Spaces:
Running
Running
File size: 15,311 Bytes
d57737f | 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 | """
Verifier execution for AWM environments.
Two verification modes:
- "sql": Verifier extracts DB state, then an LLM judges the trajectory.
- "code": Verifier deterministically returns {"result": "complete"|"others"}.
"""
import json
import logging
import os
import re
import subprocess
import sys
import textwrap
from typing import Any
from openai import AsyncOpenAI
logger = logging.getLogger(__name__)
# Wall-clock backstop for the verifier subprocess
VERIFIER_WALL_TIMEOUT_S = float(os.environ.get("OPENENV_AWM_VERIFIER_TIMEOUT", "20.0"))
# isolate runner script
_RUNNER_PATH = os.path.join(
os.path.dirname(os.path.abspath(__file__)), "_verifier_runner.py"
)
def _run_verifier_subprocess(payload: dict[str, Any]) -> dict[str, Any]:
"""Invoke the runner. Returns a dict; never raises."""
try:
proc = subprocess.run(
[sys.executable, _RUNNER_PATH],
input=json.dumps(payload),
capture_output=True,
text=True,
timeout=VERIFIER_WALL_TIMEOUT_S,
check=False,
)
except subprocess.TimeoutExpired:
return {
"execution_status": "error",
"error_message": (f"Verifier timed out after {VERIFIER_WALL_TIMEOUT_S}s"),
}
except (OSError, ValueError) as e:
return {
"execution_status": "error",
"error_message": f"Failed to launch verifier subprocess: {e}",
}
if proc.returncode != 0 and not proc.stdout.strip():
stderr = (proc.stderr or "").strip()
return {
"execution_status": "error",
"error_message": (
f"Verifier subprocess exited {proc.returncode}: {stderr[:500]}"
),
}
if not proc.stdout.strip():
return {
"execution_status": "error",
"error_message": "Verifier subprocess produced no output",
}
try:
return json.loads(proc.stdout)
except json.JSONDecodeError as e:
return {
"execution_status": "error",
"error_message": (
f"Bad JSON from verifier subprocess ({e}): {proc.stdout[:500]}"
),
}
def execute_sql_verifier(
verifier_code: str,
function_name: str,
initial_db_path: str,
final_db_path: str,
) -> dict[str, Any]:
"""Execute SQL-mode verifier code that compares initial vs final DB state.
Returns:
Verifier result dict, or error dict with "execution_status": "error".
"""
if not os.path.exists(initial_db_path):
return {
"execution_status": "error",
"error_message": f"Initial DB not found: {initial_db_path}",
}
if not os.path.exists(final_db_path):
return {
"execution_status": "error",
"error_message": f"Final DB not found: {final_db_path}",
}
original_modes = {}
for path in [initial_db_path, final_db_path]:
try:
original_modes[path] = os.stat(path).st_mode
os.chmod(path, 0o444)
except Exception:
pass
try:
return _run_verifier_subprocess(
{
"verifier_code": verifier_code,
"function_name": function_name,
"initial_db_path": initial_db_path,
"final_db_path": final_db_path,
"final_answer": None,
"mode": "sql",
}
)
finally:
for path, mode in original_modes.items():
try:
os.chmod(path, mode)
except Exception:
pass
def execute_code_verifier(
verifier_code: str,
function_name: str,
initial_db_path: str,
final_db_path: str,
final_answer: str | None = None,
) -> dict[str, Any]:
"""Execute code-mode verifier that returns {"result": "complete"|"others"}.
Returns:
Dict with "result" key ("complete" or "others") and "execution_status".
"""
if not os.path.exists(initial_db_path):
return {
"result": "others",
"execution_status": "error",
"error_message": f"Initial DB not found: {initial_db_path}",
}
if not os.path.exists(final_db_path):
return {
"result": "others",
"execution_status": "error",
"error_message": f"Final DB not found: {final_db_path}",
}
original_modes = {}
for path in [initial_db_path, final_db_path]:
try:
original_modes[path] = os.stat(path).st_mode
os.chmod(path, 0o444)
except Exception:
pass
try:
result = _run_verifier_subprocess(
{
"verifier_code": verifier_code,
"function_name": function_name,
"initial_db_path": initial_db_path,
"final_db_path": final_db_path,
"final_answer": final_answer,
"mode": "code",
}
)
finally:
for path, mode in original_modes.items():
try:
os.chmod(path, mode)
except Exception:
pass
# Normalize: code-mode callers expect "result" key on every return.
if "result" not in result:
result.setdefault("execution_status", "error")
result["result"] = "others"
return result
def run_verifier(
verifier_entry: dict,
verifier_mode: str,
initial_db_path: str,
final_db_path: str,
final_answer: str | None = None,
) -> tuple[str, dict]:
"""
Run the appropriate verifier and return (reward_type, details).
Returns:
(reward_type, verify_result_dict)
"""
verification = verifier_entry.get("verification", {})
code = verification.get("code", "")
if not code or not isinstance(code, str) or len(code.strip()) < 10:
return "judge_error", {"error": "No valid verifier code found"}
default_func = (
"verify_task_completion" if verifier_mode == "code" else "verify_task"
)
func_name = default_func
for line in code.split("\n"):
line = line.strip()
if line.startswith("def verify_") and "(" in line:
func_name = line.split("(")[0].replace("def ", "").strip()
break
if verifier_mode == "code":
result = execute_code_verifier(
code, func_name, initial_db_path, final_db_path, final_answer
)
if result.get("execution_status") == "error":
return "judge_error", result
return result.get("result", "others"), result
result = execute_sql_verifier(code, func_name, initial_db_path, final_db_path)
if isinstance(result, dict) and result.get("execution_status") == "error":
return "judge_error", result
return "incomplete", result
def _normalize_azure_url(base_url: str) -> str:
"""Normalize Azure OpenAI base URLs to include /openai/v1 suffix.
The Azure OpenAI v1 API uses the standard OpenAI client with
base_url set to https://RESOURCE.openai.azure.com/openai/v1/.
This helper auto-appends /openai/v1 when it detects an Azure endpoint.
"""
if "openai.azure.com" in base_url or "services.ai.azure.com" in base_url:
stripped = base_url.rstrip("/")
if not stripped.endswith("/openai/v1"):
return stripped + "/openai/v1"
return base_url
async def run_llm_judge(
task: str,
verifier_result: dict,
llm_base_url: str | None = None,
llm_api_key: str | None = None,
llm_model: str | None = None,
trajectory: list | None = None,
verifier_reasoning: str = "",
success_criteria: str = "",
failure_criteria: str = "",
) -> tuple[str, dict]:
"""Run LLM-as-a-judge combining agent trajectory AND SQL verification results.
Aligned with llm_judge_with_sql in agent-world-model-rl-rewards/task_reward.py.
Works with both OpenAI and Azure OpenAI (v1 API). For Azure, set
llm_base_url to https://YOUR-RESOURCE.openai.azure.com and the URL
will be auto-normalized to append /openai/v1.
Returns:
(classification, judge_result_dict)
"""
if not llm_base_url or not llm_model:
return "judge_error", {
"error": "LLM endpoint not configured for sql verifier mode"
}
llm_base_url = _normalize_azure_url(llm_base_url)
try:
client = AsyncOpenAI(base_url=llm_base_url, api_key=llm_api_key or "EMPTY")
system_prompt = textwrap.dedent("""\
You are an impartial evaluator of automated agent task results with access to database verification. Based on the provided JSON trajectory AND the Python verification results from querying the database, decide the task outcome. This trajectory is generated by an MCP agent on a simulated simplified environment. The environment provides a set of MCP tools to help the agent complete the task.
Input:
task_json: dict containing task fields such as user task and agent execution_history.
verification_json: dict containing Python code, reasoning, success_criteria, failure_criteria, and execution results that verified the database state from the Python function
Output:
You only output UTF-8 encoded string, avoid any emoji or special characters. You only output English text.
Classification categories:
- complete: all required steps and closure actions were successfully executed, AND the database state confirms the task was completed
- incomplete: partial progress or the database state shows the task is not fully completed
- server_error: the agent is blocked by MCP server/environment error, e.g., 5xx errors such as "Internal Server Error". Or the MCP server cannot process the valid tool call and return valid results. This can block the agent from completing the task.
- agent_error: the agent made mistakes, invalid parameters, or missing required data without recovery, failed to complete the user's instruction.
Priority order for classification:
1) complete (trajectory shows success AND database confirms it)
2) server_error (due to the MCP server/environment error)
3) agent_error (agent-side issue, e.g., invalid mcp_tool_call arguments, hallucination, agent mistakes)
4) incomplete (everything else unfinished or database state doesn't match expected outcome)
Key considerations:
- The verification_json contains checks performed on the database states. You can read the verification code to understand what checks were performed on the database states.
- The verification_json contains the execution results of the verification code. You can use the execution results to help you judge the task completion.
- The verification_json contains guidance for how to utilize the code execution results to help you judge the task completion, also including success_criteria and failure_criteria.
- The verification results may be empty or error, or even the verification code itself is inaccurate. You should not fully rely on the verification results. You need to comprehensively consider the trajectory information to help you judge the task completion.
Output format (must be valid JSON, no markdown fences, no additional commentary):
{
"reasoning": "<concise explanation considering both trajectory and verification code execution results, the confidence score and considerations for each classification category>",
"confidence_score": [0-100, 0-100, 0-100, 0-100] for complete, incomplete, server_error, agent_error respectively,
"classification": "<one_of_[complete, incomplete, server_error, agent_error]>",
"evidence": {
"status": "<original result.status>",
"iterations": <int>,
"error_signals": ["<important error messages or codes>"],
"last_actions": ["<summaries of last few actions>"],
"last_status_codes": [<recent status codes>],
"database_verification": "<summary of what the database state changed based on code execution results of verification>"
}
}""")
task_payload = {
"user_task": task,
"actual_execution_steps": len(trajectory) if trajectory else 0,
"trajectory": trajectory or [],
}
verification_json = _sanitize_for_json(
{
"reasoning": verifier_reasoning,
"success_criteria": success_criteria,
"failure_criteria": failure_criteria,
"code_execution_result": verifier_result,
}
)
try:
verification_json_str = json.dumps(
verification_json, ensure_ascii=False, indent=2, default=str
)
except Exception:
verification_json_str = str(verification_json)
user_prompt = (
f"task_json:\n"
f"{json.dumps(task_payload, ensure_ascii=False, indent=2, default=str)}\n\n"
f"verification_json:\n"
f"{verification_json_str}"
)
response = await client.chat.completions.create(
model=llm_model,
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
],
temperature=1.0,
max_completion_tokens=4096,
)
content = response.choices[0].message.content or ""
try:
result = json.loads(content)
except json.JSONDecodeError:
match = re.search(r"\{[\s\S]*\}", content)
if match:
try:
result = json.loads(match.group())
except json.JSONDecodeError:
return "judge_error", {
"error": f"Failed to parse LLM response: {content}"
}
else:
return "judge_error", {
"error": f"Failed to parse LLM response: {content}"
}
classification = result.get("classification", "judge_error").lower().strip()
valid = {"complete", "incomplete", "server_error", "agent_error", "judge_error"}
if classification not in valid:
classification = "judge_error"
return classification, result
except Exception as e:
logger.error(f"LLM judge failed: {e}")
return "judge_error", {"error": str(e)}
def _sanitize_for_json(obj: Any) -> Any:
if isinstance(obj, dict):
return {
str(k) if isinstance(k, tuple) else k: _sanitize_for_json(v)
for k, v in obj.items()
}
elif isinstance(obj, (list, tuple)):
return [_sanitize_for_json(item) for item in obj]
elif isinstance(obj, bytes):
return obj.decode("utf-8", errors="ignore")
else:
try:
json.dumps(obj)
return obj
except TypeError:
return str(obj)
|