The codex-daochi-client-mesh-client-rename integration, committed as that
task staged it: mesh_client replaces daochi_client, account storage and
tests move to the Mesh names, and docs, locales and CI follow, with
existing accounts, encrypted records and local storage kept compatible.
The task stopped before its dependency versions were pushed; the next
commit moves Breeze to the pushed pins it needs.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Photos downloads from Apps instead of shipping in Breeze, practice and
guide voice clips move to Ogg Opus, dependencies update (Ziran, Kryon,
Apps, Arcade, practices), and the 2.1.63 version and store notes are
prepared. The Practice tab note is translated in every locale.
Custom RELEASE_READY_OUTPUT names now map to build/release-ready, and
only build/release-ready-phone keeps a separate purpose, so task and
version names no longer leave a new folder per run. Both share
build/release-ready/home under an exclusive cache lease held through
container exit. The release command uses the same mapping. Phone
installs take a lock, select only an APK this run built and reuse the
default signed path. Nothing prunes existing output.
A run without RELEASE_READY_INCREMENTAL rebuilds outputs from fresh times;
dropping the record keeps a later incremental run from trusting them against
times recorded for another commit.
Generated tracked files such as src/apps/app_codec.zi are rewritten with the
same content during a build. Recording their final times after a pass keeps
the next no-change phone install from rebuilding the app packages and
reconfiguring CMake.
phone-install's container now keeps the times of files the commit did not
change and the Gradle and CMake build directories between runs, and the
Android source generator keeps the times of C files it emits unchanged, so
a small change no longer recompiles the whole app.
RELEASE_READY_OUTPUT=build/release-ready-NAME keeps a second run's build,
home and lock apart, so a debug build from a fixed commit can run while
another check holds build/release-ready.
Each container's session keyring counts against the account's 200 kernel keys, which long-lived sessions fill; then no container starts (crun: Disk quota exceeded).
The first full run showed what the hosted runner has beyond the apt list:
Go for the zi2go tests, Java 17 for the HTTP transport test, D-Bus for the
Lumi UI tests and the libsodium runtime for the Daochi client test. The
image now has them.
The container reads files nobody changed since HEAD live from this
checkout, so another session's edits during a run reached it and failed the
visual gate's source check. The run now compares the checkout before and
after, names the files that changed while the container could read them,
and records no pass then. Only one run uses build/release-ready/ at a time.
Last week a release needed a commit for every phone check and CI repair:
the release commit came first, then fixes, then "Record" commits, and CI
went red after phone checks because it only ran on GitHub.
make release-ready runs CI's checks before pushing: in the Test job's
Ubuntu 24.04 with its apt packages (.github/local-ci/Containerfile), on
HEAD's committed content with the locked packages, through an overlay of
this checkout so nothing reaches the working tree. It covers Secret Scan,
package URLs, the large-file guard, make test, make visual-test and make
package-compat-test, and records a stamp for that content.
make release then, with the development Moto connected, refuses
uncommitted app files and package overrides that are not clean on the
lock, syncs ./update_version.sh output, requires the release-ready stamp
for this content, builds the debug APK for the phone's ABI, runs the phone
check and makes the one "Release Inner Breeze X.Y.Z" commit with the
receipt. After a late fix, running it again moves the unpushed release
commit above the fix (commits are replayed with merge-tree; only refs
move) instead of adding another. It never pushes or tags.