[WIP] Add warp environment docs and timer alignment by hujc7 · Pull Request #4995 · isaac-sim/IsaacLab

hujc7 · 2026-03-13T09:26:15Z

Summary

Add warp environment overview doc (warp-environments.rst)
Add stable-to-warp migration guide (warp-env-migration.rst)
Align step timer setup across all 4 env base classes (stable + warp, direct + manager)

Dependencies

Must be merged after (validated against develop at 9720047):

Status

Performance comparison data included in docs.

hujc7 · 2026-03-14T21:46:16Z

@greptileai Review

greptile-apps · 2026-03-14T21:59:07Z

Greptile Summary

This PR adds warp environment documentation (warp-environments.rst, warp-env-migration.rst) and aligns the @Timer / granular sub-step timer setup across all four environment base classes (DirectRLEnv, ManagerBasedRLEnv, DirectRLEnvWarp, ManagerBasedRLEnvWarp). It also introduces the new ManagerBasedEnvWarp and ManagerBasedRLEnvWarp base classes, extends all RL-library wrappers to accept the new warp env types, fixes exit-code propagation in the CLI, and adds a large collection of warp task configurations and parity tests.

Key changes and issues found:

Documentation build will fail: index.rst references training-environments and visualization in its toctree, but neither file exists in the repository at this SHA. Sphinx will emit document not found errors for both entries.
Dead CLI flags in train.py: --step-timer, --reset-timer, and --manager_call_config are added to argparse but never consumed in main(). Because the timer enable constants (DEBUG_TIMER_STEP, DEBUG_TIMERS) are module-level booleans evaluated at import time (before main() runs), these flags cannot influence timer behavior even in principle without setting os.environ before the env modules are imported.
Import ordering in manager_based_rl_env.py: DEBUG_TIMER_STEP / DEBUG_TIMERS constants are placed between two from … import statements, breaking PEP 8 / isort convention.
The timer alignment itself (outer step → DEBUG_TIMERS, inner sub-step → DEBUG_TIMER_STEP) is consistent across all four env variants and is a clear improvement over the previous state.

Confidence Score: 3/5

The runtime code changes are safe, but a documentation build breakage and non-functional CLI flags should be resolved before merging.
The warp env and timer-alignment logic is solid and consistent. However, two concrete issues lower confidence: the missing RST files (training-environments, visualization) will break the Sphinx docs CI, and the three new CLI flags in train.py are wired to nothing at runtime. Neither issue is a simulation-correctness regression, but the docs break is a CI failure and the dead flags are confusing UX.
docs/source/experimental-features/newton-physics-integration/index.rst (missing referenced files) and scripts/reinforcement_learning/rsl_rl/train.py (unused CLI flags) need attention before merging.

Important Files Changed

Filename	Overview
docs/source/experimental-features/newton-physics-integration/index.rst	Adds `warp-environments`, `warp-env-migration`, `training-environments`, and `visualization` to the toctree, but `training-environments.rst` and `visualization.rst` do not exist — Sphinx build will fail.
scripts/reinforcement_learning/rsl_rl/train.py	Adds `--step-timer`, `--reset-timer`, and `--manager_call_config` CLI arguments that are parsed but never consumed in `main()`, so they have no effect at runtime.
source/isaaclab/isaaclab/envs/manager_based_rl_env.py	Adds `@Timer` decorator to `step()` controlled by `DEBUG_TIMERS` env-var flag, but module-level constants are inserted between import statements violating PEP 8.
source/isaaclab/isaaclab/envs/direct_rl_env.py	Adds `@Timer` decorator to `step()` and defines `DEBUG_TIMER_STEP`/`DEBUG_TIMERS` env-var flags for timer alignment; straightforward and correct.
source/isaaclab_experimental/isaaclab_experimental/envs/direct_rl_env_warp.py	Correctly aligns timer flags — outer step timer uses `DEBUG_TIMERS`, inner sub-step timers use `DEBUG_TIMER_STEP`, fixing the previous conflation of the two flags.
source/isaaclab_experimental/isaaclab_experimental/envs/manager_based_rl_env_warp.py	New file: RL-level warp manager env subclassing `ManagerBasedEnvWarp`; timers follow the same `DEBUG_TIMERS`/`DEBUG_TIMER_STEP` convention consistently.
source/isaaclab_experimental/isaaclab_experimental/envs/manager_based_env_warp.py	New file: base warp manager env providing RNG state, pre-allocated boolean masks, and Warp-first manager loading for the experimental pipeline.
source/isaaclab_rl/isaaclab_rl/rsl_rl/vecenv_wrapper.py	Extends allowed env types to include `ManagerBasedEnvWarp` and `ManagerBasedRLEnvWarp` with graceful ImportError fallback; consistent with other wrappers in this PR.
source/isaaclab_rl/isaaclab_rl/rl_games/rl_games.py	Adds warp env types to allowed env type check with graceful ImportError fallback; straightforward and symmetric with the other wrapper changes.
docs/source/experimental-features/newton-physics-integration/warp-environments.rst	New overview doc with environment list, quick-start commands, and a complete performance comparison table; well-structured and accurate.
docs/source/experimental-features/newton-physics-integration/warp-env-migration.rst	New detailed migration guide covering design rationale, term patterns, capture safety rules, and parity testing; thorough and well-written.
source/isaaclab/isaaclab/cli/init.py	Propagates `run_python_command` return code to `sys.exit()`, fixing silent exit-code suppression when using `isaaclab.sh -p`.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    subgraph "Timer Flag Resolution (module load time)"
        ENV_VAR_STEP["os.environ DEBUG_TIMER_STEP"] --> CONST_STEP["DEBUG_TIMER_STEP: bool"]
        ENV_VAR_ALL["os.environ DEBUG_TIMERS"] --> CONST_ALL["DEBUG_TIMERS: bool"]
    end

    subgraph "Stable Envs (direct_rl_env / manager_based_rl_env)"
        CONST_ALL -->|enable=| OUTER_STABLE["@Timer step() — outer wall-clock"]
    end

    subgraph "Warp Envs (direct_rl_env_warp / manager_based_rl_env_warp)"
        CONST_ALL -->|enable=| OUTER_WARP["@Timer step() — outer wall-clock"]
        CONST_STEP -->|enable=| INNER_WARP["with Timer(...) — pre_physics, physics_loop, simulate, scene_update, etc."]
    end

    subgraph "train.py CLI (NOT wired)"
        CLI1["--step-timer (parsed, unused)"]
        CLI2["--reset-timer (parsed, unused)"]
        CLI3["--manager_call_config (parsed, unused)"]
        CLI1 -. no connection .-> CONST_STEP
        CLI2 -. no connection .-> CONST_ALL
    end

    style CLI1 fill:#ffcccc
    style CLI2 fill:#ffcccc
    style CLI3 fill:#ffcccc

Comments Outside Diff (1)

scripts/reinforcement_learning/rsl_rl/train.py, line 68-90 (link)

New CLI flags are parsed but never used

--step-timer, --reset-timer, and --manager_call_config are registered with argparse here, but args_cli.step_timer, args_cli.reset_timer, and args_cli.manager_call_config are never referenced anywhere in the main() function body.

The timer enable/disable mechanism in the env files (direct_rl_env.py, manager_based_rl_env.py, direct_rl_env_warp.py, manager_based_rl_env_warp.py) reads the DEBUG_TIMER_STEP and DEBUG_TIMERS environment variables at module-load time:
```
DEBUG_TIMER_STEP = os.environ.get("DEBUG_TIMER_STEP", "0") == "1"
DEBUG_TIMERS     = os.environ.get("DEBUG_TIMERS", "0") == "1"
```
Because these are module-level constants evaluated once during import (which happens before main() runs), even if main() set os.environ["DEBUG_TIMER_STEP"] = "1" after parsing, it would have no effect on the already-evaluated boolean constants.

To wire the flags end-to-end, os.environ must be set before any env-module import, e.g. at the top of the script right after parsing:
```
if args_cli.step_timer:
    os.environ["DEBUG_TIMER_STEP"] = "1"
if args_cli.reset_timer:
    os.environ["DEBUG_TIMERS"] = "1"
if args_cli.manager_call_config:
    os.environ["MANAGER_CALL_CONFIG"] = args_cli.manager_call_config
```
Until this is done, these three flags are dead and passing --step-timer has no observable effect.

_{Last reviewed commit: de8926b}

greptile-apps · 2026-03-14T21:59:11Z

source/isaaclab_experimental/isaaclab_experimental/managers/observation_manager.py

+    @property
+    def get_IO_descriptors(self, group_names_to_export: list[str] = ["policy"]):


@property with extra parameter is non-functional

get_IO_descriptors is decorated with @property but accepts a group_names_to_export parameter. Python properties do not accept arguments — the parameter is silently ignored and always uses the default value ["policy"]. Callers attempting mgr.get_IO_descriptors(["critic"]) will get TypeError: 'dict' object is not callable.

This should be a regular method instead:

Suggested change

@property

def get_IO_descriptors(self, group_names_to_export: list[str] = ["policy"]):

def get_IO_descriptors(self, group_names_to_export: list[str] = ["policy"]):

Pre-existing in stable — same @property with parameter pattern exists in isaaclab.managers.observation_manager. Not fixing here.

greptile-apps · 2026-03-14T21:59:12Z

source/isaaclab_experimental/isaaclab_experimental/managers/action_manager.py

+        for item in data:
+            name = item.pop("name")
+            formatted_item = {"name": name, "extras": item.pop("extras")}
+            print(item["export"])


Debug print left in production code

This print(item["export"]) appears to be leftover debug output. It will print to stdout on every call to get_IO_descriptors, which is unintended in production. Should be removed or converted to a logger.debug() call.

Suggested change

print(item["export"])

logger.debug(item["export"])

Done. Removed debug print statement.

greptile-apps · 2026-03-14T21:59:13Z

source/isaaclab_experimental/isaaclab_experimental/envs/mdp/events.py

+@wp.kernel
+def _reset_joints_by_scale_kernel(
+    env_mask: wp.array(dtype=wp.bool),
+    joint_ids: wp.array(dtype=wp.int32),
+    rng_state: wp.array(dtype=wp.uint32),
+    default_joint_pos: wp.array(dtype=wp.float32, ndim=2),
+    default_joint_vel: wp.array(dtype=wp.float32, ndim=2),
+    joint_pos: wp.array(dtype=wp.float32, ndim=2),
+    joint_vel: wp.array(dtype=wp.float32, ndim=2),
+    soft_joint_pos_limits: wp.array(dtype=wp.vec2f, ndim=2),
+    soft_joint_vel_limits: wp.array(dtype=wp.float32, ndim=2),
+    pos_lo: float,
+    pos_hi: float,
+    vel_lo: float,
+    vel_hi: float,
+):
+    env_id = wp.tid()
+    if not env_mask[env_id]:
+        return
+
+    state = rng_state[env_id]
+    for joint_i in range(joint_ids.shape[0]):
+        joint_id = joint_ids[joint_i]
+
+        # scale samples in the provided ranges
+        pos_scale = wp.randf(state, pos_lo, pos_hi)
+        vel_scale = wp.randf(state, vel_lo, vel_hi)
+
+        pos = default_joint_pos[env_id, joint_id] * pos_scale
+        vel = default_joint_vel[env_id, joint_id] * vel_scale
+
+        lim = soft_joint_pos_limits[env_id, joint_id]
+        pos = wp.clamp(pos, lim.x, lim.y)
+        vmax = soft_joint_vel_limits[env_id, joint_id]
+        vel = wp.clamp(vel, -vmax, vmax)
+
+        # write into sim
+        joint_pos[env_id, joint_id] = pos
+        joint_vel[env_id, joint_id] = vel
+
+    rng_state[env_id] = state


Duplicate _reset_joints_by_scale_kernel definition

This kernel is defined identically at lines 394-434 and again at lines 540-580. The first definition here is stranded — no function between this point and line 437 (# Reset joints by offset) calls it. The actual caller reset_joints_by_scale() (line 583) uses the second definition at line 540.

This looks like a copy-paste artifact. The first definition should be removed to avoid confusion. While Warp handles duplicate kernel defs (the second shadows the first), this is dead code.

Done. Removed the dead first definition (same as PR #4945).

greptile-apps · 2026-03-14T21:59:14Z

source/isaaclab_experimental/isaaclab_experimental/managers/event_manager.py

+            if inspect.isclass(term_cfg.func):
+                self._mode_class_term_cfgs[term_cfg.mode].append(term_cfg)


Class-based event terms never registered for reset() dispatch

inspect.isclass(term_cfg.func) is checked here, but for non-prestartup modes, _resolve_common_term_cfg → _process_term_cfg_at_play (in manager_base.py:452-454) has already instantiated the class, replacing term_cfg.func with an instance. So inspect.isclass(term_cfg.func) always returns False for non-prestartup class-based terms.

This means class-based event terms for modes like reset or interval will never be added to _mode_class_term_cfgs, and their reset() method will never be called by EventManager.reset().

For prestartup, the class is instantiated at line 432-434 before reaching this check, so those also won't be registered.

Consider checking before instantiation, or using isinstance(term_cfg.func, ManagerTermBase) instead:

if inspect.isclass(term_cfg.func) or isinstance(term_cfg.func, ManagerTermBase): self._mode_class_term_cfgs[term_cfg.mode].append(term_cfg)

False positive — class-based terms are registered in both _mode_term_cfgs and _mode_class_term_cfgs. _apply_reset dispatches via _mode_term_cfgs and reset() calls .reset() via _mode_class_term_cfgs.

greptile-apps · 2026-03-14T21:59:15Z

...tasks_experimental/isaaclab_tasks_experimental/manager_based/classic/humanoid/mdp/rewards.py

+class power_consumption(ManagerTermBase):
+    """Penalty for the power consumed by the actions to the environment."""
+
+    def __init__(self, env: ManagerBasedRLEnv, cfg: RewardTermCfg):


__init__ parameter order inconsistent with base class

The __init__ signature here is (self, env, cfg) but ManagerTermBase.__init__ expects (self, cfg, env). The same issue exists in joint_pos_limits_penalty_ratio.__init__ at line 233.

The manager framework calls term_cfg.func(cfg=term_cfg, env=self._env) using keyword arguments, so this happens to work by keyword matching. However, super().__init__(cfg, env) is never called, so ManagerTermBase attributes like self.num_envs, self.device, and self._env are never set. If any code path accesses these attributes (e.g., self.num_envs or self.device), it will raise AttributeError.

Compare with progress_reward.__init__ at line 159 which correctly uses (self, cfg, env) and calls super().__init__(cfg, env).

Done. Swapped parameter order to (self, cfg, env) to match ManagerTermBase.__init__.

greptile-apps · 2026-03-14T21:59:16Z

source/isaaclab_experimental/isaaclab_experimental/managers/observation_manager.py

+                elif circular_buffer._buffer is None:
+                    # because circular buffer only exits after the simulation steps,
+                    # this guards history buffer from corruption by external calls before simulation start
+                    circular_buffer = CircularBuffer(
+                        max_len=circular_buffer.max_length,
+                        batch_size=circular_buffer.batch_size,
+                        device=circular_buffer.device,
+                    )
+                    circular_buffer.append(wp.to_torch(term_cfg.out_wp))


Circular buffer reassignment lost to local variable

When circular_buffer._buffer is None, a new CircularBuffer is created and assigned to the local variable circular_buffer, but is never written back to self._group_obs_term_history_buffer[group_name][term_name]. On the next call, the stored buffer still has _buffer is None, so a new buffer is recreated every time — the local reassignment is silently lost.

This is currently unreachable because history_length > 0 raises NotImplementedError during _prepare_terms (line 788), but should be fixed before history support is enabled:

elif circular_buffer._buffer is None: circular_buffer = CircularBuffer(...) circular_buffer.append(wp.to_torch(term_cfg.out_wp)) self._group_obs_term_history_buffer[group_name][term_name] = circular_buffer # persist

Done. New CircularBuffer is now written back to self._group_obs_term_history_buffer.

greptile-apps · 2026-03-14T21:59:17Z

source/isaaclab/isaaclab/utils/warp/kernels.py

    # set the torques to the composed torque
    if torques:
-        composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = cast_torque_to_link_frame(
-            torques[tid_env, tid_body],
-            wp.transform_get_rotation(link_poses[env_ids[tid_env], body_ids[tid_body]]),
-            is_global,
+        composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = cast_torque_to_com_frame(
+            torques[tid_env, tid_body], com_pose, is_global
        )
    # set the forces to the composed force, if the positions are provided, adds a torque to the composed torque
    # from the force at that position.
    if forces:
        # set the forces to the composed force
-        composed_forces_b[env_ids[tid_env], body_ids[tid_body]] = cast_force_to_link_frame(
-            forces[tid_env, tid_body],
-            wp.transform_get_rotation(link_poses[env_ids[tid_env], body_ids[tid_body]]),
-            is_global,
+        composed_forces_b[env_ids[tid_env], body_ids[tid_body]] = cast_force_to_com_frame(
+            forces[tid_env, tid_body], com_pose, is_global
        )
        # if there is a position offset, set the torque from the force at that position.
+        # NOTE: this overwrites any explicit torque set above. If both torques and
+        # forces+positions are provided, the correct result should be τ_explicit + r × F
+        # (i.e. += instead of =). Pre-existing behavior — no caller currently passes both.
        if positions:
-            composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = wp.skew(
-                cast_to_link_frame(
-                    positions[tid_env, tid_body],
-                    wp.transform_get_translation(link_poses[env_ids[tid_env], body_ids[tid_body]]),
-                    is_global,
-                )
-            ) @ cast_force_to_link_frame(
-                forces[tid_env, tid_body],
-                wp.transform_get_rotation(link_poses[env_ids[tid_env], body_ids[tid_body]]),
-                is_global,
+            composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = wp.cross(
+                cast_to_com_frame(positions[tid_env, tid_body], com_pose, is_global),
+                cast_force_to_com_frame(forces[tid_env, tid_body], com_pose, is_global),
            )


set_* variants overwrite explicit torque when forces+positions coexist

The NOTE comment on lines 507-509 acknowledges this: if both explicit torques and forces+positions are provided, the composed_torques_b set from explicit torques is overwritten by the cross product r x F on line 513 (= instead of +=). The add_* variants handle this correctly using +=.

While you note that no caller currently passes both, this is a correctness landmine. Consider using += here as well to match the add_* variants and future-proof the API:

Suggested change

# set the torques to the composed torque

if torques:

composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = cast_torque_to_link_frame(

torques[tid_env, tid_body],

wp.transform_get_rotation(link_poses[env_ids[tid_env], body_ids[tid_body]]),

is_global,

composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = cast_torque_to_com_frame(

torques[tid_env, tid_body], com_pose, is_global

)

# set the forces to the composed force, if the positions are provided, adds a torque to the composed torque

# from the force at that position.

if forces:

# set the forces to the composed force

composed_forces_b[env_ids[tid_env], body_ids[tid_body]] = cast_force_to_link_frame(

forces[tid_env, tid_body],

wp.transform_get_rotation(link_poses[env_ids[tid_env], body_ids[tid_body]]),

is_global,

composed_forces_b[env_ids[tid_env], body_ids[tid_body]] = cast_force_to_com_frame(

forces[tid_env, tid_body], com_pose, is_global

)

# if there is a position offset, set the torque from the force at that position.

# NOTE: this overwrites any explicit torque set above. If both torques and

# forces+positions are provided, the correct result should be τ_explicit + r × F

# (i.e. += instead of =). Pre-existing behavior — no caller currently passes both.

if positions:

composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = wp.skew(

cast_to_link_frame(

positions[tid_env, tid_body],

wp.transform_get_translation(link_poses[env_ids[tid_env], body_ids[tid_body]]),

is_global,

)

) @ cast_force_to_link_frame(

forces[tid_env, tid_body],

wp.transform_get_rotation(link_poses[env_ids[tid_env], body_ids[tid_body]]),

is_global,

composed_torques_b[env_ids[tid_env], body_ids[tid_body]] = wp.cross(

cast_to_com_frame(positions[tid_env, tid_body], com_pose, is_global),

cast_force_to_com_frame(forces[tid_env, tid_body], com_pose, is_global),

)

pos_b = cast_to_com_frame(positions[env_id, body_id], com_pose_w, is_global)

force_b = cast_force_to_com_frame(forces[env_id, body_id], com_pose_w, is_global)

composed_torques_b[env_id, body_id] += wp.cross(pos_b, force_b)

Out of scope — this is stable isaaclab code, not modified by this PR. The behavior is documented in the existing NOTE comment.

Add experimental warp-compatible manager implementations, MDP terms, utilities (buffers, modifiers, noise, warp kernels), ManagerCallSwitch for eager/captured dispatch, and manager-based env orchestration. Includes RL library wrapper updates (rsl_rl, rl_games, sb3, skrl) to accept warp env types, and minor stable fixes (settings_manager RuntimeError handling, observation_manager comment cleanup).

Add an experimental manager-based Cartpole environment using the warp manager infrastructure as a reference task for testing and benchmarking.

Add warp-first MDP terms (observations, rewards, events, terminations, actions) for manager-based envs. Update manager infrastructure with ManagerCallSwitch max_mode cap, Scene capture config, body_ids_wp resolution, and capture safety annotations. Add newton state kernels for body-frame computations used by MDP terms.

Fix WrenchComposer to use Tier 1 sim-bind buffers (body_link_pose_w, body_com_pos_b) instead of caching the lazy Tier 2 body_com_pose_w property, which became stale after the first step. Add capture_unsafe decorator for lazy-evaluated derived properties in articulation and rigid object data. Update wrench kernels to compute COM pose inline from link pose and body-frame COM offset.

Add manager-based warp env configs for classic (Ant, Humanoid), locomotion velocity (A1, AnymalB/C/D, Cassie, G1, Go1/2, H1), and manipulation reach (Franka, UR10). Include task-specific MDP terms for humanoid observations/rewards and locomotion rewards/terminations/ curriculums.

Add parity test infrastructure and tests organized by MDP category: observations, rewards, terminations, events, and actions. Tests verify warp-first implementations match stable torch baselines across three execution modes (stable, warp uncaptured, warp CUDA-graph captured). Extract shared fixtures and mock objects to parity_helpers.py.

hujc7 · 2026-03-15T07:40:01Z

@greptileai Review

greptile-apps · 2026-03-15T07:51:33Z

docs/source/experimental-features/newton-physics-integration/index.rst

+  warp-environments
+  warp-env-migration
+  training-environments
+  visualization


Missing RST files referenced in toctree

The toctree in this file references training-environments and visualization, but neither training-environments.rst nor visualization.rst exists in the newton-physics-integration/ directory (confirmed via git ls-tree at the PR's head SHA).

The directory contains only:

installation.rst

warp-environments.rst ✅ (added by this PR)

warp-env-migration.rst ✅ (added by this PR)

limitations-and-known-bugs.rst

solver-transitioning.rst

sim-to-sim.rst

sim-to-real.rst

Sphinx will fail with a WARNING: document not found: training-environments (and visualization) error during the documentation build, which typically promotes to a build error in CI. Either create the missing stub files or remove these entries from the toctree until the pages are ready.

github-actions bot added documentation Improvements or additions to documentation isaac-lab Related to Isaac Lab team labels Mar 13, 2026

hujc7 force-pushed the jichuanh/warp-env-docs branch 5 times, most recently from 57f5fed to 0a3a02b Compare March 14, 2026 21:45

greptile-apps bot reviewed Mar 14, 2026

View reviewed changes

hujc7 added 6 commits March 15, 2026 00:34

Add warp Cartpole task configuration

5108ca8

Add an experimental manager-based Cartpole environment using the warp manager infrastructure as a reference task for testing and benchmarking.

hujc7 force-pushed the jichuanh/warp-env-docs branch from 0a3a02b to de8926b Compare March 15, 2026 07:38

greptile-apps bot reviewed Mar 15, 2026

View reviewed changes

Add warp environment docs and align step timer across all env variants

7ee1f43

hujc7 force-pushed the jichuanh/warp-env-docs branch from de8926b to 7ee1f43 Compare March 15, 2026 08:58

hujc7 marked this pull request as ready for review March 16, 2026 17:34

hujc7 requested review from ClemensSchwarke, Mayankm96, Toni-SM, jtigue-bdai, kellyguo11, ooctipus and pascal-roth as code owners March 16, 2026 17:34

		@property
		def get_IO_descriptors(self, group_names_to_export: list[str] = ["policy"]):

		if inspect.isclass(term_cfg.func):
		self._mode_class_term_cfgs[term_cfg.mode].append(term_cfg)

Conversation

hujc7 commented Mar 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Dependencies

Status

Uh oh!

hujc7 commented Mar 14, 2026

Uh oh!

greptile-apps bot commented Mar 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 3/5

Important Files Changed

Flowchart

Comments Outside Diff (1)

Uh oh!

greptile-apps bot Mar 14, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Mar 14, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Mar 14, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Mar 14, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Mar 14, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Mar 14, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Mar 14, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

hujc7 commented Mar 15, 2026

Uh oh!

greptile-apps bot Mar 15, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

hujc7 commented Mar 13, 2026 •

edited

Loading

greptile-apps bot commented Mar 14, 2026 •

edited

Loading