Cache

Karva keeps only reusable history on disk: per-test durations for parallel scheduling and the test names used by --last-failed. Worker test selections, live state, and complete results travel over loopback IPC and remain in memory.

The cache lives in .karva_cache under the project root. Coverage runs also write per-worker coverage artifacts there.

Re-running just the failures

--last-failed (or --lf) restricts the run to whichever tests failed in the previous invocation:

1

uv run karva test --last-failed

A typical fix-it-up loop:

1
2
3

uv run karva test                # see the failures
uv run karva test --last-failed  # iterate on just those
uv run karva test                # confirm the full suite passes again

Combine with --watch to keep iterating until they all pass:

1

uv run karva test --watch --last-failed

If the last run had no failures, --last-failed runs nothing.

Disabling the cache

--no-cache disables reading and writing reusable test history for the current run. Tests are scheduled without duration hints and --last-failed becomes a no-op. Coverage artifacts are still written when coverage is enabled.

1

uv run karva test --no-cache

Managing the cache

Two karva cache subcommands manage cache contents directly:

1
2

uv run karva cache prune  # keep only the newest coverage/legacy run directory
uv run karva cache clean  # remove the cache directory entirely

prune preserves reusable history. Reach for clean if the cache gets corrupted or after a cache-format change.