Add Greptile review config with UW Lab coding conventions

.greptile/config.json

@@ -0,0 +1,123 @@
+{
+  "commentTypes": [
+    "logic",
+    "syntax"
+  ],
+  "instructions": "Review against the conventions in .greptile/rules.md. Formatting, import order, line length, license headers, spelling, and debug statements are already enforced by pre-commit (black/isort/flake8/pyupgrade/codespell) - do not comment on them, except under scripts/ which pre-commit skips. Prioritize, in order: (1) runtime bugs and cross-script CLI/config mismatches, (2) unvectorized, device-mismatched, or non-deterministic tensor code in mdp terms, (3) gym registration and cfg entry points that do not resolve, (4) missing extension version + CHANGELOG bumps, (5) machine-specific paths or committed artifacts. Be concise; include a code suggestion when the fix is local.",
+  "ignorePatterns": [
+    "docs/licenses/**",
+    "docs/source/_static/**",
+    ".github/workflows/license-exceptions.json",
+    "CONTRIBUTORS.md"
+  ],
+  "rules": [
+    {
+      "id": "extension-version-changelog",
+      "rule": "Any change to code under source/<ext>/<ext>/ must bump `version` in source/<ext>/config/extension.toml (semver) and add a matching entry to source/<ext>/docs/CHANGELOG.rst with the same version, a date, and Added/Changed/Deprecated/Removed/Fixed sub-sections written in past tense. Flag PRs that touch an extension's package without both, or where the two versions disagree.",
+      "severity": "high",
+      "scope": [
+        "source/**"
+      ]
+    },
+    {
+      "id": "cli-flag-consistency",
+      "rule": "When a script launches another script (subprocess, os.system, shell), every flag it passes must be defined by the target's argparse parser or Hydra config, and opt-out flags such as --no_video, --headless, or --disable_* must actually change behavior in every code path that reads them. Flag passed-but-undefined flags and flags that are parsed but ignored.",
+      "severity": "high",
+      "scope": [
+        "scripts/**",
+        "scripts_v2/**",
+        "source/**"
+      ]
+    },
+    {
+      "id": "vectorized-mdp-terms",
+      "rule": "Observation, reward, event, termination, curriculum, and command terms under **/mdp/ must be batched over the env dimension: operate on tensors shaped (num_envs, ...) or on the given env_ids, allocate tensors with device=env.device, and never loop in Python over environments, call .item()/.cpu()/.numpy() inside per-step logic, or mix CPU and CUDA tensors. Flag per-env Python loops and host round-trips in these files.",
+      "severity": "high",
+      "scope": [
+        "source/**/mdp/**"
+      ]
+    },
+    {
+      "id": "per-env-sampling-uses-torch",
+      "rule": "Per-environment randomness in mdp terms (anything that differs across env_ids: poses, velocities, physics parameters, noise) must be sampled with torch on env.device (torch.rand/randn/randint, math_utils.sample_uniform) in one batched call, not with Python `random` or `np.random` in a loop. Scalar scene-wide choices shared by all envs (e.g. picking one HDRI or a camera jitter applied to the whole scene) may use `random`, which the env seed also seeds.",
+      "severity": "medium",
+      "scope": [
+        "source/uwlab_tasks/**",
+        "source/uwlab/**"
+      ]
+    },
+    {
+      "id": "task-registration",
+      "rule": "A new environment cfg class in uwlab_tasks must be registered with gym.register in its package __init__.py, and every entry-point string in the registration (env_cfg_entry_point, *_cfg_entry_point) must resolve to an existing module:attribute or agents/ file. Flag registrations whose targets do not exist, duplicate task ids, and env cfgs that are never registered.",
+      "severity": "high",
+      "scope": [
+        "source/uwlab_tasks/**"
+      ]
+    },
+    {
+      "id": "task-package-self-contained",
+      "rule": "A new task lives in its own package under source/uwlab_tasks/uwlab_tasks/<manager_based|direct>/<category>/<task>/ (reference layout: manipulation/omnireset with __init__.py doing gym.register, *_env_cfg.py, mdp/, config/<robot>/ or agents/). By default everything the task needs - mdp terms, recorders, utilities, cfgs - stays inside that package. Changes outside it (the uwlab core extension, uwlab_rl, uwlab_assets, shared uwlab_tasks utilities, or another task's package) are sometimes correct but must be justified: the PR description or code comment should say why the code cannot live in the task package (generic and used by more than one task, a framework bug fix, a new asset). Flag out-of-package changes that carry no such justification, task-specific code added to another task's package (e.g. extending omnireset/mdp for a different task), and behaviour flags added to another task's env cfg instead of a separate env cfg.",
+      "severity": "high",
+      "scope": [
+        "source/**"
+      ]
+    },
+    {
+      "id": "new-task-docs",
+      "rule": "When a PR registers a new task id, it should also list it in docs/source/overview/uw_environments.rst. Work that ships as a publication (own Quick Start, checkpoints, paper) additionally gets docs/source/publications/<name>/index.rst and a bullet under 'Getting Started' in README.md. Mention this once as a note; do not block on it.",
+      "severity": "medium",
+      "scope": [
+        "source/uwlab_tasks/**"
+      ]
+    },
+    {
+      "id": "optional-research-dependencies",
+      "rule": "Research-only or heavy dependencies (e.g. diffusion_policy, robomimic, policy-training frameworks) must not be added to any extension's install_requires or extension.toml [python.pipapi] requirements, so a default UWLab install stays light. Ship them as an extras_require group (pattern: EXTRAS_REQUIRE in source/uwlab_rl/setup.py), a git submodule, or documented install steps on the relevant docs/source/publications/<name>/ page. Nothing under source/ may import them at module import time.",
+      "severity": "high"
+    },
+    {
+      "id": "submodule-policy",
+      "rule": "A new .gitmodules entry must point at a lab-owned repository (github.com/UW-Lab or github.com/WEIRDLabUW, not a personal fork), pin a branch or commit, be optional to initialize (nothing under source/ imports it at module import time), and come with docs saying when to run `git submodule update --init`. If docs previously told users to clone that repo manually (e.g. docs/source/publications/omnireset/distillation.rst), the PR must update them.",
+      "severity": "high"
+    },
+    {
+      "id": "no-machine-specific-paths",
+      "rule": "Committed code must not contain absolute filesystem paths (/home/<user>, /mnt, /data, /scratch), personal usernames, hard-coded conda env names, wandb entities, or local checkpoint and dataset paths. Use CLI arguments, environment variables, Hydra config, or paths relative to the repository or the extension's data directory.",
+      "severity": "high"
+    },
+    {
+      "id": "no-committed-artifacts",
+      "rule": "Do not commit training outputs or generated data: checkpoints (*.pt, *.pth, *.jit, *.onnx), logs, videos, hdf5/npz datasets, USD files, __pycache__, or wandb/tensorboard directories. Binary assets that legitimately belong in uwlab_assets must go through git LFS per .gitattributes. Trained checkpoints and datasets are published to the Hugging Face dataset UW-Lab/uwlab-assets (checkpoints under Policies/) via a fork + PR and linked from the docs, never stored in this repo.",
+      "severity": "medium"
+    },
+    {
+      "id": "type-hints-signature-only",
+      "rule": "Public functions and methods in source/ carry type hints in the signature using PEP 604 syntax (`int | None`, not `Optional[int]`), omit `-> None`, and use Google-style docstrings whose Args entries are `name: description` with no type repeated in the docstring. Flag missing hints on new public APIs and types duplicated in docstrings.",
+      "severity": "medium",
+      "scope": [
+        "source/**"
+      ]
+    },
+    {
+      "id": "type-checking-imports-only-for-cfg",
+      "rule": "Imports live at module top. `typing.TYPE_CHECKING` guarded imports are allowed only to break the cfg <-> implementation cycle (a cfg referencing its class_type/func, or an implementation annotating its cfg). The only other accepted exception is standalone scripts that must import isaaclab/uwlab modules after AppLauncher has started the app. Flag other function-local or TYPE_CHECKING imports.",
+      "severity": "medium",
+      "scope": [
+        "source/**"
+      ]
+    },
+    {
+      "id": "shell-scripts-need-bash",
+      "rule": "Shell scripts that use bash-only syntax (arrays, [[ ]], ${!arr[@]}, (( )), `local`) must start with `#!/usr/bin/env bash`, not `#!/bin/sh`.",
+      "severity": "medium",
+      "scope": [
+        "**/*.sh"
+      ]
+    },
+    {
+      "id": "new-dependencies-need-justification",
+      "rule": "Adding a package to any setup.py install_requires, extension.toml [python.pipapi] requirements, environment.yml, or pyproject.toml needs a stated reason in the PR and a license on the allowlist (MIT, Apache, BSD, ISC, zlib, PSF, BSL, MPL) or an entry in .github/workflows/license-exceptions.json. Prefer implementing with torch, numpy, or existing dependencies. Research-only or heavy dependencies (diffusion_policy, robomimic, policy-training frameworks) must not go in install_requires at all - see optional-research-dependencies.",
+      "severity": "medium"
+    }
+  ]
+}

.greptile/files.json

@@ -0,0 +1,16 @@
+{
+  "files": [
+    {
+      "path": ".github/PULL_REQUEST_TEMPLATE.md",
+      "description": "PR template. The items that matter for review are the pre-commit run and the CHANGELOG + extension.toml version bump; tests are optional and should not be requested."
+    },
+    {
+      "path": ".pre-commit-config.yaml",
+      "description": "Formatting and lint checks that CI already enforces; do not duplicate these in review comments. Note scripts/ is excluded from pre-commit."
+    },
+    {
+      "path": ".gitattributes",
+      "description": "File types that must be stored via git LFS."
+    }
+  ]
+}

.greptile/rules.md

@@ -0,0 +1,192 @@
+# UW Lab review conventions
+
+UW Lab extends [Isaac Lab](https://github.com/isaac-sim/IsaacLab) and follows its coding
+conventions (Google Python style, PEP 8, PEP 484/585 type hints, Google docstrings). This file
+is review context for Greptile. Rules that pre-commit already enforces are listed so the reviewer
+does not repeat them.
+
+## Already enforced by pre-commit (do not comment on these)
+
+`./uwlab.sh -f` runs: black (line length 120), isort (black profile, custom sections in
+`pyproject.toml`), flake8 (+simplify, +return; see `.flake8` for ignored codes), pyupgrade
+(`--py310-plus`), codespell, trailing whitespace / EOF fixes, license header insertion,
+debug-statement check, and a 2 MB file-size cap. Everything under `scripts/` is **excluded**
+from pre-commit, so basic hygiene issues there are fair game.
+
+## Repository layout
+
+Four Omniverse-style extensions live under `source/`, each with the same shape:
+
+```
+source/<ext>/
+├── config/extension.toml   # version (semver) + metadata + pip requirements
+├── docs/CHANGELOG.rst      # one entry per version, must match extension.toml
+├── <ext>/                  # the python package
+├── test/                   # pytest tests
+└── setup.py
+```
+
+- `uwlab` – framework core (envs, mdp terms, sensors, utilities). Depends on `isaaclab`.
+- `uwlab_assets` – robot / object / sensor cfg instances. Large binaries go through git LFS.
+- `uwlab_tasks` – task definitions: `manager_based/` and `direct/`, each task is a package with
+  `__init__.py` (gym registration), `*_env_cfg.py`, `mdp/`, and `agents/` (RL library cfgs).
+- `uwlab_rl` – RL library wrappers and runners.
+
+Standalone entry points: `scripts/` mirrors the Isaac Lab layout (`reinforcement_learning/`,
+`imitation_learning/`, `tools/`, `tutorials/`); `scripts_v2/tools/` holds newer standalone tools.
+
+## Standalone scripts
+
+Any script that touches Isaac Sim must create the `AppLauncher` first and import `isaaclab` /
+`uwlab*` modules only after `simulation_app` exists:
+
+```python
+"""Launch Isaac Sim Simulator first."""
+from isaaclab.app import AppLauncher
+
+app_launcher = AppLauncher(args_cli)
+simulation_app = app_launcher.app
+
+"""Rest everything follows."""
+import torch
+from isaaclab.envs import ManagerBasedRLEnv
+```
+
+Scripts take configuration from argparse / Hydra, never from hard-coded absolute paths.
+When one script shells out to another, the flags it passes must exist in the callee.
+
+## Python file and class structure
+
+Within a file: imports → constants → public functions → public classes → private functions →
+private classes. Within a class: constants → `ClassVar` attributes → `__init__`/`__del__` →
+`__repr__`/`__str__` → properties → public instance/class/static methods → private methods,
+ordered the way a user would call them (`initialize`, `reset`, `update`, `close`).
+
+Private helpers are prefixed with `_`. Imports stay at module top; the only sanctioned
+`typing.TYPE_CHECKING` use is breaking the cfg ↔ implementation cycle (a cfg's `class_type` /
+`func` default referencing the implementation, or an implementation annotating its cfg type).
+
+## Type hints and docstrings
+
+- Type hints in the signature, PEP 604 unions (`torch.Tensor | None`), no `-> None`.
+- Google docstrings. `Args:` entries are `name: description` — the type is never repeated.
+- Physical quantities state SI units and shape: `"""Joint positions [rad], shape (num_envs, num_joints)."""`
+- Docstrings explain *why* and non-obvious design choices, not just *what*.
+
+## Manager-based environments
+
+- Configs are `@configclass` dataclasses (`*Cfg`), usually alongside the implementation.
+- MDP terms are plain functions with the Isaac Lab signature:
+
+  ```python
+  def term(env: ManagerBasedRLEnv, ..., asset_cfg: SceneEntityCfg = SceneEntityCfg("robot")) -> torch.Tensor:
+  ```
+
+  Observation / reward / termination terms return a tensor of shape `(num_envs,)` or
+  `(num_envs, D)`. Event terms receive `env_ids: torch.Tensor | None` and must only touch those
+  environments. Stateful terms subclass `ManagerTermBase`.
+- Everything is batched over the env dimension on `env.device`. No Python loops over envs, no
+  `.item()` / `.cpu()` / numpy round-trips in per-step code, no CPU/CUDA tensor mixing.
+- Per-env randomness (anything that differs across `env_ids`) is sampled with torch on
+  `env.device` in one batched call (`torch.rand(..., device=env.device)`,
+  `math_utils.sample_uniform`). Python `random` is fine for scalar scene-wide choices shared by
+  every env (one HDRI, one camera jitter); the env seed seeds it too. Either way results must be
+  reproducible under `env_cfg.seed`, which `test_environment_determinism.py` checks.
+- Scene entities are resolved through `SceneEntityCfg` (joint/body ids are resolved once, then
+  indexed), not by string lookups every step.
+
+## Task registration
+
+New environments are registered in the task package's `__init__.py`:
+
+```python
+gym.register(
+    id="UW-<Task>-<Robot>-v0",
+    entry_point="isaaclab.envs:ManagerBasedRLEnv",
+    disable_env_checker=True,
+    kwargs={
+        "env_cfg_entry_point": f"{__name__}.<task>_env_cfg:<Task>EnvCfg",
+        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:<Task>PPORunnerCfg",
+    },
+)
+```
+
+Every entry-point string must resolve; `-Play-v0` variants normally reuse the training cfg with
+`num_envs`, randomization, and curriculum reduced. New tasks are picked up by
+`source/uwlab_tasks/test/test_environments.py`, so they must construct and step headless.
+
+## Adding a new task or publication
+
+- One package per task under `source/uwlab_tasks/uwlab_tasks/manager_based/<category>/<task>/`
+  (`direct/` for direct-workflow envs). `manipulation/omnireset/` is the reference layout:
+  `__init__.py` (gym registration), `*_env_cfg.py`, `mdp/` (terms and utils; its `__init__`
+  re-exports `isaaclab.envs.mdp` and `uwlab.envs.mdp`, then the local modules), and
+  `config/<robot>/` or `agents/` for robot-specific cfgs and RL library cfgs.
+- By default everything the task needs - new mdp terms, recorders, helpers, cfgs - lives inside
+  that package. Do not extend another task's package (e.g. add cube logic to `omnireset/mdp/`) and
+  do not add flags to another task's env cfg; create a separate env cfg.
+- Changes outside the task package (`uwlab` core, `uwlab_rl`, `uwlab_assets`, shared `uwlab_tasks`
+  utilities, another task's package) are sometimes the right call, but they must be justified in
+  the PR: why can this not live in the task package? Good reasons are code that is generic and used
+  by more than one task, a framework bug fix, or a new shared asset. If the justification is
+  missing, ask for it rather than rejecting the change.
+- Docs: add the task id to `docs/source/overview/uw_environments.rst`. Publication-tier work also
+  gets `docs/source/publications/<name>/index.rst` (Quick Start first; see `omnireset/index.rst`)
+  and a bullet under **Getting Started** in `README.md`.
+- Checkpoints and datasets go to the Hugging Face dataset `UW-Lab/uwlab-assets` (`Policies/` for
+  checkpoints) via fork + PR, and the docs link to them. They are never committed here.
+- Heavy or research-only dependencies (diffusion_policy, robomimic, ...) are not added to core
+  install requirements, so a default install stays light. Use an `extras_require` group
+  (`EXTRAS_REQUIRE` in `source/uwlab_rl/setup.py` is the pattern), a git submodule, or install
+  steps on the publication's docs page. Nothing under `source/` may import them at module import
+  time.
+- Submodules point at a lab-owned repo (`github.com/UW-Lab`, `github.com/WEIRDLabUW`), are pinned
+  to a branch or commit, are optional to initialize, and are documented (when to run
+  `git submodule update --init`). If docs previously said to clone the repo manually, update them
+  in the same PR.
+
+## Versioning and changelog
+
+Every PR that changes an extension's package bumps that extension's version in
+`config/extension.toml` (patch for fixes, minor for features, major for breaking changes) and adds
+a matching entry at the top of `docs/CHANGELOG.rst`:
+
+```rst
+0.13.9 (2026-08-28)
+~~~~~~~~~~~~~~~~~~~
+
+Added
+^^^^^
+
+* Added :func:`~uwlab_tasks.manager_based.manipulation.omnireset.mdp.cube_reset` event for cube reset states.
+
+Fixed
+^^^^^
+
+* Fixed ``--no_video`` being ignored in :mod:`scripts.ASTEROID.run_asteroid`.
+```
+
+Bullets are past tense, concise, and say why when it is not obvious. Sub-sections are limited to
+Added / Changed / Deprecated / Removed / Fixed. Breaking changes go under Changed, prefixed with
+`**Breaking:**`, with migration guidance.
+
+## Dependencies
+
+Prefer torch, numpy, and what is already installed. A new package must be added to the owning
+extension's `setup.py` and `extension.toml`, have an allowlisted license (MIT / Apache / BSD /
+ISC / zlib / PSF / BSL / MPL — CI fails otherwise unless listed in
+`.github/workflows/license-exceptions.json`), and be justified in the PR description.
+
+## Tests
+
+pytest under `source/<ext>/test/`, run with `./uwlab.sh -t` or
+`./uwlab.sh -p -m pytest <path>`. Tests that need the simulator create the `AppLauncher` at
+module top (headless) and are marked `@pytest.mark.isaacsim_ci`. Tests are **not required** for a
+PR — do not ask for them. Do flag changes that would break the existing suite (e.g. a new task
+that cannot be constructed headless, or a determinism regression).
+
+## Portability
+
+No absolute paths, usernames, hard-coded conda environments, wandb entities, or local
+checkpoint/dataset locations in committed code. Use CLI args, environment variables, Hydra, or
+paths relative to the repo / extension data directory.