File size: 62,283 Bytes
e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 22173b0 e0017e6 61b5f58 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 bec7c31 22173b0 61b5f58 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 22173b0 e0017e6 61b5f58 22173b0 e0017e6 22173b0 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 bec7c31 e0017e6 61b5f58 bec7c31 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 61b5f58 e0017e6 bec7c31 61b5f58 bec7c31 e0017e6 61b5f58 e0017e6 61b5f58 bec7c31 61b5f58 bec7c31 61b5f58 e0017e6 bec7c31 3da1016 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 | """
Workflow helper functions for the multi-role AI workflow.
Contains:
- WorkflowConfig: configuration flags (strict_mode, allow_persona_roles, etc.)
- Output format intent detection
- Structured QA result parsing
- Role relevance metadata and selection
- Targeted revision logic
- Final answer compression / noise stripping
- PlannerState management
"""
import json
import re
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple
# ============================================================
# Workflow Configuration
# ============================================================
@dataclass
class WorkflowConfig:
"""Runtime config flags for the multi-role workflow."""
allow_persona_roles: bool = False
max_specialists_per_task: int = 3
strict_mode: bool = True
always_include_qa: bool = True
always_include_research_for_factual_tasks: bool = True
require_evidence_for_factual_claims: bool = True
# Internal: which role keys are persona/gimmick roles
PERSONA_ROLE_KEYS: tuple = (
"mad_professor", "accountant", "artist", "lazy_slacker",
"black_metal_fundamentalist", "doris", "maga_appointee",
"chairman_of_board",
)
# Internal: core professional role keys
CORE_ROLE_KEYS: tuple = (
"creative", "technical", "research", "security",
"data_analyst", "labour_union_rep", "ux_designer", "lawyer",
)
DEFAULT_CONFIG = WorkflowConfig()
# ============================================================
# Task Classification
# ============================================================
TASK_CATEGORIES = [
"factual_question",
"comparison",
"coding_task",
"creative_writing",
"opinion_discussion",
"summarization",
"analysis",
"planning",
]
_TASK_CATEGORY_PATTERNS = [
("coding_task", [
r"\bwrite\s+(python|code|javascript|typescript|rust|java|c\+\+|go|bash|sql)\b",
r"\bcode\s+(to|for|that)\b", r"\bimplement\b",
r"\bscript\s+(to|for|that)\b", r"\bparse\s+a?\s*\w+\s+(file|data)\b",
r"\bdebug\b", r"\brefactor\b", r"\bfix\s+(the|this|my)\s+(code|bug)\b",
]),
("creative_writing", [
r"\bwrite\s+a\b.*\b(poem|story|essay|blog|article|song|haiku)\b",
r"\bcreative\s+writing\b", r"\bbrainstorm\b", r"\bimagine\b",
r"\bfiction\b", r"\bnarrative\b",
]),
("factual_question", [
r"\bwhat\s+(is|are|was|were)\b", r"\bwho\s+(is|was|are|were)\b",
r"\bwhen\s+(did|was|is)\b", r"\bwhere\s+(is|was|are)\b",
r"\bhow\s+many\b", r"\bhow\s+much\b",
r"\bnews\b", r"\brecent\b", r"\blatest\b", r"\bcurrent\b",
r"\bfact\b", r"\btrue\s+or\s+false\b",
]),
("comparison", [
r"\bcompar(e|ison|ing)\b", r"\bvs\.?\b", r"\bversus\b",
r"\bdifference\s+between\b", r"\bbetter\s+than\b",
r"\bwhich\s+is\s+(better|faster|cheaper)\b",
r"\bpros?\s+and\s+cons?\b", r"\btrade[\s-]?offs?\b",
]),
("summarization", [
r"\bsummar(y|ize|ise)\b", r"\btl;?dr\b", r"\bsynopsis\b",
r"\boverview\b", r"\brecap\b",
]),
("analysis", [
r"\banaly(sis|se|ze)\b", r"\bevaluat(e|ion)\b",
r"\bassess(ment)?\b", r"\breview\b",
r"\bexamin(e|ation)\b", r"\binvestigat(e|ion)\b",
]),
("planning", [
r"\bplan\b", r"\bstrateg(y|ic)\b", r"\broadmap\b",
r"\baction\s+items?\b", r"\bsteps?\s+to\b",
]),
("opinion_discussion", [
r"\bdiscuss\b", r"\bopinion\b", r"\bperspective\b",
r"\bpoint\s+of\s+view\b", r"\bargue\b", r"\bdebate\b",
r"\brole\s+of\b",
]),
]
def classify_task(user_request: str) -> str:
"""Classify the user's request into a task category.
Returns one of: factual_question, comparison, coding_task, creative_writing,
opinion_discussion, summarization, analysis, planning, other.
"""
lower = user_request.lower()
best_category = "other"
best_score = 0
for category, patterns in _TASK_CATEGORY_PATTERNS:
score = 0
for pat in patterns:
if re.search(pat, lower):
score += 1
if score > best_score:
best_score = score
best_category = category
return best_category
def task_needs_evidence(task_category: str) -> bool:
"""Whether this task category benefits from tool-backed evidence retrieval."""
return task_category in ("factual_question", "comparison", "analysis", "summarization")
# ============================================================
# Output Format Detection
# ============================================================
# Ordered list of (format_name, patterns) — first match wins
_FORMAT_PATTERNS = [
("single_choice", [
r"\bpick\s+one\b", r"\bchoose\s+one\b", r"\bagree\s+on\s+one\b",
r"\bselect\s+one\b", r"\bjust\s+one\b", r"\bone\s+choice\b",
r"\bwhich\s+one\b", r"\bone\s+word\b",
]),
("code", [
r"\bwrite\s+(python|code|javascript|typescript|rust|java|c\+\+|go|bash|sql)\b",
r"\bcode\s+(to|for|that)\b", r"\bimplement\b.*\b(function|class|method|script)\b",
r"\bgive\s+me\s+(the\s+)?code\b", r"\bscript\s+(to|for|that)\b",
]),
("table", [
r"\bmake\s+a\s+table\b", r"\bcreate\s+a\s+table\b", r"\btable\s+comparing\b",
r"\bcomparison\s+table\b", r"\btabular\b", r"\bin\s+table\s+form\b",
]),
("bullet_list", [
r"\bbullet\s*(ed)?\s*(point|list)\b", r"\blist\s+(the|all|some)\b",
r"\bgive\s+me\s+a\s+list\b",
]),
("short_answer", [
r"\bshort\s+answer\b", r"\bbrief(ly)?\b", r"\bconcise(ly)?\b",
r"\bin\s+one\s+sentence\b", r"\byes\s+or\s+no\b", r"\bquick\s+answer\b",
r"\banswer\s+briefly\b", r"\bkeep\s+it\s+short\b",
]),
("report", [
r"\breport\b", r"\banalysis\b", r"\bin[\s-]depth\b", r"\bdetailed\b",
r"\bcomprehensive\b", r"\btrade[\s-]?offs?\b", r"\bpros?\s+and\s+cons?\b",
]),
("paragraph", [
r"\bexplain\b", r"\bdescribe\b", r"\bparagraph\b",
]),
]
def detect_output_format(user_request: str) -> str:
"""Classify the expected output format from the user's request text.
Returns one of: single_choice, short_answer, paragraph, bullet_list,
table, report, code, other.
"""
lower = user_request.lower()
for fmt, patterns in _FORMAT_PATTERNS:
for pat in patterns:
if re.search(pat, lower):
return fmt
return "other"
def detect_brevity_requirement(user_request: str) -> str:
"""Detect how brief the answer should be.
Returns: 'minimal', 'short', 'normal', or 'verbose'.
"""
lower = user_request.lower()
minimal_signals = [
r"\bjust\s+(one|the)\b", r"\bone\s+word\b", r"\byes\s+or\s+no\b",
r"\bpick\s+one\b", r"\bchoose\s+one\b", r"\bagree\s+on\s+one\b",
]
for pat in minimal_signals:
if re.search(pat, lower):
return "minimal"
short_signals = [
r"\bshort\b", r"\bbrief(ly)?\b", r"\bconcise(ly)?\b",
r"\bquick\b", r"\bsimple\b", r"\bkeep\s+it\s+short\b",
]
for pat in short_signals:
if re.search(pat, lower):
return "short"
verbose_signals = [
r"\bdetailed\b", r"\bin[\s-]depth\b", r"\bcomprehensive\b",
r"\bthorough(ly)?\b", r"\bfull\s+report\b",
]
for pat in verbose_signals:
if re.search(pat, lower):
return "verbose"
return "normal"
# ============================================================
# Structured QA Result
# ============================================================
@dataclass
class QAIssue:
type: str # format | brevity | constraint | consistency | directness | other
message: str
owner: str # "Synthesizer" | "Planner" | specialist role display name
@dataclass
class QAResult:
status: str # "PASS" | "PASS_WITH_WARNINGS" | "FAIL"
reason: str = ""
issues: List[QAIssue] = field(default_factory=list)
warnings: List[str] = field(default_factory=list)
correction_instruction: str = ""
@property
def passed(self) -> bool:
return self.status in ("PASS", "PASS_WITH_WARNINGS")
@property
def passed_with_warnings(self) -> bool:
return self.status == "PASS_WITH_WARNINGS"
def owners(self) -> List[str]:
"""Return unique owner labels from issues."""
return list(dict.fromkeys(issue.owner for issue in self.issues))
def to_dict(self) -> dict:
return {
"status": self.status,
"reason": self.reason,
"issues": [
{"type": i.type, "message": i.message, "owner": i.owner}
for i in self.issues
],
"warnings": self.warnings,
"correction_instruction": self.correction_instruction,
}
def parse_structured_qa(qa_text: str) -> QAResult:
"""Parse QA output into a structured QAResult.
Tries JSON first (if QA produced structured output),
then falls back to the legacy text format.
"""
# Try to extract JSON from the QA output
json_match = re.search(r'\{[\s\S]*"status"\s*:', qa_text)
if json_match:
# Find the matching closing brace
start = json_match.start()
brace_count = 0
end = start
for i, ch in enumerate(qa_text[start:], start):
if ch == '{':
brace_count += 1
elif ch == '}':
brace_count -= 1
if brace_count == 0:
end = i + 1
break
try:
data = json.loads(qa_text[start:end])
issues = []
for item in data.get("issues", []):
issues.append(QAIssue(
type=item.get("type", "other"),
message=item.get("message", ""),
owner=item.get("owner", "Synthesizer"),
))
return QAResult(
status=data.get("status", "FAIL"),
reason=data.get("reason", ""),
issues=issues,
warnings=[str(w) for w in data.get("warnings", [])],
correction_instruction=data.get("correction_instruction", ""),
)
except (json.JSONDecodeError, KeyError):
pass
# Fallback: parse from legacy text format
status = "FAIL"
lower = qa_text.lower()
if "result: pass_with_warnings" in lower:
status = "PASS_WITH_WARNINGS"
elif "result: pass" in lower:
status = "PASS"
reason = ""
if "ISSUES FOUND:" in qa_text:
section = qa_text.split("ISSUES FOUND:", 1)[1]
for header in ("ROLE-SPECIFIC FEEDBACK:", "RESULT:", "RECOMMENDED FIXES:"):
if header in section:
section = section.split(header, 1)[0]
break
reason = section.strip()
correction = ""
if "RECOMMENDED FIXES:" in qa_text:
correction = qa_text.split("RECOMMENDED FIXES:", 1)[1].strip()
# Build issues from role-specific feedback section
issues = []
if "ROLE-SPECIFIC FEEDBACK:" in qa_text:
fb_section = qa_text.split("ROLE-SPECIFIC FEEDBACK:", 1)[1]
for header in ("RESULT:", "RECOMMENDED FIXES:"):
if header in fb_section:
fb_section = fb_section.split(header, 1)[0]
break
for line in fb_section.strip().splitlines():
line = line.strip().lstrip("•-* ")
if ":" not in line:
continue
role_label, _, feedback = line.partition(":")
feedback = feedback.strip()
if feedback and feedback.lower() not in ("satisfactory", "n/a", "none"):
issues.append(QAIssue(
type="other",
message=feedback,
owner=role_label.strip(),
))
# If no role-specific issues but we have a general reason, attribute to Synthesizer
if not issues and reason and reason.lower() not in ("none", "n/a", "none."):
issues.append(QAIssue(type="other", message=reason, owner="Synthesizer"))
return QAResult(
status=status,
reason=reason,
issues=issues,
correction_instruction=correction,
)
# ============================================================
# Role Relevance Metadata and Selection
# ============================================================
# Each role has keywords/domains that indicate when it's relevant
ROLE_RELEVANCE: Dict[str, Dict[str, Any]] = {
"creative": {
"keywords": ["brainstorm", "ideas", "creative", "naming", "slogan", "marketing",
"framing", "wording", "concept", "design", "brand"],
"domains": ["marketing", "content", "writing", "communication"],
"description": "Ideas, framing, wording, brainstorming",
"role_type": "creative",
"task_types": ["creative_writing", "opinion_discussion"],
},
"technical": {
"keywords": ["code", "implement", "build", "architecture", "api", "database",
"debug", "software", "programming", "algorithm", "system", "deploy",
"python", "javascript", "rust", "java", "react", "vue", "svelte",
"framework", "library", "performance", "faster"],
"domains": ["engineering", "development", "devops", "infrastructure"],
"description": "Code, architecture, implementation, technical solutions",
"role_type": "factual",
"task_types": ["coding_task", "analysis"],
},
"research": {
"keywords": ["research", "study", "evidence", "literature", "paper", "facts",
"history", "compare", "comparison", "analysis", "data", "statistics",
"science", "scientific", "information"],
"domains": ["academia", "science", "fact-finding"],
"description": "Information gathering, literature review, fact-finding",
"role_type": "factual",
"task_types": ["factual_question", "comparison", "analysis", "summarization"],
},
"security": {
"keywords": ["security", "vulnerability", "attack", "encryption", "auth",
"password", "exploit", "firewall", "compliance", "gdpr", "privacy"],
"domains": ["cybersecurity", "infosec", "compliance"],
"description": "Security analysis, vulnerability checks, best practices",
"role_type": "safety",
"task_types": ["analysis"],
},
"data_analyst": {
"keywords": ["data", "analytics", "statistics", "pattern", "trend", "metric",
"dashboard", "visualization", "dataset", "csv", "spreadsheet"],
"domains": ["analytics", "business intelligence"],
"description": "Data analysis, statistics, pattern recognition, insights",
"role_type": "analytical",
"task_types": ["analysis", "comparison", "factual_question"],
},
"labour_union_rep": {
"keywords": ["worker", "wages", "union", "labor", "labour", "employment",
"rights", "workplace", "collective", "bargaining", "fair"],
"domains": ["labor relations", "HR", "workplace policy"],
"description": "Worker rights, fair wages, job security",
"role_type": "analytical",
"task_types": ["opinion_discussion", "analysis"],
},
"ux_designer": {
"keywords": ["user", "usability", "accessibility", "interface", "ux", "ui",
"design", "wireframe", "prototype", "user experience", "user-friendly"],
"domains": ["design", "product", "UX"],
"description": "User needs, usability, accessibility",
"role_type": "analytical",
"task_types": ["analysis", "planning"],
},
"lawyer": {
"keywords": ["legal", "law", "contract", "liability", "compliance", "regulation",
"patent", "copyright", "trademark", "lawsuit", "litigation"],
"domains": ["law", "compliance", "governance"],
"description": "Legal compliance, liability, contracts, risk management",
"role_type": "analytical",
"task_types": ["analysis"],
},
# Persona roles — only active when allow_persona_roles is True
"mad_professor": {
"keywords": ["crazy", "radical", "hypothesis", "experiment", "breakthrough"],
"domains": ["speculation"],
"description": "Radical scientific hypotheses, extreme speculation",
"role_type": "persona",
"task_types": [],
"is_persona": True,
},
"accountant": {
"keywords": ["cost", "budget", "expense", "cheap", "price", "financial"],
"domains": ["finance"],
"description": "Cost scrutiny, budget optimization",
"role_type": "persona",
"task_types": [],
"is_persona": True,
},
"artist": {
"keywords": ["art", "inspiration", "vision", "aesthetic", "beauty"],
"domains": ["art"],
"description": "Unhinged creative vision, cosmic vibes",
"role_type": "persona",
"task_types": ["creative_writing"],
"is_persona": True,
},
"lazy_slacker": {
"keywords": ["lazy", "shortcut", "easy", "simple", "quick"],
"domains": [],
"description": "Minimum viable effort, shortcuts",
"role_type": "persona",
"task_types": [],
"is_persona": True,
},
"black_metal_fundamentalist": {
"keywords": ["metal", "kvlt", "underground", "nihilism"],
"domains": [],
"description": "Nihilistic kvlt critique",
"role_type": "persona",
"task_types": [],
"is_persona": True,
},
"doris": {
"keywords": [],
"domains": [],
"description": "Well-meaning but clueless observations",
"role_type": "persona",
"task_types": [],
"is_persona": True,
},
"chairman_of_board": {
"keywords": ["shareholder", "board", "governance", "strategic", "corporate"],
"domains": ["corporate governance"],
"description": "Corporate governance, shareholder value",
"role_type": "persona",
"task_types": [],
"is_persona": True,
},
"maga_appointee": {
"keywords": ["america", "patriot", "deregulation"],
"domains": [],
"description": "America First perspective",
"role_type": "persona",
"task_types": [],
"is_persona": True,
},
}
# ============================================================
# Role Capability Metadata and Task Feature Extraction
# ============================================================
# Simple capability tags per role — used for transparent scoring
ROLE_CAPABILITIES: Dict[str, List[str]] = {
"creative": ["creative", "design", "ideas", "writing", "brainstorm", "opinion"],
"technical": ["technical", "analysis", "engineering", "calculation", "code", "implementation"],
"research": ["research", "facts", "evidence", "information", "comparison", "history"],
"security": ["risk", "safety", "compliance", "security", "vulnerability"],
"data_analyst": ["data", "statistics", "analysis", "metrics", "patterns"],
"labour_union_rep": ["labor", "policy", "workplace", "rights", "fairness"],
"ux_designer": ["design", "usability", "interface", "user_experience", "accessibility"],
"lawyer": ["legal", "compliance", "contracts", "liability", "regulation"],
"mad_professor": ["persona", "speculation", "radical", "humor"],
"accountant": ["persona", "cost", "budget", "financial"],
"artist": ["persona", "creative", "aesthetic", "vision"],
"lazy_slacker": ["persona", "simple_answer", "minimal"],
"black_metal_fundamentalist": ["persona", "stylistic", "humor", "music"],
"doris": ["persona", "humor"],
"chairman_of_board": ["persona", "strategy", "corporate", "governance"],
"maga_appointee": ["persona", "political", "deregulation"],
}
# Keywords in the user request that map to task features
TASK_FEATURE_KEYWORDS: Dict[str, List[str]] = {
"analysis": ["analy", "evaluate", "assess", "review", "examine", "investigate"],
"creative": ["creative", "brainstorm", "ideas", "imagine", "invent", "story", "write a"],
"design": ["design", "wireframe", "prototype", "layout", "visual", "ui", "ux", "interface", "usability",
"user experience", "accessibility", "login page", "user interface"],
"technical": ["code", "implement", "build", "architecture", "api", "debug", "software", "program",
"algorithm", "system", "deploy", "python", "javascript", "rust", "java", "react",
"framework", "performance"],
"research": ["research", "study", "evidence", "literature", "paper", "facts", "history",
"information", "find out", "look up"],
"policy": ["policy", "regulation", "law", "compliance", "legal", "rights", "labor", "labour",
"union", "worker", "employment", "workplace"],
"simple_answer": ["yes or no", "pick one", "choose", "which one", "red or blue", "agree on one"],
"opinion": ["opinion", "perspective", "viewpoint", "discuss", "debate", "pros and cons",
"should i", "what do you think", "agree", "disagree"],
"comparison": ["compare", "comparison", "versus", "vs", "difference", "better"],
"data": ["data", "statistics", "metric", "trend", "pattern", "chart", "dashboard", "csv",
"spreadsheet", "dataset"],
"security": ["security", "vulnerability", "attack", "encryption", "password", "exploit",
"firewall", "gdpr", "privacy"],
"cost": ["cost", "budget", "expense", "cheap", "price", "financial", "roi"],
"humor": ["funny", "joke", "humorous", "kvlt", "metal", "nihil"],
"music": ["music", "metal", "band", "song", "album", "guitar"],
}
# Generalist fallback roles used when no capability matches
_GENERALIST_ROLES = ("creative", "technical", "research")
def extract_task_features(user_request: str, task_category: str = "other") -> List[str]:
"""Derive task features from the user request and task category.
Returns a deduplicated list of feature tags like ["design", "opinion"].
"""
lower = user_request.lower()
features: List[str] = []
for feature, keywords in TASK_FEATURE_KEYWORDS.items():
for kw in keywords:
if kw in lower:
features.append(feature)
break # one match per feature is enough
# Add features implied by the task category
category_features: Dict[str, List[str]] = {
"coding_task": ["technical", "code"],
"creative_writing": ["creative"],
"factual_question": ["research"],
"comparison": ["comparison", "research"],
"analysis": ["analysis"],
"summarization": ["research"],
"opinion_discussion": ["opinion"],
"planning": ["analysis"],
}
for f in category_features.get(task_category, []):
if f not in features:
features.append(f)
return features
@dataclass
class RoleScore:
"""Scoring details for a single role — used for transparent logging."""
role_key: str
role_label: str
score: int
matched_capabilities: List[str]
is_persona: bool
filtered_reason: str = "" # why it was excluded, if any
def score_roles(
task_features: List[str],
active_role_keys: List[str],
config: WorkflowConfig,
task_category: str = "other",
) -> List[RoleScore]:
"""Score each active role by capability overlap with task features.
Returns all RoleScore objects (including filtered ones) for transparency.
"""
feature_set = set(task_features)
results: List[RoleScore] = []
# Import here to avoid circular — role labels come from the caller
for role_key in active_role_keys:
capabilities = ROLE_CAPABILITIES.get(role_key, [])
meta = ROLE_RELEVANCE.get(role_key, {})
is_persona = meta.get("is_persona", False)
role_label = meta.get("description", role_key)
# Capability overlap score
matched = [cap for cap in capabilities if cap in feature_set]
score = len(matched)
# Task-category affinity bonus (from ROLE_RELEVANCE)
role_tasks = meta.get("task_types", [])
if task_category in role_tasks:
score += 2
rs = RoleScore(
role_key=role_key,
role_label=role_label,
score=score,
matched_capabilities=matched,
is_persona=is_persona,
)
# Filter personas unless allowed
if is_persona and not config.allow_persona_roles:
rs.filtered_reason = "persona role not allowed"
results.append(rs)
return results
def select_relevant_roles(
user_request: str,
active_role_keys: List[str],
config: WorkflowConfig,
task_category: str = "other",
) -> List[str]:
"""Select the most relevant specialist roles for a given request.
Uses capability-based scoring: extracts task features from the request,
scores each active role by capability overlap, and returns the top roles
up to ``config.max_specialists_per_task``.
Returns a ``_SelectionResult`` (list subclass) so callers can also access
``.scoring_info`` for transparent trace logging.
"""
task_features = extract_task_features(user_request, task_category)
role_scores = score_roles(task_features, active_role_keys, config, task_category)
# Separate eligible from filtered
eligible = [rs for rs in role_scores if not rs.filtered_reason]
eligible.sort(key=lambda rs: (-rs.score, active_role_keys.index(rs.role_key)))
selected: List[str] = []
for rs in eligible:
if len(selected) >= config.max_specialists_per_task:
break
# In strict mode, only include roles with score > 0 (unless we have none yet)
if config.strict_mode and rs.score == 0 and selected:
continue
selected.append(rs.role_key)
# Ensure at least one specialist is selected
if not selected and eligible:
selected.append(eligible[0].role_key)
# Generalist fallback when nothing matched
if not selected:
for rk in _GENERALIST_ROLES:
if rk in active_role_keys:
selected.append(rk)
break
# Auto-include research for factual tasks
if (config.always_include_research_for_factual_tasks
and task_needs_evidence(task_category)
and "research" in active_role_keys
and "research" not in selected):
selected.append("research")
return _SelectionResult(selected, role_scores, task_features)
class _SelectionResult(list):
"""A list of role keys with attached scoring metadata.
Behaves exactly like a ``list[str]`` so existing code continues to work,
but also carries ``scoring_info`` and ``task_features`` for trace logging.
"""
def __init__(
self,
selected: List[str],
scoring_info: List[RoleScore],
task_features: List[str],
):
super().__init__(selected)
self.scoring_info = scoring_info
self.task_features = task_features
def format_trace(self, role_labels: Optional[Dict[str, str]] = None) -> str:
"""Return a human-readable ROLE SCORING trace block."""
lines = ["── ROLE SCORING ──"]
lines.append(f"Task features: {self.task_features}")
for rs in sorted(self.scoring_info, key=lambda r: -r.score):
label = (role_labels or {}).get(rs.role_key, rs.role_key)
status = "SELECTED" if rs.role_key in self else "skipped"
if rs.filtered_reason:
status = f"FILTERED ({rs.filtered_reason})"
caps = ", ".join(rs.matched_capabilities) if rs.matched_capabilities else "none"
lines.append(
f" {label}: score={rs.score} caps=[{caps}] → {status}"
)
lines.append("──────────────────")
return "\n".join(lines)
# ============================================================
# Targeted Revision Logic
# ============================================================
def identify_revision_targets(
qa_result: QAResult,
role_label_to_key: Dict[str, str],
) -> List[str]:
"""Given a QAResult, return the list of role keys that need rerunning.
Rules:
- Format/brevity issues → Synthesizer only (returned as "synthesizer")
- Issues owned by a specific specialist → that specialist key
- Issues owned by Planner → "planner"
- If no clear owner → "synthesizer" (default)
"""
targets = []
for issue in qa_result.issues:
owner = issue.owner.strip()
if owner.lower() in ("synthesizer", "synthesis"):
if "synthesizer" not in targets:
targets.append("synthesizer")
elif owner.lower() == "planner":
if "planner" not in targets:
targets.append("planner")
else:
# Try to resolve the owner label to a role key
key = role_label_to_key.get(owner)
if key and key not in targets:
targets.append(key)
elif "synthesizer" not in targets:
# Unrecognised owner — attribute to synthesizer
targets.append("synthesizer")
# Format/brevity issues → always include synthesizer
for issue in qa_result.issues:
if issue.type in ("format", "brevity", "directness") and "synthesizer" not in targets:
targets.append("synthesizer")
if not targets:
targets.append("synthesizer")
return targets
# ============================================================
# Final Answer Compression / Noise Stripping
# ============================================================
# Sections that are internal workflow noise and should never appear in final output
_INTERNAL_NOISE_HEADERS = [
"TASK BREAKDOWN:", "ROLE TO CALL:", "SUCCESS CRITERIA:",
"GUIDANCE FOR SPECIALIST:", "PERSPECTIVES SUMMARY:",
"COMMON GROUND:", "TENSIONS AND TRADE-OFFS:",
"REQUIREMENTS CHECKED:", "ISSUES FOUND:",
"ROLE-SPECIFIC FEEDBACK:", "RESULT:", "RECOMMENDED FIXES:",
"DECISION: APPROVED", "DECISION: REVISE",
"REVISED INSTRUCTIONS:", "REVISION",
"SOURCES CONSULTED:", "SECURITY ANALYSIS:",
"VULNERABILITIES FOUND:", "COST ANALYSIS:",
"COST-CUTTING MEASURES:", "CHEAPEST VIABLE APPROACH:",
"KVLT VERDICT:", "WHAT THE MAINSTREAM GETS WRONG:",
"COSMIC VISION:", "DO WE EVEN NEED TO DO THIS:",
"WORKER IMPACT:", "UNION CONCERNS:",
"BOARD PERSPECTIVE:", "STRATEGIC CONCERNS:",
"AMERICA FIRST ANALYSIS:", "DEEP STATE CONCERNS:",
"LEGAL ANALYSIS:", "LIABILITIES AND RISKS:",
"WHAT DORIS THINKS IS HAPPENING:",
]
def strip_internal_noise(text: str) -> str:
"""Remove internal workflow headers/sections from text intended for the user."""
lines = text.split("\n")
cleaned = []
skip_until_next = False
for line in lines:
stripped = line.strip()
# Check if this line is an internal header
is_noise = False
for header in _INTERNAL_NOISE_HEADERS:
if stripped.startswith(header):
is_noise = True
skip_until_next = True
break
if is_noise:
continue
# If we were skipping noise, stop when we hit a non-empty non-header line
# that looks like actual content (not a sub-bullet of the skipped section)
if skip_until_next:
if stripped == "":
continue
# New section header that is NOT noise means we stop skipping
if stripped.endswith(":") and stripped == stripped.upper():
skip_until_next = False
elif not stripped.startswith("•") and not stripped.startswith("-") and not stripped.startswith("*"):
skip_until_next = False
if not skip_until_next:
cleaned.append(line)
result = "\n".join(cleaned).strip()
return result if result else text
def compress_final_answer(
draft: str,
output_format: str,
brevity: str,
user_request: str,
) -> str:
"""Apply rule-based compression to the final answer.
This does NOT call the LLM — it applies deterministic rules to trim
the answer. The LLM-based compression happens in the synthesizer.
"""
# Strip internal noise
answer = strip_internal_noise(draft)
# For single_choice: try to extract just the choice
if output_format == "single_choice" and brevity == "minimal":
# Look for a UNIFIED RECOMMENDATION section or similar
for marker in ("UNIFIED RECOMMENDATION:", "RECOMMENDED DRAFT:", "FINAL ANSWER:"):
if marker in answer:
answer = answer.split(marker, 1)[1].strip()
# Take only the first paragraph
paragraphs = answer.split("\n\n")
if paragraphs:
answer = paragraphs[0].strip()
break
# For short_answer: limit length
if output_format == "short_answer" or brevity in ("minimal", "short"):
# If the answer is very long relative to what was requested, truncate sensibly
if len(answer) > 500 and brevity == "minimal":
# Take the first meaningful paragraph
paragraphs = [p.strip() for p in answer.split("\n\n") if p.strip()]
if paragraphs:
answer = paragraphs[0]
return answer
def postprocess_format_fixes(text: str) -> str:
"""Apply lightweight format fixes before QA evaluation.
Converts common formatting artefacts so QA can focus on content quality
rather than failing for cosmetic issues.
"""
# Remove markdown headings (# / ## / ###)
text = re.sub(r'^#{1,4}\s+', '', text, flags=re.MULTILINE)
# Convert bullet-list lines to flowing sentences
def _bullets_to_sentences(m: re.Match) -> str:
lines = m.group(0).strip().splitlines()
sentences = []
for line in lines:
cleaned = re.sub(r'^\s*[-•*]\s+', '', line).strip()
if cleaned:
# Ensure it ends with a full stop
if cleaned[-1] not in '.!?':
cleaned += '.'
sentences.append(cleaned)
return ' '.join(sentences)
text = re.sub(
r'(?:^\s*[-•*]\s+.+\n?){2,}',
_bullets_to_sentences,
text,
flags=re.MULTILINE,
)
# Collapse runs of 3+ blank lines into 2
text = re.sub(r'\n{3,}', '\n\n', text)
# Remove leftover JSON-like traces (e.g. {"status": ...} blocks)
text = re.sub(r'\{[^{}]*"status"\s*:[^{}]*\}', '', text)
return text.strip()
# ============================================================
# Planner State
# ============================================================
@dataclass
class FailureRecord:
"""Record of a single failure in the workflow."""
revision: int
owner: str # role key or "synthesizer"
issue_type: str # from QAIssue.type
message: str
correction: str
def to_dict(self) -> dict:
return {
"revision": self.revision,
"owner": self.owner,
"issue_type": self.issue_type,
"message": self.message,
"correction": self.correction,
}
@dataclass
class PlannerState:
"""Persistent state object that tracks the planner's decisions through revisions.
This is the central working memory for the workflow.
All stages read from and write to this shared state.
"""
user_request: str = ""
task_summary: str = ""
task_category: str = "other"
success_criteria: List[str] = field(default_factory=list)
output_format: str = "other"
brevity_requirement: str = "normal"
selected_roles: List[str] = field(default_factory=list)
specialist_outputs: Dict[str, str] = field(default_factory=dict)
evidence: Optional[Dict] = None # serialised EvidenceResult
task_assumptions: Dict[str, str] = field(default_factory=dict)
current_draft: str = ""
qa_result: Optional[QAResult] = None
revision_count: int = 0
max_revisions: int = 3
revision_instruction: str = "" # latest revision instruction from planner
failure_history: List[FailureRecord] = field(default_factory=list)
history: List[Dict[str, str]] = field(default_factory=list)
final_answer: str = ""
def record_event(self, event_type: str, detail: str):
self.history.append({"type": event_type, "detail": detail[:500]})
def record_failure(self, qa_result: QAResult):
"""Record QA failures into the failure history."""
for issue in qa_result.issues:
self.failure_history.append(FailureRecord(
revision=self.revision_count,
owner=issue.owner,
issue_type=issue.type,
message=issue.message[:200],
correction=qa_result.correction_instruction[:200],
))
def has_repeated_failure(self, owner: str, issue_type: str) -> bool:
"""Check if the same owner+issue_type has failed in a previous revision."""
past = [
f for f in self.failure_history
if f.owner == owner and f.issue_type == issue_type
and f.revision < self.revision_count
]
return len(past) >= 1
def get_repeat_failures(self) -> List[Tuple[str, str]]:
"""Return (owner, issue_type) pairs that have failed more than once."""
counts: Dict[Tuple[str, str], int] = {}
for f in self.failure_history:
key = (f.owner, f.issue_type)
counts[key] = counts.get(key, 0) + 1
return [k for k, v in counts.items() if v >= 2]
def get_escalation_strategy(self) -> str:
"""Determine escalation strategy when failures repeat.
Returns:
'narrow_scope' — reduce role count and simplify
'rewrite_from_state' — synthesizer should rewrite from state, not reuse draft
'suppress_role' — a specific role keeps introducing unsupported content
'none' — no escalation needed
"""
repeats = self.get_repeat_failures()
if not repeats:
return "none"
synth_repeats = [(o, t) for o, t in repeats if o.lower() in ("synthesizer", "synthesis")]
role_repeats = [(o, t) for o, t in repeats if o.lower() not in ("synthesizer", "synthesis", "planner")]
if synth_repeats:
return "rewrite_from_state"
if role_repeats:
return "suppress_role"
return "narrow_scope"
def get_roles_to_suppress(self) -> List[str]:
"""Return role owners that keep introducing repeated failures."""
repeats = self.get_repeat_failures()
return list({owner for owner, _ in repeats
if owner.lower() not in ("synthesizer", "synthesis", "planner")})
def to_context_string(self) -> str:
"""Produce a compact summary string for inclusion in LLM prompts."""
lines = [
f"Task category: {self.task_category}",
f"Output format required: {self.output_format}",
f"Brevity requirement: {self.brevity_requirement}",
f"Revision: {self.revision_count}/{self.max_revisions}",
f"Selected roles: {', '.join(self.selected_roles)}",
]
if self.success_criteria:
lines.append(f"Success criteria: {'; '.join(self.success_criteria)}")
if self.task_assumptions:
assumptions_str = "; ".join(f"{k}: {v}" for k, v in self.task_assumptions.items())
lines.append(f"Shared assumptions: {assumptions_str}")
if self.evidence:
conf = self.evidence.get("confidence", "unknown")
n_items = len(self.evidence.get("results", []))
lines.append(f"Evidence: {n_items} items (confidence: {conf})")
if self.revision_instruction:
lines.append(f"Revision instruction: {self.revision_instruction}")
if self.qa_result and not self.qa_result.passed:
lines.append(f"QA status: FAIL — {self.qa_result.reason}")
if self.qa_result.correction_instruction:
lines.append(f"Correction needed: {self.qa_result.correction_instruction}")
if self.failure_history:
lines.append(f"Previous failures: {len(self.failure_history)}")
strategy = self.get_escalation_strategy()
if strategy != "none":
lines.append(f"Escalation strategy: {strategy}")
return "\n".join(lines)
def to_state_dict(self) -> dict:
"""Serialise the full state to a dictionary."""
return {
"user_request": self.user_request,
"task_summary": self.task_summary,
"task_category": self.task_category,
"success_criteria": self.success_criteria,
"output_format": self.output_format,
"brevity_requirement": self.brevity_requirement,
"selected_roles": self.selected_roles,
"specialist_outputs": self.specialist_outputs,
"evidence": self.evidence,
"task_assumptions": self.task_assumptions,
"current_draft": self.current_draft[:500],
"revision_count": self.revision_count,
"max_revisions": self.max_revisions,
"revision_instruction": self.revision_instruction,
"failure_history": [f.to_dict() for f in self.failure_history],
"final_answer": self.final_answer[:500] if self.final_answer else "",
}
# ============================================================
# Format-specific Synthesizer Instructions
# ============================================================
def get_synthesizer_format_instruction(output_format: str, brevity: str) -> str:
"""Return format-specific instructions to append to the synthesizer prompt."""
instructions = {
"single_choice": (
"CRITICAL FORMAT RULE: The user wants ONE SINGLE CHOICE.\n"
"Output ONLY the chosen option and at most one short justification sentence.\n"
"Do NOT include perspectives summary, common ground, trade-offs, or any multi-section structure.\n"
"Example: 'Veggie — it accommodates the widest range of dietary needs.'"
),
"short_answer": (
"CRITICAL FORMAT RULE: The user wants a SHORT, DIRECT answer.\n"
"Output 1-3 sentences maximum. No sections, no headers, no perspectives summary.\n"
"Answer the question directly and stop."
),
"code": (
"CRITICAL FORMAT RULE: The user wants CODE output.\n"
"Output the code directly. Only include a brief explanation if explicitly requested.\n"
"Do NOT include perspectives summary, trade-offs, or multi-section structure."
),
"table": (
"CRITICAL FORMAT RULE: The user wants a TABLE.\n"
"Output a properly formatted markdown table.\n"
"Do NOT include perspectives summary or prose-only answers."
),
"bullet_list": (
"CRITICAL FORMAT RULE: The user wants a BULLET LIST.\n"
"Output a clean bullet list. Do NOT wrap it in prose or add unnecessary sections."
),
"paragraph": (
"Output a clear, well-structured paragraph. Keep it focused and avoid unnecessary sections."
),
"report": (
"The user wants a detailed report. You may include sections like summary, "
"trade-offs, and perspectives, but keep each section concise."
),
}
base = instructions.get(output_format, "Match the output format to what the user requested.")
if brevity == "minimal":
base += "\nBREVITY: Absolute minimum. Fewer words is better."
elif brevity == "short":
base += "\nBREVITY: Keep it concise. No unnecessary elaboration."
return base
def get_qa_format_instruction(output_format: str, brevity: str) -> str:
"""Return format-specific validation rules for the QA prompt."""
rules = []
if output_format == "single_choice":
rules.append("FAIL if the output contains more than one choice or a long multi-section answer.")
rules.append("FAIL if the output includes perspectives summary, common ground, or trade-offs.")
elif output_format == "short_answer":
rules.append("FAIL if the output is longer than 3-4 sentences.")
rules.append("FAIL if the output includes unnecessary sections or headers.")
elif output_format == "code":
rules.append("FAIL if the output is mostly prose with no code.")
elif output_format == "table":
rules.append("FAIL if the output does not contain a markdown table.")
if brevity in ("minimal", "short"):
rules.append("FAIL if the output is excessively verbose for a brevity requirement.")
return "\n".join(rules) if rules else ""
# ============================================================
# Output Format Validation (pre-QA structural check)
# ============================================================
def validate_output_format(text: str, output_format: str, brevity: str) -> List[str]:
"""Check structural format constraints before QA.
Returns a list of violation descriptions. Empty list means the output is valid.
This catches common structural problems that the synthesizer repeatedly ignores
(e.g., bullet lists when paragraph-only was requested).
"""
violations: List[str] = []
stripped = text.strip()
if not stripped:
violations.append("Output is empty.")
return violations
has_bullets = bool(re.search(r"^[\s]*[-•*]\s", stripped, re.MULTILINE))
has_numbered = bool(re.search(r"^[\s]*\d+[.)]\s", stripped, re.MULTILINE))
has_headings = bool(re.search(r"^#{1,4}\s", stripped, re.MULTILINE))
has_table = bool(re.search(r"\|.*\|.*\|", stripped))
has_code_block = "```" in stripped
line_count = len([ln for ln in stripped.splitlines() if ln.strip()])
if output_format == "paragraph":
if has_bullets or has_numbered:
violations.append("Paragraph format requested but output contains bullet/numbered lists.")
if has_headings:
violations.append("Paragraph format requested but output contains markdown headings.")
if has_table:
violations.append("Paragraph format requested but output contains a table.")
elif output_format == "code":
if not has_code_block and not re.search(r"(?:def |class |import |function |const |let |var )", stripped):
violations.append("Code format requested but output contains no code block or recognisable code.")
elif output_format == "table":
if not has_table:
violations.append("Table format requested but output contains no markdown table.")
elif output_format == "single_choice":
if line_count > 5:
violations.append("Single choice requested but output is multi-section (too many lines).")
if has_bullets and line_count > 3:
violations.append("Single choice requested but output contains a bullet list.")
# Brevity checks
if brevity == "minimal" and line_count > 8:
violations.append(f"Minimal brevity requested but output has {line_count} lines.")
elif brevity == "short" and line_count > 20:
violations.append(f"Short brevity requested but output has {line_count} lines.")
return violations
def format_violations_instruction(violations: List[str]) -> str:
"""Turn format violation descriptions into a synthesis rewrite instruction."""
return (
"FORMAT VIOLATIONS DETECTED — you MUST fix these before QA:\n"
+ "\n".join(f"- {v}" for v in violations)
+ "\nRewrite the output to satisfy the required format strictly."
)
# ============================================================
# Shared Assumptions Parsing
# ============================================================
def parse_task_assumptions(plan_text: str) -> Dict[str, str]:
"""Extract TASK ASSUMPTIONS from planner output.
Looks for lines like 'key: value' under a TASK ASSUMPTIONS: header.
Returns a dict of assumption key → value.
"""
assumptions: Dict[str, str] = {}
if "TASK ASSUMPTIONS:" not in plan_text:
return assumptions
section = plan_text.split("TASK ASSUMPTIONS:", 1)[1]
# Section ends at the next header (a line that ends with ':' and starts with caps)
for header in (
"TASK BREAKDOWN:", "ROLE TO CALL:", "SUCCESS CRITERIA:",
"GUIDANCE FOR SPECIALIST:", "REVISED INSTRUCTIONS:",
):
if header in section:
section = section.split(header, 1)[0]
break
for line in section.strip().splitlines():
line = line.strip().lstrip("•-* ")
if ":" not in line:
continue
key, _, value = line.partition(":")
key = key.strip().lower().replace(" ", "_")
value = value.strip()
if key and value:
assumptions[key] = value
return assumptions
def format_assumptions_for_prompt(assumptions: Dict[str, str]) -> str:
"""Format shared assumptions for injection into specialist prompts."""
if not assumptions:
return ""
lines = ["SHARED TASK ASSUMPTIONS (use these — do NOT invent your own):"]
for key, value in assumptions.items():
lines.append(f" - {key}: {value}")
return "\n".join(lines)
# ============================================================
# Structured Expert Contributions
# ============================================================
# Suffix appended to every specialist system prompt to require JSON output
STRUCTURED_OUTPUT_SUFFIX = """
IMPORTANT — OUTPUT FORMAT:
After your analysis above, you MUST also output a JSON block at the end of your response,
wrapped in ```json ... ``` fences, with this exact structure:
```json
{
"role": "<your role name>",
"main_points": ["point 1", "point 2"],
"recommendations": ["recommendation 1"],
"evidence": ["supporting evidence or examples"],
"assumptions": ["assumption 1"],
"confidence": "high | medium | low"
}
```
- "main_points": your key substantive contributions to the answer (2-4 points)
- "recommendations": specific actionable recommendations (0-3)
- "evidence": facts, data, or examples that support your points (0-3)
- "assumptions": assumptions you relied on (0-2)
- "confidence": how confident you are in your contribution
This JSON block is REQUIRED. The Synthesizer will use it to build the final answer.
Do NOT write a complete final answer — focus on your domain-specific contribution.
"""
@dataclass
class StructuredContribution:
"""Structured output from an expert specialist."""
role: str
main_points: List[str] = field(default_factory=list)
recommendations: List[str] = field(default_factory=list)
evidence: List[str] = field(default_factory=list)
assumptions: List[str] = field(default_factory=list)
confidence: str = "medium"
raw_output: str = ""
def to_dict(self) -> dict:
return {
"role": self.role,
"main_points": self.main_points,
"recommendations": self.recommendations,
"evidence": self.evidence,
"assumptions": self.assumptions,
"confidence": self.confidence,
}
def has_substance(self) -> bool:
"""Check if this contribution has at least one substantive point."""
return bool(self.main_points or self.recommendations)
def parse_structured_contribution(text: str, role: str) -> StructuredContribution:
"""Parse a StructuredContribution from specialist LLM output.
Tries to extract a JSON block from the text. Falls back to heuristic
extraction from section headers if JSON is missing or malformed.
"""
contribution = StructuredContribution(role=role, raw_output=text)
# Try JSON extraction first — look for ```json ... ``` block
json_match = re.search(r"```json\s*(\{.*?\})\s*```", text, re.DOTALL)
if not json_match:
# Also try bare JSON object
json_match = re.search(r'(\{\s*"role"\s*:.*\})', text, re.DOTALL)
if json_match:
try:
data = json.loads(json_match.group(1))
contribution.main_points = data.get("main_points", [])
contribution.recommendations = data.get("recommendations", [])
contribution.evidence = data.get("evidence", [])
contribution.assumptions = data.get("assumptions", [])
contribution.confidence = data.get("confidence", "medium")
if data.get("role"):
contribution.role = data["role"]
return contribution
except (json.JSONDecodeError, AttributeError):
pass
# Fallback: heuristic extraction from section-based output
_extract_section_points(text, contribution)
return contribution
def _extract_section_points(text: str, contribution: StructuredContribution):
"""Heuristic fallback: extract key points from section-based specialist output."""
lines = text.strip().splitlines()
current_section = ""
buffer: List[str] = []
# Map known section headers to contribution fields
section_map = {
# Core roles
"ideas": "main_points", "rationale": "main_points",
"technical approach": "main_points", "implementation notes": "recommendations",
"evidence summary": "evidence", "key findings": "evidence",
"security analysis": "main_points", "vulnerabilities found": "main_points",
"recommendations": "recommendations",
"data overview": "main_points", "analysis": "main_points",
"insights": "recommendations",
# Persona roles
"wild hypothesis": "main_points", "scientific rationale": "evidence",
"groundbreaking implications": "main_points",
"cost analysis": "main_points", "cost-cutting measures": "recommendations",
"cosmic vision": "main_points", "wild storm of ideas": "main_points",
"minimum viable effort": "main_points",
"kvlt verdict": "main_points", "the grim truth": "main_points",
"worker impact": "main_points", "union concerns": "main_points",
"collective bargaining position": "recommendations",
"user needs analysis": "main_points", "pain points": "main_points",
"ux recommendations": "recommendations",
"what doris thinks is happening": "main_points",
"doris's thoughts": "main_points",
"board perspective": "main_points", "strategic concerns": "main_points",
"shareholder value": "recommendations",
"america first analysis": "main_points",
"making it great again": "recommendations",
"legal analysis": "main_points", "liabilities and risks": "main_points",
"legal recommendations": "recommendations",
}
def flush_buffer():
if current_section and buffer:
field_name = section_map.get(current_section.lower().rstrip(":"), "")
if field_name:
combined = " ".join(ln.strip().lstrip("•-*0123456789.) ") for ln in buffer if ln.strip())
if combined:
target = getattr(contribution, field_name)
target.append(combined[:300])
for line in lines:
header_match = re.match(r"^([A-Z][A-Z\s\'']+):?\s*$", line.strip())
if header_match:
flush_buffer()
current_section = header_match.group(1).strip()
buffer = []
else:
# Skip lines that look like "RECOMMENDED DRAFT:", "FINAL TECHNICAL DRAFT:", etc.
if re.match(r"^[A-Z][A-Z\s]+DRAFT:?\s*$", line.strip()):
flush_buffer()
current_section = "" # ignore draft sections
buffer = []
elif current_section:
buffer.append(line)
flush_buffer()
def format_contributions_for_synthesizer(
contributions: Dict[str, "StructuredContribution"],
) -> str:
"""Format structured expert contributions for the Synthesizer prompt.
Presents each expert's key points, recommendations, and evidence
so the Synthesizer can build the final answer from them.
"""
if not contributions:
return ""
parts = ["STRUCTURED EXPERT CONTRIBUTIONS:"]
for role_key, contrib in contributions.items():
role_label = contrib.role
section = [f"\n=== {role_label} (confidence: {contrib.confidence}) ==="]
if contrib.main_points:
section.append("Main points:")
for i, pt in enumerate(contrib.main_points):
section.append(f" [{i}] {pt}")
if contrib.recommendations:
section.append("Recommendations:")
for i, rec in enumerate(contrib.recommendations):
section.append(f" [{i}] {rec}")
if contrib.evidence:
section.append("Evidence:")
for ev in contrib.evidence:
section.append(f" - {ev}")
if contrib.assumptions:
section.append("Assumptions:")
for a in contrib.assumptions:
section.append(f" - {a}")
parts.append("\n".join(section))
return "\n\n".join(parts)
def format_contributions_for_qa(
contributions: Dict[str, "StructuredContribution"],
used_contributions: Dict[str, List[str]],
) -> str:
"""Format contribution data for QA to verify expert influence."""
if not contributions:
return ""
parts = ["EXPERT CONTRIBUTION TRACEABILITY:"]
for role_key, contrib in contributions.items():
role_label = contrib.role
used = used_contributions.get(role_key, [])
section = [f"\n=== {role_label} ==="]
if contrib.main_points:
for i, pt in enumerate(contrib.main_points):
tag = "USED" if f"main_points[{i}]" in used else "NOT USED"
section.append(f" main_points[{i}] [{tag}]: {pt}")
if contrib.recommendations:
for i, rec in enumerate(contrib.recommendations):
tag = "USED" if f"recommendations[{i}]" in used else "NOT USED"
section.append(f" recommendations[{i}] [{tag}]: {rec}")
parts.append("\n".join(section))
used_count = sum(len(v) for v in used_contributions.values())
total_points = sum(
len(c.main_points) + len(c.recommendations) for c in contributions.values()
)
parts.append(f"\nSummary: {used_count}/{total_points} expert contributions marked as used.")
return "\n".join(parts)
def parse_used_contributions(text: str) -> Dict[str, List[str]]:
"""Parse the Synthesizer's USED_CONTRIBUTIONS JSON block from its output.
Returns a dict mapping role_key → list of contribution references
like ["main_points[0]", "recommendations[1]"].
"""
# Look for ```json block containing "used_contributions"
json_match = re.search(r"```json\s*(\{.*?\})\s*```", text, re.DOTALL)
if json_match:
try:
data = json.loads(json_match.group(1))
if "used_contributions" in data:
return data["used_contributions"]
except (json.JSONDecodeError, AttributeError):
pass
# Look for a USED_CONTRIBUTIONS: section
if "USED_CONTRIBUTIONS:" in text:
section = text.split("USED_CONTRIBUTIONS:", 1)[1]
# Try to find JSON in the section
json_match = re.search(r"(\{.*?\})", section, re.DOTALL)
if json_match:
try:
return json.loads(json_match.group(1))
except (json.JSONDecodeError, AttributeError):
pass
return {}
def check_expert_influence(
contributions: Dict[str, "StructuredContribution"],
used_contributions: Dict[str, List[str]],
final_answer: str,
) -> List[str]:
"""Check whether the final answer materially uses expert contributions.
Returns a list of influence issues (empty = influence is adequate).
"""
issues: List[str] = []
if not contributions:
return issues
# Check 1: Are any contributions marked as used?
total_used = sum(len(refs) for refs in used_contributions.values())
total_available = sum(
len(c.main_points) + len(c.recommendations)
for c in contributions.values() if c.has_substance()
)
if total_available > 0 and total_used == 0:
issues.append(
"Final answer does not materially incorporate any specialist contributions."
)
return issues
# Check 2: For each contributing expert, is at least one point used?
for role_key, contrib in contributions.items():
if not contrib.has_substance():
continue
role_refs = used_contributions.get(role_key, [])
if not role_refs:
issues.append(
f"Expert '{contrib.role}' provided substantive points but none were used."
)
# Check 3: Do used points appear to influence the final answer?
# (Lightweight check: verify at least some expert vocabulary appears)
answer_lower = final_answer.lower()
expert_words_found = 0
for contrib in contributions.values():
for pt in contrib.main_points:
# Extract key content words (3+ chars)
words = [w for w in re.findall(r"\b\w{3,}\b", pt.lower())
if w not in ("the", "and", "for", "that", "this", "with", "from", "are", "was")]
matches = sum(1 for w in words if w in answer_lower)
if matches >= 2:
expert_words_found += 1
if expert_words_found == 0 and total_available > 0:
issues.append(
"Final answer appears to not reflect expert contribution content."
)
return issues
|