Local -> server, then push-then-pull as the only ordering the UI can invoke.
LOCAL TOMBSTONES (schema v2). Found while writing push: delete_forever and
remove_label just DROPPED the row, leaving no record it existed. Offline that
means the delete can never be pushed — and the next pull faithfully
resurrects the note from the server. A deletion that undoes itself is about
the worst thing sync can do, so deletes now record into pending_deletes until
the server acknowledges them. merge_labels had the same hole.
merge_labels also moved memberships without marking the affected notes dirty.
A note's label set only reaches the server via the note itself, so a merge
looked done locally and never synced. Now marked before the delete cascades
the rows away.
Result handling, per status:
created/applied -> clear dirty, store the returned sync_revision
noop -> clear dirty, drop the tombstone (a row the server never
saw, created and deleted entirely offline)
kept -> clear dirty WITHOUT touching content. Re-pushing would
lose the same last-write-wins comparison forever; the
following pull adopts the server's version.
rejected -> stay dirty and surface the reason. A duplicate label name
is the realistic case and only a human can resolve it.
The subtle one is `kept` plus a skewed clock. Normally the server's kept
revision sits above our cursor, so the next pull fetches it anyway. If the
clock makes a genuinely later local edit look older, that revision can be
BELOW the cursor — the pull skips it and the stale local copy stays on screen
with nothing marking it wrong. So a kept result at or below the cursor
rewinds the cursor to re-fetch that note. Both directions tested.
label_ids carries MANUAL memberships only. Tag-sourced ones are re-derived
server-side from the body; sending them would convert them into manual
assignments that no longer disappear when the #tag is deleted from the text.
engine::run_cycle is push-then-pull, and a failed push ABORTS before the
pull — pulling anyway would overwrite the exact rows we just failed to save,
turning a recoverable network error into lost work. sync_pull is removed from
the command surface accordingly: offering a bare pull would hand the UI a way
to discard unsent edits. sync_now and sync_has_pending replace it.
Both loops have anti-spin guards: push stops when a batch clears nothing,
pull stops when the cursor doesn't advance.
15 push tests against an in-memory database.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
Two macro-argument splits and a stray blank line, applied verbatim from run
2900's cargo fmt --check diff.
Also reorders the Linux job so `cargo fmt --check` runs AFTER clippy and the
tests. Fail-fast ordering would normally put the cheapest check first, but
there is no Rust toolchain on the workstation, so this lane is verified
entirely in CI — and a formatting nit failing first SKIPS clippy and the
tests, making a whole cycle teach nothing but whitespace. That has now cost
four cycles in this session alone. It still runs before the 20-40 minute
bundle build, so a fmt failure doesn't burn that either.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
Server -> local. sync/wire.rs mirrors the delta-feed JSON exactly as
notes/serialize.py sends it; sync/pull.rs applies it.
ATOMICITY IS THE POINT. The cursor is written in the SAME transaction as the
page it describes. A cursor committed ahead of its data would skip those rows
forever while reporting a clean sync — the worst kind of failure, because
nothing looks wrong. A test forces a mid-page failure and asserts the cursor
stayed put.
Every degradation leans toward re-downloading rather than skipping: an
unparseable cursor means full sync, wire fields are all defaulted so a newer
server adding a field (or an older one omitting one) yields a partial note
instead of a rejected page, and a page that fails rolls back whole.
Labels are applied before notes so a membership never references a row that
doesn't exist. A note also carries enough of its labels to materialize them,
because notes and labels page from ONE shared sequence and a note can arrive
referencing a label whose own delta landed in an earlier page.
via_tag is applied verbatim rather than re-deriving #tags from the body. The
server already reconciled them on save, and re-deriving would go through the
local find-or-create path, which marks new labels dirty — pushing them
straight back. Sync churn manufactured out of nothing.
Duplicate-label merge, the subtle one: a label created offline can collide by
name with one the server already had under a different id. Both sides enforce
one label per name, so the server's row has to win — but simply deleting the
local duplicate would CASCADE its note_labels away, stripping the label off
notes this pull never mentions, with no later page to repair it. So we free
the name, insert the server's row, re-point the memberships, then drop the
husk. Tested.
Children (items/attachments/previews/labels) are replaced wholesale rather
than diffed: a delta carries the note's FULL state, so what arrived IS the
complete set, and diffing could strand a row the server no longer has.
The loop trusts the data over the flag — a server claiming has_more without
advancing its cursor stops with an error instead of spinning forever.
Pull can overwrite a row with unpushed local edits. The documented cycle is
push-then-pull (M10.7c), so that should never happen; when it does it's
counted as clobbered_dirty and logged rather than hidden.
17 tests, all against an in-memory database.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
The pairing step. Nothing else in the sync arc can move until this works.
sync/state.rs owns the link record in the sync_state row M10.4 already put
in the local schema. Two safety properties are the reason it isn't just
three setters:
- Linking a DIFFERENT server resets the change-feed cursor. A cursor is only
meaningful against the server that issued it; carrying one across would
silently skip every change on the new server below that watermark — data
loss wearing the costume of a successful sync. Re-linking the SAME server
(a token refresh) keeps it, so a routine re-auth doesn't force a full
re-download.
- Unlink clears the cursor too, so a later link can't inherit a watermark
from a server that never issued it.
An unparseable or absent cursor reads as 0 (full sync). That direction is
always safe: a redundant re-sync costs time, a too-high cursor costs notes.
Likewise a half-written row (server but no token) reports NOT linked.
state::Status deliberately has no device_token field — it crosses into the
webview, and a long-lived bearer token has no business reachable from page
scripts. A test asserts the token never appears in its serialization.
Token lives in the app-data SQLite file, not an OS keyring: the keyring
crate needs libsecret/DBus on Linux, which adds a C dependency to a binary
that has to cross-compile and fails outright on headless/minimal-WM setups —
the same class of environment assumption behind the black-window bug.
sync_link runs the M10.6 handshake FIRST and refuses an incompatible server
before any credential is sent. Two credential paths, because neither covers
everyone: device-login (a fresh install has no session to mint a token from)
and a pasted token (some users would rather not type a password into a
desktop app). A pasted token is verified against /api/auth/me before being
stored — auth.py's login_required accepts bearer — since an unverified paste
would turn a copy/paste slip into a failure surfacing at the next sync, far
from its cause.
The store lock is taken only after all network work: a std MutexGuard isn't
Send so it cannot cross an await, and holding the store for a round-trip
would freeze every note operation in the UI.
Unlink is LOCAL only — the token stays valid server-side until revoked under
Account -> Linked devices. A pasted token arrives without its device id, so
a reliable remote revoke isn't possible from here; the UI must say so rather
than imply a revoke that didn't happen. Follow-up filed.
No UI yet — that's M10.7e.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
Adds the client's first outbound call: GET {server}/api/config, carrying
X-ThoughtSync-Client and X-ThoughtSync-Protocol, feeding compat::evaluate.
Deliberately its OWN commit. This introduces the first HTTP+TLS stack into a
crate that cross-compiles to Windows from Linux via cargo-xwin — the lane
that has already broken once on a transitive C dependency (libsqlite3-sys
needing llvm-lib). Landing it alone means a failure here has exactly one
possible cause, instead of surfacing mid-way through M10.7's much larger
change where it would be expensive to bisect.
TLS backend is native-tls, NOT rustls, and that is the whole point of the
choice: on x86_64-pc-windows-msvc native-tls resolves to `schannel`, which
is pure-Rust bindings to the OS TLS stack, so nothing C or assembly has to
cross-compile on the fragile lane. rustls would pull in ring/aws-lc-rs and
their assembler. On Linux native-tls uses OpenSSL, whose headers ci-tauri
already ships (libssl-dev, part of Tauri's own Linux prerequisites).
Verified from run 2884's log rather than assumed: tokio and http are already
in the Windows tree via tauri, but no HTTP client and no TLS stack were —
so this genuinely is new surface there, not a no-op.
probe() distinguishes "never got a usable answer" (Err) from "answered, but
we can't work with it" (Ok + verdict). Those need very different messages:
one is "check what you typed", the other is "update something". Transport
errors are translated out of reqwest's Display, which is accurate but reads
like a stack trace.
Still no UI — M10.7 owns the link/settings surface that calls server_probe.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
Version the sync WIRE PROTOCOL separately from either program's release
version, so a self-hosted server and the desktop app can sit on different
releases and still work out whether they can talk.
Each side declares two numbers — what it speaks, and the oldest counterpart
it accepts. Either side can therefore mark a change breaking without the
other shipping in step, which is the whole point: no app↔server lockstep.
Server advertises on the existing public /api/config (a client must be able
to ask "can I talk to you?" before it holds a device token, or even has an
account): sync_protocol_version, min_client_protocol_version, sync_features.
sync_features exists because a version number can only say newer/older. An
ADDITIVE change earns a capability name instead of a minimum bump, so a
newer client meeting an older server drops that one feature and syncs the
rest, rather than refusing. Raising a minimum is reserved for genuinely
breaking changes — it's the switch that hard-blocks the other side.
Client half is pure decision logic (sync/compat.rs), no I/O, so every branch
is unit-testable — there's no live-server lane in CI. Three outcomes: ok /
degraded{unavailable} / incompatible{reason, client_must_update}. The last
names which side can fix it, so the message is actionable. A server that
predates the handshake sends no protocol fields at all; that reads as
"update the server", deliberately not as a parse error, which would look to
the user like they mistyped the URL.
normalize_base_url defaults a bare host to https://, never http:// —
silently downgrading would put a long-lived device token on the wire in
cleartext because someone omitted five characters. Plain HTTP on a trusted
LAN stays supported; the user types http:// and thereby chooses it.
Transport (the actual fetch) lands next, separately: it needs an HTTP/TLS
stack, and that's a real risk to the Windows cross-compile lane, so it gets
its own CI run to bisect against rather than riding along with this.
No UI here by design — the link/settings surface it feeds is M10.7's, per
this task's own sequencing.
Policy documented in docs/sync.md.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
Follow-up to 8a8b2b1, driven by what run 2872's new checks actually printed.
The .deb verification did its job on its first run: tauri already infers
exactly libwebkit2gtk-4.1-0 + libgtk-3-0, so declaring the same two in
tauri.conf.json produced a control file listing each of them twice. Removed
the declaration — verify.sh is the real guard, and it fails the build if
inference ever stops covering what the binary links.
The pacman step revealed ci-tauri carries neither zstd nor bsdtar, so packages
currently ship as .pkg.tar.xz with no .MTREE. Both are working outcomes
(pacman reads xz; only `pacman -Qkk` needs .MTREE), but the docs promised
.zst, so the README, the release notes and ci-requirements.md now describe
what the build actually produces.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
Native packages the installer can actually fetch, before task 2014 wires up
the fetching.
Arch (task 2022, re-scoped): the source PKGBUILD is gone — asking every user
to install rust+node and compile for minutes isn't distribution. Replaced by
desktop/packaging/arch/package-prebuilt.sh, which wraps the binary the Linux
job already built into a .pkg.tar.zst. No second Rust build, no Arch CI image:
the binary bundles nothing and resolves webkit/gtk/soup by soname, identical
on both distros, with SQLite compiled in and glibc used in the safe
built-old/run-new direction. CI is Debian and has no pacman, so the step logs
.PKGINFO plus the full file listing for audit instead of pretending to verify.
Debian (task 2074): install.sh hands the .deb to every Debian/Ubuntu user and
nothing had ever inspected it. tauri.conf.json now declares
libwebkit2gtk-4.1-0 + libgtk-3-0 explicitly rather than trusting inference —
and deliberately declares no appindicator or sqlite dep, since tauri is built
with features=[] and rusqlite is "bundled". desktop/packaging/deb/verify.sh
prints the generated control file, cross-checks it against what the ELF
actually needs via dpkg-shlibdeps, confirms every declared dep exists in apt,
and clean-container installs when a docker CLI is available.
Both artifacts join the run artifact and the tagged release; install.sh grows
a pacman branch so Arch/CachyOS gets a native install instead of the AppImage
fallback. Still no release cut (rule 2).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SreJkbxB4gx8pPsu8QbLPi
There was essentially no logging — useless for proving the app renders across
different environments. Add real observability:
- tauri-plugin-log -> stdout (so `2>&1 | tee` captures a run) AND a persistent
file in the app log dir (grabbable after the fact on any machine). Level Info.
- Startup diagnostics: app version, OS/arch, the Linux display/session stack
(XDG_SESSION_TYPE, desktop, Wayland/X11, GDK_BACKEND), the WebKit render-
hardening vars actually in effect, resolved log + data dirs, DB open/migrate
result, and note/label counts.
- log_event command + a frontend logEvent() helper: boot line (data source +
WebKit user-agent) from main.ts, first-route config/session/destination from
the router guard, and — via the bridge invoke() wrapper — every failed Tauri
command named with its error, so a broken basic function is self-identifying.
Task 2040.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
Both chained `stmt.query_map(...)?.collect()?` as the block's tail expression,
so the prepared statement (a block local) was dropped before the borrow held by
the mapped rows ended (E0597). Bind `let rows` first, matching every other query
in the file. rustc error, so this also unblocks test + the release build.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
The on-device core that makes the desktop app work with no server and no login.
- rusqlite (bundled SQLite, so no system libsqlite dependency to vary across
builds); uuid v4 ids; RFC3339/Date.toISOString-compatible timestamps.
- Schema mirroring the note model: notes, labels, note_labels (with via_tag),
checklist_items, attachments, link_previews, note_revisions, saved_filters,
plus per-row sync_revision/dirty + a sync_state row for the M10.7 engine.
user_version-gated migrations.
- derive.rs: pure [[wiki-link]] + #tag scanners (mirror the frontend inline
rules, no regex dep) with unit tests; #tags re-sync via_tag labels on save,
[[links]] drive backlinks at query time (derived, never stored).
- store.rs: the full repository surface (facet/label/date/text list, create,
PATCH-semantics update, pin/archive/color/kind, checklist items, labels CRUD
+ merge, reminders complete/snooze, reorder, trash/restore/delete, revisions
+ restore, titles/search/backlinks/link-search, saved filters).
- commands.rs: ~38 #[tauri::command]s over a Mutex<Connection> in managed state.
- lib.rs: opens the DB in the platform app-data dir on setup; synthetic offline
config/user so the auth-gated router resolves with no login.
Attachment upload / URL unfurl / import are intentionally deferred (network/file
concerns); adapters/local.ts (M10.5) wires all of the above via invoke.
Task 1993.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
Native-first fix for the AppImage black window: a PKGBUILD that builds from source,
linked against SYSTEM graphics libs, so it renders on the host driver.
- desktop/packaging/arch/{PKGBUILD,thoughtsync.desktop,README.md}: `makepkg -si`
installs /usr/bin/thoughtsync + .desktop + icon; deps webkit2gtk-4.1/gtk3/...;
builds the frontend + `cargo build --release` (no tauri-cli). Uses the host
graphics stack -> avoids EGL_BAD_PARAMETER.
- Commit the icon set (desktop/src-tauri/icons/*.png, un-gitignored) so BOTH
`cargo build` (pacman) and `cargo tauri build` (deb/appimage) work without a
generate step; bundle.icon -> the 4 committed PNGs; drop the `cargo tauri icon`
step from desktop.yml.
- Broaden the WebKit software-render hardening (lib.rs) to ALL Linux (was
AppImage-scoped) so the native build also renders if system WebKit is finicky.
Can't CI-test the PKGBUILD (Arch-only; CI is Debian) -- operator builds locally.
desktop.yml re-verifies the deb+AppImage build with the committed icons.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
WebKitGTK's DMA-BUF/EGL renderer fails to init on many Linux GPU/driver/Wayland
setups -> 'EGL_BAD_PARAMETER' -> black window (known WebKitGTK issue, not app code).
Per Tauri's Linux-graphics guidance, set the software-fallback env vars at startup
before the webview is created, scoped to AppImage launches (native installs keep GPU
accel): __NV_DISABLE_EXPLICIT_SYNC / WEBKIT_DISABLE_DMABUF_RENDERER /
WEBKIT_DISABLE_COMPOSITING_MODE, each only if the user already set it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
The Linux AppImage can now install itself into the applications menu, so it behaves
like an installed app instead of a loose file.
- Rust (desktop/src-tauri/src/integration.rs): integration_status / integrate_desktop
/ unintegrate_desktop commands — detect $APPIMAGE, copy the AppImage to
~/Applications, write ~/.local/share/applications/thoughtsync.desktop + embedded
icon, update-desktop-database. Registered in lib.rs.
- Frontend: withGlobalTauri exposes window.__TAURI__.core.invoke; desktop/bridge.ts
(isDesktop + typed invoke, NO @tauri-apps/api dep -> web bundle unaffected);
DesktopIntegrationPrompt (first-run OOBE, remembered) mounted in App.vue;
AccountView "Desktop app" add/remove control. All desktop-guarded -> no-ops on web.
- desktop.yml: upload the .deb + .AppImage as a run artifact (continue-on-error) so
the build is downloadable for hand-testing.
Verified by CI: ci.yml (vue-tsc) for the frontend, desktop.yml (cargo + tauri build)
for the Rust + AppImage.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
Separate workflow (not ci.yml) so the ~20-40min Rust/AppImage build only runs on
desktop/** changes, not every backend/frontend push. Builds on the ci-tauri:1.97
image: frontend build -> `cargo tauri icon app-icon.png` (from a committed 1024px
source PNG, sidestepping SVG-input questions) -> cargo fmt --check -> clippy -D
warnings -> cargo test -> `cargo tauri build` (deb + AppImage).
APPIMAGE_EXTRACT_AND_RUN=1 for FUSE-less CI containers. Verifies the M10.2 scaffold
end-to-end and is the foundation for the integrated-AppImage work (task 2013).
Also updates ci-requirements.md with the desktop lane.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
New desktop/ Tauri v2 project (Linux-first, cross-platform-ready):
- src-tauri: Cargo.toml (lib + thin main.rs shim), build.rs, lib.rs (Builder
entry point), tauri.conf.json (frontendDist -> ../../frontend/dist, devUrl
:5173, deb+appimage bundles), capabilities/default.json (core:default),
.gitignore.
- The shared Vue 3 frontend is the sibling ../frontend; before-commands cd via
"$(git rev-parse --show-toplevel)/frontend" since frontend and src-tauri are
siblings, not nested.
- Icons generated from frontend/public/icon.svg via `cargo tauri icon` in CI
(M10.8), not committed.
Boots the shared UI in a native window. The local data adapter (M10.3/M10.5)
and CI build verification (M10.8) follow. desktop/** is not yet in the CI paths
filter — added with the desktop lane in M10.8.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRgehjoz7Yv8LkUfADxACm
