File size: 17,890 Bytes
803cb67 | 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 | """
LLM Judge grading engine for per-task container mode.
Runs the judge agent inside the same container where the task executed,
reusing the already-running OpenClaw gateway.
"""
from __future__ import annotations
import json
import logging
import os
import re
import subprocess
import tempfile
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, List, Optional
from dataclaw.utils.docker_utils import (
TMP_WORKSPACE,
collect_transcript,
detect_transcript_errors,
resolve_qualified_model,
)
logger = logging.getLogger(__name__)
DEFAULT_JUDGE_MODEL = os.environ.get(
"JUDGE_MODEL", "openrouter/anthropic/claude-opus-4.5"
)
MAX_JUDGE_PARSE_ATTEMPTS = 5
JUDGE_TIMEOUT_SECONDS = 180
@dataclass
class GradeResult:
task_id: str
score: float
max_score: float
grading_type: str
breakdown: Dict[str, float]
notes: str
def to_dict(self) -> Dict[str, Any]:
return {
"task_id": self.task_id,
"score": self.score,
"max_score": self.max_score,
"grading_type": self.grading_type,
"breakdown": self.breakdown,
"notes": self.notes,
}
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def grade_task(
*,
container_id: str,
task_id: str,
task_prompt: str,
expected_behavior: str,
grading_criteria: List[str],
llm_judge_rubric: Optional[str],
agent_transcript_path: Path,
output_dir: Path,
judge_model: str = DEFAULT_JUDGE_MODEL,
) -> GradeResult:
"""Grade a task using the LLM judge inside the task's container."""
final_text = _final_assistant_text(agent_transcript_path)
rubric = llm_judge_rubric or _format_grading_criteria(grading_criteria)
base_prompt = _build_judge_prompt(task_prompt, expected_behavior, final_text, rubric)
# Create judge agent inside the container
qualified_judge = resolve_qualified_model(container_id, judge_model)
agent_create = subprocess.run(
["docker", "exec", container_id, "openclaw", "agents", "add", "judge",
"--model", qualified_judge, "--non-interactive",
"--workspace", "/root/.openclaw/workspace"],
capture_output=True, text=True, encoding="utf-8",
)
if agent_create.returncode != 0:
logger.warning(
"[%s] Judge agent creation failed (exit=%s): %s",
task_id, agent_create.returncode,
(agent_create.stderr or agent_create.stdout or "").strip()[:500],
)
parsed: Dict[str, Any] = {"scores": {}, "total": None, "notes": ""}
last_err: Optional[str] = None
success = False
for attempt in range(1, MAX_JUDGE_PARSE_ATTEMPTS + 1):
if attempt > 1:
# Clear previous judge sessions before retry
subprocess.run(
["docker", "exec", container_id, "/bin/bash", "-c",
"rm -rf /root/.openclaw/agents/judge/sessions/*"],
capture_output=True, text=True, encoding="utf-8",
)
# Enhance prompt only when the previous failure was a format issue
if attempt > 1 and last_err and last_err.startswith("format"):
prompt = (
base_prompt
+ "\n\n---\nYour previous reply was not a valid scoring JSON with a numeric "
"`total`. Reply with ONLY a single JSON object, no markdown fences, no other "
"text, exactly this shape:\n"
'{"scores": {"criterion_name": 0.0}, "total": 0.0, "notes": "brief justification"}\n'
)
else:
prompt = base_prompt
# L1: runtime (subprocess / timeout)
try:
_run_judge_in_container(container_id, prompt)
except RuntimeError as exc:
last_err = f"runtime: {exc}"
logger.warning("[%s] Judge attempt %d/%d failed — %s",
task_id, attempt, MAX_JUDGE_PARSE_ATTEMPTS, last_err)
continue
transcript_path = collect_transcript(container_id, output_dir, agent_id="judge")
# L2: transcript validity
if not transcript_path.exists():
last_err = "transcript: file missing"
logger.warning("[%s] Judge attempt %d/%d failed — %s",
task_id, attempt, MAX_JUDGE_PARSE_ATTEMPTS, last_err)
continue
tr_err = detect_transcript_errors(transcript_path)
if tr_err:
last_err = f"transcript: {tr_err}"
logger.warning("[%s] Judge attempt %d/%d failed — %s",
task_id, attempt, MAX_JUDGE_PARSE_ATTEMPTS, last_err)
continue
try:
final_text = _final_assistant_text(transcript_path)
except OSError as exc:
last_err = f"transcript: read failed: {exc}"
logger.warning("[%s] Judge attempt %d/%d failed — %s",
task_id, attempt, MAX_JUDGE_PARSE_ATTEMPTS, last_err)
continue
if not final_text:
last_err = "transcript: empty assistant response"
logger.warning("[%s] Judge attempt %d/%d failed — %s",
task_id, attempt, MAX_JUDGE_PARSE_ATTEMPTS, last_err)
continue
# L3: format (required fields present)
try:
raw_parsed = _parse_judge_response_from_file(transcript_path)
except OSError as exc:
last_err = f"transcript: read failed: {exc}"
logger.warning("[%s] Judge attempt %d/%d failed — %s",
task_id, attempt, MAX_JUDGE_PARSE_ATTEMPTS, last_err)
continue
parsed = _normalize_judge_response(raw_parsed)
if _has_total_field(parsed):
logger.info("[%s] Judge returned valid score on attempt %d", task_id, attempt)
success = True
break
last_err = "format: missing or non-numeric total field"
logger.warning("[%s] Judge attempt %d/%d failed — %s",
task_id, attempt, MAX_JUDGE_PARSE_ATTEMPTS, last_err)
if not success:
raise RuntimeError(
f"Judge failed after {MAX_JUDGE_PARSE_ATTEMPTS} attempts: {last_err}"
)
breakdown = parsed.get("scores", {})
total = parsed.get("total")
notes = parsed.get("notes", "")
return GradeResult(
task_id=task_id,
score=float(total) if total is not None else 0.0,
max_score=1.0,
grading_type="llm_judge",
breakdown=_normalize_score_dict(breakdown),
notes=str(notes) if notes is not None else "",
)
# ---------------------------------------------------------------------------
# Container interaction
# ---------------------------------------------------------------------------
def _run_judge_in_container(container_id: str, message: str) -> None:
"""Send a message to the judge agent inside the container.
Uses docker cp + file read to bypass Windows command-line length limits.
Raises RuntimeError on docker cp failure, judge subprocess non-zero exit,
or timeout — so callers can detect and retry.
"""
fd, tmp_path = tempfile.mkstemp(suffix=".txt")
try:
with os.fdopen(fd, "w", encoding="utf-8") as f:
f.write(message)
cp_result = subprocess.run(
["docker", "cp", tmp_path,
f"{container_id}:/tmp/judge_prompt.txt"],
capture_output=True, text=True, encoding="utf-8",
)
if cp_result.returncode != 0:
raise RuntimeError(
f"docker cp failed (exit={cp_result.returncode}): "
f"{(cp_result.stderr or '').strip()[:300]}"
)
bash_cmd = (
f"prompt=$(cat /tmp/judge_prompt.txt) && "
f"cd {TMP_WORKSPACE} && "
f"openclaw agent --agent judge --session-id judge_chat "
f"--message \"$prompt\""
)
try:
result = subprocess.run(
["docker", "exec", container_id, "/bin/bash", "-c", bash_cmd],
capture_output=True, text=True, encoding="utf-8",
timeout=JUDGE_TIMEOUT_SECONDS,
)
except subprocess.TimeoutExpired as exc:
raise RuntimeError(
f"judge timed out after {JUDGE_TIMEOUT_SECONDS}s"
) from exc
if result.returncode != 0:
stderr_snippet = (result.stderr or result.stdout or "").strip()[:300]
raise RuntimeError(
f"judge subprocess exit={result.returncode}: {stderr_snippet}"
)
finally:
try:
os.unlink(tmp_path)
except OSError:
pass
# ---------------------------------------------------------------------------
# Transcript parsing
# ---------------------------------------------------------------------------
def _final_assistant_text(transcript_path: Path) -> str:
"""Extract text from the most recent assistant message that contains
text content.
Agents (especially tool-heavy ones) may have their last message be a
pure toolCall with no text — particularly when interrupted mid-call by
a timeout. In that case, fall back to the most recent assistant
message that actually has text content so the judge still sees the
agent's reasoning.
"""
if not transcript_path.exists():
return ""
last_text_message: Optional[Dict[str, Any]] = None
for line in transcript_path.read_text(encoding="utf-8").splitlines():
line = line.strip()
if not line:
continue
try:
entry = json.loads(line)
except json.JSONDecodeError:
continue
if entry.get("type") != "message":
continue
msg = entry.get("message", {})
if msg.get("role") != "assistant":
continue
content = msg.get("content", []) or []
has_text = any(
isinstance(item, dict)
and item.get("type") == "text"
and item.get("text")
for item in content
)
if has_text:
last_text_message = msg
if last_text_message is None:
return ""
parts: List[str] = []
for item in last_text_message.get("content", []) or []:
if isinstance(item, dict) and item.get("type") == "text" and item.get("text"):
parts.append(item["text"])
return "\n".join(parts).strip()
def _parse_judge_response_from_file(transcript_path: Path) -> Dict[str, Any]:
"""Parse the judge's JSON response from transcript JSONL."""
if not transcript_path.exists():
return {}
content_chunks: List[str] = []
for line in transcript_path.read_text(encoding="utf-8").splitlines():
line = line.strip()
if not line:
continue
try:
entry = json.loads(line)
except json.JSONDecodeError:
continue
if entry.get("type") != "message":
continue
msg = entry.get("message", {})
if msg.get("role") != "assistant":
continue
for item in msg.get("content", []):
if item.get("type") == "text":
content_chunks.append(item.get("text", ""))
raw_text = "\n".join(content_chunks).strip()
if not raw_text:
return {}
return _parse_json_from_text(raw_text)
def _parse_json_from_text(raw_text: str) -> Dict[str, Any]:
"""Extract a JSON object from free-form text."""
# Try code blocks first
code_block_match = re.search(r"```json\s*(.*?)\s*```", raw_text, re.DOTALL)
if code_block_match:
try:
parsed = json.loads(code_block_match.group(1))
if isinstance(parsed, dict):
return parsed
except json.JSONDecodeError:
pass
# Find balanced-brace JSON candidates
json_candidates: List[str] = []
brace_depth = 0
current_json: List[str] = []
for char in raw_text:
if char == "{":
if brace_depth == 0:
current_json = []
brace_depth += 1
if brace_depth > 0:
current_json.append(char)
if char == "}":
brace_depth -= 1
if brace_depth == 0 and current_json:
json_candidates.append("".join(current_json))
for candidate in reversed(json_candidates):
try:
parsed = json.loads(candidate)
if isinstance(parsed, dict) and "scores" in parsed:
return parsed
except json.JSONDecodeError:
continue
for candidate in reversed(json_candidates):
try:
parsed = json.loads(candidate)
if isinstance(parsed, dict):
return parsed
except json.JSONDecodeError:
continue
# Fallback: regex for "total: 0.XX"
score_pattern = re.search(
r"(?:total|overall|final)\s*(?:score)?[:\s]*(0\.\d+|1\.0+)",
raw_text,
re.IGNORECASE,
)
if score_pattern:
try:
total = float(score_pattern.group(1))
if 0.0 <= total <= 1.0:
return {"scores": {}, "total": total, "notes": "Score extracted from prose"}
except ValueError:
pass
logger.warning("Failed to parse judge JSON response")
return {}
# ---------------------------------------------------------------------------
# Normalization helpers
# ---------------------------------------------------------------------------
def _has_total_field(normalized: Dict[str, Any]) -> bool:
"""Format check: total field exists and is a numeric value (range not enforced)."""
if not isinstance(normalized, dict):
return False
total = normalized.get("total")
if total is None:
return False
try:
t = float(total)
except (TypeError, ValueError):
return False
if t != t: # NaN
return False
return True
def _normalize_judge_response(parsed: Dict[str, Any]) -> Dict[str, Any]:
"""Normalize judge response to {scores, total, notes}."""
result: Dict[str, Any] = {"scores": {}, "total": None, "notes": ""}
if "scores" in parsed:
scores_data = parsed["scores"]
if isinstance(scores_data, dict):
for key, value in scores_data.items():
if isinstance(value, dict) and "score" in value:
try:
result["scores"][key] = float(value["score"])
except (TypeError, ValueError):
pass
elif isinstance(value, (int, float)):
result["scores"][key] = value
elif "criteria_scores" in parsed:
criteria = parsed["criteria_scores"]
if isinstance(criteria, dict):
for key, value in criteria.items():
if isinstance(value, dict) and "score" in value:
result["scores"][key] = value["score"]
elif isinstance(value, (int, float)):
result["scores"][key] = value
if "total" in parsed and parsed["total"] is not None:
try:
result["total"] = float(parsed["total"])
except (TypeError, ValueError):
pass
elif "score" in parsed and isinstance(parsed["score"], (int, float)):
result["total"] = float(parsed["score"])
elif "overall_score" in parsed and isinstance(parsed["overall_score"], (int, float)):
result["total"] = float(parsed["overall_score"])
elif result["scores"]:
values = [v for v in result["scores"].values() if isinstance(v, (int, float))]
if values:
result["total"] = sum(values) / len(values)
if "notes" in parsed:
result["notes"] = str(parsed["notes"])
elif "justification" in parsed:
result["notes"] = str(parsed["justification"])
elif "reasoning" in parsed:
result["notes"] = str(parsed["reasoning"])
return result
def _normalize_score_dict(scores: Dict[str, Any]) -> Dict[str, float]:
normalized: Dict[str, float] = {}
for key, value in scores.items():
try:
normalized[str(key)] = float(value)
except (TypeError, ValueError):
continue
return normalized
# ---------------------------------------------------------------------------
# Prompt building
# ---------------------------------------------------------------------------
def _format_grading_criteria(criteria: List[str]) -> str:
if not criteria:
return ""
return "\n".join(f"- {c}" for c in criteria)
def _build_judge_prompt(
task_prompt: str,
expected_behavior: str,
agent_final_text: str,
rubric: str,
) -> str:
return (
"You are a grading function. Your ONLY job is to output a single JSON object.\n\n"
"CRITICAL RULES:\n"
"- Do NOT use any tools (no Read, Write, exec, or any other tool calls)\n"
"- Do NOT create files or run commands\n"
"- Do NOT write any prose, explanation, or commentary outside the JSON\n"
"- Respond with ONLY a JSON object — nothing else\n\n"
"Be a strict evaluator. Judge the final assistant message against the task and rubric.\n\n"
"## Task\n"
f"{task_prompt}\n\n"
"## Expected Behavior\n"
f"{expected_behavior}\n\n"
"## Agent final answer\n"
f"{agent_final_text}\n\n"
"## Grading Rubric\n"
f"{rubric}\n\n"
"Score each criterion from 0.0 to 1.0.\n\n"
"Respond with ONLY this JSON structure (no markdown, no code fences, no extra text):\n"
'{"scores": {"criterion_name": 0.0}, "total": 0.0, "notes": "brief justification"}'
)
|