Standalone access and Docker operations#

Opening a control application, running an inference server and dispatching an OmniLAB job are distinct actions. An installed application can be opened while its server is stopped; its browser URL becomes available only when that server is serving. Loading a model still requires its existing preflight and authorization.

Open the component you need#

An agent-server container is not evidence that a complete OpenHands browser UI is installed. Consult that application's actual setup before advertising a URL. Do not start an old agent-server container merely to obtain a browser interface. Its network and permission settings differ from LAB's restricted executor.

The personal Strata Control has no qualified live launch action; its Start path currently reports readiness. It requires a separately qualified supervisor, new live approval and passing safety gates. Its frozen configuration is not lazy-loading; starting its server may immediately load weights. A shortcut to port 8080 must not bypass that control. The target Coder shards are already present; see the standalone acceptance guide. Browser availability, successful model inference and OpenHands integration must be reported separately. The revised private control identifies historical stopped sessions, retains logs/configuration access and offers the existing browser only when an owned live listener is reported; it never starts inference from that button.

Independent experimental applications outside this deployment retain their own entry points and models; see research scope.

Understand LAB containers before cleanup#

LAB creates short-lived containers for an agent attempt, protected baseline and candidate verification. Names normally start with lab-; the authoritative ownership label is org.omnilab.owner, and org.omnilab.mode identifies the role. Old build/qualification containers can have other org.omnilab.* labels. Names alone are insufficient evidence that deletion is safe.

Source: DockerSandbox.create/inspect/stop, implementation/verify/collect_then_stop. The qualified transport uses no external network, drops capabilities, runs as a non-root user and enforces memory/CPU/process/storage constraints. Read-only tmpfs artifacts must be collected before stop; an already stopped container cannot recover lost tmpfs evidence by restarting it.

For each cleanup candidate, inspect exact ID, labels, state, mounts, image digest, writable-layer changes and related job/reservation state. Archive inspect data, logs and any persistent unique files before removing an owned stopped container. Retain the image needed by qualified profiles and any image that preserves build results. Keep models, named volumes, bind-mounted evidence and unrelated services.

Do not run blanket docker system prune, volume pruning, or force-remove running containers. A stopped UI with a named volume is not disposable test debris. Record what was removed and retained, then recheck inventory. Docker cleanup does not change lane qualifications or authorize inference startup.

Documentation maintenance and publication#

The documentation index is the navigation entry. Keep implementation claims in Workflow, current provider plans in the roadmap, and dated findings in existing qualification reports. Preserve historical URLs and evidence instead of deleting old reports.

The website is generated from allowlisted Markdown by website/build.mjs. It includes source links, Mermaid diagrams, search and a visible development status. A documentation publication does not merge the development PR, release a package or install a new controller. Run the documentation and public-source checks before publication. Public pages must not include private configs, model artifacts, local paths, job databases or unsanitized operational logs.