andras/vault-link
andras/vault-link
Watch
1
0
Fork
You've already forked vault-link
 0
Code Issues Pull requests 18 Projects Releases 55 Packages Wiki Activity Actions 36
vault-link/frontend/test-support/README.md
Andras Schmelczer e151208e15 mor
2026-09-27 13:58:54 +01:00

5.7 KiB
Raw Blame History

Shared sync harness support

Test-only code shared by the deterministic and seeded clients.

Suite Coverage
harness.test.ts, hardening.test.ts, oracle-regressions.test.ts Independent content/identity oracles, corruption/duplication mutation probes, exercised faults, exclusive operations, persistence ownership, cleanup and seeds
file-application.test.ts Swaps, file/directory transitions, visible conflict files, live edits, exclusions, and one metadata save per application
recovery.test.ts Interrupt every application filesystem/save boundary, reopen metadata, modify disk, and verify stable ordinary scans; manifest policy and notification identity
engine-regressions.test.ts, sync-audit.test.ts Offline identities, unknown network outcomes, rejection/retry, exclusions, interrupted in-memory pagination, namespace races and lifecycle changes
event-delivery.test.ts, polling.test.ts WebSocket hints with HTTP-only event reads, missed notifications, opt-in polling, shutdown and retry
protocol.test.ts Abrupt worker termination after accepted requests, exact retries, real server SIGKILL, CAS and event ordering
restart-reconciliation.test.ts Interrupted file application followed by user edits, two-client/server convergence and a second restart
server-restore.test.ts History checkpoints, server backup restores and interrupted metadata resets
content-limits.test.ts, protocol-audit.test.ts Content validation and API invariants
real-disk.test.ts Production Node adapter direct writes, no-replace moves, unsafe links, and rename cycles
production-adapters.test.ts Actual CLI processes, raw watchers, Obsidian adapter, complete metadata and restart after SIGKILL

The filesystem interruption tests require complete metadata to remain usable, ordinary scans to observe current files, and a second scan to remain stable. They also carry independent preservation obligations across every interruption: unreplaced/moved bytes must survive exactly once, unrelated durable identities must survive, and edits made after interruption must remain intact. The two-client restart regressions check the explicit USER EDIT and AFTER INTERRUPTION markers exactly once on both replicas through a second restart. An interrupted authorized replacement need not retain the replaced bytes or replay unfinished operations; there is no filesystem journal or hidden recovery archive.

MemoryDisk exposes before:, visible: and durable: boundaries. A process interruption retains visible changes; a simulated power loss rolls unflushed changes back. This provides different partial-application states for testing; it is not an APFS/ext4 writeback model or a production durability requirement. MemoryPersistence atomically replaces complete objects and can reject before or after replacement. Abandoned filesystem sessions are fenced against later writes.

crash-worker.ts runs a complete client in a worker, which the parent terminates at accepted-response or filesystem/persistence checkpoints without graceful shutdown. A new runtime receives the saved disk and metadata. Real-disk tests exercise the production Node adapter. The production integration suite launches built CLI processes and an Obsidian filesystem adapter against a real server. Obsidian unit tests exercise unsaved editor text, selections and mobile storage contracts; they do not launch the Obsidian app. Filesystem path checks are not a sandbox against hostile concurrent changes to ancestor directories.

Scripted and seeded suites run multiple clients against real servers. They check visible bytes, UUID mappings, canonical server state, content heads and event cursors. Polling is disabled except in tests specifically exercising it. Faults must fire; missed injections fail cleanup rather than silently passing. Workload traces, per-worker results and logs are saved by scripts/e2e.sh for reproduction.

Run from the repository root:

E2E_WORKERS=32 E2E_SEED=12 E2E_TIMEOUT_SECONDS=9000 E2E_ITERATIONS=100 scripts/e2e.sh

The command rebuilds peers and both production integrations, runs core and adapter unit tests plus harness/protocol/scripted suites, and then runs seeds 12 through 43. E2E_ARTIFACTS optionally selects the artifact directory. The runner tests also verify subprocess exit codes, log errors, spawn failures, timeouts, workspace dependency consistency and CI artifact wiring.