SystemsLab

Appearance

SystemsLab 162.0.0 ​

Added ​

  • systemslab experiment monitor <id> and systemslab context monitor <id> block until the experiment or context finishes, printing a line as each job or experiment completes. They exit 0 only on success, and non-zero on any other terminal state or when the optional --timeout <seconds> elapses.
  • systemslab logs gains --grep, --tail, --step, and --source stdout|stderr|agent filters, matching the MCP get_logs tool, and --job now also accepts a job name.
  • Experiment steps accept a reserved label field for a human-readable display name, shown in the experiment spec viewer. Unlike name, it is never passed to the action, so it is safe on every step type.
  • Hosts now report the EC2 instance behind them: GET /api/v1/host and the GraphQL host fields carry instance_id, instance_type, market, and price_per_hour_usd (absent when the host has no live instance). systemslab host show prints them, systemslab host show --log [N] shows the host's recent machine-log events, and systemslab host list gains opt-in instance-type, market, and price-per-hour columns.
  • A new /admin/infra page shows the EC2 fleet (instances up, warm pool, burn rate, today's spend) and lets an operator approve or deny pending instance requests. The admin hosts page gains instance, market, and price columns, and a drawer showing each host's machine log.
  • Metric charts now draw stretches the recorder never sampled as a dashed overlay, and keep the error bars on the rest of the series instead of dropping them for the whole series.

Changed ​

  • start-metrics now fails the step when rezolus is not installed or its scrape address is unreachable. Previously it reported success and recorded nothing.
  • The MCP server_version tool now reports the SystemsLab instance URL (systemslabUrl) and the authenticated identity along with the version. systemslab mcp also names the instance in its server instructions and in connection and authentication errors.
  • Metrics dashboards load faster, especially for long recordings: plots are fetched in a compact binary format, and zooming, panning, or revisiting a section reuses already-fetched data instead of refetching it.

Removed ​

  • Removed the "agents out of date" banner from the admin UI.

Fixed ​

  • Fixed per-step cgroups (the shell action's cgroup field) having no resource controllers enabled, which made Rezolus and cpu.stat attribute every step's usage to the systemslab-agent.service cgroup. Each job's runner now lives in its own runner leaf cgroup alongside the step cgroups.
  • Fixed KPIs missing for recordings that capture several workloads (e.g. llm-perf + vllm), and fixed KPIs from workloads with overlapping metric names being merged into one value. Affected cached KPIs are cleared by a migration on upgrade and recomputed on next view.
  • Fixed metrics being double-counted when two recordings of the same host overlap in time, such as the automatic whole-job recording and a start-metrics/stop-metrics recording.
  • Recordings from sglang-* workloads and multi-workload recordings now open their metrics dashboard instead of the query explorer, and the experiment page no longer links such artifacts to a nonexistent /artifact/{id}/explore page.
  • Fixed experiments stalling for up to two minutes after a server restart while their work was routed to the previous, now-dead server process.
  • Fixed EC2-launched agents being unable to load their configuration or register: they now start on instances that require IMDSv2, register into the host row their launch created instead of a second one, and are no longer reaped while still booting.
  • Fixed GET /api/v1/host and systemslab host list failing entirely while any host was still provisioning.
  • The admin hosts page now shows a status badge for draining and drained hosts.
  • Fixed systemslab mcp getting stuck returning expected value at line 1 column 1 for every proxied tool after the server restarted; it now re-establishes its session automatically.
  • Opening the root URL of systemslab-viewer now goes straight to the export's context page instead of showing an error.
  • Fixed systemslab-viewer failing to open an export whose metrics recording is in rezolus's .rez format, with all 1 artifact(s) failed to load. This affected exports from agents running rezolus 5.19 or later.
  • An artifact upload larger than the server would accept surfaced a raw HTML error page, with a JSON parse failure stapled to the front. systemslab artifact upload now reports artifact exceeds the maximum upload size instead. (#6124)