Record a baseline trace
Step
- Does
- freezes a working PhysX project and records a deterministic per-step trace, keyed by body and joint names, to compare everything else against.
- Used in
- Port from PhysX
- Passes when
- the frozen project runs from its own launcher, and two recordings with the same seed agree within a repeatability envelope you declared before recording.
Inputs. The original project at a commit; its assets, checkpoints or scripted trajectories; its dependency versions. Tools. The project’s own environment and launcher, plus a recorder hooked into its step loop.
- Inventory. Record commits of the application, every nested repository, assets, checkpoints and trajectory data, and hash the data files. Branch names are hints, not pins.
- Freeze. Put the project in its own checkout and environment with its own installer and launcher. Keep
PYTHONPATH, editable installs, generated assets and caches from crossing into any other tree; print the imported module roots at startup (child processes too) and check them. - Record. Fixed seeds; no domain randomization; no disturbances or observation noise. Per step, keyed by semantic body and joint names (never by index; backends order bodies differently), save: observations, commands, actions, joint targets, applied torques, body and joint state, contact and collision metadata, rewards and terminations. Include the reset state and the first observation and action.
- Repeat the same recording and diff the two. Declare the envelope first (exact equality where the baseline is deterministic, otherwise per-quantity tolerances). Differences inside it are the baseline’s own variability and set the floor for every later comparison; a difference outside it is a baseline problem to fix first.
Outputs. The frozen checkout and launcher, the pins and hashes, and the trace files (one per run, with the command that made them).
Source: agents/steps/record-baseline-trace/README.md, the copy agents read.
Steel docs and tutorials are licensed under Apache-2.0. Third-party notices. "Skild", "Skild AI" and the Skild logo are trademarks of Skild AI and are not licensed under Apache-2.0.