test_phase_two_candidates.py 31 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886
  1. from __future__ import annotations
  2. from dataclasses import replace
  3. from datetime import UTC, datetime
  4. from hashlib import sha256
  5. from types import SimpleNamespace
  6. from typing import Any, cast
  7. import pytest
  8. from agent.orchestration import ArtifactRef
  9. from script_build_host.application.phase_two_candidates import (
  10. PhaseTwoCandidateError,
  11. PhaseTwoCandidateService,
  12. )
  13. from script_build_host.application.phase_two_inputs import (
  14. ActiveFrontierResolver,
  15. PhaseTwoInputError,
  16. )
  17. from script_build_host.domain.artifacts import (
  18. ArtifactKind,
  19. ArtifactState,
  20. ArtifactVersion,
  21. DirectionArtifact,
  22. DirectionGoal,
  23. EvidenceRecordV1,
  24. )
  25. from script_build_host.domain.errors import ProtocolViolation
  26. from script_build_host.domain.phase_two_artifacts import (
  27. CandidatePortfolioArtifactV1,
  28. ComparisonArtifactV1,
  29. GoalCoverage,
  30. ScriptParagraphV1,
  31. StructuredScriptArtifactV1,
  32. )
  33. from script_build_host.domain.task_contracts import (
  34. AcceptedDecisionRef,
  35. AcceptedInput,
  36. AcceptedInputBundleV1,
  37. ScriptCriterion,
  38. ScriptIntentClass,
  39. ScriptTaskBudget,
  40. ScriptTaskContractV1,
  41. ScriptTaskKind,
  42. )
  43. from script_build_host.domain.workspaces import CandidateWriteContext, WorkspaceError
  44. from script_build_host.infrastructure.canonical_json import canonical_sha256
  45. from script_build_host.repositories.sqlalchemy import (
  46. SqlAlchemyScriptBusinessArtifactRepository,
  47. )
  48. from script_build_host.repositories.workspace import (
  49. SqlAlchemyCandidateWorkspaceRepository,
  50. )
  51. DIGEST = "sha256:" + "a" * 64
  52. SCOPE = "script-build://scopes/opening"
  53. WRITE = "script-build://writes/paragraphs/opening"
  54. def _contract(
  55. kind: ScriptTaskKind,
  56. *,
  57. refs: tuple[AcceptedDecisionRef, ...] = (),
  58. comparison_refs: tuple[AcceptedDecisionRef, ...] = (),
  59. adopted: tuple[str, ...] = (),
  60. held: tuple[str, ...] = (),
  61. order: tuple[str, ...] = (),
  62. ) -> ScriptTaskContractV1:
  63. schemas = {
  64. ScriptTaskKind.PARAGRAPH: "paragraph-artifact/v1",
  65. ScriptTaskKind.DECODE_RETRIEVAL: "evidence-record/v1",
  66. ScriptTaskKind.COMPARE: "comparison-artifact/v1",
  67. ScriptTaskKind.COMPOSE: "structured-script/v1",
  68. ScriptTaskKind.CANDIDATE_PORTFOLIO: "candidate-portfolio/v1",
  69. }
  70. return ScriptTaskContractV1(
  71. task_kind=kind,
  72. scope_ref=SCOPE,
  73. intent_class=(
  74. ScriptIntentClass.COMPOSE
  75. if kind is ScriptTaskKind.COMPOSE
  76. else ScriptIntentClass.COMPARE
  77. if kind is ScriptTaskKind.COMPARE
  78. else ScriptIntentClass.EXPLORE
  79. ),
  80. objective="produce one bounded candidate increment",
  81. input_decision_refs=refs if kind is ScriptTaskKind.PARAGRAPH else (),
  82. base_artifact_ref=None,
  83. write_scope=(WRITE,),
  84. gap_ref=None,
  85. output_schema=schemas[kind],
  86. criteria=(ScriptCriterion("closed", "output is concretely complete"),),
  87. budget=ScriptTaskBudget(),
  88. goal_ids=("goal-1",),
  89. candidate_closure_decision_refs=(
  90. refs if kind in {ScriptTaskKind.COMPOSE, ScriptTaskKind.CANDIDATE_PORTFOLIO} else ()
  91. ),
  92. adopted_decision_ids=adopted,
  93. held_or_rejected_decision_ids=held,
  94. compose_order=order,
  95. comparison_decision_refs=comparison_refs,
  96. )
  97. class _Bindings:
  98. async def get_by_root(self, root_trace_id: str) -> Any:
  99. return SimpleNamespace(
  100. root_trace_id=root_trace_id,
  101. script_build_id=7,
  102. input_snapshot_id=11,
  103. active_direction_artifact_version_id=10,
  104. )
  105. class _TaskStore:
  106. def __init__(self, ledger: Any) -> None:
  107. self.ledger = ledger
  108. async def load(self, root_trace_id: str) -> Any:
  109. assert root_trace_id == "root"
  110. return self.ledger
  111. class _AcceptedInputs:
  112. def __init__(
  113. self, contracts: dict[str, ScriptTaskContractV1], bundle: AcceptedInputBundleV1
  114. ) -> None:
  115. self.contracts = contracts
  116. self.bundle = bundle
  117. async def contract_for_task(self, *, task: Any, **_: Any) -> ScriptTaskContractV1:
  118. return self.contracts[task.task_id]
  119. async def resolve(self, **_: Any) -> AcceptedInputBundleV1:
  120. return self.bundle
  121. class _FrameworkArtifacts:
  122. async def get(self, *_: Any) -> Any:
  123. raise AssertionError("framework snapshot should not be read")
  124. class _SnapshotArtifacts:
  125. def __init__(self, ref: ArtifactRef) -> None:
  126. self.ref = ref
  127. async def get(self, root_trace_id: str, snapshot_id: str) -> Any:
  128. assert (root_trace_id, snapshot_id) == ("root", "validation-snapshot")
  129. return SimpleNamespace(artifact_refs=[self.ref], evidence_refs=[])
  130. class _RawStore:
  131. def __init__(self, content: bytes) -> None:
  132. self.content = content
  133. async def read_bytes(self, ref: str) -> bytes:
  134. assert ref.endswith(sha256(self.content).hexdigest())
  135. return self.content
  136. class _Artifacts:
  137. def __init__(self, versions: dict[str, ArtifactVersion]) -> None:
  138. self.versions = versions
  139. self.frozen: ArtifactVersion | None = None
  140. async def read_by_ref(self, ref: ArtifactRef, **_: Any) -> ArtifactVersion:
  141. return self.versions[ref.uri]
  142. async def get_by_attempt(self, **owners: Any) -> ArtifactVersion:
  143. for value in self.versions.values():
  144. if value.attempt_id == owners["attempt_id"]:
  145. return value
  146. raise AssertionError("Attempt artifact not found")
  147. async def get_by_id(self, identifier: int, **_: Any) -> ArtifactVersion:
  148. for value in self.versions.values():
  149. if value.artifact_version_id == identifier:
  150. return value
  151. if identifier != 10:
  152. raise AssertionError("Artifact version not found")
  153. direction = DirectionArtifact(
  154. goals=(
  155. DirectionGoal(
  156. "goal-1",
  157. "complete the script",
  158. "the delivery needs one complete script",
  159. success_criteria=("the script is complete",),
  160. ),
  161. ),
  162. evidence_refs=("script-build://artifact-versions/90",),
  163. )
  164. return ArtifactVersion(
  165. 10,
  166. 7,
  167. "direction",
  168. "direction-attempt",
  169. 1,
  170. ArtifactKind.DIRECTION,
  171. DIGEST,
  172. ArtifactState.FROZEN,
  173. direction,
  174. datetime.now(UTC),
  175. datetime.now(UTC),
  176. )
  177. async def freeze(self, *, artifact: Any, **owners: Any) -> tuple[ArtifactVersion, ArtifactRef]:
  178. digest = canonical_sha256(artifact.content_payload()).wire
  179. artifact = replace(artifact, canonical_sha256=digest)
  180. artifact_kind = (
  181. ArtifactKind.COMPARISON
  182. if isinstance(artifact, ComparisonArtifactV1)
  183. else ArtifactKind.CANDIDATE_PORTFOLIO
  184. )
  185. version = ArtifactVersion(
  186. artifact_version_id=99,
  187. script_build_id=owners["script_build_id"],
  188. task_id=owners["task_id"],
  189. attempt_id=owners["attempt_id"],
  190. spec_version=owners["spec_version"],
  191. artifact_type=artifact_kind,
  192. canonical_sha256=digest,
  193. state=ArtifactState.FROZEN,
  194. artifact=artifact,
  195. created_at=datetime.now(UTC),
  196. frozen_at=datetime.now(UTC),
  197. )
  198. self.frozen = version
  199. ref = ArtifactRef(
  200. "script-build://artifact-versions/99",
  201. artifact_kind.value,
  202. "99",
  203. digest,
  204. )
  205. self.versions[ref.uri] = version
  206. return version, ref
  207. def _scope_fixture(
  208. contract: ScriptTaskContractV1,
  209. *,
  210. bundle: AcceptedInputBundleV1 | None = None,
  211. task_id: str = "task",
  212. attempt_id: str = "attempt",
  213. ) -> tuple[Any, dict[str, Any]]:
  214. task = SimpleNamespace(task_id=task_id, attempt_ids=[attempt_id])
  215. attempt = SimpleNamespace(
  216. attempt_id=attempt_id,
  217. task_id=task_id,
  218. spec_version=1,
  219. accepted_child_decision_ids=(),
  220. )
  221. ledger = SimpleNamespace(
  222. tasks={task_id: task},
  223. attempts={attempt_id: attempt},
  224. decisions={},
  225. )
  226. accepted = _AcceptedInputs(
  227. {task_id: contract},
  228. bundle or AcceptedInputBundleV1((), DIGEST),
  229. )
  230. return (ledger, accepted), {
  231. "root_trace_id": "root",
  232. "task_id": task_id,
  233. "attempt_id": attempt_id,
  234. "spec_version": 1,
  235. }
  236. @pytest.mark.asyncio
  237. async def test_phase_one_manifest_does_not_require_candidate_workspace() -> None:
  238. contract = _contract(ScriptTaskKind.DECODE_RETRIEVAL)
  239. (ledger, accepted), context = _scope_fixture(contract)
  240. evidence = EvidenceRecordV1(
  241. evidence_id="evidence",
  242. source_type="decode",
  243. tool_name="retrieve_decode",
  244. query={},
  245. source_refs=("source",),
  246. raw_artifact_ref=None,
  247. summary="bounded evidence",
  248. supports=("goal",),
  249. confidence="high",
  250. limitations=(),
  251. content_sha256=DIGEST,
  252. created_at=datetime.now(UTC),
  253. )
  254. version = ArtifactVersion(
  255. 1,
  256. 7,
  257. "task",
  258. "attempt",
  259. 1,
  260. ArtifactKind.EVIDENCE,
  261. DIGEST,
  262. ArtifactState.FROZEN,
  263. evidence,
  264. datetime.now(UTC),
  265. datetime.now(UTC),
  266. )
  267. service = PhaseTwoCandidateService(
  268. bindings=cast(Any, _Bindings()),
  269. task_store=_TaskStore(ledger),
  270. framework_artifact_store=_FrameworkArtifacts(),
  271. artifacts=cast(Any, _Artifacts({"script-build://artifact-versions/1": version})),
  272. accepted_inputs=accepted,
  273. active_frontier=ActiveFrontierResolver(),
  274. workspaces=None,
  275. )
  276. manifest = await service.resolve_attempt_manifest(context=context)
  277. assert manifest.artifact_ref.kind == "evidence"
  278. assert manifest.scope_ref == SCOPE
  279. @pytest.mark.asyncio
  280. async def test_paragraph_workspace_uses_positive_branch_and_freezes_on_manifest(
  281. database: Any,
  282. ) -> None:
  283. _, sessions = database
  284. repository = SqlAlchemyScriptBusinessArtifactRepository(sessions)
  285. workspaces = SqlAlchemyCandidateWorkspaceRepository(sessions, repository)
  286. contract = _contract(ScriptTaskKind.PARAGRAPH)
  287. (ledger, accepted), context = _scope_fixture(contract)
  288. service = PhaseTwoCandidateService(
  289. bindings=cast(Any, _Bindings()),
  290. task_store=_TaskStore(ledger),
  291. framework_artifact_store=_FrameworkArtifacts(),
  292. artifacts=repository,
  293. accepted_inputs=accepted,
  294. active_frontier=ActiveFrontierResolver(),
  295. workspaces=workspaces,
  296. )
  297. created = await service.create_script_paragraph(
  298. payload={"paragraph_index": 1, "name": "opening", "content_range": {}},
  299. context=context,
  300. )
  301. manifest = await service.resolve_attempt_manifest(context=context)
  302. version = await repository.get_by_attempt(
  303. script_build_id=7, task_id="task", attempt_id="attempt"
  304. )
  305. assert created["paragraph_id"] > 0
  306. assert version.artifact_version_id > 0
  307. assert manifest.artifact_ref.digest == version.canonical_sha256
  308. assert manifest.artifact_ref.kind == "paragraph"
  309. with pytest.raises(PhaseTwoCandidateError, match="WRITE_SCOPE_VIOLATION"):
  310. await service.create_script_paragraph(
  311. payload={
  312. "paragraph_index": 2,
  313. "name": "forged",
  314. "content_range": {},
  315. "branch_id": 0,
  316. },
  317. context=context,
  318. )
  319. @pytest.mark.asyncio
  320. async def test_paragraph_batch_maps_parent_keys_and_rolls_back_atomically(database: Any) -> None:
  321. _, sessions = database
  322. repository = SqlAlchemyScriptBusinessArtifactRepository(sessions)
  323. workspaces = SqlAlchemyCandidateWorkspaceRepository(sessions, repository)
  324. contract = _contract(ScriptTaskKind.PARAGRAPH)
  325. (ledger, accepted), context = _scope_fixture(contract)
  326. service = PhaseTwoCandidateService(
  327. bindings=cast(Any, _Bindings()),
  328. task_store=_TaskStore(ledger),
  329. framework_artifact_store=_FrameworkArtifacts(),
  330. artifacts=repository,
  331. accepted_inputs=accepted,
  332. active_frontier=ActiveFrontierResolver(),
  333. workspaces=workspaces,
  334. )
  335. created = await service.create_script_paragraphs(
  336. paragraphs=(
  337. {
  338. "client_key": "root",
  339. "paragraph_index": 1,
  340. "name": "opening",
  341. "content_range": {"beats": ["setup", "turn"]},
  342. },
  343. {
  344. "client_key": "child",
  345. "parent_client_key": "root",
  346. "paragraph_index": 2,
  347. "level": 2,
  348. "name": "turn",
  349. "content_range": {"beats": ["turn"]},
  350. },
  351. ),
  352. context=context,
  353. )
  354. assert created["created"] == 2
  355. assert set(created["paragraph_ids_by_client_key"]) == {"root", "child"}
  356. contract2 = replace(contract, objective="second isolated workspace")
  357. (ledger2, accepted2), context2 = _scope_fixture(
  358. contract2, task_id="task-2", attempt_id="attempt-2"
  359. )
  360. service2 = PhaseTwoCandidateService(
  361. bindings=cast(Any, _Bindings()),
  362. task_store=_TaskStore(ledger2),
  363. framework_artifact_store=_FrameworkArtifacts(),
  364. artifacts=repository,
  365. accepted_inputs=accepted2,
  366. active_frontier=ActiveFrontierResolver(),
  367. workspaces=workspaces,
  368. )
  369. with pytest.raises(WorkspaceError, match="child content range"):
  370. await service2.create_script_paragraphs(
  371. paragraphs=(
  372. {
  373. "client_key": "root",
  374. "paragraph_index": 1,
  375. "name": "root",
  376. "content_range": {"beats": ["setup"]},
  377. },
  378. {
  379. "client_key": "child",
  380. "parent_client_key": "root",
  381. "paragraph_index": 2,
  382. "level": 2,
  383. "name": "child",
  384. "content_range": {"beats": ["missing"]},
  385. },
  386. ),
  387. context=context2,
  388. )
  389. write_context = CandidateWriteContext(
  390. script_build_id=7,
  391. task_id="task-2",
  392. attempt_id="attempt-2",
  393. spec_version=1,
  394. objective=contract2.objective,
  395. input_refs=(),
  396. write_scope=contract2.write_scope,
  397. )
  398. assert not (await workspaces.snapshot(write_context)).paragraphs
  399. @pytest.mark.asyncio
  400. async def test_candidate_service_discards_existing_draft_workspace_on_abandoned_attempt(
  401. database: Any,
  402. ) -> None:
  403. _, sessions = database
  404. repository = SqlAlchemyScriptBusinessArtifactRepository(sessions)
  405. workspaces = SqlAlchemyCandidateWorkspaceRepository(sessions, repository)
  406. contract = _contract(ScriptTaskKind.PARAGRAPH)
  407. (ledger, accepted), context = _scope_fixture(contract)
  408. service = PhaseTwoCandidateService(
  409. bindings=cast(Any, _Bindings()),
  410. task_store=_TaskStore(ledger),
  411. framework_artifact_store=_FrameworkArtifacts(),
  412. artifacts=repository,
  413. accepted_inputs=accepted,
  414. active_frontier=ActiveFrontierResolver(),
  415. workspaces=workspaces,
  416. )
  417. await service.create_script_paragraph(
  418. payload={"paragraph_index": 1, "name": "draft", "content_range": {}},
  419. context=context,
  420. )
  421. await service.discard_attempt(
  422. root_trace_id="root",
  423. task_id="task",
  424. attempt_id="attempt",
  425. reason="validator blocked the bounded increment",
  426. )
  427. write_context = CandidateWriteContext(
  428. script_build_id=7,
  429. task_id="task",
  430. attempt_id="attempt",
  431. spec_version=1,
  432. objective=contract.objective,
  433. input_refs=(),
  434. write_scope=contract.write_scope,
  435. )
  436. assert (await workspaces.require(write_context)).state is ArtifactState.DISCARDED
  437. with pytest.raises(WorkspaceError, match="ATTEMPT_WORKSPACE_FROZEN"):
  438. await service.create_script_paragraph(
  439. payload={"paragraph_index": 2, "name": "late", "content_range": {}},
  440. context=context,
  441. )
  442. @pytest.mark.asyncio
  443. async def test_frozen_image_must_be_accepted_and_digest_and_mime_are_rechecked() -> None:
  444. content = b"\x89PNG\r\n\x1a\n" + b"safe"
  445. raw_ref = f"script-build://raw-artifacts/sha256/{sha256(content).hexdigest()}"
  446. ref = ArtifactRef("script-build://artifact-versions/2", "evidence", "2", DIGEST)
  447. item = AcceptedInput("d1", ref, ScriptTaskKind.DECODE_RETRIEVAL, SCOPE, "explicit")
  448. bundle = AcceptedInputBundleV1((item,), DIGEST)
  449. contract = _contract(ScriptTaskKind.PARAGRAPH)
  450. (ledger, accepted), context = _scope_fixture(contract, bundle=bundle)
  451. evidence = EvidenceRecordV1(
  452. "e",
  453. "decode",
  454. "retrieve_decode",
  455. {},
  456. ("source",),
  457. raw_ref,
  458. "summary",
  459. ("goal",),
  460. "high",
  461. (),
  462. DIGEST,
  463. datetime.now(UTC),
  464. )
  465. version = ArtifactVersion(
  466. 2,
  467. 7,
  468. "child",
  469. "child-attempt",
  470. 1,
  471. ArtifactKind.EVIDENCE,
  472. DIGEST,
  473. ArtifactState.FROZEN,
  474. evidence,
  475. datetime.now(UTC),
  476. datetime.now(UTC),
  477. )
  478. service = PhaseTwoCandidateService(
  479. bindings=cast(Any, _Bindings()),
  480. task_store=_TaskStore(ledger),
  481. framework_artifact_store=_FrameworkArtifacts(),
  482. artifacts=cast(Any, _Artifacts({ref.uri: version})),
  483. accepted_inputs=accepted,
  484. active_frontier=ActiveFrontierResolver(),
  485. workspaces=None,
  486. raw_artifacts=_RawStore(content),
  487. )
  488. images = await service.view_frozen_images(raw_artifact_refs=[raw_ref], context=context)
  489. assert images[0]["media_type"] == "image/png"
  490. assert "url" not in images[0]
  491. with pytest.raises(PhaseTwoCandidateError, match="outside the accepted"):
  492. await service.view_frozen_images(raw_artifact_refs=[_RAW_OTHER], context=context)
  493. @pytest.mark.asyncio
  494. async def test_active_frontier_write_conflict_is_not_resolved_by_completion_order() -> None:
  495. ref1 = ArtifactRef("script-build://artifact-versions/1", "paragraph", "1", DIGEST)
  496. ref2 = ArtifactRef("script-build://artifact-versions/2", "paragraph", "2", DIGEST)
  497. refs = (
  498. AcceptedDecisionRef("d1", ref1, SCOPE, ScriptTaskKind.PARAGRAPH),
  499. AcceptedDecisionRef("d2", ref2, SCOPE, ScriptTaskKind.PARAGRAPH),
  500. )
  501. contract = _contract(
  502. ScriptTaskKind.COMPOSE, refs=refs, adopted=("d1", "d2"), order=("d1", "d2")
  503. )
  504. bundle = AcceptedInputBundleV1(
  505. (
  506. AcceptedInput("d1", ref1, ScriptTaskKind.PARAGRAPH, SCOPE, "explicit"),
  507. AcceptedInput("d2", ref2, ScriptTaskKind.PARAGRAPH, SCOPE, "explicit"),
  508. ),
  509. DIGEST,
  510. )
  511. (ledger, accepted), context = _scope_fixture(contract, bundle=bundle)
  512. child_contract = _contract(ScriptTaskKind.PARAGRAPH)
  513. for decision_id in ("d1", "d2"):
  514. task_id = f"task-{decision_id}"
  515. ledger.tasks[task_id] = SimpleNamespace(task_id=task_id, attempt_ids=[f"a-{decision_id}"])
  516. ledger.attempts[f"a-{decision_id}"] = SimpleNamespace(
  517. attempt_id=f"a-{decision_id}", task_id=task_id, spec_version=1
  518. )
  519. ledger.decisions[decision_id] = SimpleNamespace(
  520. task_id=task_id, attempt_id=f"a-{decision_id}"
  521. )
  522. accepted.contracts[task_id] = child_contract
  523. service = PhaseTwoCandidateService(
  524. bindings=cast(Any, _Bindings()),
  525. task_store=_TaskStore(ledger),
  526. framework_artifact_store=_FrameworkArtifacts(),
  527. artifacts=cast(Any, _Artifacts({})),
  528. accepted_inputs=accepted,
  529. active_frontier=ActiveFrontierResolver(),
  530. workspaces=None,
  531. )
  532. with pytest.raises(PhaseTwoInputError, match="WRITE_SCOPE_CONFLICT"):
  533. await service.read_active_frontier(context=context)
  534. @pytest.mark.asyncio
  535. async def test_comparison_candidates_are_contract_derived_and_pass_fairness_precheck() -> None:
  536. candidate_refs = (
  537. ArtifactRef("script-build://artifact-versions/1", "paragraph", "1", DIGEST),
  538. ArtifactRef("script-build://artifact-versions/2", "paragraph", "2", DIGEST),
  539. )
  540. decision_refs = tuple(
  541. AcceptedDecisionRef(
  542. f"candidate-{index}",
  543. ref,
  544. SCOPE,
  545. ScriptTaskKind.PARAGRAPH,
  546. )
  547. for index, ref in enumerate(candidate_refs, start=1)
  548. )
  549. bundle = AcceptedInputBundleV1(
  550. tuple(
  551. AcceptedInput(
  552. decision_ref.decision_id,
  553. decision_ref.artifact_ref,
  554. ScriptTaskKind.PARAGRAPH,
  555. SCOPE,
  556. "explicit",
  557. )
  558. for decision_ref in decision_refs
  559. ),
  560. DIGEST,
  561. )
  562. contract = _contract(ScriptTaskKind.COMPARE, comparison_refs=decision_refs)
  563. (ledger, accepted), context = _scope_fixture(contract, bundle=bundle)
  564. artifacts = _Artifacts({})
  565. service = PhaseTwoCandidateService(
  566. bindings=cast(Any, _Bindings()),
  567. task_store=_TaskStore(ledger),
  568. framework_artifact_store=_FrameworkArtifacts(),
  569. artifacts=cast(Any, artifacts),
  570. accepted_inputs=accepted,
  571. active_frontier=ActiveFrontierResolver(),
  572. workspaces=None,
  573. )
  574. criterion_results = [
  575. {
  576. "criterion_id": "closed",
  577. "candidate_results": [
  578. {"artifact_ref": candidate_refs[0].uri, "reason": "uses one visible detail"},
  579. {"artifact_ref": candidate_refs[1].uri, "reason": "remains more abstract"},
  580. ],
  581. }
  582. ]
  583. frozen = await service.save_comparison_candidate(
  584. payload={
  585. "criterion_results": criterion_results,
  586. "conflicts": ["both candidates write the opening scope"],
  587. "recommendation": candidate_refs[0].uri,
  588. },
  589. context=context,
  590. )
  591. assert artifacts.frozen is not None
  592. comparison = artifacts.frozen.artifact
  593. assert isinstance(comparison, ComparisonArtifactV1)
  594. assert comparison.candidate_artifact_refs == tuple(ref.uri for ref in candidate_refs)
  595. assert comparison.criterion_results == tuple(criterion_results)
  596. frozen_ref = ArtifactRef.from_dict(frozen["artifact_ref"])
  597. service._framework_artifact_store = _SnapshotArtifacts(frozen_ref)
  598. rules = await service.deterministic_precheck(
  599. context={**context, "snapshot_id": "validation-snapshot"}
  600. )
  601. assert {str(item["rule_id"]): item["verdict"] for item in rules} == {
  602. "business-artifact-owner": "passed",
  603. "realized-content": "passed",
  604. "candidate-lineage": "passed",
  605. "comparison-fairness": "passed",
  606. }
  607. with pytest.raises(PhaseTwoCandidateError, match="derived from the frozen contract"):
  608. await service.save_comparison_candidate(
  609. payload={
  610. "candidate_artifact_refs": [candidate_refs[1].uri],
  611. "criterion_results": criterion_results,
  612. "recommendation": candidate_refs[1].uri,
  613. },
  614. context=context,
  615. )
  616. @pytest.mark.asyncio
  617. async def test_portfolio_adoption_and_canonical_digest_come_from_frozen_contract() -> None:
  618. ref1 = ArtifactRef("script-build://artifact-versions/1", "structured_script", "1", DIGEST)
  619. ref2 = ArtifactRef("script-build://artifact-versions/2", "structured_script", "2", DIGEST)
  620. refs = (
  621. AcceptedDecisionRef("d1", ref1, SCOPE, ScriptTaskKind.COMPOSE),
  622. AcceptedDecisionRef("d2", ref2, SCOPE, ScriptTaskKind.COMPOSE),
  623. )
  624. contract = _contract(
  625. ScriptTaskKind.CANDIDATE_PORTFOLIO,
  626. refs=refs,
  627. adopted=("d1",),
  628. held=("d2",),
  629. order=("d1",),
  630. )
  631. bundle = AcceptedInputBundleV1(
  632. (
  633. AcceptedInput("d1", ref1, ScriptTaskKind.COMPOSE, SCOPE, "explicit"),
  634. AcceptedInput("d2", ref2, ScriptTaskKind.COMPOSE, SCOPE, "explicit"),
  635. ),
  636. DIGEST,
  637. )
  638. (ledger, accepted), context = _scope_fixture(contract, bundle=bundle)
  639. paragraph = ScriptParagraphV1(1, 1, 1, None, "opening", {})
  640. def structured(identifier: int) -> ArtifactVersion:
  641. artifact = StructuredScriptArtifactV1(
  642. direction_ref="script-build://artifact-versions/10",
  643. input_closure_digest=DIGEST,
  644. paragraphs=(paragraph,),
  645. elements=(),
  646. paragraph_element_links=(),
  647. source_artifact_refs=(f"script-build://artifact-versions/{20 + identifier}",),
  648. goal_coverage=(
  649. GoalCoverage(
  650. "goal-1", (f"script-build://artifact-versions/{20 + identifier}",)
  651. ),
  652. ),
  653. evidence_refs=(),
  654. acceptance_notes=("accepted",),
  655. canonical_sha256=DIGEST,
  656. )
  657. return ArtifactVersion(
  658. identifier,
  659. 7,
  660. f"compose-{identifier}",
  661. f"compose-attempt-{identifier}",
  662. 1,
  663. ArtifactKind.STRUCTURED_SCRIPT,
  664. DIGEST,
  665. ArtifactState.FROZEN,
  666. artifact,
  667. datetime.now(UTC),
  668. datetime.now(UTC),
  669. )
  670. artifacts = _Artifacts({ref1.uri: structured(1), ref2.uri: structured(2)})
  671. service = PhaseTwoCandidateService(
  672. bindings=cast(Any, _Bindings()),
  673. task_store=_TaskStore(ledger),
  674. framework_artifact_store=_FrameworkArtifacts(),
  675. artifacts=cast(Any, artifacts),
  676. accepted_inputs=accepted,
  677. active_frontier=ActiveFrontierResolver(),
  678. workspaces=None,
  679. )
  680. result = await service.save_candidate_portfolio(
  681. payload={
  682. "unresolved_defects": [
  683. {
  684. "severity": "warning",
  685. "defect_code": "MINOR_STYLE",
  686. "criterion_id": "closed",
  687. "scope_ref": SCOPE,
  688. "observed_excerpt": "opening cadence is slightly repetitive",
  689. "evidence_refs": [],
  690. "invalidated_inputs": [],
  691. "recommended_action_class": "revise",
  692. }
  693. ],
  694. },
  695. context=context,
  696. )
  697. assert artifacts.frozen is not None
  698. portfolio = artifacts.frozen.artifact
  699. assert isinstance(portfolio, CandidatePortfolioArtifactV1)
  700. assert portfolio.accepted_decision_ids == ("d1",)
  701. assert portfolio.rejected_or_held_decision_ids == ("d2",)
  702. assert result["artifact_ref"]["digest"] == canonical_sha256(portfolio.content_payload()).wire
  703. with pytest.raises(PhaseTwoCandidateError, match="outside the accepted closure"):
  704. await service.save_candidate_portfolio(
  705. payload={
  706. "unresolved_defects": [
  707. {
  708. "severity": "warning",
  709. "defect_code": "UNTRUSTED_EVIDENCE",
  710. "criterion_id": "closed",
  711. "scope_ref": SCOPE,
  712. "observed_excerpt": "unsupported warning",
  713. "evidence_refs": ["script-build://artifact-versions/999"],
  714. "invalidated_inputs": [],
  715. "recommended_action_class": "revise",
  716. }
  717. ],
  718. },
  719. context=context,
  720. )
  721. with pytest.raises(ProtocolViolation, match=r"unsupported enum|must not be blank"):
  722. await service.save_candidate_portfolio(
  723. payload={
  724. "unresolved_defects": [
  725. {"severity": "warning", "defect_code": "INCOMPLETE_WARNING"}
  726. ],
  727. },
  728. context=context,
  729. )
  730. with pytest.raises(PhaseTwoCandidateError, match="derived from the frozen contract"):
  731. await service.save_candidate_portfolio(
  732. payload={"adopted_structured_script_ref": ref2.uri}, context=context
  733. )
  734. @pytest.mark.asyncio
  735. async def test_candidate_portfolio_command_replay_reuses_one_frozen_version(database: Any) -> None:
  736. _, sessions = database
  737. repository = SqlAlchemyScriptBusinessArtifactRepository(sessions)
  738. paragraph = ScriptParagraphV1(1, 1, 1, None, "opening", {})
  739. async def freeze_structured(identifier: int) -> tuple[ArtifactVersion, ArtifactRef]:
  740. return await repository.freeze(
  741. script_build_id=7,
  742. task_id=f"compose-{identifier}",
  743. attempt_id=f"compose-attempt-{identifier}",
  744. spec_version=1,
  745. artifact=StructuredScriptArtifactV1(
  746. direction_ref="script-build://artifact-versions/10",
  747. input_closure_digest=DIGEST,
  748. paragraphs=(paragraph,),
  749. elements=(),
  750. paragraph_element_links=(),
  751. source_artifact_refs=(f"script-build://artifact-versions/{20 + identifier}",),
  752. goal_coverage=(
  753. GoalCoverage(
  754. "goal-1", (f"script-build://artifact-versions/{20 + identifier}",)
  755. ),
  756. ),
  757. evidence_refs=(),
  758. acceptance_notes=("accepted",),
  759. ),
  760. )
  761. (_, stored_ref1), (_, stored_ref2) = await freeze_structured(1), await freeze_structured(2)
  762. ref1 = ArtifactRef(
  763. stored_ref1.uri,
  764. stored_ref1.kind,
  765. stored_ref1.version,
  766. stored_ref1.digest,
  767. )
  768. ref2 = ArtifactRef(
  769. stored_ref2.uri,
  770. stored_ref2.kind,
  771. stored_ref2.version,
  772. stored_ref2.digest,
  773. )
  774. refs = (
  775. AcceptedDecisionRef("d1", ref1, SCOPE, ScriptTaskKind.COMPOSE),
  776. AcceptedDecisionRef("d2", ref2, SCOPE, ScriptTaskKind.COMPOSE),
  777. )
  778. contract = _contract(
  779. ScriptTaskKind.CANDIDATE_PORTFOLIO,
  780. refs=refs,
  781. adopted=("d1",),
  782. held=("d2",),
  783. order=("d1",),
  784. )
  785. bundle = AcceptedInputBundleV1(
  786. (
  787. AcceptedInput("d1", ref1, ScriptTaskKind.COMPOSE, SCOPE, "explicit"),
  788. AcceptedInput("d2", ref2, ScriptTaskKind.COMPOSE, SCOPE, "explicit"),
  789. ),
  790. DIGEST,
  791. )
  792. (ledger, accepted), context = _scope_fixture(
  793. contract,
  794. bundle=bundle,
  795. task_id="portfolio",
  796. attempt_id="portfolio-attempt",
  797. )
  798. service = PhaseTwoCandidateService(
  799. bindings=cast(Any, _Bindings()),
  800. task_store=_TaskStore(ledger),
  801. framework_artifact_store=_FrameworkArtifacts(),
  802. artifacts=repository,
  803. accepted_inputs=accepted,
  804. active_frontier=ActiveFrontierResolver(),
  805. workspaces=None,
  806. )
  807. first = await service.save_candidate_portfolio(
  808. payload={"unresolved_defects": []}, context=context
  809. )
  810. replay = await service.save_candidate_portfolio(
  811. payload={"unresolved_defects": []}, context=context
  812. )
  813. frozen = await repository.get_by_attempt(
  814. script_build_id=7,
  815. task_id="portfolio",
  816. attempt_id="portfolio-attempt",
  817. )
  818. assert first == replay
  819. assert first["artifact_ref"]["version"] == str(frozen.artifact_version_id)
  820. assert frozen.artifact_type is ArtifactKind.CANDIDATE_PORTFOLIO
  821. _RAW_OTHER = "script-build://raw-artifacts/sha256/" + "b" * 64