The 15 November scheduled dry-run wrote x86_64 archives in the aarch64
job because an unset PKG_CONFIG_PATH in the cross env heredoc left
CARGO_BUILD_TARGET unset, and FreeBSD sh did not abort. #242 already
stops the heredoc from expanding that unset variable. This checks that
the env file actually exported the target, and that the build and
package steps still have it, so a similar miss fails before packaging.
* ci: preserve every main-branch Sonar scan
* ci: never cancel a running main Sonar scan, as the app repos do
Per-run concurrency groups let main scans run side by side, so an older
commit's analysis can reach SonarCloud after a newer one. Keep one group
per ref and only stop cancelling on main, matching every app repo's
sonarcloud.yml.
Tests that run a workspace binary through env!("CARGO_BIN_EXE_<name>")
hold the build job's absolute target path. nextest extracts the archive
to a temporary directory, so in every coverage shard those tests failed
with "No such file or directory" while the gate passed them. Since #222
made shard test failures fail the scan, that turned SonarCloud red on
designcraft, cadcraft, pdfcraft and wordcraft main.
The build job now records cargo's target directory next to the archive.
Each shard extracts the archive to a known directory and links the
recorded path to it before running its partition.
* release.yml: a publish input publishes a complete draft at the end of the run
Part E of the faster-release plan: no one waits by the run to publish. Only on the release
branch, not in a dry run, only when no build job failed (the job already requires that) or was
cancelled, and only when the draft has at least as many assets as the last published release, so
a skipped build job can't publish an incomplete release. Off by default; apps opt in.
* publish: compare with the most recently published release, and say why in the summary
gh release list orders by creation, so the newest-created release was the baseline; an older
draft published later would have been missed. Every refusal now also writes its reason to the
job summary, not only an error annotation.
* tests: publish doesn't change the FreeBSD build, so it stays out of the cache key
* fix(ci): provide pinned translations to coverage tests
* fix(ci): keep translation checkout out of scan
---------
Co-authored-by: FredApps <FredApps@users.noreply.github.com>
Closes SonarCloud's path-injection alert on the merged release_train.py (pythonsecurity:S8707):
an absolute path, .., or anything outside the working directory is refused before the folder is
made.
The release FreeBSD jobs keyed their Cargo cache on toJSON(inputs), and
contributors-artifact is contributor-snapshot-<run id>-<attempt>, so every run
wrote an entry no other run could restore: on 11 Nov every release built FreeBSD
cold. The key now holds only the inputs that change the FreeBSD build (app name,
architecture, craft-fonts ref, FreeBSD release); dry-run is left out too, so a dry
run warms the cache a release restores. A test makes every new input pick a side.
The aarch64 job cross-compiles, and Cargo writes that output to
target/aarch64-unknown-freebsd/release, which the cache action never saved: only
the host build scripts in target/release were kept. The action takes the triple as
a new target input and saves and restores that directory too.
Each coverage shard now merges its own raw profiles into one sparse
.profdata with the archive toolchain's llvm-profdata and uploads only
that. The scan job merges the shard profiles instead of downloading
every raw profile beside the test archive, which on 11 Nov 2026 filled
the runner for vectorcraft (four ~7 GB zipped profraw artifacts).
A raw profile llvm-profdata cannot read, such as one cut short by a
test process that crashed while writing it (lightcraft, 11 Nov 2026),
now fails its shard with an error naming the file, instead of being
uploaded and failing the scan's merge.
Single-shard callers keep the cargo llvm-cov path unchanged.
* Release: scripts/cut_release.py cuts a release from green main without a bump pull request
On 11 Nov 2026 the release of 14 apps took most of the day. A large part went to version-bump pull
requests that changed only version strings yet waited on the full CI of a pull request (1-2 h of
FreeBSD, sharded SonarCloud coverage), and to rerunning it. The script commits the bump on the
newest main commit whose CI passed, merges it into release (which starts release.yml), opens the
same bump for main as its own pull request, and with --publish publishes the draft once the run
passed and has assets. Tests cover the bump: workspace and path-dependency versions, only the
workspace's own Cargo.lock entries, CRLF files, the repository hook.
* cut_release: review fixes
- A trailing comment after the workspace version stays; the value alone is replaced.
- Cargo.lock: only the workspace members' entries (cargo metadata --no-deps) change, not a local
path dependency outside the workspace that shares the version.
- Cargo.lock is checked with a resolving cargo metadata --locked, on the bump, on release after
the merge and on the sync branch; --no-deps never compared it with path dependency versions.
- --publish refuses a draft with fewer assets than the last release (a skipped build job leaves
the run green), unless --allow-fewer-assets.
- The sync PR never moves main backwards: versions compare as numbers, and a main at or past the
release gets none.
- A failed auto-merge says so, with the PR to merge by hand.
- Arguments that reach git or gh are checked first (owner/name, version, commit hash, workflow,
title), so none can be read as an option. git, gh and cargo are required up front.
* Release: scripts/release_train.py releases many apps at once with cut_release.py
Every app in parallel, a log per app, and one summary line each (published, pushed, or stopped
and why); one app stopping doesn't stop the rest. Tested on figcraft and soundcraft with
--dry-run (both done in under a minute).
* tests: release_train options before -- in the test
* release_train: only owner/name repos and known cut_release.py options reach the subprocess
* release.yml: a dry-run input builds and checks without releasing
For a nightly run on main that finds release-only breakage before release day. On 11 Nov 2026
gridcraft's riscv64 build, materialcraft's private-repo FreeBSD jobs, wordcraft's Print job and
the riscv64 apt mirrors each surfaced only during the release, one fix PR and release run at a
time. With dry-run: no job enters the release environment (no signing secrets; the environment
admits only the release branch), the macOS job and the draft are skipped, and the Flatpak OCI
layouts are not pushed to GHCR or Docker Hub.
* tests: the Flatpak OCI push and the release environment also stay out of a dry run
During the ubuntu.com outage, cloud-init's apt_configure module (once per
instance, at boot) spent two 10-second DNS timeouts looking for a mirror
and then wrote ubuntu.sources from its template, after the job had
started. On those VMs configure-apt-mirrors had already pointed the
sources at its list: apt-get update fetched every index for that list,
cloud-init put the image's mirror list back, and apt-get install found
no indexes for it, so no Ubuntu package had a candidate (materialcraft's
Flatpak x86_64 job, the FreeBSD tests job on 26.04). A repro on 12
runners failed on 8, every round, and a file watcher caught cloud-init's
write landing mid-update.
wait-for-cloud-init.sh runs `cloud-init status --wait` (at most 300 s,
never failing the job) before configure-apt-mirrors and
install-riscv64-cross's select-mirrors.sh touch anything.
The release builds' Linux riscv64 job read amd64 from archive and
security.ubuntu.com and riscv64 from ports.ubuntu.com only. With no
ubuntu.com name resolving on the runners, apt spent 3m20s on failed
lookups and then found no libc6-dev:riscv64, so every app's release
failed there.
select-mirrors.sh now chooses one pair of mirrors for both architectures:
Ubuntu's own, then dotsrc (it carries both trees), then kernel.org with
TU Chemnitz's ports. It takes the first pair whose amd64 and riscv64
InRelease files answer for every suite with the same Date (Ubuntu signs
one Release for archive and ports), so Multi-Arch: same packages still
match. The image's own Ubuntu sources are disabled in that job: they
only added a second, lagging snapshot and, during the outage, the stall.
The riscv64 job runs Harden-Runner in audit mode, so release.yml needs
no allowlist change.
One run on 11 Nov 2026 found no candidate for tcsh or debsigs right after
an update that fetched every index from kernel.org; the runs before and
after it, under the same ubuntu.com outage, passed. Print apt-cache policy,
the architecture settings, the universe index files and whether the CPU
supports x86-64-v3, so the next failure shows the cause.
On 11 Nov 2026 no ubuntu.com name resolved on the runners for about
half an hour. The images' mirror list (azure.archive, archive and
security.ubuntu.com) sits entirely in that zone, so every retry spent
its five minutes on 20-second lookups of the same dead names and the
SonarCloud archive builds failed with "downloading the packages failed
3 times".
configure-apt-mirrors points the image's Ubuntu sources at
/etc/apt/craft-apt-mirrors.txt: the image's mirrors in the image's
order, then mirrors.edge.kernel.org and mirrors.dotsrc.org (arm:
mirrors.dotsrc.org and ftp.tu-chemnitz.de). Each mirror is checked
first, in parallel; those that do not serve this release's InRelease
move to the end. install-apt-packages runs it before every network
attempt, and the plain apt-get steps (Flatpak, release, packaging lint,
RPM signing, Linux deps, FreeBSD test) run it before apt-get update.
Block-mode Harden-Runner lists allow the new hosts.
* fix(freebsd): give the VM a swap file and cap the ZFS ARC
photocraft's FreeBSD job was SIGKILLed compiling photocraft-ui-egui's lib tests late in the run (release run 42609972735, dispatch 42805955303). Add an 8 GB swap file when the guest has none, cap the ARC at 2 GB when the image uses ZFS, and print disk, swap and memory so the next OOM is diagnosable.
* fix(freebsd): only touch swap and the ARC inside a FreeBSD guest
The workflow tests run the prepare script on Linux under sh -eu, where swapctl doesn't exist. Run the ARC cap, swap file and diagnostics only when uname is FreeBSD, and don't let the diagnostics fail the step.
* fix(freebsd): add the swap file only with room to spare, and remove a partial one
Per review: a dd that runs out of space must not leave a partial /var/swap0 that then starves the build. Require 16 GB free on /var first and delete the file if any step fails.
* fix(freebsd): keep only the ZFS ARC cap; drop the swap-file fallback
Per review: swap on a ZFS-backed file can hang FreeBSD under memory pressure, and today's image already has a swap partition, so the fallback never ran. The verified run passed on the ARC cap alone.
Bound archive and checksum download retries, including refused connections, while retaining HTTPS and SHA-256 validation. Report missing checksum entries using an exact archive filename lookup. Add nine real-curl regression tests for recovery, permanent failures, retry bounds and checksum validation.
The path is resolved and has to stay inside the named directory. The existing file's text is written back as text, so the path-traversal finding on that write is suppressed.
Co-authored-by: Cursor <cursoragent@cursor.com>
A checkout path and a recorded commit have to match a fixed pattern, and the standards path has to be one of the two known files. The per-block check is its own function.
Co-authored-by: Cursor <cursoragent@cursor.com>
The writer and the checker share one normalization. A rewritten digest does not make an edit pass, an intact older copy warns only when main has moved on, and a source that cannot be read is reported as could not check.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add changed-paths.yml: run a job only when its paths change or a label asks
* changed-paths: read paths NUL-separated, keep the reason on one line; README: what the checkout holds
Anyone can commit under an agent email, so a free-form Git name such as
"Private Person <noreply@anthropic.com>" became a published model title,
which the apps would ship in their About windows. Each provider now lists
the words a model title may use. A name with any other word credits the
agent's default model and repeats none of it; versions are bounded too.
Across the co-author trailers in the 13 apps' history, 18,099 keep their
model row and 7 unusual names now fall back to the default. Also pin that
the PR lookup reads the merge commit rather than the PR head.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
effectcraft's 4,000-commit history made 50-commit association queries time
out on GitHub, so batches are smaller and a failed batch is retried once.
GraphQL names bot openers without the [bot] suffix, which turned dependabot
into a person; only User accounts now count as PR openers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GitHub links agent trailers such as noreply@anthropic.com to the claude and
codex accounts, so they appeared as the top human contributors. Agents are now
recognized by exact Git email or GitHub account, never by name, and credited
only in the snapshot's models (About > Models), including commits an agent
authored itself. Bots are no longer listed as people. Model and version are
read from the agent's Git name, covering every provider the apps' history
uses.
Commits merged with a merge commit or rebase never equal a PR's merge commit,
so self-added names lost verification and unlinked commits lost their author.
GitHub's commit-to-PR association now supplies the opener of the one merged PR
containing the commit. An integration PR opened by a writer that carries the
person's own linked commit also verifies their entry.
Against the shipped snapshots of effectcraft, pdfcraft, photocraft and
lightcraft at their recorded heads, no person loses a row, a name or a commit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
v2.87.29 knows nextest only up to 0.9.148, so every gate and coverage run that installs the
pinned cargo-nextest@0.9.150 fell back to cargo-binstall@latest: an unpinned download, then a
binstall of nextest. v2.87.33 is the oldest release whose manifest has 0.9.150; the releases
since v2.87.29 change only tool manifests.
Co-authored-by: FredApps <FredApps@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(install-apt-packages): time-limit and retry mirror downloads
A SonarCloud job in vectorcraft (run 41253493430) sat in this step for 58 minutes: the Azure
Ubuntu mirror timed out, apt fell back to archive.ubuntu.com, then a connection went silent and
apt printed nothing until the run was cancelled. Acquire::http::Timeout doesn't end a connection
that stays open, so the job's 90-minute timeout was the only bound.
apt-get update and a new --download-only pass now each run under timeout(1), 300 s per attempt,
up to 3 attempts. The install itself then works from downloaded files only, so dpkg is never
killed mid-configure. Acquire::https::Timeout is set alongside the http one for the HTTPS fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(install-apt-packages): retry an update that fetched only some indexes
apt-get update warns and exits 0 when an index fails to download, so the retry never ran for
it. APT::Update::Error-Mode=any makes that a failure, which is retried. If indexes are still
missing after the last attempt the step continues with a warning, as it did before: they often
belong to a repository the packages don't need, and a package that needs one fails the download.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: FredApps <FredApps@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Pull requests that touch Windows packaging saved their own windows-arm64-package cache (about
0.7 GB), visible only to that pull request. Every app repository is at the 10 GB cache limit,
so each one evicted a cache main's gate restores. The nightly run on main still saves it.
Co-authored-by: FredApps <FredApps@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>