$APP/"