Spaces:
Running
Running
File size: 22,149 Bytes
9a3b69b 43f41de 9a3b69b eb1ebe6 43f41de eb1ebe6 9a3b69b 5869d56 9a3b69b 43f41de eb1ebe6 43f41de 9a3b69b 5869d56 9a3b69b 43f41de eb1ebe6 43f41de 9a3b69b 8fa7af1 9a3b69b eb1ebe6 9a3b69b eb1ebe6 9a3b69b eb1ebe6 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 9a3b69b eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 9a3b69b eb1ebe6 8fa7af1 43f41de eb1ebe6 43f41de 9a3b69b 43f41de 9a3b69b 43f41de 9a3b69b b12f1bd eb1ebe6 43f41de eb1ebe6 43f41de 9a3b69b eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 9a3b69b eb1ebe6 43f41de eb1ebe6 9a3b69b 43f41de 8fa7af1 43f41de 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de 8fa7af1 eb1ebe6 9a3b69b eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de b12f1bd eb1ebe6 43f41de eb1ebe6 b12f1bd eb1ebe6 43f41de b12f1bd 43f41de eb1ebe6 9a3b69b eb1ebe6 43f41de eb1ebe6 9a3b69b eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 8fa7af1 eb1ebe6 5869d56 eb1ebe6 43f41de 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 43f41de 8fa7af1 eb1ebe6 43f41de eb1ebe6 9a3b69b 43f41de 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 43f41de 8fa7af1 eb1ebe6 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de eb1ebe6 9a3b69b eb1ebe6 8fa7af1 43f41de eb1ebe6 b12f1bd eb1ebe6 8fa7af1 eb1ebe6 43f41de eb1ebe6 9a3b69b 8fa7af1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 | """
Research -> Interactive Explainer Environment (multi-step, async).
Episode flow:
1. reset() β agent gets a topic + tier
2. step(explore) Γ 0..MAX_EXPLORE β agent calls research tools
3. step(generate) Γ 1 β agent produces marimo/manim code
4. step(repair) Γ 0..MAX_REPAIR β agent fixes lint/build errors if needed
Each step returns a per-step reward. The final generate step also includes
a generation reward that accounts for how well the code uses the research.
The environment supports async via reset_async() / step_async() overrides.
OpenEnv's HTTP server detects these and calls them directly (no thread pool).
"""
import random
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
try:
from ..constants import MAX_EXPLORE_STEPS, MAX_REPAIR_STEPS, clamp_action_reward
from ..models import ExplainerAction, ExplainerObservation
from ..research import AVAILABLE_TOOLS, run_research_tool
from ..rewards.exploration import compute_explore_reward
from ..rewards.generation import adjust_repair_reward, compute_generate_reward
from ..rewards.sandbox import validate_code
from ..task_bank import ALL_TASKS, EASY_TASKS, HARD_TASKS, MEDIUM_TASKS, Task
except ImportError:
from constants import MAX_EXPLORE_STEPS, MAX_REPAIR_STEPS, clamp_action_reward
from models import ExplainerAction, ExplainerObservation
from research import AVAILABLE_TOOLS, run_research_tool
from rewards.exploration import compute_explore_reward
from rewards.generation import adjust_repair_reward, compute_generate_reward
from rewards.sandbox import validate_code
from task_bank import ALL_TASKS, EASY_TASKS, HARD_TASKS, MEDIUM_TASKS, Task
MB002_REPAIR_HINT = (
"MB002 repair checklist: Marimo treats every non-underscore assignment as a "
"global notebook variable, including `for` loop variables. Audit the whole "
"file and rename cell-local names to private names everywhere: `arr` -> "
"`_arr`, `target` -> `_target`, `i` -> `_i`, `t` -> `_t`, `freqs` -> "
"`_freqs`, `fig` -> `_fig`, `ax` -> `_ax`. Public names should only be used "
"for values intentionally passed to later cells, and each public name may be "
"defined once globally."
)
def _render_errors_with_hints(errors: str, error_codes: list[str]) -> str:
if "MB002" not in error_codes:
return errors
return f"{errors}\n\n{MB002_REPAIR_HINT}"
class ExplainerEnvironment(Environment):
"""
Multi-step Research β Interactive Explainer environment.
Phase 1 (explore): agent issues search queries, receives papers/wiki sections.
Phase 2 (generate): agent produces marimo/manim code using the research.
Supports async via reset_async() / step_async() β OpenEnv's server detects
the overrides and awaits them directly instead of using a thread pool.
"""
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(self):
super().__init__()
self._state = State(episode_id=str(uuid4()), step_count=0)
self._current_task: Task | None = None
self._difficulty_pool: list[Task] = EASY_TASKS
self._accumulated_context: list[str] = []
self._explore_actions: list[str] = []
self._used_tools: set[str] = set()
self._explore_steps: int = 0
self._repair_steps: int = 0
self._phase: str = "explore"
self._done: bool = False
self._last_code: str = ""
self._last_format: str = "marimo"
self._last_narration: str = ""
self._last_errors: str = ""
self._last_error_codes: list[str] = []
# ------------------------------------------------------------------
# Sync interface (fallback β OpenEnv prefers async when overridden)
# ------------------------------------------------------------------
def reset(self, seed=None, episode_id=None, **kwargs) -> ExplainerObservation:
"""Sample a task and return the initial observation (sync)."""
return self._do_reset(seed=seed, episode_id=episode_id, **kwargs)
def step(self, action: ExplainerAction, timeout_s=None, **kwargs) -> ExplainerObservation:
"""Route to explore or generate handler (sync β explore uses blocking fallback)."""
import asyncio
self._state.step_count += 1
task = self._current_task
if task is None:
return ExplainerObservation(
feedback="Error: no task set. Call reset() first.",
done=True,
reward=-1.0,
)
if self._done:
return self._make_obs(
task,
phase="done",
feedback="Episode is already done. Call reset() to start a new one.",
reward=0.0,
done=True,
)
try:
if action.action_type == "explore":
# Run async explore in a new event loop for sync callers
return asyncio.run(self._handle_explore(action, task))
elif action.action_type == "generate":
return self._handle_generate(action, task)
elif action.action_type == "repair":
return self._handle_repair(action, task)
else:
return self._make_obs(
task,
phase="explore",
feedback=f"Unknown action_type: {action.action_type}",
reward=0.0,
done=True,
)
except Exception as e:
return self._make_obs(
task,
phase="done",
feedback=f"Environment error: {e}",
reward=0.0,
done=True,
)
# ------------------------------------------------------------------
# Async interface (preferred β OpenEnv detects these overrides)
# ------------------------------------------------------------------
async def reset_async(self, seed=None, episode_id=None, **kwargs) -> ExplainerObservation:
"""Sample a task and return the initial observation (async)."""
return self._do_reset(seed=seed, episode_id=episode_id, **kwargs)
async def step_async(self, action: ExplainerAction, timeout_s=None, **kwargs) -> ExplainerObservation:
"""Route to explore or generate handler (async)."""
self._state.step_count += 1
task = self._current_task
if task is None:
return ExplainerObservation(
feedback="Error: no task set. Call reset() first.",
done=True,
reward=-1.0,
)
if self._done:
return self._make_obs(
task,
phase="done",
feedback="Episode is already done. Call reset() to start a new one.",
reward=0.0,
done=True,
)
try:
if action.action_type == "explore":
return await self._handle_explore(action, task)
elif action.action_type == "generate":
return self._handle_generate(action, task)
elif action.action_type == "repair":
return self._handle_repair(action, task)
else:
return self._make_obs(
task,
phase="explore",
feedback=f"Unknown action_type: {action.action_type}",
reward=0.0,
done=True,
)
except Exception as e:
return self._make_obs(
task,
phase="done",
feedback=f"Environment error: {e}",
reward=0.0,
done=True,
)
# ------------------------------------------------------------------
# Internal
# ------------------------------------------------------------------
def _do_reset(self, seed=None, episode_id=None, **kwargs) -> ExplainerObservation:
"""Shared reset logic (no I/O, so sync is fine)."""
self._state = State(
episode_id=episode_id or str(uuid4()), step_count=0
)
self._accumulated_context = []
self._explore_actions = []
self._used_tools = set()
self._explore_steps = 0
self._repair_steps = 0
self._phase = "explore"
self._done = False
self._last_code = ""
self._last_format = "marimo"
self._last_narration = ""
self._last_errors = ""
self._last_error_codes = []
# Allow selecting a specific task by topic name
topic = kwargs.get("topic", None)
if topic:
match = next((t for t in ALL_TASKS if t.topic == topic), None)
if match:
self._current_task = match
else:
# Fallback to random if topic not found
rng = random.Random(seed) if seed is not None else random.Random()
self._current_task = rng.choice(ALL_TASKS)
else:
difficulty = kwargs.get("difficulty", None)
if difficulty == "medium":
pool = MEDIUM_TASKS
elif difficulty == "hard":
pool = HARD_TASKS
elif difficulty == "easy":
pool = EASY_TASKS
else:
pool = self._difficulty_pool
rng = random.Random(seed) if seed is not None else random.Random()
self._current_task = rng.choice(pool) if pool else rng.choice(ALL_TASKS)
t = self._current_task
return ExplainerObservation(
topic=t.topic,
content=t.content,
tier=t.tier,
keywords=t.keywords,
data_available=t.data_available,
difficulty=t.difficulty,
phase="explore",
feedback=(
"Research phase: choose a tool and query relevant to the topic. "
f"Available tools: {', '.join(AVAILABLE_TOOLS)}."
),
search_results="",
explored_context="",
explore_steps_left=MAX_EXPLORE_STEPS,
repair_attempts_left=MAX_REPAIR_STEPS,
available_tools=list(AVAILABLE_TOOLS),
done=False,
reward=0.0,
)
async def _handle_explore(self, action: ExplainerAction, task: Task) -> ExplainerObservation:
"""Process an explore action: call a research tool and score the result."""
if self._phase not in {"explore", "generate"}:
return self._make_obs(
task,
phase=self._phase,
feedback=f"Cannot explore during phase '{self._phase}'.",
reward=0.0,
)
if self._explore_steps >= MAX_EXPLORE_STEPS:
self._phase = "generate"
return self._make_obs(
task,
phase="generate",
feedback="Max explore steps reached. You must now generate.",
reward=0.0,
)
self._explore_steps += 1
query = action.query.strip()
intent = action.intent.strip()
tool = action.tool or "search_wikipedia"
if not query:
return self._make_obs(
task,
phase="explore",
feedback="Empty query. Provide a search query.",
reward=0.0,
)
previous_context = list(self._accumulated_context)
previous_actions = list(self._explore_actions)
used_tools = set(self._used_tools)
result = await run_research_tool(tool, query, intent)
results_text = result.render()
self._explore_actions.append(_explore_action_text(tool, query, intent))
if result.ok:
self._accumulated_context.append(result.text)
self._used_tools.add(tool)
# Compute per-step exploration reward
reward, components = compute_explore_reward(
query=query,
tool=tool,
intent=intent,
result=result,
topic=task.topic,
keywords_csv=task.keywords,
task_content=task.content,
difficulty=task.difficulty,
previous_context=previous_context,
accumulated_context=self._accumulated_context,
used_tools=used_tools,
previous_actions=previous_actions,
)
steps_left = MAX_EXPLORE_STEPS - self._explore_steps
if steps_left > 1:
phase = "explore"
hint = f"Research going well β {steps_left} more steps available. Keep searching or move to generation."
elif steps_left == 1:
phase = "explore"
hint = "Last research step available. Search for any missing context, or proceed to generate."
else:
phase = "generate"
hint = "Research phase complete. Time to generate your explanation."
self._phase = phase
top_chunks = _top_chunks_payload(result.chunks)
return self._make_obs(
task,
phase=phase,
feedback=f"{hint}\nTool: {tool}\nReward: {components}",
search_results=results_text,
top_chunks=top_chunks,
reward=reward,
metadata={
"step": self._state.step_count,
"phase": "explore",
"tool": tool,
"source_count": len(result.chunks),
"top_chunks": top_chunks,
"error": result.error,
**components,
},
)
def _handle_generate(self, action: ExplainerAction, task: Task) -> ExplainerObservation:
"""Process a generate action: run sandbox, maybe open repair phase."""
if self._phase not in {"explore", "generate"}:
return self._make_obs(
task,
phase=self._phase,
feedback=f"Cannot generate during phase '{self._phase}'.",
reward=0.0,
)
fmt = action.format or "marimo"
code = action.code
narration = action.narration
# Penalise generating without any exploration
if self._explore_steps == 0:
skip_penalty = -0.1
penalty_msg = "Warning: generating without any research. -0.1 penalty."
else:
skip_penalty = 0.0
penalty_msg = ""
sandbox = validate_code(fmt, code)
# Generation reward
reward, components = compute_generate_reward(
code=code,
fmt=fmt,
narration=narration,
task=task,
exec_success=sandbox.exec_success,
accumulated_context=self._accumulated_context,
static_check_passed=sandbox.check_passed,
error_codes=sandbox.error_codes,
)
reward = clamp_action_reward(reward + skip_penalty)
components["generate_total"] = round(reward, 4)
self._last_code = code
self._last_format = fmt
self._last_narration = narration
rendered_errors = _render_errors_with_hints(sandbox.render_errors(), sandbox.error_codes)
self._last_errors = rendered_errors
self._last_error_codes = sandbox.error_codes
# Feedback
parts = []
if penalty_msg:
parts.append(penalty_msg)
if not sandbox.parses:
parts.append("SYNTAX ERROR: code does not parse.")
elif not sandbox.exec_success:
parts.append(f"EXECUTION FAILED: {rendered_errors}")
else:
parts.append(f"EXECUTION OK: {sandbox.message}")
parts.append(
f"Reward: {', '.join(f'{k}={v}' for k, v in components.items())}"
)
done = sandbox.exec_success or self._repair_steps >= MAX_REPAIR_STEPS
phase = "done" if done else "repair"
self._phase = phase
self._done = done
if not done:
parts.append(
f"Repair phase: {MAX_REPAIR_STEPS} attempts available. "
"Submit a revised artifact using the error feedback."
)
return self._make_obs(
task,
phase=phase,
feedback="\n".join(parts),
reward=reward,
done=done,
last_errors="" if sandbox.exec_success else rendered_errors,
metadata={
"step": self._state.step_count,
"phase": "generate",
"explore_steps_used": self._explore_steps,
"sandbox_message": sandbox.message,
"error_codes": sandbox.error_codes,
**components,
},
)
def _handle_repair(self, action: ExplainerAction, task: Task) -> ExplainerObservation:
"""Process one repair attempt after a failed generate action."""
if self._phase != "repair":
return self._make_obs(
task,
phase=self._phase,
feedback="Repair is only available after a failed generate step.",
reward=0.0,
done=self._done,
)
if self._repair_steps >= MAX_REPAIR_STEPS:
self._phase = "done"
self._done = True
return self._make_obs(
task,
phase="done",
feedback="No repair attempts left.",
reward=0.0,
done=True,
)
self._repair_steps += 1
fmt = action.format or self._last_format or "marimo"
code = action.code
narration = action.narration or self._last_narration
previous_code = self._last_code
previous_errors = list(self._last_error_codes)
sandbox = validate_code(fmt, code)
base_reward, components = compute_generate_reward(
code=code,
fmt=fmt,
narration=narration,
task=task,
exec_success=sandbox.exec_success,
accumulated_context=self._accumulated_context,
static_check_passed=sandbox.check_passed,
error_codes=sandbox.error_codes,
)
repair_reward, repair_components = adjust_repair_reward(
base_reward,
repair_success=sandbox.exec_success,
previous_error_codes=previous_errors,
new_error_codes=sandbox.error_codes,
previous_code=previous_code,
repaired_code=code,
)
components.update(repair_components)
self._last_code = code
self._last_format = fmt
self._last_narration = narration
rendered_errors = _render_errors_with_hints(sandbox.render_errors(), sandbox.error_codes)
self._last_errors = rendered_errors
self._last_error_codes = sandbox.error_codes
attempts_left = MAX_REPAIR_STEPS - self._repair_steps
done = sandbox.exec_success or attempts_left <= 0
phase = "done" if done else "repair"
self._phase = phase
self._done = done
status = "REPAIR OK" if sandbox.exec_success else "REPAIR FAILED"
feedback_parts = [
f"{status}: {sandbox.message if sandbox.exec_success else rendered_errors}",
f"Reward: {', '.join(f'{k}={v}' for k, v in components.items())}",
]
if not done:
feedback_parts.append(
f"Repair phase continues: {attempts_left} repair attempts left. "
"Submit another corrected artifact using the latest error feedback."
)
feedback = "\n".join(feedback_parts)
return self._make_obs(
task,
phase=phase,
feedback=feedback,
reward=repair_reward,
done=done,
last_errors="" if sandbox.exec_success else rendered_errors,
metadata={
"step": self._state.step_count,
"phase": "repair",
"explore_steps_used": self._explore_steps,
"repair_steps_used": self._repair_steps,
"sandbox_message": sandbox.message,
"error_codes": sandbox.error_codes,
**components,
},
)
def _make_obs(
self,
task: Task,
*,
phase: str,
feedback: str,
reward: float = 0.0,
done: bool = False,
search_results: str = "",
top_chunks: list[dict] | None = None,
last_errors: str | None = None,
metadata: dict | None = None,
) -> ExplainerObservation:
"""Helper to build a consistent observation."""
return ExplainerObservation(
topic=task.topic,
content=task.content,
tier=task.tier,
keywords=task.keywords,
data_available=task.data_available,
difficulty=task.difficulty,
phase=phase,
feedback=feedback,
search_results=search_results,
top_chunks=top_chunks or [],
explored_context="\n---\n".join(self._accumulated_context),
explore_steps_left=MAX_EXPLORE_STEPS - self._explore_steps,
repair_attempts_left=MAX_REPAIR_STEPS - self._repair_steps,
last_errors=self._last_errors if last_errors is None else last_errors,
available_tools=list(AVAILABLE_TOOLS),
done=done,
reward=reward,
metadata=metadata or {},
)
@property
def state(self) -> State:
return self._state
def _explore_action_text(tool: str, query: str, intent: str) -> str:
return f"{tool} {query.strip()} {intent.strip()}".strip()
def _top_chunks_payload(chunks) -> list[dict]:
return [
{
"rank": chunk.rank,
"source": chunk.source,
"title": chunk.title,
"url": chunk.url,
"score": round(chunk.score, 4),
"snippet": chunk.text,
}
for chunk in chunks[:5]
]
|