1. use_place_action is the invalid param.
2. duplicate names of atomic actions under action_agent_pipeline/env_adapters/tableware and action_agent_pipeline/runtime. only keep one implement of atomic actions.
3. invalid code like sync_agent_state_from_robot, extract_drive_calls, etc.
4. the action edge contains the non-atomic action.
5. invalid functions like _build_action_cfg_and_start() in action_agent_pipeline/runtime/atom_actions.py.
6. orientation seems like no use in prompt.

Thanks for the detailed review. I updated the PR branch and addressed the action-agent cleanup items:

• Removed use_place_action and verified it no longer appears in action_agent_pipeline.
• Kept only the runtime atomic-action implementation; the tableware adapter now imports the current env registration path, and the Demo3 semantic grasp integration test imports agent_env instead of the removed tableware.atomic_actions module.
• Removed stale/invalid leftovers including _build_action_cfg_and_start, monitor return fields, monitor runtime args, and unused compile LLM wiring.
• The graph compiler and runtime now reject legacy/non-atomic action schemas (action, fn/kwargs, unsupported edge fields).
• Removed orientation from the supported target-pose schema and added regression coverage that rejects it as an unsupported field.

Validation in embodichain040:

conda run -n embodichain040 python -m pytest \
  tests/gen_sim/action_agent_pipeline/test_backend_atomic_runtime.py \
  tests/gen_sim/action_agent_pipeline/test_graph_spec_backend_atomic.py \
  tests/gen_sim/action_agent_pipeline/test_llm_usage.py \
  -q

Result: 20 passed.

I also updated the PR description because the PR branch has now been synced to the current local branch, so the scope is broader than the original runtime-only slice and includes generated config / demo pipeline support.

Code review

Branch: 17 commits, 57 files, +17,269 LOC. The PR adds the action-agent runtime + generated-config pipeline for image-to-tabletop manipulation. The cleanup items from the prior review (duplicate atomic-actions, use_place_action, stale helpers) were verified as resolved.

Good design patterns

1. Clean three-layer separation in runtime/ — graph_compiler.py (JSON→graph), task_graph.py (pure graph + linear traversal), atom_actions.py (execution) each own one concern; the compiler injects graph_cls/action_module so layers are independently testable.
2. Single schema boundary — AtomicActionSpec (frozen dataclass) + normalize_atomic_action_spec is the one validation entry; legacy schemas (fn, action, target) are rejected with actionable messages.
3. Compile-time path validation — _validate_nominal_path rejects duplicate node/edge ids, dangling refs, cycles, and unreachable goals before runtime, so AgentTaskGraph.run() can assume a clean linear path.
4. Compiled-graph cache with hash + schema version — SHA-256 of canonical sorted-key JSON + COMPILED_GRAPH_SCHEMA_VERSION; stale artifacts self-heal.
5. Content-addressed CoACD mesh cache — keys cache on source_sha256 + policy_version + dexsim_engine_version + transform, so policy bumps auto-invalidate.
6. Consistent HTTP error envelope across SAM3/SAM3D/LLM clients — explicit timeout_s, typed errors, JSON-shape validation.
7. Zip-slip protection in _safe_extract_zip.
8. Declarative compositional success evaluator — recursive all/any/not over a leaf-predicate catalog.
9. Non-invasive LLM usage tracking — thin proxy over LangChain, process-local env vars, scrubbed from subprocess envs.
10. Network-gated tests — pytest.mark.skipif(... RUN_DEXSIM_GRASP_TESTS ...) and pytest.importorskip(...) keep the suite green without services.

Points to improve

Correctness / robustness

1. Schema/sentinel mismatch in graph_compiler.py — _validate_task_spec only treats JSON null as empty, but _compile_action also coerces ""/"none"/"null". An edge like {"left_arm_action": "", "right_arm_action": "none"} passes validation, compiles to both-None, and fails later in execute_parallel_atomic_actions with a worse error.

2. success.py::_pose() crashes opaquely on bad uid — env.sim.get_rigid_object(uid) returns None for a typoed object id, then .get_local_pose(...) raises AttributeError. Every leaf predicate funnels through _pose. Add an explicit None check raising ValueError(f"Unknown rigid object uid: {uid!r}").

3. Non-atomic cache writes — coacd_cache_bridge.py (pickle.dump to final path) and coacd_cache.py (CoACD .obj write). An interrupt leaves a corrupt file that future readers silently consume. Use a sibling temp + os.replace().

4. Unbounded retries / polls — prompt2geometry/pipeline.py (while True with bare except Exception on JSON extraction) and image2tabletop_client.py::wait_for_job (polls forever with no deadline). Add max-attempts + backoff + timeout.

5. Silent exception swallowing — atom_actions.py:957 (except Exception: return on grasp-collision cache prep — bugs get masked, collision checking silently degrades); agents/llm.py::_create_llm_safe returns None with no log — root cause lost. Narrow clauses and log.

6. resolve_arm_side doesn't raise — calls log_error(error_type=ValueError) but log_error only logs; the invalid side is returned and downstream _select_arm_parts produces confusing shape errors.

7. task_agent.py cache has no input hash — unlike CompileAgent, a stale agent_task_graph.json is loaded silently when prompt/task_name changes but the file path doesn't. Add a prompt-version or input-hash check.

8. env=None defaults on required params — execute_parallel_atomic_actions and AgentTaskGraph.run. Env is dereferenced immediately; None produces an opaque AttributeError. Make required.

9. utils/mllm.py::apply_proxy_env mutates os.environ globally as a side effect of create_openai_client/create_chat_openai — leaks across tests and parallel workers. Pass http_client=/proxies to constructors instead.

Structure / contracts

10. ur5_basket_config.py is 3,665 lines holding ≥6 separable concerns (scene parsing, target replacement, LLM role refinement, relative-placement geometry, GLB/glTF parsing, static robot/sensor/light templates). At minimum split out a glb_io.py (note: _read_glb in this file duplicates mesh_frame_normalization.py::_read_glb) and move static config templates to YAML/JSON data files.

https://github.com/DexForce/EmbodiChain/blob/24901bddef40c7fe69cae4e79d4d235aa4d73462/embodichain/gen_sim/action_agent_pipeline/generation/ur5_basket_config.py

11. run_agent_pipeline.py is 1,329 lines mixing arg parsing, two subprocess stages, rigid-object alias resolution, history lookup, and manifest writing. Split: target-replacement resolution, image2scene stage, arg parser + _DEFAULT_* constants into separate modules.

https://github.com/DexForce/EmbodiChain/blob/24901bddef40c7fe69cae4e79d4d235aa4d73462/embodichain/gen_sim/action_agent_pipeline/cli/run_agent_pipeline.py

12. base_agent_env.py::_init_agents fragile to reserved keys — _agent_config_with_prompt_keys strips prompt_kwargs but not task_name/config_dir; an Agent config block with either key raises TypeError: got multiple values for keyword argument.

13. agent_env.py::__init__ double-consumes **kwargs — forwards to both EmbodiedEnv.__init__ and _init_agents. The contract is fragile and undocumented.

14. execute_parallel_atomic_actions couples build + step + UI — root of the previously-flagged "executes immediately" concern. Split into build_parallel_action_stream(...) (pure) and step_env_with_actions(env, actions).

15. Redundant double-normalization — _compile_action normalizes via normalize_atomic_action_spec and stores a dict; execute_atomic_action re-normalizes on every execution via AtomicActionSpec.from_mapping. Cache the AtomicActionSpec at compile time.

16. Hardcoded intranet IPs as defaults — 192.168.3.23:{5013,5015,5016,4523} in pipeline.py, image2tabletop_client.py, zimage_client.py, run_agent_pipeline.py. Make required or env-only.

17. utils/__init__.py is empty and missing __all__ — AGENTS.md requires __all__ in every public module. All other __init__.py files are populated correctly; every source file has the Apache 2.0 header and from __future__ import annotations.

18. Inconsistent timestamps — pipeline records use local time / second precision; usage logs use UTC / millisecond precision. Pick one convention.

19. test_backend_atomic_runtime.py::_FakeBackendAction.capture is a class-level mutable shared across instances. Reset via fixture or make it instance-scoped.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

Module structure: agents/hierarchy/ vs gen_sim/action_agent_pipeline/agents/

The two modules are partially redundant — not duplicates of the same feature, but an in-progress migration from "LLM generates Python code that is exec()'d" (old agents/hierarchy/) to "LLM emits a JSON task graph that is compiled and executed as atomic actions" (new gen_sim/.../agents/).

Redundancy map

External usage

• agents/hierarchy/ is imported only from lab/gym/envs/tasks/tableware/base_agent_env.py:23-26 (old tableware task).
• gen_sim/.../agents/ is imported only from gen_sim/action_agent_pipeline/env_adapters/tableware/base_agent_env.py:30-36 (new gen_sim path).

There's a parallel old/new split at the env-adapter layer too. Note agents/hierarchy/code_agent.py:216 imports embodichain.lab.sim.atom_actions (singular, old) while the new pipeline uses embodichain.lab.sim.atomic_actions/ (plural package, new) — same generational split exists at the sim layer.

Improvement options

1. Finish the migration, delete agents/hierarchy/. Migrate the one consumer (lab/gym/envs/tasks/tableware/base_agent_env.py) to the gen_sim pipeline, then remove agents/hierarchy/ entirely. ValidationAgent (vision-LLM success checking) has no replacement, so either port it to env_adapters/.../success.py or accept losing that capability.
2. Extract shared base to agents/hierarchy/_base.py, have both pipelines inherit. Removes the AgentBase + _resolve_prompt_path + LLM-factory duplication. Tradeoff: introduces cross-package coupling (gen_sim would depend on agents/hierarchy/ or vice versa); only worth it if Option 1 is blocked.
3. Leave as-is, document the split. Add a deprecation docstring at the top of agents/hierarchy/ marking it as superseded-by-gen_sim.

Recommendation: Option 1 if the old tableware task is still actively used; Option 3 as a holding pattern if it's already abandoned. Option 2 is a trap — it locks in the duplication by making it structural rather than temporary.

The real question for the PR author: is the old code_agent (LLM-writes-Python) path still needed for anything the JSON-graph path can't do yet? If not, finish the migration in a follow-up and delete agents/hierarchy/.

🤖 Generated with Claude Code

Code review

Branch: 17 commits, 57 files, +17,269 LOC. The PR adds the action-agent runtime + generated-config pipeline for image-to-tabletop manipulation. The cleanup items from the prior review (duplicate atomic-actions, use_place_action, stale helpers) were verified as resolved.

Good design patterns

1. Clean three-layer separation in runtime/ — graph_compiler.py (JSON→graph), task_graph.py (pure graph + linear traversal), atom_actions.py (execution) each own one concern; the compiler injects graph_cls/action_module so layers are independently testable.
2. Single schema boundary — AtomicActionSpec (frozen dataclass) + normalize_atomic_action_spec is the one validation entry; legacy schemas (fn, action, target) are rejected with actionable messages.
3. Compile-time path validation — _validate_nominal_path rejects duplicate node/edge ids, dangling refs, cycles, and unreachable goals before runtime, so AgentTaskGraph.run() can assume a clean linear path.
4. Compiled-graph cache with hash + schema version — SHA-256 of canonical sorted-key JSON + COMPILED_GRAPH_SCHEMA_VERSION; stale artifacts self-heal.
5. Content-addressed CoACD mesh cache — keys cache on source_sha256 + policy_version + dexsim_engine_version + transform, so policy bumps auto-invalidate.
6. Consistent HTTP error envelope across SAM3/SAM3D/LLM clients — explicit timeout_s, typed errors, JSON-shape validation.
7. Zip-slip protection in _safe_extract_zip.
8. Declarative compositional success evaluator — recursive all/any/not over a leaf-predicate catalog.
9. Non-invasive LLM usage tracking — thin proxy over LangChain, process-local env vars, scrubbed from subprocess envs.
10. Network-gated tests — pytest.mark.skipif(... RUN_DEXSIM_GRASP_TESTS ...) and pytest.importorskip(...) keep the suite green without services.

Points to improve

Correctness / robustness

1. Schema/sentinel mismatch in graph_compiler.py — _validate_task_spec only treats JSON null as empty, but _compile_action also coerces ""/"none"/"null". An edge like {"left_arm_action": "", "right_arm_action": "none"} passes validation, compiles to both-None, and fails later in execute_parallel_atomic_actions with a worse error.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/runtime/graph_compiler.py

Lines 156 to 160 in 24901bd

	edge_ids.add(edge_id)
	if edge.get("left_arm_action") is None and edge.get("right_arm_action") is None:
	raise ValueError(f"Nominal edge '{edge_id}' must define an arm action.")
	for node_key in ("source", "target"):

2. success.py::_pose() crashes opaquely on bad uid — env.sim.get_rigid_object(uid) returns None for a typoed object id, then .get_local_pose(...) raises AttributeError. Every leaf predicate funnels through _pose. Add an explicit None check raising ValueError(f"Unknown rigid object uid: {uid!r}").

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/env_adapters/tableware/success.py

Lines 100 to 103 in 24901bd

	def _pose(env, uid: str) -> torch.Tensor:
	return env.sim.get_rigid_object(uid).get_local_pose(to_matrix=True)

3. Non-atomic cache writes — coacd_cache_bridge.py (pickle.dump to final path) and coacd_cache.py (CoACD .obj write). An interrupt leaves a corrupt file that future readers silently consume. Use a sibling temp + os.replace().

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/runtime/coacd_cache_bridge.py

Lines 174 to 183 in 24901bd

	with cache_path.open("wb") as cache_file:
	pickle.dump(
	{
	"plane_equations": plane_equations,
	"plane_equation_counts": plane_equation_counts,
	},
	cache_file,
	)

4. Unbounded retries / polls — prompt2geometry/pipeline.py (while True with bare except Exception on JSON extraction) and image2tabletop_client.py::wait_for_job (polls forever with no deadline). Add max-attempts + backoff + timeout.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/gym_project_api/prompt2geometry/pipeline.py

Lines 334 to 342 in 24901bd

	]
	while True:
	try:
	raw = client.chat_json(messages=messages)
	return _validate_glb_stem_output(raw)
	except Exception:
	time.sleep(1.0)
	continue

5. Silent exception swallowing — atom_actions.py:957 (except Exception: return on grasp-collision cache prep — bugs get masked, collision checking silently degrades); agents/llm.py::_create_llm_safe returns None with no log — root cause lost. Narrow clauses and log.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/runtime/atom_actions.py

Lines 956 to 958 in 24901bd

	)
	except Exception:
	return

6. resolve_arm_side doesn't raise — calls log_error(error_type=ValueError) but log_error only logs; the invalid side is returned and downstream _select_arm_parts produces confusing shape errors.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/runtime/atom_action_utils.py

Lines 40 to 48 in 24901bd

	if side not in _available_arm_sides(env):
	log_error(
	f"Requested {side}_arm for robot_name='{robot_name}', but available "
	f"control parts are {getattr(env.robot, 'control_parts', None)}.",
	error_type=ValueError,
	)
	return side

7. task_agent.py cache has no input hash — unlike CompileAgent, a stale agent_task_graph.json is loaded silently when prompt/task_name changes but the file path doesn't. Add a prompt-version or input-hash check.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/agents/task_agent.py

Lines 55 to 59 in 24901bd

	if not kwargs.get("regenerate", False) and file_path.exists():
	print(f"Task graph already exists at {file_path}.")
	return load_txt(file_path)

8. env=None defaults on required params — execute_parallel_atomic_actions and AgentTaskGraph.run. Env is dereferenced immediately; None produces an opaque AttributeError. Make required.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/runtime/atom_actions.py

Lines 436 to 440 in 24901bd

	right_arm_action=None,
	env=None,
	return_result: bool = False,
	**runtime_kwargs,
	):

9. utils/mllm.py::apply_proxy_env mutates os.environ globally as a side effect of create_openai_client/create_chat_openai — leaks across tests and parallel workers. Pass http_client=/proxies to constructors instead.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/utils/mllm.py

Lines 39 to 46 in 24901bd

	def apply_proxy_env(proxy_url: str | None) -> None:
	"""Apply an optional proxy URL for OpenAI-compatible clients."""
	if not proxy_url:
	return
	os.environ["HTTP_PROXY"] = proxy_url
	os.environ["HTTPS_PROXY"] = proxy_url

Structure / contracts

10. ur5_basket_config.py is 3,665 lines holding ≥6 separable concerns (scene parsing, target replacement, LLM role refinement, relative-placement geometry, GLB/glTF parsing, static robot/sensor/light templates). At minimum split out a glb_io.py (note: _read_glb in this file duplicates mesh_frame_normalization.py::_read_glb) and move static config templates to YAML/JSON data files.

https://github.com/DexForce/EmbodiChain/blob/24901bddef40c7fe69cae4e79d4d235aa4d73462/embodichain/gen_sim/action_agent_pipeline/generation/ur5_basket_config.py

11. run_agent_pipeline.py is 1,329 lines mixing arg parsing, two subprocess stages, rigid-object alias resolution, history lookup, and manifest writing. Split: target-replacement resolution, image2scene stage, arg parser + _DEFAULT_* constants into separate modules.

https://github.com/DexForce/EmbodiChain/blob/24901bddef40c7fe69cae4e79d4d235aa4d73462/embodichain/gen_sim/action_agent_pipeline/cli/run_agent_pipeline.py

12. base_agent_env.py::_init_agents fragile to reserved keys — _agent_config_with_prompt_keys strips prompt_kwargs but not task_name/config_dir; an Agent config block with either key raises TypeError: got multiple values for keyword argument.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/env_adapters/tableware/base_agent_env.py

Lines 47 to 61 in 24901bd

	)
	self.task_agent = TaskAgent(
	task_llm,
	**task_agent_config,
	**agent_config["TaskAgent"],
	task_name=task_name,
	config_dir=agent_config_path,
	)
	self.compile_agent = CompileAgent(
	**compile_agent_config,
	**agent_config["CompileAgent"],
	task_name=task_name,
	config_dir=agent_config_path,
	)

13. agent_env.py::__init__ double-consumes **kwargs — forwards to both EmbodiedEnv.__init__ and _init_agents. The contract is fragile and undocumented.

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/env_adapters/tableware/agent_env.py

Lines 36 to 42 in 24901bd

	def __init__(self, cfg: EmbodiedEnvCfg = None, **kwargs):
	super().__init__(cfg, **kwargs)
	if bool(getattr(self, "ignore_terminations_during_agent", False)):
	self.cfg.ignore_terminations = True
	super()._init_agents(**kwargs)

14. execute_parallel_atomic_actions couples build + step + UI — root of the previously-flagged "executes immediately" concern. Split into build_parallel_action_stream(...) (pure) and step_env_with_actions(env, actions).

EmbodiChain/embodichain/gen_sim/action_agent_pipeline/runtime/atom_actions.py

Lines 487 to 491 in 24901bd

	for action in tqdm(actions):
	env.step(action)
	env.update_obj_info()

15. Redundant double-normalization — _compile_action normalizes via normalize_atomic_action_spec and stores a dict; execute_atomic_action re-normalizes on every execution via AtomicActionSpec.from_mapping. Cache the AtomicActionSpec at compile time.
16. Hardcoded intranet IPs as defaults — 192.168.3.23:{5013,5015,5016,4523} in pipeline.py, image2tabletop_client.py, zimage_client.py, run_agent_pipeline.py. Make required or env-only.
17. utils/__init__.py is empty and missing __all__ — AGENTS.md requires __all__ in every public module. All other __init__.py files are populated correctly; every source file has the Apache 2.0 header and from __future__ import annotations.
18. Inconsistent timestamps — pipeline records use local time / second precision; usage logs use UTC / millisecond precision. Pick one convention.
19. test_backend_atomic_runtime.py::_FakeBackendAction.capture is a class-level mutable shared across instances. Reset via fixture or make it instance-scoped.

🤖 Generated with Claude Code

• If this code review was useful, please react with 👍. Otherwise, react with 👎.

Thanks for the thorough review. I addressed the cleanup items in the latest update:

• Aligned graph validation with runtime action normalization, so "", "none", and "null" are rejected at compile time when both arms are empty.
• Added an explicit unknown-object check in success._pose().
• Switched CoACD/grasp cache writes to temp-file + os.replace().
• Added bounded retry/timeout behavior for prompt2geometry naming and Image2Tabletop polling.
• Narrowed/logged grasp-cache fallback exceptions and logged LLM initialization failures.
• Made resolve_arm_side() raise clear ValueErrors.
• Added prompt-hash metadata for TaskAgent cache validation.
• Made env required for graph/action execution paths.
• Changed proxy handling to per-client httpx.Client instead of mutating os.environ.
• Split the large generation and CLI modules into focused files, including glb_io.py, config/template helpers, target replacement, image2scene stage, parser/defaults, and pipeline runner modules.
• Split action stream building from environment stepping via build_parallel_action_stream() and step_env_with_actions().
• Compile-time action normalization now stores AtomicActionSpec objects to avoid redundant normalization.
• Removed hardcoded intranet service defaults and kept service URLs env/argument driven.
• Added __all__ and aligned timestamp records to UTC milliseconds.
• Reset the fake backend capture through an autouse fixture.

Module structure: agents/hierarchy/ vs gen_sim/action_agent_pipeline/agents/

The two modules are partially redundant — not duplicates of the same feature, but an in-progress migration from "LLM generates Python code that is exec()'d" (old agents/hierarchy/) to "LLM emits a JSON task graph that is compiled and executed as atomic actions" (new gen_sim/.../agents/).

Redundancy map

File agents/hierarchy/ gen_sim/.../agents/ Verdict
agent_base.py AgentBase ABC + _resolve_prompt_path (95 lines) AgentBase ABC + _resolve_prompt_path (95 lines) Near-verbatim duplicate. Only diffs: gen_sim adds from __future__ import annotations + PEP 604 union, and get_composed_observations() drops the env.get_obs_for_agent() coupling. _resolve_prompt_path is byte-identical.
task_agent.py Emits free-text plan → agent_generated_plan.txt Emits JSON task graph → agent_task_graph.json Same skeleton (constructor, generate() flow, act() pass-through), different output contract.
llm.py Azure factory + 4 module-level instances OpenAI factory via shared utils/mllm.py + 1 instance Redundant factory. Gen_sim already factored the shared factory out to utils/mllm.py; hierarchy still hardcodes Azure.
code_agent.py LLM writes Python → exec() with AST kwargs injection (289 lines) — Unique to old design; replaced by CompileAgent + graph_compiler.
validation_agent.py Vision-LLM execution validator + view selector (241 lines) — Unique to old design; new pipeline uses declarative env_adapters/.../success.py predicates instead.

External usage

• agents/hierarchy/ is imported only from lab/gym/envs/tasks/tableware/base_agent_env.py:23-26 (old tableware task).
• gen_sim/.../agents/ is imported only from gen_sim/action_agent_pipeline/env_adapters/tableware/base_agent_env.py:30-36 (new gen_sim path).

There's a parallel old/new split at the env-adapter layer too. Note agents/hierarchy/code_agent.py:216 imports embodichain.lab.sim.atom_actions (singular, old) while the new pipeline uses embodichain.lab.sim.atomic_actions/ (plural package, new) — same generational split exists at the sim layer.

Improvement options

1. Finish the migration, delete agents/hierarchy/. Migrate the one consumer (lab/gym/envs/tasks/tableware/base_agent_env.py) to the gen_sim pipeline, then remove agents/hierarchy/ entirely. ValidationAgent (vision-LLM success checking) has no replacement, so either port it to env_adapters/.../success.py or accept losing that capability.
2. Extract shared base to agents/hierarchy/_base.py, have both pipelines inherit. Removes the AgentBase + _resolve_prompt_path + LLM-factory duplication. Tradeoff: introduces cross-package coupling (gen_sim would depend on agents/hierarchy/ or vice versa); only worth it if Option 1 is blocked.
3. Leave as-is, document the split. Add a deprecation docstring at the top of agents/hierarchy/ marking it as superseded-by-gen_sim.

Recommendation: Option 1 if the old tableware task is still actively used; Option 3 as a holding pattern if it's already abandoned. Option 2 is a trap — it locks in the duplication by making it structural rather than temporary.

The real question for the PR author: is the old code_agent (LLM-writes-Python) path still needed for anything the JSON-graph path can't do yet? If not, finish the migration in a follow-up and delete agents/hierarchy/.

🤖 Generated with Claude Code

Thanks for calling this out. I finished the migration and removed the legacy Python-code-generation action-agent path.
What changed:

• Deleted embodichain/agents/hierarchy/ and the old prompt files used only by that path.
• Removed the legacy singular embodichain.lab.sim.atom_actions module.
• Removed the old PourWaterAgent-v3 and RearrangementAgent-v3 registrations that depended on BaseAgentEnv.
• Kept the non-agent task envs (PourWater-v3, Rearrangement-v3) intact.
• Updated API docs so Sphinx no longer imports the removed legacy modules.
  The supported demos now go only through embodichain.gen_sim.action_agent_pipeline and AtomicActionsAgent-v3.