---
title: Testing
description: Prove filesystem safety across restarts, platforms, and real data
sidebarTitle: Testing
---

Spacedrive tests behavior at four scales: pure logic, one daemon, several
processes, and real filesystems. A test suite that never touches real storage
cannot support the release claim.

## Commands

~~~bash
cargo test
cargo test <TEST_NAME>
cargo test --lib
cargo test -- --nocapture
cargo clippy
cargo fmt --check
~~~

Frontend changes also run the relevant package type checks and production
build. Public Rust type changes regenerate TypeScript before those checks.

## Test layers

### Unit tests

Keep pure rules beside their implementation:

- identity evidence ordering;
- capture-policy matching;
- content evidence promotion;
- sync conflict rules;
- path and URI parsing;
- permission decisions.

Unit tests should be deterministic and avoid global daemon state.

### Single-daemon integration tests

Use an isolated temporary data directory and real SQLite stores. Cover complete
operation paths, transaction behavior, jobs, events, and restart reconciliation.

Tests inspect public operations and durable state. They do not rely only on log
messages.

### Multi-process tests

Use independent daemon processes for pairing, peer transport, remote operations,
relay behavior, disconnects, and protocol compatibility. Separate processes
catch assumptions hidden by an in-memory harness.

Use a deterministic transport harness for merge algorithms and precise packet
ordering, then keep at least one real-network test for the protocol.

### Filesystem matrix

Run destructive and metadata-sensitive tests on every claimed platform and
filesystem. Include:

- macOS APFS;
- Windows NTFS;
- Linux ext4 and supported network filesystems;
- removable drives;
- case-sensitive and case-insensitive names;
- symlinks and permission denial;
- names that are not valid UTF-8;
- sparse, large, and changing files.

An emulated filesystem is useful for injection but cannot replace native
metadata verification.

## Source ingest regression suite

Every source ingest must prove:

- the daemon data directory is excluded;
- a failed batch salvages valid records and identifies the bad record;
- capture policy matches between the initial walk and watcher;
- final writes commit before job completion;
- interrupted walks restart from zero;
- incomplete evidence cannot sweep unseen records;
- a surprising mass sweep trips the safety fuse;
- volume-relative roots survive remount;
- unreadable content remains represented with a reason;
- reindex preserves assertions.

Compare final record counts with an independent filesystem walk. Sample stored
metadata against platform APIs.

## Content verification suite

Test:

- deterministic candidate identity;
- candidate collisions;
- full-byte confirmation;
- files changing during a read;
- final batch flush;
- permission and I/O failures;
- duplicate grouping across records;
- failure-domain counts across physical storage.

No test may treat a sampled candidate as authorization to delete bytes.

## Job recovery suite

Interrupt each job kind:

- before its first checkpoint;
- during an external write;
- after data commit but before progress update;
- during final output;
- during cancellation.

Restart the daemon and assert that every queued or running row becomes resumed,
failed with a reason, cancelled, or completed. No zombie state is acceptable.

## File operation suite

Copy, move, sync, archive, and delete tests inject:

- destination collision;
- source change after preview;
- short read and short write;
- disk full;
- network loss;
- provider throttling;
- process exit;
- checksum mismatch;
- permission revocation;
- cross-device rename fallback.

Assert both bytes and durable operation evidence. For destructive work, assert
that failed verification leaves the source copy intact.

## Sync and networking suite

Test local and relayed pairing, device revocation, scoped source grants, remote
query completeness, range reads, reconnect watermarks, source authority, and
multi-writer assertions.

Library sync, source sync, and file sync have separate assertions. A passing
metadata sync test does not prove that bytes were copied.

## Extension suite

Run a real WASM extension that:

- requests and is denied a capability;
- registers a file kind and preview;
- checkpoints a job;
- writes each supported model scope;
- migrates and rolls back;
- survives reindex;
- degrades safely after uninstall.

The host must reject database, filesystem, and network access outside the grant.

## Real-data qualification

Release candidates run against datasets large enough to expose queue races,
filesystem churn, and query-plan failures. Keep the origins backed up and
perform independent acceptance checks.

Record:

- source and independent counts;
- pending content work;
- unreadable objects and reasons;
- verification results;
- frozen store checks;
- restart and recovery observations.

Fixtures prove known cases. Real data discovers cases the suite does not yet
know to model.

## Test hygiene

- Use unique temporary directories.
- Never point destructive tests at a home directory or workspace root.
- Keep timeouts bounded.
- Assert structured output before logs.
- Preserve failed test artifacts when they contain useful evidence.
- Avoid sleeps when an event or state query can prove readiness.
- Add every production data-loss bug as a regression test.

See [Jobs](/core/jobs), [Indexing](/core/indexing), and [Releases](/core/releases)
for the contracts these tests enforce.