File size: 10,868 Bytes
31715b5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 | """OpenEnv-conformant adapter for OpenSleuthEnv.
Wraps the existing multi-episode :class:`OpenSleuthEnv` registry as a
single-episode-per-session :class:`openenv.core.env_server.interfaces.Environment`
so the canonical OpenEnv HTTP / WebSocket protocol can be served alongside
the legacy ``/reset`` + ``/step`` endpoints the in-flight trainer uses.
This module is *additive*. It does not touch the legacy server contract;
``server.py`` mounts the OpenEnv-style sub-application at ``/openenv/*`` so the
trainer (which talks to the bare ``/reset`` and ``/step``) is unaffected.
The adapter conforms to OpenEnv 0.2.x:
* ``Environment.reset(seed, episode_id, **kwargs) -> Observation``
* ``Environment.step(action, timeout_s, **kwargs) -> Observation``
* ``Environment.state -> State``
* ``Environment.get_metadata() -> EnvironmentMetadata``
See https://github.com/meta-pytorch/OpenEnv (v0.2.3, BSD-3) for the spec.
"""
from __future__ import annotations
from typing import Any, List, Literal, Optional
from uuid import uuid4
from pydantic import Field
try:
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import (
Action as OEAction,
EnvironmentMetadata,
Observation as OEObservation,
State as OEState,
)
OPENENV_AVAILABLE = True
except ImportError: # pragma: no cover - openenv is required at runtime in the Space
OPENENV_AVAILABLE = False
OEAction = object # type: ignore[assignment, misc]
OEObservation = object # type: ignore[assignment, misc]
OEState = object # type: ignore[assignment, misc]
Environment = object # type: ignore[assignment, misc]
EnvironmentMetadata = object # type: ignore[assignment, misc]
from .env import OpenSleuthEnv
from .models import ProbeAction, SubmitAction
if OPENENV_AVAILABLE:
class OpenSleuthAction(OEAction):
"""Unified OpenEnv-style action.
The OpenEnv spec wants a single concrete Action subclass per
environment; we encode the probe / submit choice via the
``action_type`` discriminator field. Internally we still translate
to the original :class:`ProbeAction` / :class:`SubmitAction` so the
legacy reward shaping is preserved bit-for-bit.
"""
action_type: Literal["probe", "submit"] = Field(
..., description="Either 'probe' (with input_repr) or 'submit' (with code)."
)
input_repr: Optional[str] = Field(
default=None,
description="Python literal repr of the probe input. Required when action_type='probe'.",
)
code: Optional[str] = Field(
default=None,
description="Python source defining the target function. Required when action_type='submit'.",
)
class OpenSleuthObservation(OEObservation):
"""OpenEnv observation wrapper.
OpenEnv's ``Observation`` base class supplies ``done``, ``reward``,
and ``metadata``. We add OpenSleuth-specific fields for the agent
(target signature, probe history, etc.). Trainer-facing structured
info is also surfaced via ``info`` for backwards compat.
"""
episode_id: str = Field(default="", description="Per-session episode id.")
target_function_name: str = Field(default="")
target_function_signature: str = Field(
default="", description="Public signature + docstring for the target."
)
probe_history: List[dict] = Field(
default_factory=list,
description="Recent probe records (input_repr, output_repr, is_error, ...).",
)
last_error: str = Field(default="", description="Last error string, if any.")
steps_taken: int = Field(default=0)
max_steps: int = Field(default=25)
difficulty: Optional[str] = Field(default=None)
coverage_buckets_seen: int = Field(default=0)
seen_outputs_count: int = Field(default=0)
seen_error_types_count: int = Field(default=0)
info: dict = Field(
default_factory=dict,
description="Structured info from the underlying step (matches the legacy info dict).",
)
class OpenSleuthState(OEState):
"""OpenEnv-style episode state."""
target_function_name: Optional[str] = Field(default=None)
max_steps: int = Field(default=25)
finished: bool = Field(default=False)
class OpenSleuthEnvironment(Environment):
"""OpenEnv-conformant adapter around :class:`OpenSleuthEnv`.
One adapter instance == one episode (one WebSocket session). Inside,
we keep a single :class:`OpenSleuthEnv` registry but only ever populate
a single episode at a time.
``SUPPORTS_CONCURRENT_SESSIONS = True`` is safe because each WebSocket
connection in OpenEnv's :class:`HTTPEnvServer` instantiates its own
:class:`OpenSleuthEnvironment`, and our underlying registries are
per-instance.
"""
SUPPORTS_CONCURRENT_SESSIONS = True
def __init__(self) -> None:
super().__init__()
self._env = OpenSleuthEnv()
self._episode_id: Optional[str] = None
self._target_function_name: Optional[str] = None
self._max_steps: int = 25
self._step_count: int = 0
self._done: bool = False
def reset( # type: ignore[override]
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
target_name: Optional[str] = None,
target_code: Optional[str] = None,
target_function_name: Optional[str] = None,
max_steps: int = 25,
edge_cases: Optional[list] = None,
fuzz_spec: Optional[dict] = None,
**kwargs: Any,
) -> "OpenSleuthObservation":
# Default to a builtin so a bare reset() still produces a valid
# episode (per OpenEnv spec, reset() with no args must work).
if not target_name and not target_code:
target_name = "fibonacci"
obs = self._env.reset(
target_name=target_name,
seed=seed if seed is not None else 0,
max_steps=max_steps,
target_code=target_code,
target_function_name=target_function_name,
edge_cases=edge_cases,
fuzz_spec=fuzz_spec,
)
self._episode_id = episode_id or obs.episode_id
self._target_function_name = obs.target_function_name
self._max_steps = max_steps
self._step_count = 0
self._done = False
return self._wrap_obs(obs, reward=None, done=False, info={})
def step( # type: ignore[override]
self,
action: "OpenSleuthAction",
timeout_s: Optional[float] = None,
**kwargs: Any,
) -> "OpenSleuthObservation":
if self._episode_id is None:
# Auto-reset on first step with the default target so HTTP /step
# smoke tests don't 500 just because /reset wasn't called first.
self.reset()
internal_action: Any
if action.action_type == "probe":
if action.input_repr is None:
raise ValueError(
"OpenSleuthAction(action_type='probe') requires input_repr."
)
internal_action = ProbeAction(input_repr=action.input_repr)
elif action.action_type == "submit":
if action.code is None:
raise ValueError(
"OpenSleuthAction(action_type='submit') requires code."
)
internal_action = SubmitAction(code=action.code)
else: # pragma: no cover - Pydantic Literal already constrains this
raise ValueError(f"Unknown action_type: {action.action_type!r}")
assert self._episode_id is not None
resp = self._env.step(self._episode_id, internal_action)
self._step_count += 1
self._done = resp.done
return self._wrap_obs(
resp.observation, reward=resp.reward, done=resp.done, info=resp.info
)
@property
def state(self) -> "OpenSleuthState": # type: ignore[override]
return OpenSleuthState(
episode_id=self._episode_id,
step_count=self._step_count,
target_function_name=self._target_function_name,
max_steps=self._max_steps,
finished=self._done,
)
def get_metadata(self) -> "EnvironmentMetadata": # type: ignore[override]
return EnvironmentMetadata(
name="OpenSleuth",
description=(
"Algorithmic detective: probe a hidden Python function then submit "
"code that reproduces it. Used for GRPO RL training on Qwen-2.5."
),
version="0.4.1",
author="OpenSleuth team",
documentation_url=(
"https://huggingface.co/spaces/anugrah55/opensleuth-env-gemini-cli"
),
)
def close(self) -> None: # type: ignore[override]
self._episode_id = None
self._target_function_name = None
self._step_count = 0
self._done = False
def _wrap_obs(
self,
internal_obs: Any,
*,
reward: Optional[float],
done: bool,
info: dict,
) -> "OpenSleuthObservation":
return OpenSleuthObservation(
done=done,
reward=reward,
episode_id=internal_obs.episode_id,
target_function_name=internal_obs.target_function_name,
target_function_signature=internal_obs.target_function_signature,
probe_history=[r.model_dump() for r in internal_obs.probe_history],
last_error=internal_obs.last_error,
steps_taken=internal_obs.steps_taken,
max_steps=internal_obs.max_steps,
difficulty=internal_obs.difficulty,
coverage_buckets_seen=internal_obs.coverage_buckets_seen,
seen_outputs_count=internal_obs.seen_outputs_count,
seen_error_types_count=internal_obs.seen_error_types_count,
info=info,
metadata={"info": info},
)
__all__ = ["OPENENV_AVAILABLE"]
if OPENENV_AVAILABLE:
__all__ += [
"OpenSleuthAction",
"OpenSleuthObservation",
"OpenSleuthState",
"OpenSleuthEnvironment",
]
|