Skip to content

Steps

A step is one self-contained procedure: its inputs, the exact commands, when it passes and what it leaves behind. Workflows put steps in order; a step calls tools in the repository and never calls a workflow. Each lives in agents/steps/<name>/README.md.

StepDoesPasses whenUsed in
Autotune settingsrepeats trials until no fix is left, then keeps only cheaper settings that stay healthy.tune.json has final_healthy: true.Add an example, Tune a scene
Benchmark throughputmeasures simulation and application throughput and memory for two configurations (two backends, or two settings) under one timing contract, and tunes Steel in a fixed order.both configurations ran the same workload under the same contract, every plotted point passed the behavior gates and showed no overflow or dropped work, and invalid points are marked.Port from PhysX, Tune a scene, Validate a solver change
Check an exampleruns an example's own test headless with a preset and records the result in presets.toml.shelf check reports works for every preset you changed (status written to presets.toml).Add an example, Tune a scene, Validate a solver change
Check paritycompares two traces of the same task step by step by name, and audits each difference down to a cause.every difference is within the declared tolerance or has a named cause that is fixed or accepted in writing.Debug instability, Port from PhysX
Randomize one factorturns on domain randomization one factor at a time and compares the backends under each.each factor's samples and deltas stay within the tolerance set at nominal parity, or the factor's difference has a named cause.Port from PhysX
Record a baseline tracefreezes a working PhysX project and records a deterministic per-step trace, keyed by body and joint names, to compare everything else against.the frozen project runs from its own launcher, and two recordings with the same seed agree within a repeatability envelope you declared before recording.Port from PhysX
Render and graderenders examples with their Steel presets on one or more GPUs, records stability metrics, and grades each scene ok, suspect, unstable or failed.every planned scene has a grade other than failed, or failed twice and is reported as broken.Add an example, Debug instability, Tune a scene, Validate a solver change
Run a trialruns one example's whole episode with Steel at given settings and lists what is wrong, in fix order.trial.json has status: "ok", no advice.actions and no critical flag, and the check passed.Add an example, Debug instability, Tune a scene
Size capacitiessets Steel's constraint-row capacities so no row is ever dropped, with headroom, from measured demand.every row family reports zero overflowing world-steps and check_constraint_capacity() passes on full episodes at the target world count.Debug instability, Port from PhysX, Tune a scene
Tune one knobchanges a single setting, holding everything else fixed, and keeps it only if every gate still passes.the changed setting passes the same behavior gates as the reference and, for a speed change, is faster beyond noise.Debug instability, Port from PhysX, Tune a scene
Visual reviewlooks at frames across each episode and records a grade and a note per scene the metrics may have misjudged.every rendered scene's contact sheet has been looked at and visual_review.json covers every scene you have an opinion on.Add an example, Debug instability, Tune a scene, Validate a solver change