Agentic skills#

COMPASS ships three Claude Code skills that take you from a natural-language prompt to a runnable, deployed policy. Each skill is a focused recipe the agent follows on your behalf.

Pick the right skill#

If you want to…	Use this skill	Lives at
Train a policy, search SAGE-10k scenes, register a new scene, evaluate a checkpoint, submit to OSMO, deploy a checkpoint	`/compass`	umbrella
Diagnose vague COMPASS errors — “training won’t start”, “what’s wrong”	`/compass-doctor`	specialty
Onboard a new robot platform (cfg files, EmbodimentEnvCfgMap registration)	`/compass-newembodiment`	specialty

/compass is the default front door — if you’re not sure, start there and it’ll point you at the matching specialty when your intent is debugging or onboarding a new robot.

How a skill gets invoked#

Inside a Claude Code session at the COMPASS repo root:

/skill compass <natural-language task>
/skill compass-doctor
/skill compass-newembodiment add Galileo robot, USD at ./assets/usd/galileo.usd, like Spot

Or just describe the task in plain English; Claude Code auto-routes when the intent matches the skill’s description. The descriptions are written to be non-overlapping, so a diagnostic-flavored question routes to /compass-doctor even without the slash command.

What each skill does#

`/compass` — umbrella front door#

End-to-end training pipeline. Covers:

Scene search against the SAGE-10k catalogue (scripts/sage10k_search.py).
SAGE → USD conversion via scripts/sage10k_to_usd.py.
Scene registration in compass/rl_env/exts/mobility_es/mobility_es/config/environments.py and EnvSceneAssetCfgMap in run.py.
OMap generation for the new scene (see Auto OMap from USDs).
Training launch via run.py — preview run with one env first, then full-scale.
Evaluation via run.py with eval_config.gin.
OSMO cloud submission via osmo/run_osmo.py.

Auto-OMap is the SAGE-driven training smoothness win: SAGE scenes don’t ship with hand-authored occupancy maps, and the auto OMap from USDs generator closes that gap so every new scene the agent registers comes with a ready-to-use OMap, no manual UI step.

`/compass-doctor` — one-shot diagnostics#

Read-only health check. Bundles a scripts/compass_status.sh helper that runs eight diagnostic checks in parallel and prints a markdown table:

Check	What it verifies
Container	`compass-rl` is up
Activated shell	`./docker/activate` shim is on PATH
GPU	`nvidia-smi` responds
Base ckpt	`./assets/x_mobility.ckpt` exists
USDs	`./assets/usd/` has scene content
Recent log	a training run has produced log output
(–deep) Isaac Sim init	headless `SimulationApp` starts and closes
(–ckpt PATH) Ckpt load	`torch.load(...)` succeeds on the file

The skill reports the table, identifies the root cause, and points you at the specialty skill that owns the fix — but does not auto-fix the issue. Understanding the root cause is more useful than a silent recovery.

`/compass-newembodiment` — onboard a new robot#

Interactive walkthrough of adding a new robot platform. Parses what you already supplied (robot name, USD path, base-class hint) and only asks about what’s missing.

Ask the four onboarding questions (skipping any you already answered).
Read the matching reference (carter for wheeled, spot for quadrupeds, h1 for humanoids).
Show a diff of the proposed changes — new ArticulationCfg block in robots.py, new <robot>_env_cfg.py, edits to run.py.
Apply after your confirmation.
Smoke-test with --num_envs 1 and the GUI on so you can watch the robot spawn.

The handbook page Adding a new embodiment or scene is the canonical written guide; the skill is the operational sibling.

Reading the skill source#

Skills are plain markdown with YAML frontmatter under .claude/skills/. Each SKILL.md documents every sub-flow, the prompts the agent uses, and the verification steps it runs. Treat it as the spec for what the agent will do on your behalf — useful both for understanding what happens and for editing the skill if you want to fork the workflow for your own purposes.