First minor release. Bundles every Sprint 2 deliverable (T-219 → T-224) plus the Sprint 1 follow-ups merged since v0.0.2. See per-task entries below for details. Highlights:

• Kanban project picker (T-215, T-216, T-217, T-218) — multi-project workspace with ProjectTabBar, persisted DB-backed projects table, drag-and-drop kanban columns.
• Task creation + edit + delete UX (T-219, T-220) — Radix Dialog wizard (3-step), TaskEditDialog with all status / priority fields, ConfirmDeleteDialog with focus-trap hardening.
• Auto-select first project on hydration (T-221) — useEffect boot effect picks the first project when none is persisted, with refetch on switch.
• i18n EN+FR parity test (T-222) — vitest hard-fails any new key added to one locale and not mirrored in the other.
• Settings KV table + persisted current_project_id (T-223) — new domain/settings/ bounded context, V002 migration, IPC commands settings_get / settings_set, App.tsx boot flow restores the user's last selected project across restarts.
• Sprint 2 smoke E2E + persistence-after-restart (T-224) — scripts/smoke-sprint-2.sh boots pnpm tauri dev twice and asserts (1) project + task creation through the wizard, (2) drag-to-Code persists to libsql, (3) the same row + DOM selection both restore on cold boot. Shipped alongside the Sprint 1 smoke under a single workflow matrix with per-shard artifacts.
• Smoke E2E unblocked on GH Actions (PR #119) — tauri_plugin_single_instance gated off under --features pilot to remove the D-Bus session bus dependency that was hanging every CI run; workflow now installs dbus-x11 and wraps the smoke under dbus-run-session -- as belt-and-suspenders.

Fixed

• src-tauri/src/lib.rs + .github/workflows/smoke-e2e.yml — Sprint 2 smoke E2E (T-224) was hanging on every GH Actions run since the merge of #118: tauri-pilot ping responded but tauri-pilot wait --selector timed out at 60s, snapshot artifact was 0 bytes, and the daily app log file was never created. Boot-stage instrumentation (debug/smoke-webview-blank branch, runs 25864797891 / 25865187570 / 25865551513) localised the hang to inside builder.run(generate_context!()) BEFORE the user setup() closure runs — tauri_plugin_single_instance uses zbus on Linux to register a well-known name on the D-Bus session bus and (a) xvfb does not provide a session bus and (b) even with dbus-run-session the plugin races against the bus daemon's name-release bookkeeping when the second consecutive pnpm tauri dev lifecycle starts. Fix: gate the plugin out under cfg(not(feature = "pilot")) so the smoke build skips the D-Bus dependency entirely (production unchanged: every shipped Tauri bundle keeps single_instance because --features pilot is OFF). Added dbus-x11 to the workflow apt list and wrapped the smoke under dbus-run-session -- as belt-and-suspenders against future plugin additions that rediscover a session-bus dependency. Verified by debug-branch run 25866097897: both pnpm tauri dev smoke (sprint-1) and pnpm tauri dev smoke (sprint-2) jobs green end-to-end.

• src/App.tsx + src/features/projects/persist-selected-project-id.ts (+ tests) — drive pnpm exec oxlint . --max-warnings=0 (CI gate) to 0 warnings on the T-223 surface. (1) Convert multi-line // comment blocks to /* ... */ so eslint(capitalized-comments) does not trigger on continuation lines. (2) Extract parsePersistedId from loadPersistedSelectedProjectId so the entrypoint stays under the max-statements: 10 ceiling. (3) Extract resolveBootSelection from App.tsx's boot effect for the same reason. (4) Replace every magic literal in the test file with named *_ID / *_CALL / *_DELAY_MS constants (no-magic-numbers). (5) Reorder imports so Multiple-syntax declarations precede Single-syntax ones and members are alphabetical (sort-imports). (6) Rename T generic to TValue (id-length). (7) Tighten the persist subscription useEffect arrow body so it satisfies arrow-body-style. CI frontend job now matches local oxlint . --max-warnings=0.

• src/features/projects/persist-selected-project-id.ts (+ tests) — address PR #117 bot review for T-223. (1) Serialise concurrent persistSelectedProjectId writes through a module-scoped promise chain so rapid select(a) then select(b) clicks cannot overwrite the persisted id with a stale value when libsql races on per-call connections (codex finding). (2) Switch Number.isInteger → Number.isSafeInteger in loadPersistedSelectedProjectId so an i64 id beyond Number.MAX_SAFE_INTEGER cannot silently round and restore the wrong project (cubic finding). (3) Add afterAll + mockRestore on the console.warn spy so the mock does not leak across vitest worker files (CodeRabbit finding). Test seam resetWriteChainForTesting exposed for beforeEach isolation. New regression cases: serialises concurrent writes in submission order via the module write chain, keeps subsequent writes flowing when an earlier write rejects, returns undefined for ids beyond Number.MAX_SAFE_INTEGER. 14/14 module tests, 333/333 vitest, 0 oxlint errors.

• src/features/kanban/components/KanbanColumn.test.tsx + src/features/kanban/components/KanbanBoard.test.tsx + src/features/kanban/components/__snapshots__/KanbanColumn.test.tsx.snap + src/features/kanban/components/__snapshots__/KanbanBoard.test.tsx.snap — stabilize relative-time snapshots in kanban tests. Both test files render TaskCard whose created_at_label computes "X days/min/h ago" against Date.now(). With hardcoded created_at: "2026-05-12T00:00:00Z" in the in-test buildTask / DEMO_PROJECT builders, the snapshot output drifted by one day every 24 hours of wall-clock time — failing the populated-board snapshot and seven col-<phase>-3-tasks snapshots every morning regardless of code changes. Wire vi.useFakeTimers({ shouldAdvanceTime: true }) + vi.setSystemTime(new Date("2026-05-14T00:00:00Z")) into the beforeEach of both files and vi.useRealTimers() into the afterEach. shouldAdvanceTime: true keeps async ticks (await, userEvent.setup() v14, Radix transition timers) responsive while the wall-clock value is pinned. Regenerate the eight affected snapshots once at the canonical "2 days ago" / "in 2 days" delta so they stay deterministic regardless of when CI runs. The fix unblocks the T-222 PR (#90) push which was failing the pnpm exec vitest run pre-push hook from the day-1 drift.

Added

• T-224 — Sprint 2 smoke E2E (closes #92). New scripts/smoke-sprint-2.sh boots pnpm tauri dev --features pilot twice; lifecycle 1 drives the full UI flow (ProjectTabBar + → create smoke-sprint-2 project pointing at $REPO_ROOT → IPC + sqlite cross-check for the new projects row → Backlog + Add task → TaskCreationWizard title "smoke task" → step-1 → step-2 → step-3 advance → submit → IPC + sqlite cross-check for the new tasks row with phase='Backlog' → tauri-pilot drag from [data-testid="task-card-<id>"] to [data-testid="kanban-column-code"] with an IPC tasks_move fallback armed because @dnd-kit PointerSensor (activationConstraint: { distance: 6 } in src/features/kanban/hooks/use-kanban-dnd.ts:28) listens to pointerdown/move/up while tauri-pilot drag dispatches the HTML5 drag* family — the AC is tasks.phase persistence, not the input vector → sqlite3 -readonly probe SELECT phase FROM tasks WHERE id = '<id>' returns 'Code' → DOM snapshot anchored on kanban-task-<id> confirms card under kanban-column-code-list); shutdown_app() then issues kill -TERM -- "-$DEV_PID" with SIGKILL fallback (process-group discipline copied verbatim from scripts/smoke-e2e.sh:87-107) and unlinks the tauri-pilot socket so lifecycle 2 binds cleanly; lifecycle 2 boots a fresh pnpm tauri dev instance, waits for the Kanban board to mount, re-probes the sqlite phase (still 'Code' — persistence AC), and waits for the same kanban-task-<id> card to reappear inside kanban-column-code-list (proves T-223 persisted current_project_id AND T-221's tasks-list refetch hits the right project). Three PNG screenshots (target/smoke-sprint-2-{pre-drag,post-drag,post-restart}.png) attach to the PR description for visual review. Hardening reuse from PR #66 / T-134: identical setsid + negative-PID kill process-group discipline so unrelated target/debug/forgent or node ...vite processes are never reaped (CodeRabbit #3217663820); identical plugin_socket_dir() helper mirroring tauri-plugin-pilot/src/server/unix.rs:80 so the CLI and the plugin never disagree on whether $XDG_RUNTIME_DIR is a private directory (CodeRabbit #3217860886); per-lifecycle freshness gate (LIFECYCLE_LOG_SIZE_BEFORE = $(wc -c <"$LOG_FILE") + tail -c "+N" slice + process-substitution grep -q "Forgent ready") so each boot proves it actually mounted the bootstrap layer (CodeRabbit #3217663815 + cubic-dev-ai #3217772727); same redacted *_DISPLAY paths for every [smoke] ... echo line; same cargo install tauri-pilot-cli --version "=0.5.1" --locked pin (workflow). DB + WAL/SHM/journal sidecars are wiped at script start (rm -f "$DB_PATH" "$DB_PATH-wal" "$DB_PATH-shm" "$DB_PATH-journal") because Sprint 2 asserts on specific rows (projects.name, tasks.title) — Sprint 1's file-existence + mtime gate is not strong enough. IPC poller (poll_ipc_for_id) walks .. | objects so the four wrap shapes Sprint 1 normalises against ([], {"result":...}, {"data":...}, {"ok":...}) are all accepted without per-shape branching. Workflow update (.github/workflows/smoke-e2e.yml): single job restructured into a strategy.matrix.scenario of two entries (sprint-1 → scripts/smoke-e2e.sh, sprint-2 → scripts/smoke-sprint-2.sh) with fail-fast: false so a sprint-2 regression cannot mask a sprint-1 regression; apt install list extends to sqlite3 + jq (required by smoke-sprint-2.sh's sqlite pr...

Sprint 1 — Hexagonal foundations. First feature release after the v0.0.1-bootstrap scaffold tag. Closes the Sprint 1 task list (T-101 → T-135, 35 tasks, GH issues #1 → #49 closed). Sprint goal met: pnpm tauri dev boots the app with a Sidebar landmark, libsql migration V001 (projects, tasks, runs, phase_events) runs on startup, IPC tasks_list returns [], architecture test enforces forbidden imports in domain/, smoke E2E gate green on main. v0.1 MVP remains the post-Sprint-10 target per PRD §1.4 / ARCHI §22.

Fixed

• scripts/smoke-e2e.sh + .github/workflows/smoke-e2e.yml — three hardening fixes from CodeRabbit's PR #66 review (T-134). (1) The dev server is now launched under setsid so cleanup uses kill -- -$DEV_PID (process group leader) instead of the previous broad pkill -f "target/debug/forgent" / pkill -f "node .*vite" fallbacks that would have reaped any matching user process outside the repo (CodeRabbit #3217663820); the negative-PID signal targets every descendant (pnpm, the tauri-cli wrapper, cargo run, the forgent binary and vite) without false positives. (2) DB and log assertions are now freshness-gated against SMOKE_START_EPOCH (stat -c %Y "$DB_PATH" >= SMOKE_START_EPOCH) and LOG_SIZE_BEFORE (tail -c "+$((LOG_SIZE_BEFORE + 1))" | grep) — a leftover tasks.db from yesterday's green run, or a stale Forgent ready line in the daily-rolled log file, no longer satisfies the gate when bootstrap silently regresses (CodeRabbit #3217663815). (3) cargo install tauri-pilot-cli in the CI workflow is pinned to --version "=0.5.1" --locked to match tauri-plugin-pilot = "0.5.1" in src-tauri/Cargo.toml; running unpinned would silently break the smoke gate the moment 0.6 lands an incompatible ipc / wait / screenshot signature (CodeRabbit #3217663799). Verified locally on 2026年05月11日: [smoke] PASS — Sprint 1 acceptance gate met, including the new mtime ≥ run start and after run start markers in the assertion log.

• src/shared/observability/sentry.ts — drop pilot.* permission-denied promise rejections at the Sentry SDK boundary via Sentry.init({ ignoreErrors: [/pilot\./] }) (T-134 follow-up, fixes RUST-2). The tauri-plugin-pilot JS bridge fires pilot.__callback IPCs whose rejection is unhandled while build.rs rewrites capabilities/pilot.json between dev rebuilds, surfacing as UnhandledRejection: pilot.__callback not allowed in Sentry. The plugin is gated behind the pilot cargo feature and never ships in release bundles, so these events have zero production value — filtering them at SDK init keeps the dashboard signal-to-noise high during smoke-E2E iteration without touching the live capture path for any other unhandled rejection. ignoreErrors was chosen over a beforeSend filter because the latter requires a null return to drop (forbidden by repo-wide oxlint unicorn/no-null), whereas ignoreErrors matches on the event title / message via regex and short-circuits before any Sentry transport.

• src/App.tsx + src/shared/i18n/locales/{en,fr}.json — PR #64 review feedback (CodeRabbit + cubic-dev-ai, both inline). The <main aria-label="Main content"> shipped in T-130 hardcoded an English screen-reader label while every other landmark in the shell (sidebar.landmark, sidebar.navigation, onboarding.landmark) routes through react-i18next — under FR locale the main region therefore announced as "Main content" instead of "Contenu principal", breaking the accessibility contract from CLAUDE.md (i18n is mandatory for end-user UI text). Added common.main_landmark key in en ("Main content") and fr ("Contenu principal"), introduced const translate = useT("common") at the top of App(), and bound aria-label={translate("main_landmark")}. The common namespace was the natural home (shell-level landmark, not feature-scoped) — no new namespace registration needed since common is already in NAMESPACES. Existing tests assert by getByRole("complementary", { name: "Sidebar" }) (sidebar landmark) and don't query the main region by accessible name, so no test changes were needed; the new key inherits coverage from the i18n suite's namespace-resolution check. pnpm exec tsc -b clean. pnpm exec oxlint . 0 warnings / 0 errors. pnpm exec oxfmt clean. pnpm exec vitest run 106/106 still pass.

• src/features/onboarding/components/WelcomeScreen.tsx + WelcomeScreen.test.tsx — PR #63 review feedback (CodeRabbit Major). next() previously used if (!isLast) setStepIndex((current) => current + STEP_DELTA), where isLast came from the render-snapshot closure. On the penultimate step, two rapid clicks both saw isLast = false, both queued current + 1 updaters, and React composed them to push stepIndex past STEPS.length - 1 — STEPS[stepIndex] then resolved to undefined and the next render crashed on translate("step." + undefined + ".title"). Fix introduces a module-level LAST_INDEX = STEPS.length - STEP_DELTA constant, switches the guard from the stale isLast closure to a fresh stepIndex < LAST_INDEX read, and clamps the updater via Math.min(LAST_INDEX, current + STEP_DELTA) so even racing batched updates cannot overshoot. Added regression test clamps stepIndex to the final step when Next is clicked rapidly on the penultimate step advancing to step 6 then triple-clicking through the boundary, asserting data-step="success" and aria-valuenow <= aria-valuemax. pnpm exec vitest run src/features/onboarding 18/18 pass. pnpm exec oxlint src/features/onboarding clean. pnpm exec tsc -b clean.

• src-tauri/src/infrastructure/observability/sentry_init.rs + src/shared/observability/sentry.ts + scripts/no-secrets.sh — PR #60 review feedback (CodeRabbit + cubic-dev-ai). Backend sentry_init::init now eagerly parses SENTRY_DSN via sentry::types::Dsn::from_str and returns None (with an eprintln! diagnostic — tracing is not yet wired) when parsing fails, instead of silently building ClientOptions { dsn: None } and still returning a guard. Previous behaviour masked telemetry misconfiguration: a malformed DSN would compile, return Some(guard), and Sentry would no-op without warning. Frontend initSentry now .trim()s import.meta.env.VITE_SENTRY_DSN so whitespace-only values are treated as unset rather than initialising Sentry with a blank string. scripts/no-secrets.sh filename allowlist tightened from \.example$ (which would have let prod.key.example through) to (^|/)\.env(\.[^/]+)?\.example$, scoping the bypass to the .env.example family only. Added init_returns_none_when_dsn_malformed regression test (5/5 sentry_init tests pass). cargo clippy --lib -- -D warnings clean. pnpm exec tsc -b / oxlint / oxfmt / vitest 61 clean.

Added

• scripts/smoke-e2e.sh + .github/workflows/smoke-e2e.yml + src-tauri/build.rs + src-tauri/Cargo.toml + src-tauri/src/lib.rs + src-tauri/.gitignore — Sprint 1 acceptance gate (T-134, GH issue #48, ARCHI §21 livrable Sprint 1, depends on T-130 / T-120 / T-121 already on main). The smoke flow boots pnpm tauri dev --features pilot, polls tauri-pilot ping (default 180 s local / 300 s in CI), then asserts the five Sprint 1 acceptance criteria in order: the Sidebar landmark (<aside aria-label> from T-128) is mounted in the WebView via tauri-pilot wait --selector "aside[aria-label]" followed by a snapshot grep that accepts either Main navigation (EN), Navigation principale (FR) or the locale-free data-testid="usage-indicator" footer pill (so the assertion does not break when the WebView falls back to the host locale — i18next-browser-languagedetector returns FR on a French Linux session even with EN-only fallback configured), tauri-pilot ipc tasks_list --args '{}' returns the literal [] (also accepts envelope-wrapped {"result":[]} / {"data":[]} / {"ok":[]} to stay compatible with any future tauri-pilot output-format change; the v0.5 --args flag replaced the v0.4 positional '{}' form referenced in CLAUDE.md / T-134.md), <repo>/.claude/forgent/tasks.db exists on disk (libsql sqlite created by AppContainer::build), ~/.forgent/logs/forgent.log.$(date -u +%Y-%m-%d) contains the Forgent ready line emitted by bootstrap() after app.manage, and tauri-pilot screenshot --output target/smoke-screenshot.png (falls back to the positional form for older pilot builds) produces a non-empty PNG. Cleanup traps EXIT/INT/TERM, gives pnpm tauri dev 5 s SIGTERM grace then SIGKILLs the top-level wrapper plus orphaned target/debug/forgent / node ...vite children (the dev-server double-fork was the root cause of port 1420 staying claimed across reruns) and dumps the last 40 lines of target/smoke-logs/tauri-dev.log to stderr on failure. The script exports FORGENT_PROJECT_ROOT="$REPO_ROOT" before launching the dev command: tauri-cli cds into src-tauri/ before invoking cargo run, so without that env var AppConfig::load's current_dir() call resolves to the cargo manifest dir and the DB lands at src-tauri/.claude/forgent/tasks.db instead of the repo root — failing the acceptance criterion silently. TAURI_PILOT_SOCKET="$XDG_RUNTIME_DIR/tauri-pilot-com.forgent.app.sock" is also exported up-front: tauri-pilot defaults to com.pilot.test-app for the socket name (auto-detection only works when exactly one pilot-enabled binary is on the system), so without the override the CLI talks to a stale socket from another project. Three companion infra pieces make the script actually runnable end-to-end:
  1. Cargo.toml gains an optional tauri-plugin-pilot = "0.5.1" dependency from crates.io (added via cargo add --optional tauri-plugin-pilot) and a corresponding [features] entry pilot = ["dep:tauri-plugin-pilot"] plus the auto-generated alias tauri-plugin-pilot = ["pilot"] (kept for tooling that infers the feature name from the dep name). The flag is off by default — release ...