| 1 | """Tests for turn bootstrap and workflow preparation.""" |
| 2 | |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | from pathlib import Path |
| 6 | |
| 7 | import pytest |
| 8 | |
| 9 | from loader.agent.loop import AgentConfig |
| 10 | from loader.llm.base import CompletionResponse, ToolCall |
| 11 | from loader.runtime.completion_trace import CompletionTraceEntry |
| 12 | from loader.runtime.conversation import ConversationRuntime |
| 13 | from loader.runtime.dod import DefinitionOfDoneStore, create_definition_of_done |
| 14 | from loader.runtime.runtime_handle import RuntimeHandle |
| 15 | from tests.helpers.runtime_harness import ScriptedBackend |
| 16 | |
| 17 | |
| 18 | def non_streaming_config() -> AgentConfig: |
| 19 | """Shared config for direct turn-preparation tests.""" |
| 20 | |
| 21 | return AgentConfig(auto_context=False, stream=False, max_iterations=8) |
| 22 | |
| 23 | |
| 24 | @pytest.mark.asyncio |
| 25 | async def test_turn_preparation_bootstraps_execute_turn_state( |
| 26 | temp_dir: Path, |
| 27 | ) -> None: |
| 28 | backend = ScriptedBackend() |
| 29 | handle = RuntimeHandle( |
| 30 | backend=backend, |
| 31 | config=non_streaming_config(), |
| 32 | project_root=temp_dir, |
| 33 | ) |
| 34 | runtime = ConversationRuntime(handle) |
| 35 | events = [] |
| 36 | task = "Update README.md heading." |
| 37 | handle.session.last_completion_decision_code = "verification_passed" |
| 38 | handle.session.last_completion_decision_summary = ( |
| 39 | "accepted the response after verification evidence passed" |
| 40 | ) |
| 41 | handle.session.append_completion_trace_entry( |
| 42 | CompletionTraceEntry( |
| 43 | stage="definition_of_done", |
| 44 | outcome="complete", |
| 45 | decision_code="verification_passed", |
| 46 | decision_summary="accepted the response after verification evidence passed", |
| 47 | ) |
| 48 | ) |
| 49 | |
| 50 | async def capture(event) -> None: |
| 51 | events.append(event) |
| 52 | |
| 53 | prepared = await runtime.turn_preparation.prepare( |
| 54 | task=task, |
| 55 | emit=capture, |
| 56 | requested_mode="execute", |
| 57 | original_task=None, |
| 58 | on_user_question=None, |
| 59 | ) |
| 60 | |
| 61 | assert prepared.task == task |
| 62 | assert prepared.effective_task == task |
| 63 | assert prepared.executor is not None |
| 64 | assert prepared.rollback_plan is not None |
| 65 | assert prepared.effective_max_tokens >= 512 |
| 66 | assert prepared.summary.workflow_mode == "execute" |
| 67 | assert prepared.definition_of_done.current_mode == "execute" |
| 68 | assert prepared.definition_of_done.storage_path is not None |
| 69 | assert handle.session.current_task == task |
| 70 | assert handle.session.active_dod_path == prepared.definition_of_done.storage_path |
| 71 | assert handle.session.workflow_mode == "execute" |
| 72 | assert handle.session.last_completion_decision_code is None |
| 73 | assert handle.session.completion_trace == [] |
| 74 | assert any(event.type == "dod_status" for event in events) |
| 75 | assert any( |
| 76 | event.type == "workflow_mode" and event.workflow_mode == "execute" |
| 77 | for event in events |
| 78 | ) |
| 79 | |
| 80 | |
| 81 | @pytest.mark.asyncio |
| 82 | async def test_turn_preparation_can_bootstrap_clarify_handoff( |
| 83 | temp_dir: Path, |
| 84 | ) -> None: |
| 85 | backend = ScriptedBackend( |
| 86 | completions=[ |
| 87 | CompletionResponse( |
| 88 | content="I need one clarification before I proceed.", |
| 89 | tool_calls=[ |
| 90 | ToolCall( |
| 91 | id="ask-1", |
| 92 | name="AskUserQuestion", |
| 93 | arguments={ |
| 94 | "question": ( |
| 95 | "Should I keep the runtime change inside " |
| 96 | "src/loader/runtime/workflow_lanes.py?" |
| 97 | ), |
| 98 | }, |
| 99 | ) |
| 100 | ], |
| 101 | ), |
| 102 | CompletionResponse( |
| 103 | content="\n".join( |
| 104 | [ |
| 105 | "## Task Statement", |
| 106 | "Tighten Loader runtime clarify behavior.", |
| 107 | "", |
| 108 | "## Desired Outcome", |
| 109 | "- Keep the clarify handoff grounded in one runtime seam.", |
| 110 | "", |
| 111 | "## Non Goals", |
| 112 | "- Do not broaden into unrelated CLI changes.", |
| 113 | "", |
| 114 | "## Decision Boundaries", |
| 115 | "- Stop and confirm before crossing into other runtime modules.", |
| 116 | "", |
| 117 | "## Constraints", |
| 118 | "- Stay within the current repository.", |
| 119 | "", |
| 120 | "## Likely Touchpoints", |
| 121 | "- src/loader/runtime/workflow_lanes.py", |
| 122 | "", |
| 123 | "## Acceptance Criteria", |
| 124 | "- workflow_lanes.py remains the primary touchpoint.", |
| 125 | ] |
| 126 | ) |
| 127 | ), |
| 128 | ] |
| 129 | ) |
| 130 | handle = RuntimeHandle( |
| 131 | backend=backend, |
| 132 | config=non_streaming_config(), |
| 133 | project_root=temp_dir, |
| 134 | ) |
| 135 | runtime = ConversationRuntime(handle) |
| 136 | events = [] |
| 137 | asked_questions: list[str] = [] |
| 138 | |
| 139 | async def capture(event) -> None: |
| 140 | events.append(event) |
| 141 | |
| 142 | async def answer(question: str, _: list[str] | None) -> str: |
| 143 | asked_questions.append(question) |
| 144 | return "Yes, keep it inside workflow_lanes.py." |
| 145 | |
| 146 | prepared = await runtime.turn_preparation.prepare( |
| 147 | task="Tighten Loader runtime clarify behavior.", |
| 148 | emit=capture, |
| 149 | requested_mode="clarify", |
| 150 | original_task=None, |
| 151 | on_user_question=answer, |
| 152 | ) |
| 153 | |
| 154 | assert asked_questions == [ |
| 155 | "Should I keep the runtime change inside src/loader/runtime/workflow_lanes.py?" |
| 156 | ] |
| 157 | assert prepared.summary.workflow_mode == "execute" |
| 158 | assert prepared.definition_of_done.current_mode == "execute" |
| 159 | assert prepared.definition_of_done.clarify_brief is not None |
| 160 | assert Path(prepared.definition_of_done.clarify_brief).exists() |
| 161 | assert handle.session.workflow_mode == "execute" |
| 162 | assert any( |
| 163 | entry.kind == "clarify_exit" for entry in prepared.summary.workflow_timeline |
| 164 | ) |
| 165 | assert any( |
| 166 | entry.reason_code.startswith("post_clarify_") |
| 167 | for entry in prepared.summary.workflow_timeline |
| 168 | ) |
| 169 | assert [ |
| 170 | event.workflow_mode |
| 171 | for event in events |
| 172 | if event.type == "workflow_mode" and event.workflow_mode |
| 173 | ] == ["clarify", "execute"] |
| 174 | |
| 175 | |
| 176 | @pytest.mark.asyncio |
| 177 | async def test_turn_preparation_does_not_resume_latest_dod_from_older_session( |
| 178 | temp_dir: Path, |
| 179 | ) -> None: |
| 180 | backend = ScriptedBackend() |
| 181 | handle = RuntimeHandle( |
| 182 | backend=backend, |
| 183 | config=non_streaming_config(), |
| 184 | project_root=temp_dir, |
| 185 | ) |
| 186 | runtime = ConversationRuntime(handle) |
| 187 | task = "Update /tmp/fortran/index.html so the chapter list matches the real files." |
| 188 | |
| 189 | stale_dod = create_definition_of_done(task) |
| 190 | stale_dod.status = "fixing" |
| 191 | stale_dod.touched_files.append("/tmp/fortran/index.html") |
| 192 | stale_dod.mutating_actions.append("edit") |
| 193 | stale_path = DefinitionOfDoneStore(temp_dir).save(stale_dod) |
| 194 | |
| 195 | events = [] |
| 196 | |
| 197 | async def capture(event) -> None: |
| 198 | events.append(event) |
| 199 | |
| 200 | prepared = await runtime.turn_preparation.prepare( |
| 201 | task=task, |
| 202 | emit=capture, |
| 203 | requested_mode="execute", |
| 204 | original_task=None, |
| 205 | on_user_question=None, |
| 206 | ) |
| 207 | |
| 208 | assert prepared.definition_of_done.storage_path != str(stale_path) |
| 209 | assert prepared.definition_of_done.touched_files == [] |
| 210 | assert prepared.definition_of_done.mutating_actions == [] |
| 211 | assert prepared.definition_of_done.pending_items == ["Complete the requested work"] |
| 212 | |
| 213 | |
| 214 | @pytest.mark.asyncio |
| 215 | async def test_turn_preparation_resumes_active_session_dod( |
| 216 | temp_dir: Path, |
| 217 | ) -> None: |
| 218 | backend = ScriptedBackend() |
| 219 | handle = RuntimeHandle( |
| 220 | backend=backend, |
| 221 | config=non_streaming_config(), |
| 222 | project_root=temp_dir, |
| 223 | ) |
| 224 | runtime = ConversationRuntime(handle) |
| 225 | task = "Keep repairing the runtime state controller." |
| 226 | |
| 227 | existing_dod = create_definition_of_done(task) |
| 228 | existing_dod.status = "fixing" |
| 229 | existing_dod.pending_items.append("Collect verification evidence") |
| 230 | existing_dod.touched_files.append(str(temp_dir / "index.html")) |
| 231 | existing_path = DefinitionOfDoneStore(temp_dir).save(existing_dod) |
| 232 | handle.session.active_dod_path = str(existing_path) |
| 233 | |
| 234 | events = [] |
| 235 | |
| 236 | async def capture(event) -> None: |
| 237 | events.append(event) |
| 238 | |
| 239 | prepared = await runtime.turn_preparation.prepare( |
| 240 | task=task, |
| 241 | emit=capture, |
| 242 | requested_mode="execute", |
| 243 | original_task=None, |
| 244 | on_user_question=None, |
| 245 | ) |
| 246 | |
| 247 | assert prepared.definition_of_done.storage_path == str(existing_path) |
| 248 | assert prepared.definition_of_done.touched_files == [str(temp_dir / "index.html")] |
| 249 | assert prepared.definition_of_done.status == "fixing" |