Application behavior tests #

These tests preserve Sift's behavior across the Bun migration and cover its repaired defects. They do not modify application source files or use the repository's lists.db.

Run #

Use Node.js 24 or newer for the test runner and SQLite fixture inspection, Bun 1.4.4 for the application, and agent-browser with Chromium for browser tests. The orb already provides Node and the browser tools; .agents/setup installs Bun and locked development dependencies.

./tests/run baseline
./tests/run regression
./tests/run all

• baseline checks existing behavior and should exit successfully.
• regression covers defects reproduced before the port. These are ordinary assertions, not skipped or expected-failure tests; they must now pass.
• all runs both groups and must pass. bun run test runs this command too.

Run one area independently with Node's test runner:

node --test tests/database.baseline.test.mjs
node --test tests/protocol.regression.test.mjs
node --test tests/browser.regression.test.mjs

The runner launches Bun, which executes TypeScript without type checking. Check the server and database types separately:

bun run check

Isolation and migration #

Each fixture copies the application's server, database module, and static assets into a temporary directory. Database workers and HTTP servers run there in child processes. Tests close connections, stop the processes, and remove their temporary databases. No live service or application data is required.

Passing database, protocol, and browser tests start with an explicitly created version-4 database. This isolates their behavior from fresh-database migration; it is not evidence that fresh startup works. Migration regressions instead start from an absent database or a historical schema and exercise the migration code.

The v2 fixture comes from the historical migration SQL. The v1 fixture is a reconstruction: no original v1 schema creator survives in available history. Its documented six-character room constraint is an assumption; its real foreign keys reproduce the child-data loss independently of that constraint. Cleanup tests advance only the isolated worker's JavaScript clock, not the machine clock.

HTTP/WebSocket tests use real connections, and browser tests use Chromium against the real server. Database tests call public database operations in separate processes and inspect persisted SQLite data. No production driver or transport is mocked.

Defects reproduced before the port #

Contract	Observed failure before the port
Complete all pending migrations before use	Fresh/v1 stop at v2; v2 stops at v3; fresh reads/writes encounter missing columns
Preserve data during v1 migration	Rooms survive, but existing items and votes disappear
Return the number of expired rooms deleted	Two rooms disappear, but the helper returns zero
Expire empty rooms created after a v3 upgrade	A new room survives forty days because its activity is NULL
Enforce the WebSocket limit in bytes	Oversized UTF-8 input changes the shared title instead of closing with 1009
Navigate and vote using loaded room items	j does not advance selection and 1 does not cast a vote

Multiple tests expose the migration defects separately so one early failure does not hide missing columns or data loss. These are not ten unrelated bug fixes. The pre-port test commit retains the original Deno launcher and failing results. The Bun port changes the launcher, not those assertions. Node remains the independent test runner.

Additional regressions found during the Bun port #

These defects were reproduced against the port before applying their fixes:

Contract	Observed failure
Preserve recent activity during upgrades	Startup cleanup deletes old rooms with recent items or votes
Expire legacy rooms with NULL activity	Already-v4 rooms created by the old insertion code never expire
Filter invalid bulk entries before database writes	A NUL-prefixed item rolls back otherwise valid additions
Escape cancels inline edits	Removing the focused editor triggers blur and persists the canceled draft
Help blocks list shortcuts	Enter votes instead of activating the focused close button
Room changes reset selection	A new room retains the previous room's selected item ID and cannot receive keyboard votes
Leaving discards undo history	Rejoining a room lets undo change a vote from the previous visit
Replacing or disabling an editor never saves its draft	A rerender overwrites a collaborator's edit; disconnect queues an unfinished draft for saving on reconnect
Unrelated updates preserve item drafts, focus, and text selection	Any list update cancels the editor, including another row's vote

session.test.mjs exercises the public room-session interface against real Bun servers and WebSockets. It replays late transport events to verify disposal and socket replacement, checks reconnect/undo behavior, and checks the pending-message limit. It runs with ./tests/run all or node --test tests/session.test.mjs.