Package layout migration
The current source uses these top-level packages:
| Previous source namespace | Current namespace | Responsibility |
|---|---|---|
examples/ |
benchmarks/ |
Benchmark protocols, native environments, servers and hardware references |
harness/ |
benchkit/ |
Shared transport, case scheduling, records, reports and internal audits |
AgentPolicies/ |
Unchanged | Agent methods, trusted decision loops and model-owned results |
Update imports, deployment scripts and module invocations for new runs:
python -m benchkit.sim_server.cli list
python -m benchmarks.simBenchmarks.RoboDojo.server --help
python -m benchkit.campaigns --help
The installed commands rsi-bench and starvla-eval keep their names. Reinstall
the package after updating the checkout (pip install -e . for development).
There are no forwarding packages at the old import paths. Benchmark identifiers,
wire protocols, native action semantics, scoring and execution budgets do not
change. Central fixed-plan scheduling remains available in benchkit/campaigns/.
Existing results and frozen deployments
Do not rewrite closed results, pinned upstream inventories or frozen source directories to make old paths look current. Their paths, hashes and configuration describe the code that actually ran. Existing deployments continue to use their own frozen packages; deploy a new snapshot for a new run.
New campaign/v3 manifests declare runtime_namespace: benchkit, checked
against their captured source inventory. Unmarked historical v3 campaigns keep
their harness runtime; v1/v2 keep starVLA.evaluation. This selects each
campaign's own frozen implementation, not a forwarding module in today's tree.
Persisted harness_source record keys/directories and the
starvla-rsi.harness-source/* schema identifiers remain unchanged. They identify
an evidence format, not a current Python import. Historical reports can still
mention examples/, harness/ or starVLA/; these are not live compatibility
entrypoints. Native dependency inventories may also contain an upstream package's
own examples/ directory, which must not be renamed.
For repository documentation relocation indexes, compose their recorded
current_path with the root mapping above when browsing today's source tree.
This does not rewrite the original index or reinterpret paths inside a frozen run.
Existing deployments may still contain ignored historical local files at old paths. Prefer organizing retained evidence under the current results namespace with byte-checked relocation and an explicit path/hash receipt, rather than keeping old-root aliases. Source moves do not authorize deleting those records or restarting evaluations.
Why an old directory may still be visible
Git tracks files, not the existence of a directory. Moving tracked code does not
move ignored __pycache__/, local media or historical indexes. An examples/ or
harness/ directory on an existing deployment therefore does not establish that
the old Python entrypoint is supported. Inspect without deleting:
git ls-files -- examples harness
git status --short --ignored -- examples harness
git check-ignore -v -- /path/to/a/specific/local/file
Use current packages for new imports and launches. Move retained videos and indexes only through an explicit evidence-preserving operation: verify the new copy, retain a relocation receipt, and leave original closed-record contents unchanged. Keep anonymous evidence anonymous; do not assign it to a named model without attribution. Disposable caches are distinct from historical evidence and need their own bounded cleanup. Do not recursively remove an old root just because the tracked source has moved.