SwiftAcervo

Wire a repository's model-dependent integration test(s) to actually run in CI against cached SwiftAcervo models, the standardized way. Use this skill whenever the user wants to fix, enable, add, or standardize integration tests that need a downloaded ML model in GitHub Actions — phrases like "make the integration tests run in CI", "my model tests are disabled in CI", "cache the models in CI", "add the integration test to the pipeline", "standardize how we test real inference", or "the generation/TTS/inference test never runs on CI". It vendors a credential-free CDN primer + cache-key script into .github/scripts/, adds/updates an integration-tests workflow that restores models via actions/cache and primes on miss, wires the TEST_RUNNER_ACERVO_* env so xcodebuild forwards the cache into the xctest runner, and normalizes the test's gate to model-presence (XCTSkipUnless / @Test(.enabled(if:))). Input is the specific integration test (file/target/suite) when it can't be derived from repo context. Works one repo at

Run a queue of `acervo ship` model uploads sequentially — launch one, wait for it to complete and verify, then launch the next, until the list is exhausted. Use this skill whenever the user wants to ship, mirror, publish, or upload multiple models to the intrusive-memory CDN in one go — phrases like "ship these models", "batch upload", "queue up these acervo ships", "publish this list of models", "run acervo on all of these", "process this models.txt", or any time the user hands over more than one HuggingFace `org/repo` id and wants them shipped one after the other. Also triggers on follow-up checks like "how's the queue going", "what's left in the queue", or "is the batch done". Builds on the single-model `acervo-download-ship` skill and reuses its launcher; the queue layer just handles sequencing, per-item finalization, and skip-on-failure progression. Integrates with `/loop` dynamic mode so the agent isn't pinned waiting on hour-long uploads.

Launch a long-running `acervo ship` model download in a detached shell so the agent isn't blocked while multi-gigabyte models stream from HuggingFace to the intrusive-memory CDN. Trigger this skill whenever the user asks to ship, mirror, publish, upload, push, or download a model with acervo — phrases like "acervo ship X", "ship mlx-community/Y", "publish this model to the CDN", "mirror Z to R2", or "kick off the download for org/repo". Also use it to check on or report status for any acervo download already in flight ("how's the download", "is it done yet", "check the acervo log"). The skill captures verbose output to a log file, enforces single-download-at-a-time, and integrates with `/loop` dynamic mode so checks happen automatically without the agent having to wait.