plugins/manifest.go:37 declares manifest-schema.json as the input to a real
go:generate directive producing manifest_gen.go, so a schema-only edit skipped
the very check that catches a forgotten regeneration. Verified by adding a
property to the schema and re-running go generate, which changed the output.
All nine commits that ever touched the schema were already caught, but that is
the check working rather than evidence it is unnecessary: an author who forgets
to regenerate cannot reach merged history while the check runs.
This is the only gap of its kind. The repo has two go:generate directives, and
wire's inputs are all Go source.
`grep -q` exits on its first match, which closes the pipe under the still-
writing `printf`. Under `pipefail` the killed writer, not the successful match,
sets the exit status, so a large diff emitted `false` for an area that had
changed. It is a race rather than a buffer threshold: it starts firing at
around 1200 paths, well under the 64 KiB pipe buffer, and it kills the second
`grep` in the build pipeline too. Both pipelines are now here-strings. The
largest diff in this repo's history is 519 files, so this was latent, but a
repo-wide sweep or a mass directory move reaches it.
pipeline.yml and this script are now inputs to the go, js and i18n filters. A
pipeline-only change previously skipped every suite that pipeline.yml defines,
so a broken test command merged green and surfaced on master. Worse, a wrong
gating expression is invisible in every run: `outputs.golang` instead of
`outputs.go` reads as empty on master pushes too, and the suite disappears
silently and permanently. This change adds 63 such expressions. Of 82
historical pipeline-only commits, 38 edited a job that is now gated, and the
measured cost is about 12 extra full runs a year.
The download-link workflow returned before looking for its own previous
comment, so a PR that built binaries and then became docs-only kept advertising
artifacts from a commit that is no longer the head. That branch was unreachable
until this PR gated the build. It now deletes the stale comment. Deleting
rather than rewording is deliberate: the comment is matched by its header, so a
reworded body would either have to keep a header that contradicts it or start
posting duplicates. The lookup is also paginated now; it saw only the first 30
comments.
Rename detection makes `git diff --name-only` report only the destination, so
moving a file out of a gated area dropped the source from every filter. A
rename of ui/src/a.js to docs/a.js.md emitted all four flags as false, skipping
the build for a change that deleted UI source.
--no-renames turns the rename back into a delete plus an add, so both paths
participate. It is also about 8x faster on a large diff, since rename detection
is the expensive part, and it can only ever move a flag from false to true.
tests/ holds only Go test scaffolding: the shared mocks, tests/fixtures, and
tests/navidrome-test.toml, which tests.Init reads for every suite. A change to
any of it can alter a Go test result, and the directory can never collide with
a frontend or docs path.
This deliberately does not cover testdata/ generally. Measured over the full
history, 99 of 103 pull requests touching testdata/ or tests/fixtures/ already
matched the Go filter, and the single missed pull request was a frontend one
that a broader pattern would have made worse by running the whole Go matrix.
Four changes, three of them gaps in the filters added by the previous commit.
Go tests execute db/migrations/*.sql for real: persistence_suite_test.go calls
db.Init, which runs goose against an in-memory database. `make migration-sql`
produces a .sql-only diff, and six such commits exist in history, so those PRs
would have run zero Go tests. Added ^db/migrations/ to the Go filter.
validate-translations.sh reads ui/src/i18n/en.json as its reference, not
resources/i18n/en.json, which does not exist. A commit that only removes an
English key (dd4802c0c is one) would have skipped the only check that reports
the orphaned keys left in all 36 translations.
The coverage-artifact probe was unpaginated. A full Go pipeline run produces
exactly 30 artifacts, the API default page size, and octocov-pr sorts to index
27 because the test jobs finish first. Run 33503006884 already produced 33 and
pushed all three coverage artifacts off page one. Server-side ?name= filtering
has no count ceiling.
The script also now fails closed if the base ref cannot be resolved. The fetch
itself stays non-fatal: actions/checkout already created the ref, so a failed
refresh is harmless, and making it fatal would turn a transient blip into a red
pipeline.
Also adds workflow_dispatch, so a manual run is possible and gets every flag.
A new `changes` job diffs the PR against its base branch and emits four
flags (go, js, i18n, build) that gate the steps of the test, lint and build
jobs. A docs-only PR now runs no Go tests and no 11-platform build.
The flags gate steps rather than jobs on purpose: a job-level skip
propagates through the needs chain (actions/runner#491) and would take the
release jobs down with it on tag pushes. Only upload-packages and the
push-manifest jobs skip at the job level, where nothing depends on them.
Master pushes and tags always get every flag, so a release can never be
built from a partially validated tree.
coverage-on-pr.yml now probes for the coverage artifact before downloading
it, since a PR with no Go changes produces none.
