From d5aafafec136bb0e7fe9745873bd26cdee8f0572 Mon Sep 17 00:00:00 2001 From: Charles Yao Date: Fri, 6 Nov 2026 10:57:52 -0800 Subject: [PATCH] fix(builder): retain source requirements across task resume --- .../scripts/builder-confirmed-tool-units.py | 2 + .../dify_builder/knowledge_source_intent.py | 10 +- .../test_knowledge_source_intent.py | 152 ++++++++++++++++++ 3 files changed, 162 insertions(+), 2 deletions(-) diff --git a/.github/scripts/builder-confirmed-tool-units.py b/.github/scripts/builder-confirmed-tool-units.py index f7c4cb66e2a..01950bb09b8 100644 --- a/.github/scripts/builder-confirmed-tool-units.py +++ b/.github/scripts/builder-confirmed-tool-units.py @@ -2918,6 +2918,7 @@ knowledge_cases.update( ) # Source-bound semantic intent: exact captured-negative and safety cases. a4_expected["tests.unit_tests.services.dify_builder.test_knowledge_source_intent"] = { + "test_actual_pending_business_form_stop_resume_keeps_logical_task_requirement": 2, "test_actual_submitted_control_refreshes_intent_in_the_existing_plan_call": 2, "test_analysis_preserves_fixed_grounded_source_intent_without_an_extra_call": 1, "test_confirmed_bool_source_cannot_match_a_changed_numeric_value": 1, @@ -2928,6 +2929,7 @@ a4_expected["tests.unit_tests.services.dify_builder.test_knowledge_source_intent "test_invalid_typed_submission_does_not_refresh_or_confirm_a_default": 5, "test_missing_or_malformed_semantics_keep_the_conservative_gate": 4, "test_missing_or_unknown_refresh_cannot_drop_a_new_typed_source_answer": 2, + "test_new_logical_task_does_not_inherit_positive_and_old_exclusion_is_stale": 1, "test_old_planning_intent_cannot_be_rebound_after_a_new_real_answer": 1, "test_planning_clarification_keeps_semantic_metadata_without_extra_calls": 1, "test_retained_source_cannot_be_waived_by_semantic_exclusion": 1, diff --git a/api/services/dify_builder/knowledge_source_intent.py b/api/services/dify_builder/knowledge_source_intent.py index 2e8dd064d75..b99bb234c34 100644 --- a/api/services/dify_builder/knowledge_source_intent.py +++ b/api/services/dify_builder/knowledge_source_intent.py @@ -122,6 +122,7 @@ def user_source(context: DifyBuilderContext) -> dict | None: "goal": context.goal_text, "answers": answers, "generation": context.task_control.get("generation", 0), + "logical_task_id": context.logical_task_id, }, ensure_ascii=False, allow_nan=False, @@ -156,15 +157,20 @@ def record_source_intent(context: DifyBuilderContext, value: object) -> None: **intent, "source_digest": canonical_digest(source), "generation": context.task_control.get("generation", 0), + "logical_task_id": context.logical_task_id, "required": required, } def established_source_required(context: DifyBuilderContext) -> bool: - """Keep established positive intent or a retained discovery goal sticky.""" + """Keep required intent for its logical task, including ordinary Resume. + + Execution generation fences stale exclusions/results; it does not authorize + dropping a positive requirement. New logical tasks do not inherit it. + """ record = context.knowledge_source_intent return ( - record.get("required") is True and record.get("generation") == context.task_control.get("generation", 0) + record.get("required") is True and record.get("logical_task_id") == context.logical_task_id ) or context.resource_selection.get("blocked_capability") == "knowledge_retrieval_phase_two" diff --git a/api/tests/unit_tests/services/dify_builder/test_knowledge_source_intent.py b/api/tests/unit_tests/services/dify_builder/test_knowledge_source_intent.py index 16445e27977..24c662061c7 100644 --- a/api/tests/unit_tests/services/dify_builder/test_knowledge_source_intent.py +++ b/api/tests/unit_tests/services/dify_builder/test_knowledge_source_intent.py @@ -2,6 +2,7 @@ import json from datetime import datetime +from typing import TYPE_CHECKING import pytest @@ -16,6 +17,12 @@ from tests.unit_tests.core.dify_builder.fakes import FakeDifyPort, InMemoryRepos from tests.unit_tests.services.dify_builder.test_bound_resource_plan_reconciliation import ReconciliationAgent from tests.unit_tests.services.dify_builder.test_build_cognition import _FakeInstance +if TYPE_CHECKING: + from services.dify_builder.repository import SqlDifyBuilderRepository + from services.dify_builder.service import DifyBuilderService + +pytest_plugins = ["tests.unit_tests.services.dify_builder.test_service"] + GOAL = ( "帮我做一个文本翻译工作流：用户输入普通文本 source_text 和目标语言 target_language，通过 LLM 翻译并输出 " "translation。" @@ -431,3 +438,148 @@ def test_semantic_source_survives_the_real_context_json_boundary() -> None: legacy = context_from_dict({"goal_text": GOAL}) assert legacy.knowledge_source_intent == legacy.confirmed_requirement_sources == {} assert current_source_intent(legacy) is None + + +@pytest.mark.parametrize("replacement_state", ["excluded", "unspecified"]) +def test_actual_pending_business_form_stop_resume_keeps_logical_task_requirement( + monkeypatch: pytest.MonkeyPatch, + service: "DifyBuilderService", + repo: "SqlDifyBuilderRepository", + replacement_state: str, +) -> None: + from core.dify_builder.models import Graph + from services.dify_builder.knowledge_source_intent import SOURCE_INPUT, record_source_intent + from services.dify_builder.task_lifecycle import advance_task_identity + from tests.unit_tests.services.dify_builder.test_service import ACCOUNT_ID, APP_ID, TENANT_ID, _actor + + goal = "Use CompanyFAQ knowledge base. Do not use another knowledge base." + session = Session( + app_id=APP_ID, + tenant_id=TENANT_ID, + owner_account_id=ACCOUNT_ID, + entry_mode=EntryMode.BUILD, + current_state=PcState.BUILD_GOAL_ANALYSIS, + ) + context = DifyBuilderContext( + goal_text=goal, + confirmed_requirement_keys=[], + form_fields=[{"key": "language", "label": "Answer language", "type": "text", "required": True}], + ) + repo.create_session(session, context, []) + task_id = advance_task_identity(session, context, EntryMode.BUILD) + assert task_id == f"{session.id}:1" + record_source_intent(context, {"state": "required", "evidence": ["Use CompanyFAQ knowledge base"]}) + base = repo.compare_and_advance(session.id, 1, session.current_state, context, []) + service._read_task_dsl_fn = lambda *_: (Graph(graph=Graph(nodes=[], edges=[]), features=Graph()), "resume-revision") + stopped = service.stop_task(session.id, _actor(), base_version=base, client_action_id="source-stop") + service.resume_task(session.id, _actor(), base_version=stopped.version, client_action_id="source-resume") + resumed, context = repo.get_session(session.id) + assert context.task_control["generation"] == 1 + assert context.logical_task_id == task_id + assert resumed.current_state == PcState.BUILD_GOAL_ANALYSIS + assert context.form_fields[0]["key"] == "language" + agent = ReconciliationAgent(None) + model = _FakeInstance( + [ + json.dumps( + { + "title": "No source", + "input_summary": "Text", + "output_summary": "Answer", + "plan": ["Return text"], + "knowledge_source_intent": { + "state": replacement_state, + "evidence": ["Do not use another knowledge base"], + }, + } + ) + ] + ) + calls: list[dict[str, object]] = [] + + def propose( + requirements: dict[str, object], *, goal_text: str, reply_language: str | None = None + ) -> WorkflowPlan | BuildClarification | None: + assert SOURCE_INPUT in requirements + calls.append(requirements) + return build.propose_plan_v1(model, requirements, goal_text=goal_text, reply_language=reply_language) + + def discover(*_args: object, **_kwargs: object) -> ResourceDiscovery: + return ResourceDiscovery([]) + + monkeypatch.setattr(agent, "propose_plan_v1", propose) + monkeypatch.setattr(agent, "discover_resources", discover) + runner = Runner(Env(agent=agent, repo=repo, dify=FakeDifyPort(), now=lambda: datetime.min), build_registry()) + result = runner.advance( + session.id, + Turn( + actor=_actor(), + action=Action(kind="submit_requirements", payload={"language": "English"}, base_version=resumed.version), + ), + ) + context = repo.get_session(session.id)[1] + assert result.current_state == PcState.BUILD_RESOURCE_RECOMMENDATION + assert context.logical_task_id == task_id + assert context.knowledge_source_intent["required"] is True + assert context.plan_items == [] + assert calls == model.calls == [] + assert "knowledge_scope_decision" not in context.resource_selection + assert context.confirmed_requirement_sources["language"]["value"] == "English" + # The existing current form and explicit waiver still own the goal change + # after Resume; neither the generation bump nor model metadata replaces it. + pending = runner.advance( + session.id, + Turn( + actor=_actor(), + action=Action(kind="confirm_resources", payload={"resource_ids": []}, base_version=result.version), + ), + ) + context = repo.get_session(session.id)[1] + field = context.form_fields[0] + waived = runner.advance( + session.id, + Turn( + actor=_actor(), + action=Action( + kind="submit_requirements", payload={field["key"]: field["options"][1]}, base_version=pending.version + ), + ), + ) + context = repo.get_session(session.id)[1] + assert waived.current_state == PcState.BUILD_RESOURCE_RECOMMENDATION + assert context.resource_selection["knowledge_scope_decision"]["choice"] == "waive" + assert context.resource_selection["knowledge_scope_decision"]["original_goal"] == goal + assert context.goal_text.startswith(goal) + assert context.goal_text != goal + assert len(calls) == len(model.calls) == 1 + assert context.built_node_ids == [] + + +def test_new_logical_task_does_not_inherit_positive_and_old_exclusion_is_stale() -> None: + from services.dify_builder.knowledge_source_intent import ( + current_source_intent, + established_source_required, + record_source_intent, + ) + from services.dify_builder.task_lifecycle import advance_task_identity + + session = Session(app_id="app", entry_mode=EntryMode.BUILD, current_state=PcState.BUILD_CAPABILITY_CHECK) + context = DifyBuilderContext( + goal_text="Use CompanyFAQ knowledge base. Do not use another knowledge base.", confirmed_requirement_keys=[] + ) + repo = InMemoryRepository() + repo.create_session(session, context, []) + first_task = advance_task_identity(session, context, EntryMode.BUILD) + record_source_intent(context, {"state": "required", "evidence": ["Use CompanyFAQ knowledge base"]}) + assert established_source_required(context) + second_task = advance_task_identity(session, context, EntryMode.BUILD) + assert second_task != first_task + assert not established_source_required(context) + assert current_source_intent(context) is None + context.goal_text = GOAL + record_source_intent(context, INTENT) + assert not established_source_required(context) + assert current_source_intent(context) == INTENT + # An execution resume invalidates only the exclusion freshness, not task identity. + context.task_control["generation"] = 1 + assert current_source_intent(context) is None