HandDetect Quality Workbench

docker compose up -d starts the full local workbench stack on the host IP and ports from .env:

• Workbench: http://10.7.0.4:60050
• MLflow: http://10.7.0.4:60900
• Evidently: http://10.7.0.4:60904
• FiftyOne: http://10.7.0.4:60901
• Label Studio: http://10.7.0.4:60902
• CVAT: http://10.7.0.4:60903
• Rerun: http://10.7.0.4:60905

Use the Workbench UI to trigger a smoke run, inspect artifacts, and replay prior runs. The same smoke flow can be triggered from the repo root with one command:

docker compose exec workbench handdetect run-smoke-experiment --config configs/smoke-experiment.toml

That command runs the adapter pipeline, publishes the review platform artifacts, and prints the suite_id and run_id used by the Workbench run report.

Local Label Studio credentials:

• Email: handdetect@example.local
• Password: handdetect-local
• API token: handdetect-local-token

Local CVAT credentials:

• Username: handdetect
• Password: handdetect-local
• Email: handdetect@example.local

Local BioDock credentials:

• Username: biodock-admin
• Password: biodock-local-admin-password
• Service client: biodock-controller
• Service secret: biodock-local-controller-secret

Developer-only host commands are available after make bootstrap:

make seed
make run CONFIG=configs/smoke-experiment.toml
python -m handdetect.cli.main review open --suite-id <suite_id> --run-id <run_id>

Default local runtime state, including run artifacts and third-party tracking files, is under /tmp/handdetect. Docker service backing stores are under /tmp/handdetect-services. Generated state is pruned automatically; defaults keep 5 suites, cap runtime state at 20 GiB, and preserve 5 GiB free disk. Runtime retention runs automatically with HANDDETECT_KEEP_SUITES, HANDDETECT_MAX_RUNTIME_BYTES, and HANDDETECT_MIN_FREE_BYTES. The dvc status surfaced in run artifacts is DVCLive metrics output only; replay restore authority in this assignment remains content-hash snapshot lineage, even when a replay worktree can execute dvc pull/checkout. CVAT, Rerun, and Datumaro are Docker-managed by default. Run reports link directly to the run-scoped CVAT task/job, Rerun recording, and Datumaro diff manifest when those services are ready. No repo-local or host Python virtualenv is required for normal use; Python dependencies are isolated inside Docker images, with the workbench and Datumaro each using their own container-internal virtualenv.