Three things, all needed before the update loop can be tested.
The public signing key is committed. Verified before trusting it: algorithm
`Ed`, key ID 90E96FEA2F6D9B6A matching its own comment, 32-byte Ed25519 key.
Dev builds now carry a version that RISES. Every build took its version from
Cargo.toml, so each one was 0.1.0 — an installed 0.1.0 would read a manifest
advertising 0.1.0, conclude it was current, and never update. The rolling
channel would have looked broken while working exactly as written. Dev builds
are now 0.1.<ci-run-number>, from one helper shared by both bundle jobs and the
manifest writer, because three separate derivations of "what version is this"
is three chances for the binary and the manifest to disagree.
Plain semver, not a `-dev.N` prerelease: prerelease versions sort BELOW the
release they qualify, so a tagged build would never update to a newer dev one,
and Windows installer metadata wants a numeric X.Y.Z regardless. Bumping the
minor still beats any dev build on the old line — 0.2.0 > 0.1.2932.
The Windows job also gets the signing environment it was missing, so its NSIS
installer is signed too. Without that the manifest would have had a Linux entry
and nothing for the platform actually being tested.
docker-compose.yml is now the production stack, per request: it pulls the
published image instead of building, keeps Postgres OFF the host network, sets
restart policies, health checks and log rotation, and refuses to start without
a POSTGRES_PASSWORD rather than shipping a known one. Volume names are
deliberately unchanged so an existing deployment upgrades in place instead of
silently coming up against an empty database. Development keeps its own
clearly-named file.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
New thoughtsync.dbwait polls the DB (SELECT 1) up to 60×1s before startup,
logging each attempt, so a briefly slow/unready database no longer crash-
loops the container. Wired as `python -m thoughtsync.dbwait &&` ahead of
`alembic upgrade head` in the image CMD and the dev compose command; exits
non-zero after the window so a restart policy can take over. Prod compose
gains restart: unless-stopped as the complementary piece.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
The media/data location is no longer configurable — DATA_DIR is a fixed
constant (/var/thoughtsync) and THOUGHTSYNC_MEDIA_ROOT is removed, so a
mutable path can't drift from where the volume is mounted. media_root() =
DATA_DIR/media. Both compose files drop THOUGHTSYNC_DATA_DIR and mount the
data volume at /var/thoughtsync (was the contradictory /data). conftest
drops the stale MEDIA_ROOT monkeypatch (create_app never reads DATA_DIR).
Net env surface: THOUGHTSYNC_DATABASE_URL (required) + THOUGHTSYNC_SECRET_KEY
(optional break-glass).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
Dev stack with live reload: Postgres + backend (hypercorn --reload, src mounted)
+ Vite dev server (mounted, proxies /api to the api service). vite proxy target
is now env-configurable (VITE_API_TARGET, defaults localhost:5000 for host dev).
README documents both compose files (dev hot-reload vs prod image).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
bvandeusenandClaude Opus 5