aws_rl_env / server /aws_rl_env_environment.py
Sizzing's picture
Upload folder using huggingface_hub
e56d042 verified
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
Aws Rl Env Environment Implementation.
An RL environment backed by a simulated AWS cloud powered by MiniStack.
The agent sends AWS CLI commands as actions and receives CLI output plus
the current resource state as observations.
"""
import logging
from typing import Any, Callable, Optional
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from models import (
AwsRlAction,
AwsRlObservation,
AwsRlState,
EpisodeID,
StepCount,
Task,
TaskInfo,
TrackerState,
)
from server.services.chaos_engine import ChaosEngine
from server.services.curriculum import Curriculum
from server.services.environment_strategy import EnvironmentStrategy
from server.services.simulator_strategy import SimulatorStrategy
from server.services.environment_designer import EnvironmentDesigner
from server.services.episode_context import EpisodeContext
from server.services.episode_tracker import EpisodeTracker
from server.services.hint_provider import HintProvider, MAX_HINT_LEVEL
from server.services.task_grader import TaskGrader
logger = logging.getLogger(__name__)
class AwsRlEnvironment(Environment[AwsRlAction, AwsRlObservation, AwsRlState]):
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(self, strategy: Optional[EnvironmentStrategy] = None) -> None:
print("Initializing AWS RL Environment...")
self._state = AwsRlState(episode_id=str(uuid4()), step_count=0)
self._backend = strategy if strategy is not None else SimulatorStrategy()
self._curriculum = Curriculum()
self._grader = TaskGrader(self._backend)
self._designer = EnvironmentDesigner(self._backend)
self._tracker = EpisodeTracker()
self._chaos_engine = ChaosEngine(self._backend)
self._hint_provider = HintProvider()
self._episode: Optional[EpisodeContext] = None
self._pool_release: Optional[Callable[[], None]] = None
@property
def _current_task(self) -> Optional[Task]:
"""Convenience accessor — None until the first reset()."""
return self._episode.task if self._episode is not None else None
def _sync_state(self) -> None:
"""Sync internal state to the AwsRlState object."""
self._state.current_task = self._current_task
self._state.tracker = TrackerState(
step_count=self._tracker.step_count,
hints_used=self._tracker.hints_used,
progress=self._tracker.previous_progress,
commands_executed=[s.command for s in self._tracker.command_history],
credited_operations=[
f"{op}:{res}" for op, res in self._tracker._credited_operations
],
)
self._state.chaos_occurred = self._chaos_engine.chaos_occurred
self._state.current_tier = (
self._episode.tier.value
if self._episode is not None
else self._curriculum.current_difficulty.value
)
self._state.infra_state = self._backend.get_infra_state()
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
task: Optional[Task | dict] = None,
**kwargs: Any,
) -> AwsRlObservation:
self._backend.reset_environment()
self._state = AwsRlState(episode_id=episode_id or str(uuid4()), step_count=0)
self._tracker.reset()
self._chaos_engine.reset()
# Trainer mode: caller supplied the Task. Local curriculum stays
# untouched — the trainer owns result recording.
# Local mode: curriculum picks and records the task.
if task is not None:
# Client sends Task.model_dump() over the wire; coerce back.
task_obj = task if isinstance(task, Task) else Task(**task)
self._episode = EpisodeContext.for_external(task=task_obj)
else:
task_obj = self._curriculum.next_task()
self._episode = EpisodeContext.for_local(
task=task_obj, curriculum=self._curriculum
)
self._designer.apply(task_obj)
self._sync_state()
return AwsRlObservation(
episode_id=EpisodeID(self._state.episode_id or ""),
step_count=StepCount(self._state.step_count),
command_success=True,
command_output="Environment reset. Infra state wiped.",
task=TaskInfo.from_task(task_obj),
done=False,
reward=0.0,
)
def _intercept_command(self, command: str) -> AwsRlObservation | None:
"""Handle anti-hack validation, hint requests, and help commands.
Returns an observation if the command was intercepted, None otherwise.
"""
if not command.startswith("aws "):
return AwsRlObservation(
episode_id=EpisodeID(self._state.episode_id or ""),
step_count=StepCount(self._state.step_count),
command_success=False,
command_output="",
error="Only AWS CLI commands (starting with 'aws') are allowed.",
task=TaskInfo.from_task(self._current_task)
if self._current_task
else None,
task_achieved=False,
done=False,
reward=0.0,
)
if command == "aws help --task-hint":
hint_level = self._tracker.record_hint()
clamped_level = min(hint_level, MAX_HINT_LEVEL)
assert self._current_task is not None
hint_text = self._hint_provider.get_hint(self._current_task, clamped_level)
return AwsRlObservation(
episode_id=EpisodeID(self._state.episode_id or ""),
step_count=StepCount(self._state.step_count),
command_success=True,
command_output=hint_text,
task=TaskInfo.from_task(self._current_task)
if self._current_task
else None,
task_achieved=False,
done=False,
reward=0.0,
hints_used=self._tracker.hints_used,
hint_text=hint_text,
)
parts = command.split()
if len(parts) == 3 and parts[0] == "aws":
service_name = None
if parts[2] == "help":
service_name = parts[1]
elif parts[1] == "help":
service_name = parts[2]
if service_name is not None:
svc_success, help_text = self._backend.get_service_help(service_name)
return AwsRlObservation(
episode_id=EpisodeID(self._state.episode_id or ""),
step_count=StepCount(self._state.step_count),
command_success=svc_success,
command_output=help_text if svc_success else "",
error="" if svc_success else help_text,
task=TaskInfo.from_task(self._current_task)
if self._current_task
else None,
task_achieved=False,
done=False,
reward=0.0,
)
return None
def step(
self,
action: AwsRlAction,
timeout_s: Optional[float] = None,
**kwargs: Any,
) -> AwsRlObservation:
assert self._episode is not None, "Call reset() before step()"
episode = self._episode
task = episode.task
self._state.step_count += 1
command = action.command.strip()
intercepted = self._intercept_command(command)
if intercepted is not None:
return intercepted
success, stdout, stderr = self._backend.execute_command(command)
# Record in tracker
latest_step = self._tracker.record_step(command, success, stdout, stderr)
# Grade the task (pass cumulative chaos flag and hint count)
grade_result = self._grader.grade(
task,
self._tracker,
latest_step,
chaos_occurred=self._chaos_engine.chaos_occurred,
hints_used=self._tracker.hints_used,
)
task_achieved = grade_result.task_achieved
reward = grade_result.reward
# Terminal result recording: trainer mode has record_result=None and
# owns recording centrally; local mode wires back to self._curriculum.
if task_achieved and episode.record_result is not None:
episode.record_result(task, True, reward)
# Inject chaos AFTER grading — disrupts state for future steps.
# Chaos probability is per-task-tier, not per-curriculum-cursor.
self._chaos_engine.maybe_inject(
task,
self._tracker,
episode.chaos_probability,
)
self._sync_state()
return AwsRlObservation(
episode_id=EpisodeID(self._state.episode_id or ""),
step_count=StepCount(self._state.step_count),
command_success=success,
command_output=stdout,
error=stderr,
task=TaskInfo.from_task(task),
task_achieved=task_achieved,
partial_progress=self._tracker.previous_progress,
done=task_achieved,
reward=reward,
hints_used=self._tracker.hints_used,
)
@property
def state(self) -> AwsRlState:
return self._state
def close(self) -> None:
if self._pool_release is None:
return
try:
self._backend.reset_environment()
except Exception:
logger.exception("Failed to scrub MiniStack state during close")
try:
self._pool_release()
finally:
self._pool_release = None