Reasons stacked up:
- AV: unsigned JARs that auto-download binaries + upload files trigger
Windows Defender false-positives more often than Python scripts
invoked by code-signed python.exe.
- Qt UI option: PySide6 opens a path to a real Qt UI (matching Prism's
look) if needed later. JVM Qt bindings are abandoned.
- frazclient already needs Python; inlining as 'import cloud_sync' is
zero overhead vs the launcher always shelling out to java.
Implementation:
- cloud_sync package: cli.py (argparse), creds.py, scope.py,
restic.py (binary discovery + auto-download + sha256 verify),
sync.py (pull/push subprocess restic).
- pyproject.toml with hatchling backend; pip-installable.
- Makefile builds cloud-sync.pyz via python -m zipapp (~53 KB).
- 33 pytest tests, stdlib only on runtime.
- CI workflow runs pytest matrix (3.10/3.11/3.12) + builds pyz.
- DESIGN.md + README.md updated to reflect Python.
E2E verified against local restic-rest-server:
pull empty → push initial → rm -rf local → pull restores → modify+push
creates second snapshot → client forget --prune blocked by --append-only.
Throws away ~565 LOC of Kotlin (and 18 jar tests) committed earlier in
this same session. Net result is ~250 LOC Python + 33 tests = smaller
and more aligned with the rest of the stack.
Compiles to a 3.4 MB fat jar via shadow plugin.
Entrypoint with CLI dispatch (pull/push subcommands) + FlatLaf init.
Real sync logic stubbed; lands in subsequent commits per task plan.
Build via containerized gradle (8.10.2 + jdk21) to sidestep the
Kotlin 2.1 + JDK 26 compiler crash. Documented in README.
CI workflow tags-only release artifact upload prepared (needs
RELEASE_TOKEN secret in repo settings before first tag).
claude-timemachine