SystemsLab

Appearance

SystemsLab 166.0.0 ​

Added ​

  • (new command) systemslab export workspace saves a whole workspace as a tarball or zip archive: its contexts, notebooks, items and labels.
  • (new command) systemslab import workspace loads an archive from systemslab export workspace as a new workspace. Links between its notebooks, contexts and experiments point at the new copies.
  • The experiment page has an Execution diagram that shows the order each job's steps run in and where jobs wait for each other at barriers. For an experiment that has run, it also shows each job's host, how long each step took, which steps failed, and how long each job waited at each barrier. It warns about barriers that jobs reach in conflicting orders, which would make the experiment hang.
  • systemslab.num() and systemslab.bool() in Jsonnet specs read a parameter as a number or boolean whether it was passed as -p threads=4 or as a string.
  • EC2 hosts can now use any stock machine image with cloud-init. The agent is installed when the instance boots, so you no longer need to build a custom image.
  • Virtual tags can have a setup script that runs on each new EC2 instance before it takes work, e.g. to install benchmark tools. Edit it on the infrastructure admin page.
  • systemslab queue shows each experiment's priority.
  • systemslab mcp has a login tool, so an AI agent can ask you to sign in through the browser without leaving the MCP client.
  • The validate_spec tool in systemslab mcp checks Jsonnet and YAML specs as well as TOML, so any spec you can submit can be checked first.
  • Search understands more filters. by:<user> finds what someone else created, created:>7d or created:2026-10-01..2026-10-31 limits results to a date range, state: takes several states separated by commas, and sort: changes the order. Put - in front of a filter to exclude its matches, e.g. -state:failed. You can also search for an id, or the start of one.
  • The homepage search is easier to use. It suggests filters as you type, has menus for kind, state, workspace, author and sort order, shows how many results of each kind match, and loads more results on request. Use the arrow keys and Enter to open a result, and / to jump to the search box. The homepage opens with your contexts from the last 30 days.
  • Quick Jump in the navbar accepts the start of an id, and anything that is not an id opens a search.
  • The Contexts, Workspaces and Notebooks pages use the same search box as the homepage. Search finds contexts by outcome with status:, e.g. status:failure,error, and pinned contexts with is:pinned. Old links to these pages still work.
  • AI agents using the SystemsLab MCP server can rename a context and edit its description.

Changed ​

  • systemslab mcp now starts even when your login has expired, so the agent can ask you to sign in again instead of the MCP client reporting that the server failed to start.
  • systemslab mcp lists all of the SystemsLab server's tools even before you sign in or while the server can't be reached, so your agent no longer has to be restarted to see them. It keeps trying to reach the server in the background and tells your agent when the server's own tool list arrives.
  • systemslab queue lists experiments in the order they will run: highest priority first, then oldest first. The default columns are now id, name, priority, created-at and jobs.
  • Rezolus dashboards have new ext4 and xfs sections.
  • Agent packages are available from the public SystemsLab repositories, so installing an agent no longer needs repository credentials.
  • The job logs page stays fast for very long logs. Turning on Follow jumps straight to the end of the log instead of waiting for the whole log to load, and the page tells you if the live log connection drops. Each step's exit now shows in the log, e.g. ── Step 2 exited with status 0 ──.
  • AI agents using the SystemsLab MCP server see a short excerpt of each context description instead of the full text, which keeps their context small. They can still ask for the full description.
  • AI agents must give a description when they create a context through the MCP servers, including for sweep and matrix submissions.

Fixed ​

  • CLI warnings are labelled warning: instead of error:.
  • systemslab import context keeps the context's description.
  • -p enabled=TRUE and other uppercase booleans no longer crash the CLI.
  • The MCP list_metrics_series tool filters by its query and respects its limit. It returned every series of every metric.
  • Recording metrics with start-metrics no longer doubles the totals in experiment and job metrics queries.
  • An experiment no longer ends in Error if the server restarts while the experiment is being cancelled.
  • Cancelling an experiment no longer lets one host run two experiments' jobs at the same time.
  • An experiment no longer waits until it times out without starting an EC2 host it needs.
  • SystemsLab no longer starts an extra EC2 instance for an experiment that is already running.