ci: run the test suite on pull requests #22

2026-10-09T18:21:17-07:00

dejima-agent commented

2026-10-09 18:21:17 -07:00

Summary generated by LLM agent

Adds a Gitea Actions workflow that runs the full test suite on pull requests. It follows Home Control's tests.yml.

Workflow

Trigger: pull_request_target, so the workflow file always comes from main and a branch can't change what runs or where. The checkout names the PR head SHA explicitly, otherwise it would test main.
Runner: dejima-confined, in a node:22-bookworm container. Node is there for actions/checkout, and Debian's python3 is 3.11, matching the Dockerfile.
Reachability check: an inline step checks the six hosts the job needs and explains how to fix the runner's allowlist if one is unreachable. It's inline because under pull_request_target a script in the repo would be missing from branches that predate it.
Install: CPU-only torch/torchaudio from the PyTorch CPU index first, then pip install -e '.[dev]', which finds torch already satisfied. This avoids several GB of CUDA libraries that no test uses. Neither demucs nor torchcrepe caps the torch version.
Tests: pytest. The e2e test stays skipped because SONGTRACE_E2E is unset.

`.dejima/allowed-domains`

New file listing the hosts the job needs: Debian's two mirrors, PyPI with its file host, and the PyTorch CPU index with download-r2.pytorch.org, which is where that index redirects its downloads. It also applies to dejima dev sandboxes for this project.

Before the first run

This PR can't run its own CI: pull_request_target reads the workflow from main, which doesn't have it yet. The first run will be on the next PR after merge.
The allowlist has to be rendered onto the runner after merge (sudo dejima-ci-allowlist render).
test_web.py fails intermittently until #18 is fixed, so CI will sometimes go red for that reason alone.

Testing

Checked locally that the workflow YAML parses and the inline shell passes bash -n.
The full suite passes locally with CPU torch, demucs and torchcrepe installed, including test_audio_io.py and test_separate.py. The exceptions are the #18 flakes and test_version, which needs the package installed, as CI will do.
The job itself hasn't run anywhere yet.

**Summary generated by LLM agent** Adds a Gitea Actions workflow that runs the full test suite on pull requests. It follows Home Control's `tests.yml`. ## Workflow - **Trigger:** `pull_request_target`, so the workflow file always comes from `main` and a branch can't change what runs or where. The checkout names the PR head SHA explicitly, otherwise it would test `main`. - **Runner:** `dejima-confined`, in a `node:22-bookworm` container. Node is there for `actions/checkout`, and Debian's `python3` is 3.11, matching the Dockerfile. - **Reachability check:** an inline step checks the six hosts the job needs and explains how to fix the runner's allowlist if one is unreachable. It's inline because under `pull_request_target` a script in the repo would be missing from branches that predate it. - **Install:** CPU-only `torch`/`torchaudio` from the PyTorch CPU index first, then `pip install -e '.[dev]'`, which finds torch already satisfied. This avoids several GB of CUDA libraries that no test uses. Neither demucs nor torchcrepe caps the torch version. - **Tests:** `pytest`. The e2e test stays skipped because `SONGTRACE_E2E` is unset. ## `.dejima/allowed-domains` New file listing the hosts the job needs: Debian's two mirrors, PyPI with its file host, and the PyTorch CPU index with `download-r2.pytorch.org`, which is where that index redirects its downloads. It also applies to dejima dev sandboxes for this project. ## Before the first run - This PR can't run its own CI: `pull_request_target` reads the workflow from `main`, which doesn't have it yet. The first run will be on the next PR after merge. - The allowlist has to be rendered onto the runner after merge (`sudo dejima-ci-allowlist render`). - `test_web.py` fails intermittently until #18 is fixed, so CI will sometimes go red for that reason alone. ## Testing - Checked locally that the workflow YAML parses and the inline shell passes `bash -n`. - The full suite passes locally with CPU torch, demucs and torchcrepe installed, including `test_audio_io.py` and `test_separate.py`. The exceptions are the #18 flakes and `test_version`, which needs the package installed, as CI will do. - The job itself hasn't run anywhere yet.

dejima-agent added 1 commit 2026-10-09 18:21:18 -07:00

ci: run the test suite on pull requests 0ec00b597b

Runs on pull_request_target so the workflow always comes from main, on the
confined runner tier. Installs CPU-only torch before the package so pip
does not pull the CUDA build, then runs pytest. .dejima/allowed-domains
declares the hosts the job fetches from.

branderson approved these changes 2026-10-10 19:32:39 -07:00

branderson merged commit e107596e84 into main

2026-10-10 19:32:43 -07:00

branderson referenced this issue from a commit

2026-10-15 14:17:08 -07:00

Merge pull request 'ci: run the test suite on pull requests' (#22) from ci-tests into main

dejima-agent referenced this issue from a commit

2026-10-15 17:44:37 -07:00

Merge pull request 'ci: run the test suite on pull requests' (#22) from ci-tests into main

Sign in to join this conversation.