Tags: getopenscreen/openscreen
Tags
ci(build): do not run the release matrix for internal `v0.0.0-*` tags `v0.0.0-*` is this repository's existing marker for a tag that is not a product version — a place to hang a binary needing a permanent public URL, the way `v0.0.0-stt-models` hosts the 360 MB Whisper model. `build.yml` triggers on `v*` with no further filter, so creating one runs the entire matrix — Windows, macOS x2, Linux, installers, notarisation — to publish an archive it was never asked to build. That has already happened once: there is a `build.yml` run against a `v0.0.0` tag in the history. The reason this matters now is #591: adopting a locally built ONNX Runtime means publishing it under such a tag, and doing that should not cost a full release build every time the pinned version moves. The other five release-triggered workflows do NOT need the same guard, which was checked rather than assumed: `aur-publish`, `bump-nix-package`, `publish-winget` and `update-homebrew-cask` all gate on `!github.event.release.prerelease`, and an internal release is marked as a prerelease — so none of them can push a bogus version to a package manager. `announce-release` has no such gate and would post to Discord; that is noise rather than a wrong publication, and whether an internal release deserves an announcement is a call for whoever owns the channel, so it is left alone here.
refactor(timeline): delete the duplicate answer, route the rest throu… …gh the one mapping `resolveInsertPlacement`'s recording branch was a second answer to "where is this insertion on the timeline", written as a plain shift the clip's own insertions made wrong — and its only caller threw it away. Deleted, along with the placement variant and the two tests that pinned it. `rulerInserts` is the one place that places an insertion. The remaining hand-written shifts now go through `sourceToTimelineSec` / `timelineToSourceSec`: the added-word marks drawn inside a clip, the trim pill's ruler span, and the ruler→source map that authors a trim. Each was wrong past an insertion by exactly the inserted time — a cut authored on the wrong footage, a pill drawn left of what it removes. Left deliberately: `anchoredToRawSpanSec`, which places zoom/annotation/speed pills. Same defect, but threading the ranges through `anchorRegionsWithDerivedMs` reaches fifteen callers across migration paths that have no document in scope. Marked with a `ponytail:` comment naming the ceiling and the upgrade path rather than half-done.
PreviousNext