Recorded runs and live tracking
Follow current and past machine runs, inspect measured curves and results, and connect attempts to your Research Map.
Preview feature. This guide applies to builds with Runs, project-machine mirroring, and shared live observations. These controls may be absent from an older installed build.
Open Runs in the project sidebar to see recorded work independently of chat. Search by run name or identifier and use All, Active, and History beside the search field. Select an attempt to inspect its status, command, logs, tracking, and results.
Ask Kosmo to run a method
Start from an owned local desktop project with a configured compute mirror. The chosen remote Python environment needs a compatible Kosmico SDK. A missing SDK blocks launch; it is not installed silently.
You can focus on the research request: “Run this method on our dataset.” For iterative work, Kosmo is instructed to prepare useful measurements and diagnostic figures alongside the method. Review the actual machine and command before approving execution. Selected required inputs must finish synchronizing; paused mirroring, conflicting files, or excluded required inputs prevent launch. Save editor changes first.
A durable recorded run survives SSH disconnection and closing the desktop. Reopening the local project resumes observation of the same run, without launching the command again. An ordinary foreground SSH command does not offer that guarantee. Host reboot recovery, scheduler reservations, and automatic restart of a job are outside this flow.
Verify that tracking arrived
The launch preflight checks for instrumentation and a tracking plan with named metrics, figure paths, and source files. A later arrival check verifies real observations. Waiting, verified, and missing tracking remain different states; a promised metric is not a measured value.
The overview shows measured progress, live metric curves, and supported PNG/JPEG figures. Inspect curve values with the pointer or keyboard. A theme-aware grid helps read both individual and comparison plots. Existing Trackio instrumentation is reused where available; SDK tracking can continue with a reported fallback when Trackio is unavailable. A public Trackio or Gradio server is not required for the shared previews.
Check that metrics, units, dataset splits, and figures make sense for the method. A non-iterative utility may explain why metric tracking does not apply. Neither an exit code nor a tracking indicator proves a scientific conclusion.
Review results and failed attempts
Logs, receipts, and eligible declared outputs return under results/runs/<runId>/. Failed or cancelled runs retain available logs and partial declared outputs. Execution can finish before collection does. Oversized, unavailable, or conflicting outputs keep an explicit collection reason; inspect it before requesting another run.
Stop run on the connected owner desktop asks for confirmation and contacts the live supervisor. Closing a card or the Runs page does not cancel work. An unreachable machine cannot be reported as successfully stopped.
Compare and connect the evidence
Choose Compare runs beside search, then select two to five compatible attempts. Matching metric names share axes at their recorded steps, with separate lines and legends. Check units, datasets, objectives, and step definitions first. Missing observations are omitted, and no automatic winner is inferred.
Choose Show in Research Map from a run, or open its Runs details from the map inspector. Observed completed and failed attempts stay connected by the same run identifier, even while outputs are still collecting. Interpret whether the hypothesis was supported only after reviewing the evidence.
Collaborate without sharing machine access
Authorized web collaborators can see shared status, recent logs, curves, and plot previews when the local project is cloud-synced and its owner desktop relay is available. Keep that project open with an authenticated owner connection. Closing the project or losing connectivity pauses updates while the remote job can continue. Reconnecting catches up; stale observations do not mean failure.
Commands, logs, and result contents are project-visible, so keep secrets out of them. Web viewing does not grant machine launch or Stop controls. Kosmo and external MCP clients can query live observations and comparisons through the research-graph tools, checking freshness. One monitoring call is not persistent supervision.