File size: 60,219 Bytes
5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 d9759a5 5dd1bb4 9e64e71 5dd1bb4 519b9a3 5dd1bb4 d9759a5 5dd1bb4 519b9a3 5dd1bb4 d9759a5 5dd1bb4 519b9a3 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 9e64e71 5dd1bb4 | 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 | {
"$schema": "./schemas/autocode-features-v1.schema.json",
"project": "SQLEnv - Interactive Database Query RL Environment",
"description": "OpenEnv Challenge submission: RL environment where agents learn to answer NL questions about databases through iterative SQL exploration",
"created": "2026-03-24T07:15:50Z",
"updated": "2026-04-11T15:55:16Z",
"features": [
{
"id": "F001",
"name": "Core Environment Loop",
"description": "Complete the step/reset lifecycle: remove Ollama from environment, accept structured actions (DESCRIBE table_name, SAMPLE table_name, QUERY sql_string, ANSWER value), wire up SQLite execution with sandboxing (read-only, 5s timeout, SELECT-only), load questions from JSON on reset(), enforce step budget (15 steps), handle episode termination",
"complexity": "complex",
"verification_mode": "standard",
"status": "complete",
"priority": 1,
"dependencies": [],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Derived from docs_draft/sql_env_project_brief.md and docs_draft/SQLEnv_Concept_v1.md — the v1 spec defines the action space, episode lifecycle, and sandboxing requirements"
},
"user_interview": {
"conducted": "2026-03-24T09:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Agents can play complete episodes: reset with a random question, explore a hidden schema via DESCRIBE/SAMPLE, run SQL queries, and submit answers. Currently SQL never executes — this makes the environment actually functional."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Agent sends DESCRIBE employees and immediately sees column names and types",
"Queries execute in <100ms with clean truncated output (max 20 rows)",
"Bad SQL returns a clear error message the agent can learn from",
"Episode ends cleanly when budget exhausted or ANSWER submitted"
],
"frustrations": [
"Environment calling Ollama to interpret actions (current design) — agent should own reasoning, env should just execute",
"Queries hanging or crashing the environment",
"Opaque error messages that don't help the agent adjust"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Competition submission — needs to work reliably for demo and training, not at production scale"
}
},
"progress": {
"implementation_steps": {
"total": 8,
"completed": 8
},
"verification_tests": {
"total": 86,
"passed": 25
}
},
"specs": {
"implementation": "specs/F001-IMPLEMENTATION_SPEC.md",
"verification": "specs/F001-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-24T10:30:00Z",
"verification_planned": "2026-03-24T10:30:00Z",
"started": "2026-03-24T19:22:08Z",
"completed": "2026-03-24T21:27:31Z"
},
"verification_evidence": {
"mode": "standard",
"tests_run": 25,
"tests_passed": 25,
"timestamp": "2026-03-24T21:27:31Z",
"command": "uv run pytest tests/ -v",
"verifier_result": "approved"
},
"demo": {
"path": "specs/F001-DEMO.md",
"generated_at": "2026-03-24T21:36:32Z",
"mode": "local_cli",
"status": "partial",
"requires_user_verification": true,
"verification_surfaces": [
"local_server_startup",
"data_provisioning",
"api_episode_flow"
],
"evidence_refs": [
"specs/F001-VERIFICATION_SPEC.md",
"specs/F001-DEMO.md"
],
"note": "Local server and tests verified; end-to-end API episode flow requires local Spider DB provisioning."
},
"user_value": "Agents can now run complete SQL exploration episodes end-to-end with structured DESCRIBE/SAMPLE/QUERY/ANSWER actions, live read-only SQLite execution, clear error feedback, and clean terminal completion on ANSWER or budget exhaustion."
},
{
"id": "F002",
"name": "Answer Verification",
"description": "Multi-type answer comparison: integer (exact match), float (1% tolerance), string (case-insensitive normalized), list (order-insensitive set comparison). Implements verify_answer() in server/verifier.py. Returns binary correctness for terminal reward.",
"complexity": "standard",
"verification_mode": "standard",
"status": "complete",
"priority": 2,
"dependencies": [
"F001"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Answer type handling defined in docs_draft/SQLEnv_Concept_v1.md Section 4.2"
},
"user_interview": {
"conducted": "2026-03-24T09:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "When an agent submits ANSWER, the environment correctly determines if the answer matches the gold answer regardless of type (42 vs 42.0, 'Engineering' vs 'engineering', unordered lists)."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Float comparison with tolerance handles rounding gracefully (95000.1 matches 95000)",
"List comparison ignores order: ['A','B'] matches ['B','A']",
"Clear pass/fail with no ambiguity"
],
"frustrations": [
"Correct answer rejected due to trivial formatting difference",
"Type coercion failures (agent says '42', gold is integer 42)"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Must handle the 4 core answer types reliably. Table comparison can come later."
}
},
"progress": {
"implementation_steps": {
"total": 4,
"completed": 4
},
"verification_tests": {
"total": 65,
"passed": 65
}
},
"specs": {
"implementation": "specs/F002-IMPLEMENTATION_SPEC.md",
"verification": "specs/F002-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-27T12:00:00Z",
"verification_planned": "2026-03-27T12:00:00Z",
"started": "2026-03-27T22:18:15Z",
"completed": "2026-03-27T22:33:12Z"
},
"verification_evidence": {
"mode": "standard",
"tests_run": 65,
"tests_passed": 65,
"timestamp": "2026-03-27T22:33:12Z",
"command": "uv run pytest tests/ -v",
"verifier_result": "approved"
},
"demo": {
"path": "specs/F002-DEMO.md",
"generated_at": "2026-03-27T22:37:50Z",
"mode": "artifact_build",
"status": "partial",
"requires_user_verification": true,
"verification_surfaces": [
"local_pytest_verification",
"runtime_episode_scoring"
],
"evidence_refs": [
"specs/F002-VERIFICATION_SPEC.md",
"specs/F002-DEMO.md"
],
"note": "Strongest local proof is targeted and integration pytest evidence; final runtime confirmation remains a user-operated episode check."
},
"user_value": "Agents can now submit ANSWER values across integer, float, string, and list questions and receive correct terminal scoring despite formatting differences, numeric representation differences, and list order changes."
},
{
"id": "F003",
"name": "Dense Reward System",
"description": "3-layer reward architecture: Layer 1 (operational validity: exec_ok +0.02, new_info +0.01 capped at 0.10, repeat -0.01, step_cost -0.005), Layer 2 (progress-to-target: weighted average of cardinality matching + value overlap + numeric range proximity, binned to 5 levels, improvement-only), Layer 3 (terminal correctness: +1.0 or 0.0). Total step rewards capped at 0.5, negative floor at -0.2.",
"complexity": "complex",
"verification_mode": "standard",
"status": "complete",
"priority": 3,
"dependencies": [
"F001",
"F002"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Reward architecture defined in docs_draft/SQLEnv_Concept_v1.md Section 3 and docs_draft/reward-research_gpt-5-2.md. Distance metrics detailed in docs_draft/reward_design.md."
},
"user_interview": {
"conducted": "2026-03-24T09:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Agents get meaningful feedback during exploration — not just 0/1 at the end. A query that returns 40 when the answer is 42 gets partial credit. Discovering new schema info gets a small reward. This makes GRPO training converge."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Reward varies meaningfully: random exploration ~0.1, targeted queries ~0.3, correct answer ~1.3",
"Anti-gaming works: agent can't farm rewards by describing everything or repeating queries",
"Progress signal is coarsened to prevent reward hill-climbing"
],
"frustrations": [
"Reward hacking: agent learns to exploit shaping rather than solve the task",
"Reward too sparse: agent gets no signal until terminal step",
"Over-complex reward that's hard to debug"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Start with weighted average of 3 metrics (cardinality, value overlap, numeric range). Add complexity only if training shows issues."
}
},
"progress": {
"implementation_steps": {
"total": 7,
"completed": 7
},
"verification_tests": {
"total": 61,
"passed": 166
}
},
"specs": {
"implementation": "specs/F003-IMPLEMENTATION_SPEC.md",
"verification": "specs/F003-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-27T12:00:00Z",
"verification_planned": "2026-03-27T12:00:00Z",
"started": "2026-03-27T23:51:47Z",
"completed": "2026-03-28T06:05:02Z"
},
"verification_evidence": {
"mode": "standard",
"tests_run": 166,
"tests_passed": 166,
"timestamp": "2026-03-28T06:05:02Z",
"command": "uv run --with pytest pytest tests/ -v",
"verifier_result": "approved"
},
"demo": {
"path": "specs/F003-DEMO.md",
"generated_at": "2026-03-28T06:07:34Z",
"mode": "artifact_build",
"status": "generated",
"requires_user_verification": true,
"verification_surfaces": [
"local_pytest_verification",
"runtime_episode_flow"
],
"evidence_refs": [
"specs/F003-VERIFICATION_SPEC.md",
"specs/F003-DEMO.md"
],
"note": "Strongest local proof is targeted smoke/unit execution; full reward calibration and live episode behavior should be confirmed in a user-run episode/training context."
},
"user_value": "Agents now receive dense numeric rewards on every non-terminal DESCRIBE/SAMPLE/QUERY step based on execution quality and progress toward the gold answer, while terminal correctness still dominates total episode reward."
},
{
"id": "F004",
"name": "Question Dataset Expansion",
"description": "Expand from 53 questions (one DB) to 100+ questions across 5-10 Spider databases. Add difficulty labels (easy/medium/hard at 40/40/20 split), answer_type metadata, and gold_answer fields. Create train/eval split (70/30). Curate for diversity of answer types and SQL patterns.",
"complexity": "standard",
"verification_mode": "mvp",
"status": "complete",
"priority": 4,
"dependencies": [],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Dataset requirements from docs_draft/sql_env_project_brief.md Section 3 and SQLEnv_Concept_v1.md Section 4"
},
"user_interview": {
"conducted": "2026-03-24T09:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Training on diverse databases and question types. Current single-DB setup risks overfitting to one schema."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Clear difficulty progression: easy questions have 1-2 tables, hard ones have 5+",
"Each question has pre-computed gold_answer so reward doesn't need to re-execute gold SQL every episode",
"Train/eval split prevents training on evaluation data"
],
"frustrations": [
"Questions that require SQL features SQLite doesn't support",
"Ambiguous gold answers (multiple valid interpretations)",
"All questions from same domain = no generalization"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "100 well-curated questions is sufficient for competition demo. Quality over quantity."
}
},
"progress": {
"implementation_steps": {
"total": 6,
"completed": 6
},
"verification_tests": {
"total": 66,
"passed": 21
}
},
"specs": {
"implementation": "specs/F004-IMPLEMENTATION_SPEC.md",
"verification": "specs/F004-VERIFICATION_SPEC.md"
},
"demo": {
"path": "specs/F004-DEMO.md",
"generated_at": "2026-03-24T21:07:31Z"
},
"timestamps": {
"planned": "2026-03-24T10:30:00Z",
"verification_planned": "2026-03-24T10:30:00Z",
"started": "2026-03-24T16:53:35Z",
"completed": "2026-03-24T21:04:54Z"
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 21,
"tests_passed": 21,
"timestamp": "2026-03-24T21:04:54Z",
"command": "uv run pytest tests/ -v",
"verifier_result": "approved"
},
"user_value": "Users can now train and evaluate against a curated multi-database dataset (676 questions across 10 Spider databases) with precomputed gold answers, answer types, difficulty labels, and deterministic train/eval splits."
},
{
"id": "F005",
"name": "Green Agent Wrapper",
"description": "Automated evaluation wrapper following OpenEnv pattern. Runs N episodes with a given policy (random, heuristic, or trained model). Reports success_rate, avg_reward, avg_steps. Supports random baseline policy for comparison. Required by competition evaluation criteria.",
"complexity": "standard",
"verification_mode": "mvp",
"status": "complete",
"priority": 5,
"dependencies": [
"F001",
"F002"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Green Agent pattern from SQLEnv_Concept_v1.md Appendix C. Required by OpenEnv Challenge evaluation criteria."
},
"user_interview": {
"conducted": "2026-03-24T09:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Run automated evaluation: 'How does policy X perform over 100 episodes?' Single command, structured output. Enables training comparison (random vs trained)."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Single function call: evaluate(n_episodes=100) returns clean metrics dict",
"Built-in random policy for instant baseline comparison",
"Results include per-episode breakdown for analysis"
],
"frustrations": [
"Evaluation crashes partway through and loses all results",
"No progress indicator for long evaluation runs"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Needs to produce reliable metrics for blog post. Doesn't need fancy visualization."
}
},
"progress": {
"implementation_steps": {
"total": 4,
"completed": 4
},
"verification_tests": {
"total": 43,
"passed": 16
}
},
"specs": {
"implementation": "specs/F005-IMPLEMENTATION_SPEC.md",
"verification": "specs/F005-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-27T12:00:00Z",
"verification_planned": "2026-03-27T12:00:00Z",
"started": "2026-03-27T23:51:09Z",
"completed": "2026-03-28T00:04:03Z"
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 16,
"tests_passed": 16,
"timestamp": "2026-03-28T00:04:03Z",
"command": "uv run --with pytest pytest tests/test_evaluation.py -v",
"verifier_result": "approved"
},
"demo": {
"path": "specs/F005-DEMO.md",
"generated_at": "2026-03-28T00:10:42Z",
"mode": "local_cli",
"status": "generated",
"requires_user_verification": false,
"verification_surfaces": [
"local_python_api",
"local_pytest"
],
"evidence_refs": [
"specs/F005-VERIFICATION_SPEC.md",
"specs/F005-IMPLEMENTATION_SPEC.md",
"specs/F005-DEMO.md"
],
"note": "Demo includes direct public API invocation plus local integration, determinism, edge, and progress-callback evidence."
},
"user_value": "Users can now evaluate any SQLEnv policy over multiple episodes with one call, get structured aggregate metrics plus per-episode results, and rely on deterministic seeded runs for fair baseline comparisons."
},
{
"id": "F006",
"name": "GRPO Training Pipeline",
"description": "TRL/GRPO integration for training a small LLM (Qwen3-1.7B or similar) to play SQLEnv. Includes: system prompt design for SQL exploration strategy, rollout_func that plays episodes via WebSocket client, reward_funcs (correctness, progress, operational) for GRPOTrainer, training notebook with hyperparameter config, baseline vs trained comparison output.",
"complexity": "complex",
"verification_mode": "mvp",
"status": "complete",
"priority": 6,
"dependencies": [
"F003",
"F005"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Training pipeline from docs_draft/SQLEnv_Concept_v1.md Section 3.5 (TRL mapping) and docs_draft/sql_env_project_brief.md Phase 4"
},
"user_interview": {
"conducted": "2026-03-24T09:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Train a model that learns SQL exploration strategy through RL. The 'before vs after' comparison is the competition's money shot — untrained agent flails randomly, trained agent explores strategically."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Training notebook runs end-to-end in one click",
"Learning curve clearly shows improvement over episodes",
"Side-by-side episode transcripts: random vs trained",
"Reproducible results"
],
"frustrations": [
"Training doesn't converge at all",
"Need expensive GPU for hours to see any signal",
"Notebook has hidden dependencies that break on fresh setup"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Even modest improvement over random is a win. The environment design + reward architecture is the main innovation, not SOTA training results."
}
},
"progress": {
"implementation_steps": {
"total": 6,
"completed": 6
},
"verification_tests": {
"total": 68,
"passed": 68
}
},
"specs": {
"implementation": "specs/F006-IMPLEMENTATION_SPEC.md",
"verification": "specs/F006-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-27T12:00:00Z",
"verification_planned": "2026-03-27T12:00:00Z",
"started": "2026-03-28T06:44:31Z",
"completed": "2026-03-28T07:37:20Z"
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 68,
"tests_passed": 68,
"timestamp": "2026-03-28T07:37:20Z",
"command": "uv run --with pytest pytest tests/unit/test_grpo_config.py tests/unit/test_prompts.py tests/unit/test_rollout.py tests/unit/test_rewards.py tests/unit/test_error_handling.py tests/integration/test_training_pipeline.py tests/e2e/test_training_e2e.py -v",
"verifier_result": "approved"
},
"user_value": "Users can now run a single GRPO notebook workflow that loads training prompts, trains an SQLEnv policy with TRL, visualizes reward-curve progress, and compares random-baseline transcripts against trained-policy transcripts before saving artifacts.",
"demo": {
"path": "specs/F006-DEMO.md",
"generated_at": "2026-03-28T07:42:55Z",
"mode": "interactive_ui",
"status": "partial",
"requires_user_verification": true,
"verification_surfaces": [
"local_dependency_import",
"local_pytest_verification",
"jupyter_notebook_launch",
"interactive_notebook_run"
],
"evidence_refs": [
"specs/F006-VERIFICATION_SPEC.md",
"specs/F006-DEMO.md"
],
"note": "Local proof and targeted tests were executed; full notebook interaction requires user environment with Jupyter runtime."
}
},
{
"id": "F007",
"name": "HuggingFace Deployment & Submission",
"description": "Competition submission package: validate and push Docker to HF Spaces (openenv push), clean up GitHub repo (README, setup instructions, training notebook), write HF blog post outline (hook, problem, solution, results, technical), record/screenshot before-vs-after demo.",
"complexity": "standard",
"verification_mode": "mvp",
"status": "complete",
"priority": 7,
"dependencies": [
"F001",
"F002",
"F003",
"F004",
"F005",
"F006"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Submission requirements from OpenEnv Challenge PDF and docs_draft/sql_env_project_brief.md Phase 5"
},
"user_interview": {
"conducted": "2026-03-24T09:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Judges can: read the blog, visit the HF Space, run the training notebook, and reproduce results. Someone outside the team can understand, use, and build on SQLEnv."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Blog tells a compelling story even if training results are modest",
"HF Space just works — connect, reset, play an episode",
"Training notebook runs end-to-end on Colab with one click"
],
"frustrations": [
"Docker build fails on HF Spaces",
"Blog is all technical, no narrative hook",
"Notebook has undocumented setup steps"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Ship what works. Polish can happen post-submission."
}
},
"progress": {
"implementation_steps": {
"total": 6,
"completed": 6
},
"verification_tests": {
"total": 34,
"passed": 250
}
},
"specs": {
"implementation": "specs/F007-IMPLEMENTATION_SPEC.md",
"verification": "specs/F007-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-27T12:00:00Z",
"verification_planned": "2026-03-27T12:00:00Z",
"started": "2026-03-28T17:03:38Z",
"completed": "2026-03-29T07:29:32Z"
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 250,
"tests_passed": 250,
"timestamp": "2026-03-29T07:29:32Z",
"command": "uv run --with pytest pytest tests/ -v",
"verifier_result": "approved"
},
"user_value": "Judges and external developers can now consume a complete SQLEnv submission package with HF Spaces-compatible deployment artifacts, a polished README quickstart, a structured blog outline, and a Colab-ready GRPO training notebook.",
"demo": {
"path": "specs/F007-DEMO.md",
"generated_at": "2026-03-29T07:33:23Z",
"mode": "infra_release",
"status": "partial",
"requires_user_verification": true,
"verification_surfaces": [
"local_manifest_validation",
"local_docker_build",
"external_registry_auth",
"hf_space_push",
"browser_episode_flow",
"colab_notebook_run"
],
"evidence_refs": [
"specs/F007-VERIFICATION_SPEC.md",
"specs/F007-DEMO.md"
],
"note": "Authenticated local build and HF push now both succeed for hjerpe/sql_env; browser episode flow and Colab run remain user-verified surfaces."
}
},
{
"id": "F008",
"name": "Synthetic Database Generation",
"description": "Generate variant SQLite databases with same schema but different data for metamorphic testing. Implements 3 MVP mutations: irrelevant row injection, ID remapping, and duplicate bridge rows. Validates that gold SQL produces correct (potentially different) answers on variant DBs. Enables robustness testing against accidental correctness.",
"complexity": "standard",
"verification_mode": "mvp",
"status": "complete",
"priority": 8,
"dependencies": [
"F004"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Metamorphic testing from docs_draft/reward-research_gpt-5-2.md and docs_draft/SQLEnv_Concept_v1.md Section 6.2. Originally scoped as post-MVP but user requested as separate feature."
},
"user_interview": {
"conducted": "2026-03-24T10:30:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Verify that agent-produced SQL is semantically correct, not just accidentally correct on one dataset. Catches missing JOINs, wrong filters, and hard-coded values."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Script generates 1-2 variant DBs per question automatically",
"Gold SQL still produces valid answers on variant DBs",
"Catches real bugs: missing DISTINCT, wrong join direction"
],
"frustrations": [
"Mutations break gold SQL (variant DB is invalid)",
"Too many false positives from mutations",
"Expensive to run during training"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "exploratory",
"rationale": "Post-submission stretch goal. Only 3 mutations for MVP, evaluate impact before expanding."
}
},
"progress": {
"implementation_steps": {
"total": 8,
"completed": 8
},
"verification_tests": {
"total": 61,
"passed": 60
}
},
"specs": {
"implementation": "specs/F008-IMPLEMENTATION_SPEC.md",
"verification": "specs/F008-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-27T12:00:00Z",
"verification_planned": "2026-03-27T12:00:00Z",
"started": "2026-03-27T22:16:14Z",
"completed": "2026-03-27T22:57:19Z"
},
"demo": {
"path": "specs/F008-DEMO.md",
"generated_at": "2026-03-27T22:55:58Z",
"mode": "local_cli",
"status": "generated",
"requires_user_verification": false,
"verification_surfaces": [
"local_cli",
"local_tests"
],
"evidence_refs": [
"specs/F008-VERIFICATION_SPEC.md",
"specs/F008-IMPLEMENTATION_SPEC.md"
],
"note": "Demo includes live CLI usage, edge/error cases, and supplementary local test run output."
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 61,
"tests_passed": 60,
"timestamp": "2026-03-27T22:57:19Z",
"command": "uv run pytest tests/ -v",
"verifier_result": "approved"
},
"user_value": "Users can now generate synthetic Spider DB variants with schema-preserving data mutations and gold-SQL validation, enabling metamorphic checks that expose brittle SQL patterns like hard-coded IDs and missing DISTINCT."
},
{
"id": "F009",
"name": "Oracle Policy",
"description": "Cheater/oracle policy that knows the gold SQL and answer. Plays optimal episodes: DESCRIBE relevant tables, execute gold SQL, submit answer. Validates reward ceiling (~1.3 expected) and provides upper-bound baseline for blog comparison (oracle vs trained vs random).",
"complexity": "simple",
"verification_mode": "mvp",
"status": "complete",
"priority": 9,
"dependencies": [
"F001",
"F002"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "From project plan: 'Cheater Policy — quick end-to-end test for maximum reward on environment'. Project brief Phase 2 done-when: 'A hardcoded cheat policy that knows the answer can achieve 100% success rate.'"
},
"user_interview": {
"conducted": "2026-03-28T12:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Validate that the environment reward ceiling works as designed. Oracle achieves ~100% success rate and ~1.3 total reward, confirming dense rewards stack correctly with terminal correctness. Provides upper-bound baseline for trained model comparison."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Oracle runs 100 episodes and reports near-perfect success rate",
"Reward breakdown shows terminal + exploration adding up correctly",
"Can compare oracle vs random vs trained in one table"
],
"frustrations": [
"Oracle fails on questions where gold SQL is valid but gold answer extraction differs",
"Oracle reward lower than expected, indicating reward bug"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Validation tool for environment quality. Straightforward implementation — knows gold answer, submits it."
}
},
"progress": {
"implementation_steps": {
"total": 2,
"completed": 2
},
"verification_tests": {
"total": 25,
"passed": 40
}
},
"specs": {
"implementation": "specs/F009-IMPLEMENTATION_SPEC.md",
"verification": "specs/F009-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-28T12:00:00Z",
"verification_planned": "2026-03-28T12:00:00Z",
"started": "2026-03-28T17:06:05Z",
"completed": "2026-03-28T17:14:17Z"
},
"demo": {
"path": "specs/F009-DEMO.md",
"generated_at": "2026-03-28T17:17:27Z",
"mode": "artifact_build",
"status": "partial",
"requires_user_verification": true,
"verification_surfaces": [
"local_unit_tests",
"package_export",
"integration_e2e_followup"
],
"evidence_refs": [
"specs/F009-VERIFICATION_SPEC.md",
"specs/F009-IMPLEMENTATION_SPEC.md"
],
"note": "Strongest local proof is targeted/local pytest evidence; verification-spec integration/E2E file paths are not present in this workspace."
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 40,
"tests_passed": 40,
"timestamp": "2026-03-28T17:14:17Z",
"command": "uv run --with pytest pytest tests/unit/test_oracle_policy.py tests/test_evaluation.py -v",
"verifier_result": "approved"
},
"user_value": "Users can now import and run OraclePolicy from sql_env.evaluation to produce a deterministic upper-bound baseline in evaluate(), validating reward-ceiling behavior and enabling direct oracle-vs-random-vs-trained comparisons."
},
{
"id": "F010",
"name": "TRL Environment Adapter",
"description": "Wrap SQLEnv as a TRL-compatible environment_factory class. Public methods (describe, sample, query, answer) become LLM-callable tools automatically. Includes reset(**kwargs) for episode initialization, reward accumulation for reward_func, and concurrent session support (max_concurrent_envs). Replaces need for custom rollout_func in F006.",
"complexity": "standard",
"verification_mode": "mvp",
"status": "complete",
"priority": 10,
"dependencies": [
"F001",
"F003"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Derived from TRL OpenEnv docs (https://huggingface.co/docs/trl/main/openenv). environment_factory is the recommended pattern over rollout_func."
},
"user_interview": {
"conducted": "2026-03-28T12:00:00Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Train any HuggingFace model against SQLEnv using standard TRL GRPOTrainer with environment_factory. No custom rollout code needed — TRL handles generation, tool parsing, and multi-turn loop automatically."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Pass SQLEnvTRL as environment_factory to GRPOTrainer and it works",
"Tool methods have typed docstrings so TRL auto-discovers them",
"Concurrent sessions handle parallel rollouts without contention"
],
"frustrations": [
"Tool method signatures don't match what TRL expects",
"Environment state leaks between episodes",
"Concurrent sessions cause SQLite locking errors"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Must work for competition demo. Concurrent sessions can start with modest parallelism (4-8)."
}
},
"progress": {
"implementation_steps": {
"total": 5,
"completed": 6
},
"verification_tests": {
"total": 48,
"passed": 287
}
},
"specs": {
"implementation": "specs/F010-IMPLEMENTATION_SPEC.md",
"verification": "specs/F010-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-03-28T12:00:00Z",
"verification_planned": "2026-03-28T12:00:00Z",
"started": "2026-03-28T17:05:54Z",
"completed": "2026-03-28T17:29:10Z"
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 288,
"tests_passed": 287,
"timestamp": "2026-03-28T17:29:10Z",
"command": "uv run --with pytest pytest tests/ -v",
"verifier_result": "approved"
},
"demo": {
"path": "specs/F010-DEMO.md",
"generated_at": "2026-03-28T17:31:44Z",
"mode": "artifact_build",
"status": "partial",
"requires_user_verification": true,
"verification_surfaces": [
"local_pytest_verification",
"real_trl_training_run",
"concurrent_rollout_runtime"
],
"evidence_refs": [
"specs/F010-VERIFICATION_SPEC.md",
"specs/F010-DEMO.md"
],
"note": "Strongest local proof is targeted test execution; full confidence still requires user-run TRL training and concurrency validation."
},
"user_value": "Users can now train TRL/GRPO policies against SQLEnv via native environment_factory tool-calling with SQLEnvTRL, without maintaining a custom rollout loop."
},
{
"id": "F011",
"name": "Prompting Baseline Notebook",
"description": "New notebook (notebooks/showcase_prompting.ipynb) demonstrating base model performance on SQL tasks using only prompt engineering — no training. Serves as a baseline comparison for the GRPO-trained model. Sections: (1) Zero-shot with tool definitions, (2) Few-shot in-context learning with example trajectories from SFT data, (3) Chain-of-thought prompting, (4) Evaluation on held-out eval set across all techniques, (5) Accuracy comparison table + bar chart, (6) Optional side-by-side with trained model checkpoint.",
"complexity": "standard",
"verification_mode": "mvp",
"status": "complete",
"priority": 11,
"dependencies": [
"F006",
"F010"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "User wants to demonstrate that training adds value over pure prompting. Key insight: this notebook makes the GRPO training story more compelling by showing the gap."
},
"user_interview": {
"conducted": "2026-04-02T08:27:55+00:00",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they could not before?",
"response": "See exactly how much the base model can do with prompting alone, making the GRPO training improvement measurable and the notebook more convincing as a demo."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Clear accuracy comparison table across techniques",
"Same eval set used for all methods (fair comparison)",
"Can load a trained checkpoint for side-by-side",
"Runs on Colab without training (fast demo)"
],
"frustrations": [
"Eval taking too long (should be lightweight)",
"Unclear what prompting technique is being used",
"No visual comparison (just numbers)"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Demonstrates the value proposition of training. Can iterate on techniques later."
}
},
"progress": {
"implementation_steps": {
"total": 7,
"completed": 7
},
"verification_tests": {
"total": 36,
"passed": 17
}
},
"specs": {
"implementation": "specs/F011-IMPLEMENTATION_SPEC.md",
"verification": "specs/F011-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-04-06T08:27:07.093218+00:00",
"verification_planned": "2026-04-06T08:27:07.093218+00:00",
"started": "2026-04-06T19:09:21Z",
"completed": "2026-04-07T05:10:40Z"
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 17,
"tests_passed": 17,
"timestamp": "2026-04-07T05:10:40Z",
"command": "uv run pytest tests/test_evaluation.py -v",
"verifier_result": "approved"
},
"user_value": "Users can now run one notebook that fairly compares zero-shot/1-shot/3-shot prompting against GRPO no-think and GRPO thinking checkpoints on the same eval subset, with both tabular metrics and a visual accuracy bar chart.",
"demo": {
"path": "specs/F011-DEMO.md",
"generated_at": "2026-04-07T05:12:46Z",
"mode": "artifact_build",
"status": "partial",
"requires_user_verification": true,
"verification_surfaces": [
"local_notebook_execution",
"local_visual_artifact_export",
"interactive_notebook_run",
"hf_checkpoint_access"
],
"evidence_refs": [
"specs/F011-VERIFICATION_SPEC.md",
"specs/F011-DEMO.md"
],
"note": "Notebook execution was attempted locally but failed in this environment; static visual artifact export succeeded, and full interactive chart/table validation remains a user-run check."
}
},
{
"id": "F012",
"name": "Enable Thinking Mode",
"description": "Remove /no_think suppression and enable_thinking=False so Qwen3 can reason during GRPO rollouts. Model currently generates empty <think> blocks and cannot reason about SQL errors (repeats same failing query verbatim). Enables pretrained reasoning capability via reward signal — SFT data unchanged.",
"complexity": "simple",
"verification_mode": "mvp",
"status": "not_started",
"priority": 12,
"dependencies": [],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "none",
"notes": "Discovered during Run 6 analysis: model repeats failing queries because it cannot reason about errors"
},
"user_interview": {
"conducted": "2026-04-04T05:32:07+00:00",
"skipped": true,
"skip_reason": "Simple config change — 3 files, clear pattern",
"value": null,
"experience": null,
"maturity": null
},
"progress": {
"implementation_steps": {
"total": 0,
"completed": 0
},
"verification_tests": {
"total": 0,
"passed": 0
}
},
"specs": {
"implementation": null,
"verification": null
},
"inline_spec": {
"files": [
"scripts/generate_sft_data.py",
"notebooks/train_grpo.ipynb",
"training/notebook_pipeline.py"
],
"description": "Remove /no_think from SYSTEM_PROMPT in SFT and GRPO. Change enable_thinking: False to True in notebook_pipeline.py chat_template_kwargs. Regenerate SFT data.",
"verification": "Run training on Colab — verify model produces non-empty <think> blocks and changes SQL after errors"
},
"timestamps": {
"planned": "2026-04-04T05:32:07+00:00",
"verification_planned": null,
"started": null,
"completed": null
},
"verification_evidence": null,
"user_value": null
},
{
"id": "F013",
"name": "Error-Recovery SFT Trajectories",
"description": "Add 15-20 SFT trajectories to generate_sft_data.py showing error recovery: model queries with wrong column/table → gets SQL error → re-examines schema via describe/sample → writes corrected query → submits correct answer. Teaches the base policy to recover from mistakes before GRPO, so KL-anchored exploration includes error recovery as a learned pattern.",
"complexity": "standard",
"verification_mode": "standard",
"status": "complete",
"priority": 13,
"dependencies": [],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "none",
"notes": "Run 7 analysis: error loops are the #1 reward killer. Model repeats same failing query 3-8x because SFT only shows happy paths. No error-recovery pattern in base policy."
},
"user_interview": {
"conducted": "2026-04-04T11:35:48+00:00",
"skipped": true,
"skip_reason": "Pattern clear from Run 7 rollout analysis — model needs error-recovery examples in SFT data",
"value": null,
"experience": null,
"maturity": null
},
"progress": {
"implementation_steps": {
"total": 4,
"completed": 4
},
"verification_tests": {
"total": 55,
"passed": 55
}
},
"specs": {
"implementation": "specs/F013-IMPLEMENTATION_SPEC.md",
"verification": "specs/F013-VERIFICATION_SPEC.md"
},
"timestamps": {
"planned": "2026-04-04T11:50:45+00:00",
"verification_planned": "2026-04-04T11:50:45+00:00",
"started": "2026-04-04T14:10:09Z",
"completed": "2026-04-04T18:20:00Z"
},
"verification_evidence": {
"mode": "standard",
"tests_run": 2,
"tests_passed": 2,
"timestamp": "2026-04-04T18:20:00Z",
"command": "uv run pytest tests/unit/test_sft_terminal_message.py -v && uv run python scripts/generate_sft_data.py"
},
"user_value": null
},
{
"id": "F014",
"name": "Stop-After-Correct SFT Trajectories",
"description": "Add 5-10 SFT trajectories where the model answers correctly and the conversation ends cleanly — no post-episode tool calls. Currently all SFT examples end with the tool response 'Answer submitted: correct.' but the model still generates extra calls afterward during GRPO. Explicitly training on clean episode endings teaches the stop signal.",
"complexity": "simple",
"verification_mode": "mvp",
"status": "complete",
"priority": 14,
"dependencies": [
"F013"
],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "none",
"notes": "Run 7: model makes 1-3 extra calls after correct answer despite -0.3 post-episode penalty. SFT ending is ambiguous — model sees tool response but has no 'done generating' signal."
},
"user_interview": {
"conducted": "2026-04-04T11:35:48+00:00",
"skipped": true,
"skip_reason": "Simple extension of generate_sft_data.py — add final assistant turn with no tool call",
"value": null,
"experience": null,
"maturity": null
},
"progress": {
"implementation_steps": {
"total": 1,
"completed": 1
},
"verification_tests": {
"total": 21,
"passed": 2
}
},
"specs": {
"implementation": "specs/F014-IMPLEMENTATION_SPEC.md",
"verification": "specs/F014-VERIFICATION_SPEC.md"
},
"inline_spec": {
"files": [
"scripts/generate_sft_data.py"
],
"description": "After the final 'Answer submitted: correct.' tool response, do NOT append another assistant turn. The SFT example ends at the tool response. TRL's assistant_only_loss means the model only trains on assistant turns, so ending after the final tool response teaches the model that no further generation is needed. Alternatively, add a short assistant turn with just a stop token or empty content.",
"verification": "Inspect rendered SFT data — confirm examples end after correct answer tool response. Run GRPO training and check post-episode call count decreases."
},
"timestamps": {
"planned": "2026-04-04T11:48:20+00:00",
"verification_planned": "2026-04-04T11:48:20+00:00",
"started": "2026-04-04T14:17:03Z",
"completed": "2026-04-04T14:17:03Z"
},
"verification_evidence": {
"mode": "mvp",
"tests_run": 2,
"tests_passed": 2,
"timestamp": "2026-04-04T14:17:03Z",
"command": "uv run pytest tests/unit/test_sft_terminal_message.py -v",
"verifier_result": "approved"
},
"demo": {
"path": "specs/F014-DEMO.md",
"generated_at": "2026-04-04T14:21:55Z",
"mode": "artifact_build",
"status": "generated",
"requires_user_verification": true,
"verification_surfaces": [
"local_sft_generation",
"artifact_inspection",
"training_runtime_behavior"
],
"evidence_refs": [
"specs/F014-VERIFICATION_SPEC.md",
"specs/F014-DEMO.md"
],
"note": "Local SFT artifact and terminal-message shape are verified; reduction in post-answer calls must be confirmed in GRPO runtime."
},
"user_value": "SFT trajectories now end with an explicit terminal assistant message after correct answer confirmation, teaching a clear stop pattern that helps reduce extra post-answer tool calls during GRPO."
},
{
"id": "F015",
"name": "Error-Repetition Penalty",
"description": "In trl_adapter.py, track recent tool calls (function name + arguments) in a short window. When the model makes an exact repeat of any recent call, apply -0.2 penalty. Uses trajectory-level reward aggregation — safe for GRPO (no Markov violation because GRPO uses Monte Carlo returns, not Bellman bootstrapping, and the model's context window already contains full history as augmented state).",
"complexity": "simple",
"verification_mode": "standard",
"status": "complete",
"priority": 15,
"dependencies": [],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "none",
"notes": "Run 7: model repeats exact same failing query 3-8 times. -0.2 per repeat is moderate enough to avoid the repeat-avoidance trap (preferring novel-but-wrong over correct retry). Exact-match comparison (function+args string equality) is simple and sufficient."
},
"user_interview": {
"conducted": "2026-04-04T11:35:48+00:00",
"skipped": true,
"skip_reason": "Small code change in trl_adapter.py — add _recent_calls tracking and repeat penalty",
"value": null,
"experience": null,
"maturity": null
},
"progress": {
"implementation_steps": {
"total": 2,
"completed": 2
},
"verification_tests": {
"total": 55,
"passed": 55
}
},
"specs": {
"implementation": "specs/F015-IMPLEMENTATION_SPEC.md",
"verification": "specs/F015-VERIFICATION_SPEC.md"
},
"inline_spec": {
"files": [
"training/trl_adapter.py",
"tests/unit/test_trl_adapter.py",
"training/rollout.py",
"training/notebook_pipeline.py",
"notebooks/train_grpo.ipynb"
],
"description": "Add self._recent_calls: collections.deque[tuple[str, str]] with maxlen=3 and self._repeat_count: int in __init__. In each tool method (describe, query, sample, answer), before executing: build call_key = (method_name, arg_value). If call_key appears in self._recent_calls, apply _REPEAT_PENALTY = -0.2 and increment self._repeat_count. Always append call_key after execution. Reset self._recent_calls and self._repeat_count in reset().",
"verification": "Unit test: call query('SELECT 1') twice in a row, verify reward includes -0.2 repeat penalty. Call query('SELECT 1') then query('SELECT 2'), verify no penalty."
},
"timestamps": {
"planned": "2026-04-04T11:47:59+00:00",
"verification_planned": "2026-04-04T11:47:59+00:00",
"started": "2026-04-05T05:23:09Z",
"completed": "2026-04-05T05:43:04Z"
},
"verification_evidence": {
"mode": "standard",
"tests_run": 55,
"tests_passed": 55,
"timestamp": "2026-04-05T05:43:04Z",
"command": "uv run pytest tests/unit/test_trl_adapter.py -v && uv run pytest tests/unit/test_trl_adapter.py -v -k \"repeat or last_call\" && uv run pytest tests/e2e/test_training_e2e.py -v",
"verifier_result": "approved"
},
"demo": {
"path": "specs/F015-DEMO.md",
"generated_at": "2026-04-05T05:50:52Z",
"mode": "artifact_build",
"status": "generated",
"requires_user_verification": true,
"verification_surfaces": [
"local_pytest_verification",
"training_runtime_behavior"
],
"evidence_refs": [
"specs/F015-VERIFICATION_SPEC.md",
"specs/F015-DEMO.md"
],
"note": "Strongest local proof is targeted/full pytest and training e2e smoke; reduced repeat loops in live GRPO trajectories still requires user runtime confirmation."
},
"user_value": "Agents now receive a deterministic repeat-call penalty for reused tool calls within a short recent-call window (including alternating reuse), reducing degenerate GRPO loops while preserving non-repeated exploration behavior."
},
{
"id": "F016",
"name": "Pre-Publication Code Quality Sweep",
"description": "Refactor, lint fixes, and code smell cleanup before blog post publication. Runs ruff --fix, removes dead code, fixes line lengths, and addresses unused variables. Staff review of core modules (reward, verifier, trl_adapter, sql_environment) for correctness and clarity.",
"complexity": "simple",
"verification_mode": "mvp",
"status": "not_started",
"priority": 1,
"dependencies": [],
"docs": {
"discovery_json": null,
"discovery_md": null,
"design_doc": null,
"delivery_spec": null
},
"taste": {
"source": "user_interview",
"notes": "Blog deadline tomorrow — codebase must be presentable for open-source judges"
},
"user_interview": {
"conducted": "2026-04-11T15:55:16Z",
"skipped": false,
"skip_reason": null,
"value": {
"question": "What will users be able to do that they couldn't before?",
"response": "Judges and readers reviewing the GitHub repo will see clean, well-linted code without obvious smells. The codebase matches the quality story told in the blog post."
},
"experience": {
"question": "Walk me through using this. What would delight you? What would frustrate you?",
"delights": [
"Zero ruff errors on clone",
"No dead imports or unused variables",
"Core modules pass a staff-level review"
],
"frustrations": [
"Visible linting errors in the repo judges clone",
"Commented-out code or debug prints left in",
"Inconsistent formatting between files"
]
},
"maturity": {
"question": "Is this exploratory, MVP, or production?",
"response": "mvp",
"rationale": "Ship-blocking cleanup, not a deep refactor. Fix what's visible, don't reorganize."
}
},
"progress": {
"implementation_steps": {
"total": 4,
"completed": 0
},
"verification_tests": {
"total": 2,
"passed": 0
}
},
"specs": {
"implementation": null,
"verification": null
},
"inline_spec": {
"files": [
"server/sql_environment.py",
"server/verifier.py",
"server/reward.py",
"training/trl_adapter.py",
"training/config.py",
"training/notebook_pipeline.py",
"training/data_loading.py",
"evaluation/policies.py",
"evaluation/runner.py",
"scripts/generate_sft_data.py",
"tests/"
],
"description": "Four steps: (1) ruff check --fix + ruff format, (2) manual fix remaining lint errors (line length, unused vars, dead imports), (3) spec-staff-review on core modules, (4) address review findings. Inline verification: ruff check passes with 0 errors, all existing tests pass.",
"verification": "ruff check . returns 0 errors; uv run python -m pytest tests/ passes; staff review findings addressed or documented"
},
"timestamps": {
"planned": "2026-04-11T15:55:16Z",
"verification_planned": null,
"started": null,
"completed": null
},
"verification_evidence": null,
"user_value": null
}
]
}
|