| 1 | """Tests for response-repair helpers on RuntimeContext.""" |
| 2 | |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | import json |
| 6 | from pathlib import Path |
| 7 | from types import SimpleNamespace |
| 8 | |
| 9 | import pytest |
| 10 | |
| 11 | from loader.llm.base import Message, Role, ToolCall |
| 12 | from loader.runtime.context import RuntimeContext |
| 13 | from loader.runtime.dod import create_definition_of_done |
| 14 | from loader.runtime.path_display import display_runtime_path |
| 15 | from loader.runtime.permissions import ( |
| 16 | PermissionMode, |
| 17 | build_permission_policy, |
| 18 | load_permission_rules, |
| 19 | ) |
| 20 | from loader.runtime.recovery import RecoveryContext |
| 21 | from loader.runtime.repair import ResponseRepairer |
| 22 | from loader.tools.base import create_default_registry |
| 23 | from tests.helpers.runtime_harness import ScriptedBackend |
| 24 | |
| 25 | |
| 26 | class FakeSession: |
| 27 | def __init__(self) -> None: |
| 28 | self.messages = [] |
| 29 | |
| 30 | def append(self, message) -> None: |
| 31 | self.messages.append(message) |
| 32 | |
| 33 | |
| 34 | class FakeCodeFilter: |
| 35 | def reset(self) -> None: |
| 36 | return None |
| 37 | |
| 38 | |
| 39 | class FakeSafeguards: |
| 40 | def __init__(self) -> None: |
| 41 | self.action_tracker = object() |
| 42 | self.validator = object() |
| 43 | self.code_filter = FakeCodeFilter() |
| 44 | |
| 45 | def filter_stream_chunk(self, content: str) -> str: |
| 46 | return content |
| 47 | |
| 48 | def filter_complete_content(self, content: str) -> str: |
| 49 | return content |
| 50 | |
| 51 | def should_steer(self) -> bool: |
| 52 | return False |
| 53 | |
| 54 | def get_steering_message(self) -> str | None: |
| 55 | return None |
| 56 | |
| 57 | def record_response(self, content: str) -> None: |
| 58 | return None |
| 59 | |
| 60 | def detect_text_loop(self, content: str) -> tuple[bool, str]: |
| 61 | return False, "" |
| 62 | |
| 63 | def detect_loop(self) -> tuple[bool, str]: |
| 64 | return False, "" |
| 65 | |
| 66 | |
| 67 | def build_context( |
| 68 | *, |
| 69 | temp_dir: Path, |
| 70 | use_react: bool, |
| 71 | ) -> RuntimeContext: |
| 72 | registry = create_default_registry(temp_dir) |
| 73 | registry.configure_workspace_root(temp_dir) |
| 74 | rule_status = load_permission_rules(temp_dir) |
| 75 | policy = build_permission_policy( |
| 76 | active_mode=PermissionMode.WORKSPACE_WRITE, |
| 77 | workspace_root=temp_dir, |
| 78 | tool_requirements=registry.get_tool_requirements(), |
| 79 | rules=rule_status.rules, |
| 80 | ) |
| 81 | session = FakeSession() |
| 82 | return RuntimeContext( |
| 83 | project_root=temp_dir, |
| 84 | backend=ScriptedBackend(), |
| 85 | registry=registry, |
| 86 | session=session, # type: ignore[arg-type] |
| 87 | config=SimpleNamespace(force_react=use_react), |
| 88 | capability_profile=SimpleNamespace(supports_native_tools=not use_react), # type: ignore[arg-type] |
| 89 | project_context=None, |
| 90 | permission_policy=policy, |
| 91 | permission_config_status=rule_status, |
| 92 | workflow_mode="execute", |
| 93 | safeguards=FakeSafeguards(), |
| 94 | ) |
| 95 | |
| 96 | |
| 97 | def test_response_repairer_uses_runtime_parser_for_bracket_tool_fallback( |
| 98 | temp_dir: Path, |
| 99 | ) -> None: |
| 100 | context = build_context( |
| 101 | temp_dir=temp_dir, |
| 102 | use_react=False, |
| 103 | ) |
| 104 | repairer = ResponseRepairer(context) |
| 105 | |
| 106 | analysis = repairer.analyze_response( |
| 107 | content="I need clarification.", |
| 108 | response_content='[calls askuserquestion tool with: question="Which path?"]', |
| 109 | tool_calls=[], |
| 110 | extracted_iterations=0, |
| 111 | max_extracted_iterations=3, |
| 112 | ) |
| 113 | |
| 114 | assert analysis.tool_calls == [ |
| 115 | ToolCall( |
| 116 | id="call_0", |
| 117 | name="AskUserQuestion", |
| 118 | arguments={"question": "Which path?"}, |
| 119 | ) |
| 120 | ] |
| 121 | assert analysis.tool_source == "raw_text" |
| 122 | assert analysis.clear_stream is True |
| 123 | |
| 124 | |
| 125 | def test_response_repairer_recovers_todowrite_from_runtime_registry( |
| 126 | temp_dir: Path, |
| 127 | ) -> None: |
| 128 | context = build_context( |
| 129 | temp_dir=temp_dir, |
| 130 | use_react=False, |
| 131 | ) |
| 132 | repairer = ResponseRepairer(context) |
| 133 | |
| 134 | analysis = repairer.analyze_response( |
| 135 | content="I'll track the work first.", |
| 136 | response_content=json.dumps( |
| 137 | { |
| 138 | "name": "TodoWrite", |
| 139 | "arguments": { |
| 140 | "todos": [ |
| 141 | { |
| 142 | "content": "Run tests", |
| 143 | "active_form": "Running tests", |
| 144 | "status": "in_progress", |
| 145 | } |
| 146 | ] |
| 147 | }, |
| 148 | } |
| 149 | ), |
| 150 | tool_calls=[], |
| 151 | extracted_iterations=0, |
| 152 | max_extracted_iterations=3, |
| 153 | ) |
| 154 | |
| 155 | assert analysis.tool_source == "raw_text" |
| 156 | assert analysis.clear_stream is True |
| 157 | assert analysis.tool_calls == [ |
| 158 | ToolCall( |
| 159 | id="call_0", |
| 160 | name="TodoWrite", |
| 161 | arguments={ |
| 162 | "todos": [ |
| 163 | { |
| 164 | "content": "Run tests", |
| 165 | "active_form": "Running tests", |
| 166 | "status": "in_progress", |
| 167 | } |
| 168 | ] |
| 169 | }, |
| 170 | ) |
| 171 | ] |
| 172 | |
| 173 | |
| 174 | def test_response_repairer_fails_honestly_when_raw_tool_budget_is_exhausted( |
| 175 | temp_dir: Path, |
| 176 | ) -> None: |
| 177 | context = build_context( |
| 178 | temp_dir=temp_dir, |
| 179 | use_react=False, |
| 180 | ) |
| 181 | repairer = ResponseRepairer(context) |
| 182 | |
| 183 | analysis = repairer.analyze_response( |
| 184 | content=json.dumps( |
| 185 | { |
| 186 | "name": "read", |
| 187 | "arguments": {"file_path": "README.md"}, |
| 188 | } |
| 189 | ), |
| 190 | response_content=json.dumps( |
| 191 | { |
| 192 | "name": "read", |
| 193 | "arguments": {"file_path": "README.md"}, |
| 194 | } |
| 195 | ), |
| 196 | tool_calls=[], |
| 197 | extracted_iterations=3, |
| 198 | max_extracted_iterations=3, |
| 199 | ) |
| 200 | |
| 201 | assert analysis.should_stop is True |
| 202 | assert analysis.final_response == ( |
| 203 | "I couldn't safely continue because the model kept emitting raw-text " |
| 204 | "tool calls instead of proper tool invocations. Please try again or " |
| 205 | "switch to a different backend/model." |
| 206 | ) |
| 207 | assert analysis.failure == "raw-text tool recovery budget exhausted" |
| 208 | assert "Let me know if you'd like me to continue" not in analysis.final_response |
| 209 | |
| 210 | |
| 211 | def test_empty_response_retry_message_surfaces_missing_planned_artifacts_and_working_note( |
| 212 | temp_dir: Path, |
| 213 | ) -> None: |
| 214 | context = build_context( |
| 215 | temp_dir=temp_dir, |
| 216 | use_react=False, |
| 217 | ) |
| 218 | repairer = ResponseRepairer(context) |
| 219 | implementation_plan = temp_dir / "implementation.md" |
| 220 | implementation_plan.write_text( |
| 221 | "\n".join( |
| 222 | [ |
| 223 | "# Implementation Plan", |
| 224 | "", |
| 225 | "## File Changes", |
| 226 | f"- `{temp_dir / 'guides' / 'nginx' / 'index.html'}`", |
| 227 | f"- `{temp_dir / 'guides' / 'nginx' / 'chapters'}`", |
| 228 | "", |
| 229 | ] |
| 230 | ) |
| 231 | ) |
| 232 | first_artifact = temp_dir / "guides" / "nginx" / "index.html" |
| 233 | first_artifact.parent.mkdir(parents=True) |
| 234 | first_artifact.write_text("<html></html>\n") |
| 235 | |
| 236 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 237 | dod.implementation_plan = str(implementation_plan) |
| 238 | dod.touched_files.append(str(first_artifact)) |
| 239 | dod.completed_items.append("Create the main index.html file") |
| 240 | dod.pending_items.append("Create each chapter file in sequence") |
| 241 | |
| 242 | context.session.append( |
| 243 | SimpleNamespace( |
| 244 | role="tool", |
| 245 | content=( |
| 246 | "Observation [notepad_write_working]: Result: " |
| 247 | "- [2026-04-21T19:17:34Z] Creating fifth chapter file: Advanced configurations" |
| 248 | ), |
| 249 | ) |
| 250 | ) |
| 251 | |
| 252 | decision = repairer.handle_empty_response( |
| 253 | task="Create a multi-file nginx guide.", |
| 254 | original_task=None, |
| 255 | empty_retry_count=1, |
| 256 | max_empty_retries=2, |
| 257 | dod=dod, |
| 258 | ) |
| 259 | |
| 260 | assert decision.should_continue is True |
| 261 | assert decision.retry_message is not None |
| 262 | assert "Latest working note: Creating fifth chapter file: Advanced configurations" in decision.retry_message |
| 263 | assert "Confirmed touched files: `index.html`" in decision.retry_message |
| 264 | assert "Confirmed completed work: Create the main index.html file" in decision.retry_message |
| 265 | assert "Next pending item: Create each chapter file in sequence" in decision.retry_message |
| 266 | assert "Continue from the confirmed progress below instead of restarting." in decision.retry_message |
| 267 | |
| 268 | |
| 269 | def test_empty_response_retry_mentions_write_can_create_missing_parent_directories( |
| 270 | temp_dir: Path, |
| 271 | ) -> None: |
| 272 | context = build_context( |
| 273 | temp_dir=temp_dir, |
| 274 | use_react=False, |
| 275 | ) |
| 276 | repairer = ResponseRepairer(context) |
| 277 | |
| 278 | guide_root = temp_dir / "guides" / "nginx" |
| 279 | index_path = guide_root / "index.html" |
| 280 | |
| 281 | implementation_plan = temp_dir / "implementation.md" |
| 282 | implementation_plan.write_text( |
| 283 | "\n".join( |
| 284 | [ |
| 285 | "# Implementation Plan", |
| 286 | "", |
| 287 | "## File Changes", |
| 288 | f"- `{index_path}`", |
| 289 | "", |
| 290 | ] |
| 291 | ) |
| 292 | ) |
| 293 | |
| 294 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 295 | dod.implementation_plan = str(implementation_plan) |
| 296 | dod.pending_items.extend( |
| 297 | [ |
| 298 | "Create nginx guide directory structure", |
| 299 | "Write main index.html for nginx guide", |
| 300 | ] |
| 301 | ) |
| 302 | |
| 303 | decision = repairer.handle_empty_response( |
| 304 | task="Create a multi-file nginx guide.", |
| 305 | original_task=None, |
| 306 | empty_retry_count=1, |
| 307 | max_empty_retries=2, |
| 308 | dod=dod, |
| 309 | ) |
| 310 | |
| 311 | assert decision.should_continue is True |
| 312 | assert decision.retry_message is not None |
| 313 | assert ( |
| 314 | "Resume with this exact next step: create `index.html`." |
| 315 | in decision.retry_message |
| 316 | ) |
| 317 | assert ( |
| 318 | "Prefer one `write` call for " |
| 319 | f"`{display_runtime_path(index_path)}` before any more reference reads." |
| 320 | in decision.retry_message |
| 321 | ) |
| 322 | assert ( |
| 323 | "The `write` tool can create that file's parent directories automatically, so do the write in one step instead of stopping for a separate mkdir." |
| 324 | in decision.retry_message |
| 325 | ) |
| 326 | assert ( |
| 327 | f'Emit this tool shape now: `write(file_path="{display_runtime_path(index_path)}", content="...")`.' |
| 328 | in decision.retry_message |
| 329 | ) |
| 330 | assert ( |
| 331 | "Write a compact but real initial version of this file now, then refine or expand it in later edits." |
| 332 | in decision.retry_message |
| 333 | ) |
| 334 | assert "Do not restart discovery unless one specific missing fact blocks this step." in decision.retry_message |
| 335 | |
| 336 | |
| 337 | def test_empty_response_retry_uses_directory_creation_for_setup_targets( |
| 338 | temp_dir: Path, |
| 339 | ) -> None: |
| 340 | context = build_context( |
| 341 | temp_dir=temp_dir, |
| 342 | use_react=False, |
| 343 | ) |
| 344 | repairer = ResponseRepairer(context) |
| 345 | |
| 346 | guide_root = temp_dir / "guides" / "nginx" |
| 347 | chapters_path = guide_root / "chapters" |
| 348 | index_path = guide_root / "index.html" |
| 349 | |
| 350 | implementation_plan = temp_dir / "implementation.md" |
| 351 | implementation_plan.write_text( |
| 352 | "\n".join( |
| 353 | [ |
| 354 | "# Implementation Plan", |
| 355 | "", |
| 356 | "## File Changes", |
| 357 | f"- `{chapters_path}/`", |
| 358 | f"- `{index_path}`", |
| 359 | "", |
| 360 | ] |
| 361 | ) |
| 362 | ) |
| 363 | |
| 364 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 365 | dod.implementation_plan = str(implementation_plan) |
| 366 | dod.pending_items.extend( |
| 367 | [ |
| 368 | "Create the nginx directory structure", |
| 369 | "Create the main index.html file for nginx guide", |
| 370 | ] |
| 371 | ) |
| 372 | |
| 373 | decision = repairer.handle_empty_response( |
| 374 | task="Create a multi-file nginx guide.", |
| 375 | original_task=None, |
| 376 | empty_retry_count=1, |
| 377 | max_empty_retries=2, |
| 378 | dod=dod, |
| 379 | ) |
| 380 | |
| 381 | assert decision.should_continue is True |
| 382 | assert decision.retry_message is not None |
| 383 | assert ( |
| 384 | "Resume with this exact next step: continue `Create the nginx directory structure` " |
| 385 | "by creating `chapters/`." |
| 386 | in decision.retry_message |
| 387 | ) |
| 388 | assert ( |
| 389 | "Prefer one concrete directory-creation step for " |
| 390 | f"`{display_runtime_path(chapters_path)}` before more research." |
| 391 | in decision.retry_message |
| 392 | ) |
| 393 | expected_command = f"mkdir -p {display_runtime_path(chapters_path)}" |
| 394 | assert ( |
| 395 | f'Emit this tool shape now: `bash(command="{expected_command}")`.' |
| 396 | in decision.retry_message |
| 397 | ) |
| 398 | assert f'write(file_path="{display_runtime_path(chapters_path)}"' not in decision.retry_message |
| 399 | |
| 400 | |
| 401 | def test_empty_response_retry_uses_home_relative_path_for_home_artifacts( |
| 402 | temp_dir: Path, |
| 403 | monkeypatch: pytest.MonkeyPatch, |
| 404 | ) -> None: |
| 405 | monkeypatch.setenv("HOME", str(temp_dir.resolve(strict=False))) |
| 406 | context = build_context( |
| 407 | temp_dir=temp_dir, |
| 408 | use_react=False, |
| 409 | ) |
| 410 | repairer = ResponseRepairer(context) |
| 411 | |
| 412 | guide_root = temp_dir / "Loader" / "guides" / "nginx" |
| 413 | index_path = guide_root / "index.html" |
| 414 | |
| 415 | implementation_plan = temp_dir / "implementation.md" |
| 416 | implementation_plan.write_text( |
| 417 | "\n".join( |
| 418 | [ |
| 419 | "# Implementation Plan", |
| 420 | "", |
| 421 | "## File Changes", |
| 422 | f"- `{index_path}`", |
| 423 | "", |
| 424 | ] |
| 425 | ) |
| 426 | ) |
| 427 | |
| 428 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 429 | dod.implementation_plan = str(implementation_plan) |
| 430 | dod.pending_items.extend( |
| 431 | [ |
| 432 | "Create nginx guide directory structure", |
| 433 | "Write main index.html for nginx guide", |
| 434 | ] |
| 435 | ) |
| 436 | |
| 437 | decision = repairer.handle_empty_response( |
| 438 | task="Create a multi-file nginx guide.", |
| 439 | original_task=None, |
| 440 | empty_retry_count=1, |
| 441 | max_empty_retries=2, |
| 442 | dod=dod, |
| 443 | ) |
| 444 | |
| 445 | assert decision.should_continue is True |
| 446 | assert decision.retry_message is not None |
| 447 | assert "`~/Loader/guides/nginx/index.html`" in decision.retry_message |
| 448 | assert ( |
| 449 | 'Emit this tool shape now: `write(file_path="~/Loader/guides/nginx/index.html", content="...")`.' |
| 450 | in decision.retry_message |
| 451 | ) |
| 452 | assert ( |
| 453 | "Write a compact but real initial version of this file now, then refine or expand it in later edits." |
| 454 | in decision.retry_message |
| 455 | ) |
| 456 | |
| 457 | |
| 458 | def test_empty_response_retry_recovers_blocked_empty_file_path_to_concrete_target( |
| 459 | temp_dir: Path, |
| 460 | ) -> None: |
| 461 | context = build_context( |
| 462 | temp_dir=temp_dir, |
| 463 | use_react=False, |
| 464 | ) |
| 465 | repairer = ResponseRepairer(context) |
| 466 | |
| 467 | guide_root = temp_dir / "guides" / "nginx" |
| 468 | chapters = guide_root / "chapters" |
| 469 | chapters.mkdir(parents=True) |
| 470 | index_path = guide_root / "index.html" |
| 471 | first_chapter = chapters / "01-introduction.html" |
| 472 | second_chapter = chapters / "02-installation.html" |
| 473 | index_path.write_text("<html></html>\n") |
| 474 | first_chapter.write_text("<h1>Intro</h1>\n") |
| 475 | |
| 476 | implementation_plan = temp_dir / "implementation.md" |
| 477 | implementation_plan.write_text( |
| 478 | "\n".join( |
| 479 | [ |
| 480 | "# Implementation Plan", |
| 481 | "", |
| 482 | "## File Changes", |
| 483 | f"- `{index_path}`", |
| 484 | f"- `{first_chapter}`", |
| 485 | f"- `{second_chapter}`", |
| 486 | "", |
| 487 | ] |
| 488 | ) |
| 489 | ) |
| 490 | |
| 491 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 492 | dod.implementation_plan = str(implementation_plan) |
| 493 | dod.touched_files.extend([str(index_path), str(first_chapter)]) |
| 494 | dod.pending_items.append("Creating Chapter 2: Installation and Setup") |
| 495 | |
| 496 | context.recovery_context = RecoveryContext( |
| 497 | original_tool="write", |
| 498 | original_args={"file_path": "", "content": "<html></html>\n"}, |
| 499 | ) |
| 500 | context.recovery_context.add_attempt( |
| 501 | "write", |
| 502 | {"file_path": "", "content": "<html></html>\n"}, |
| 503 | "Empty file path", |
| 504 | ) |
| 505 | |
| 506 | decision = repairer.handle_empty_response( |
| 507 | task="Create a multi-file nginx guide.", |
| 508 | original_task=None, |
| 509 | empty_retry_count=1, |
| 510 | max_empty_retries=2, |
| 511 | dod=dod, |
| 512 | ) |
| 513 | |
| 514 | assert decision.should_continue is True |
| 515 | assert decision.retry_message is not None |
| 516 | assert ( |
| 517 | "Last tool failure: resend `write` for " |
| 518 | f"`{display_runtime_path(second_chapter)}` with a valid `file_path` and real `content`." |
| 519 | in decision.retry_message |
| 520 | ) |
| 521 | assert "Do not leave `file_path` empty" in decision.retry_message |
| 522 | assert ( |
| 523 | f'Emit this tool shape now: `write(file_path="{display_runtime_path(second_chapter)}", content="...")`.' |
| 524 | in decision.retry_message |
| 525 | ) |
| 526 | |
| 527 | |
| 528 | def test_empty_response_retry_respects_discovery_first_pending_step( |
| 529 | temp_dir: Path, |
| 530 | ) -> None: |
| 531 | context = build_context( |
| 532 | temp_dir=temp_dir, |
| 533 | use_react=False, |
| 534 | ) |
| 535 | repairer = ResponseRepairer(context) |
| 536 | |
| 537 | implementation_plan = temp_dir / "implementation.md" |
| 538 | implementation_plan.write_text( |
| 539 | "\n".join( |
| 540 | [ |
| 541 | "# Implementation Plan", |
| 542 | "", |
| 543 | "## File Changes", |
| 544 | f"- `{temp_dir / 'guides' / 'nginx' / 'index.html'}`", |
| 545 | f"- `{temp_dir / 'guides' / 'nginx' / 'chapters'}`", |
| 546 | "", |
| 547 | ] |
| 548 | ) |
| 549 | ) |
| 550 | |
| 551 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 552 | dod.implementation_plan = str(implementation_plan) |
| 553 | dod.pending_items.extend( |
| 554 | [ |
| 555 | "First, examine the existing fortran guide structure and content to understand the format", |
| 556 | "Create the nginx directory structure", |
| 557 | "Develop the main index.html file for the nginx guide", |
| 558 | ] |
| 559 | ) |
| 560 | |
| 561 | context.session.append( |
| 562 | SimpleNamespace( |
| 563 | role="tool", |
| 564 | content=( |
| 565 | "Observation [notepad_write_working]: Result: " |
| 566 | "- [2026-04-22T22:42:18Z] Analyzing the fortran guide structure before creating nginx guide" |
| 567 | ), |
| 568 | ) |
| 569 | ) |
| 570 | |
| 571 | decision = repairer.handle_empty_response( |
| 572 | task="Create a multi-file nginx guide.", |
| 573 | original_task=None, |
| 574 | empty_retry_count=1, |
| 575 | max_empty_retries=2, |
| 576 | dod=dod, |
| 577 | ) |
| 578 | |
| 579 | assert decision.should_continue is True |
| 580 | assert decision.retry_message is not None |
| 581 | assert ( |
| 582 | "Resume with this exact next step: advance `First, examine the existing fortran guide structure and content to understand the format`." |
| 583 | in decision.retry_message |
| 584 | ) |
| 585 | assert "one concrete evidence-gathering tool call" in decision.retry_message |
| 586 | assert "Resume with this exact next step: create `index.html`." not in decision.retry_message |
| 587 | |
| 588 | |
| 589 | def test_empty_response_retry_budget_extends_for_late_stage_multi_artifact_progress( |
| 590 | temp_dir: Path, |
| 591 | ) -> None: |
| 592 | context = build_context( |
| 593 | temp_dir=temp_dir, |
| 594 | use_react=False, |
| 595 | ) |
| 596 | repairer = ResponseRepairer(context) |
| 597 | |
| 598 | guide_root = temp_dir / "guides" / "nginx" |
| 599 | chapters = guide_root / "chapters" |
| 600 | chapters.mkdir(parents=True) |
| 601 | index_path = guide_root / "index.html" |
| 602 | chapter_one = chapters / "01-getting-started.html" |
| 603 | chapter_two = chapters / "02-installation.html" |
| 604 | chapter_three = chapters / "03-first-website.html" |
| 605 | chapter_four = chapters / "04-configuration-basics.html" |
| 606 | index_path.write_text("<html></html>\n") |
| 607 | chapter_one.write_text("<h1>One</h1>\n") |
| 608 | chapter_two.write_text("<h1>Two</h1>\n") |
| 609 | chapter_three.write_text("<h1>Three</h1>\n") |
| 610 | |
| 611 | implementation_plan = temp_dir / "implementation.md" |
| 612 | implementation_plan.write_text( |
| 613 | "\n".join( |
| 614 | [ |
| 615 | "# Implementation Plan", |
| 616 | "", |
| 617 | "## File Changes", |
| 618 | f"- `{guide_root}/`", |
| 619 | f"- `{chapters}/`", |
| 620 | f"- `{index_path}`", |
| 621 | f"- `{chapter_one}`", |
| 622 | f"- `{chapter_two}`", |
| 623 | f"- `{chapter_three}`", |
| 624 | f"- `{chapter_four}`", |
| 625 | "", |
| 626 | ] |
| 627 | ) |
| 628 | ) |
| 629 | |
| 630 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 631 | dod.implementation_plan = str(implementation_plan) |
| 632 | dod.touched_files.extend( |
| 633 | [str(index_path), str(chapter_one), str(chapter_two), str(chapter_three)] |
| 634 | ) |
| 635 | dod.completed_items.extend( |
| 636 | [ |
| 637 | "Create the directory structure for the new nginx guide", |
| 638 | "Create the main index.html file with proper structure", |
| 639 | ] |
| 640 | ) |
| 641 | dod.pending_items.append("Create each chapter file in sequence") |
| 642 | |
| 643 | decision = repairer.handle_empty_response( |
| 644 | task="Create a multi-file nginx guide.", |
| 645 | original_task=None, |
| 646 | empty_retry_count=3, |
| 647 | max_empty_retries=2, |
| 648 | dod=dod, |
| 649 | ) |
| 650 | |
| 651 | assert decision.should_continue is True |
| 652 | assert decision.retry_message is not None |
| 653 | assert "retry 3/4" in decision.retry_message |
| 654 | assert "Follow the same full-payload one-file-at-a-time write pattern" in decision.retry_message |
| 655 | |
| 656 | |
| 657 | def test_empty_response_retry_budget_extends_when_concrete_next_output_is_known( |
| 658 | temp_dir: Path, |
| 659 | ) -> None: |
| 660 | context = build_context( |
| 661 | temp_dir=temp_dir, |
| 662 | use_react=False, |
| 663 | ) |
| 664 | repairer = ResponseRepairer(context) |
| 665 | |
| 666 | implementation_plan = temp_dir / "implementation.md" |
| 667 | implementation_plan.write_text( |
| 668 | "\n".join( |
| 669 | [ |
| 670 | "# Implementation Plan", |
| 671 | "", |
| 672 | "## File Changes", |
| 673 | f"- `{temp_dir / 'guides' / 'nginx' / 'index.html'}`", |
| 674 | f"- `{temp_dir / 'guides' / 'nginx' / 'chapters'}`", |
| 675 | "", |
| 676 | ] |
| 677 | ) |
| 678 | ) |
| 679 | |
| 680 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 681 | dod.implementation_plan = str(implementation_plan) |
| 682 | dod.pending_items.append("Develop the main index.html file for the nginx guide") |
| 683 | |
| 684 | decision = repairer.handle_empty_response( |
| 685 | task="Create a multi-file nginx guide.", |
| 686 | original_task=None, |
| 687 | empty_retry_count=3, |
| 688 | max_empty_retries=2, |
| 689 | dod=dod, |
| 690 | ) |
| 691 | |
| 692 | assert decision.should_continue is True |
| 693 | assert decision.retry_message is not None |
| 694 | assert "retry 3/4" in decision.retry_message |
| 695 | assert "Next missing planned artifact: `index.html`" in decision.retry_message |
| 696 | assert ( |
| 697 | "Resume with this exact next step: continue `Develop the main index.html file for the nginx guide` " |
| 698 | "by creating `index.html`." |
| 699 | in decision.retry_message |
| 700 | ) |
| 701 | |
| 702 | |
| 703 | def test_empty_response_retry_budget_extends_further_after_first_output_file_exists( |
| 704 | temp_dir: Path, |
| 705 | ) -> None: |
| 706 | context = build_context( |
| 707 | temp_dir=temp_dir, |
| 708 | use_react=False, |
| 709 | ) |
| 710 | repairer = ResponseRepairer(context) |
| 711 | |
| 712 | guide_root = temp_dir / "guides" / "nginx" |
| 713 | chapters = guide_root / "chapters" |
| 714 | guide_root.mkdir(parents=True) |
| 715 | chapters.mkdir() |
| 716 | index_path = guide_root / "index.html" |
| 717 | index_path.write_text("<html></html>\n") |
| 718 | |
| 719 | implementation_plan = temp_dir / "implementation.md" |
| 720 | implementation_plan.write_text( |
| 721 | "\n".join( |
| 722 | [ |
| 723 | "# Implementation Plan", |
| 724 | "", |
| 725 | "## File Changes", |
| 726 | f"- `{chapters}/`", |
| 727 | f"- `{index_path}`", |
| 728 | "", |
| 729 | ] |
| 730 | ) |
| 731 | ) |
| 732 | |
| 733 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 734 | dod.implementation_plan = str(implementation_plan) |
| 735 | dod.touched_files.append(str(index_path)) |
| 736 | dod.completed_items.extend( |
| 737 | [ |
| 738 | "Create the new nginx guide directory structure", |
| 739 | "Develop the main index.html file with proper structure", |
| 740 | ] |
| 741 | ) |
| 742 | dod.pending_items.append("Create 01-introduction.html") |
| 743 | |
| 744 | decision = repairer.handle_empty_response( |
| 745 | task="Create a multi-file nginx guide.", |
| 746 | original_task=None, |
| 747 | empty_retry_count=5, |
| 748 | max_empty_retries=2, |
| 749 | dod=dod, |
| 750 | ) |
| 751 | |
| 752 | assert decision.should_continue is True |
| 753 | assert decision.retry_message is not None |
| 754 | assert "retry 5/6" in decision.retry_message |
| 755 | assert "Continue `Create 01-introduction.html` by creating `01-introduction.html`." in decision.retry_message |
| 756 | assert 'Emit this tool shape now: `write(file_path="' in decision.retry_message |
| 757 | assert "No narration, no TodoWrite, no rereads, and no empty response" in decision.retry_message |
| 758 | |
| 759 | |
| 760 | def test_empty_response_retry_uses_compact_prompt_after_substantial_progress( |
| 761 | temp_dir: Path, |
| 762 | ) -> None: |
| 763 | context = build_context( |
| 764 | temp_dir=temp_dir, |
| 765 | use_react=False, |
| 766 | ) |
| 767 | context.session.messages.append( |
| 768 | SimpleNamespace( |
| 769 | content=( |
| 770 | "Observation [notepad_write_working]: Result: " |
| 771 | "- [2026-04-23T19:00:00Z] Creating fifth chapter file: Advanced features" |
| 772 | ) |
| 773 | ) |
| 774 | ) |
| 775 | repairer = ResponseRepairer(context) |
| 776 | |
| 777 | guide_root = temp_dir / "guides" / "nginx" |
| 778 | chapters = guide_root / "chapters" |
| 779 | chapters.mkdir(parents=True) |
| 780 | index_path = guide_root / "index.html" |
| 781 | chapter_one = chapters / "01-getting-started.html" |
| 782 | chapter_two = chapters / "02-installation.html" |
| 783 | chapter_three = chapters / "03-first-website.html" |
| 784 | chapter_four = chapters / "04-configuration-basics.html" |
| 785 | chapter_five = chapters / "05-advanced-features.html" |
| 786 | index_path.write_text("<html></html>\n") |
| 787 | chapter_one.write_text("<h1>One</h1>\n") |
| 788 | chapter_two.write_text("<h1>Two</h1>\n") |
| 789 | chapter_three.write_text("<h1>Three</h1>\n") |
| 790 | chapter_four.write_text("<h1>Four</h1>\n") |
| 791 | |
| 792 | implementation_plan = temp_dir / "implementation.md" |
| 793 | implementation_plan.write_text( |
| 794 | "\n".join( |
| 795 | [ |
| 796 | "# Implementation Plan", |
| 797 | "", |
| 798 | "## File Changes", |
| 799 | f"- `{guide_root}/`", |
| 800 | f"- `{chapters}/`", |
| 801 | f"- `{index_path}`", |
| 802 | f"- `{chapter_one}`", |
| 803 | f"- `{chapter_two}`", |
| 804 | f"- `{chapter_three}`", |
| 805 | f"- `{chapter_four}`", |
| 806 | f"- `{chapter_five}`", |
| 807 | "", |
| 808 | ] |
| 809 | ) |
| 810 | ) |
| 811 | |
| 812 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 813 | dod.implementation_plan = str(implementation_plan) |
| 814 | dod.touched_files.extend( |
| 815 | [str(index_path), str(chapter_one), str(chapter_two), str(chapter_three)] |
| 816 | ) |
| 817 | dod.completed_items.extend( |
| 818 | [ |
| 819 | "Create the directory structure for the new nginx guide", |
| 820 | "Create the main index.html file with proper structure", |
| 821 | ] |
| 822 | ) |
| 823 | dod.pending_items.append("Create each chapter file in sequence") |
| 824 | |
| 825 | decision = repairer.handle_empty_response( |
| 826 | task="Create a multi-file nginx guide.", |
| 827 | original_task=None, |
| 828 | empty_retry_count=3, |
| 829 | max_empty_retries=2, |
| 830 | dod=dod, |
| 831 | ) |
| 832 | |
| 833 | assert decision.should_continue is True |
| 834 | assert decision.retry_message is not None |
| 835 | assert "Continue from the exact next step below." in decision.retry_message |
| 836 | assert "Latest working note:" not in decision.retry_message |
| 837 | assert "Confirmed completed work:" not in decision.retry_message |
| 838 | assert "Next pending item:" not in decision.retry_message |
| 839 | |
| 840 | |
| 841 | def test_empty_response_retry_points_at_next_output_file_when_planned_directory_is_empty( |
| 842 | temp_dir: Path, |
| 843 | ) -> None: |
| 844 | context = build_context( |
| 845 | temp_dir=temp_dir, |
| 846 | use_react=False, |
| 847 | ) |
| 848 | repairer = ResponseRepairer(context) |
| 849 | |
| 850 | guide_root = temp_dir / "guides" / "nginx" |
| 851 | chapters = guide_root / "chapters" |
| 852 | chapters.mkdir(parents=True) |
| 853 | index_path = guide_root / "index.html" |
| 854 | index_path.write_text("<html></html>\n") |
| 855 | |
| 856 | implementation_plan = temp_dir / "implementation.md" |
| 857 | implementation_plan.write_text( |
| 858 | "\n".join( |
| 859 | [ |
| 860 | "# Implementation Plan", |
| 861 | "", |
| 862 | "## File Changes", |
| 863 | f"- `{guide_root}/`", |
| 864 | f"- `{chapters}/`", |
| 865 | f"- `{index_path}`", |
| 866 | f"- `{chapters / '02-installation.html'}`", |
| 867 | "", |
| 868 | ] |
| 869 | ) |
| 870 | ) |
| 871 | |
| 872 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 873 | dod.implementation_plan = str(implementation_plan) |
| 874 | dod.touched_files.append(str(index_path)) |
| 875 | dod.pending_items.append("Write the introduction chapter") |
| 876 | |
| 877 | decision = repairer.handle_empty_response( |
| 878 | task="Create a multi-file nginx guide.", |
| 879 | original_task=None, |
| 880 | empty_retry_count=1, |
| 881 | max_empty_retries=2, |
| 882 | dod=dod, |
| 883 | ) |
| 884 | |
| 885 | assert decision.should_continue is True |
| 886 | assert decision.retry_message is not None |
| 887 | assert "Next missing planned artifact: `chapters/`" in decision.retry_message |
| 888 | assert ( |
| 889 | "Resume with this exact next step: continue `Write the introduction chapter` " |
| 890 | "by creating the next output file under `chapters/`." |
| 891 | in decision.retry_message |
| 892 | ) |
| 893 | assert ( |
| 894 | "Prefer one concrete `write` call for a file inside " |
| 895 | f"`{display_runtime_path(chapters)}` before more research." |
| 896 | in decision.retry_message |
| 897 | ) |
| 898 | |
| 899 | |
| 900 | def test_empty_response_retry_treats_develop_index_step_as_mutation_work( |
| 901 | temp_dir: Path, |
| 902 | ) -> None: |
| 903 | context = build_context( |
| 904 | temp_dir=temp_dir, |
| 905 | use_react=False, |
| 906 | ) |
| 907 | repairer = ResponseRepairer(context) |
| 908 | |
| 909 | guide_root = temp_dir / "guides" / "nginx" |
| 910 | chapters = guide_root / "chapters" |
| 911 | guide_root.mkdir(parents=True) |
| 912 | chapters.mkdir() |
| 913 | chapter_one = chapters / "01-introduction.html" |
| 914 | index_path = guide_root / "index.html" |
| 915 | |
| 916 | implementation_plan = temp_dir / "implementation.md" |
| 917 | implementation_plan.write_text( |
| 918 | "\n".join( |
| 919 | [ |
| 920 | "# Implementation Plan", |
| 921 | "", |
| 922 | "## File Changes", |
| 923 | f"- `{guide_root}/`", |
| 924 | f"- `{index_path}`", |
| 925 | f"- `{chapters}/`", |
| 926 | f"- `{chapter_one}`", |
| 927 | "", |
| 928 | ] |
| 929 | ) |
| 930 | ) |
| 931 | |
| 932 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 933 | dod.implementation_plan = str(implementation_plan) |
| 934 | dod.completed_items.extend( |
| 935 | [ |
| 936 | "First, examine the existing Fortran guide structure to understand the format and depth", |
| 937 | "Create the new nginx guide directory structure", |
| 938 | ] |
| 939 | ) |
| 940 | dod.pending_items.append("Develop the main index.html file with proper structure") |
| 941 | |
| 942 | decision = repairer.handle_empty_response( |
| 943 | task="Create a multi-file nginx guide.", |
| 944 | original_task=None, |
| 945 | empty_retry_count=2, |
| 946 | max_empty_retries=2, |
| 947 | dod=dod, |
| 948 | ) |
| 949 | |
| 950 | assert decision.should_continue is True |
| 951 | assert decision.retry_message is not None |
| 952 | assert ( |
| 953 | "Resume with this exact next step: continue `Develop the main index.html file with proper structure`" |
| 954 | in decision.retry_message |
| 955 | ) |
| 956 | assert "Next missing planned artifact: `index.html`" in decision.retry_message |
| 957 | assert "Prefer one `write(content=...)` call" in decision.retry_message |
| 958 | assert "Make the next response one concrete evidence-gathering tool call" not in decision.retry_message |
| 959 | |
| 960 | |
| 961 | def test_empty_response_retry_prefers_pending_index_over_broad_directory_headline( |
| 962 | temp_dir: Path, |
| 963 | ) -> None: |
| 964 | context = build_context( |
| 965 | temp_dir=temp_dir, |
| 966 | use_react=False, |
| 967 | ) |
| 968 | repairer = ResponseRepairer(context) |
| 969 | |
| 970 | guide_root = temp_dir / "guides" / "nginx" |
| 971 | chapters = guide_root / "chapters" |
| 972 | guide_root.mkdir(parents=True) |
| 973 | chapters.mkdir() |
| 974 | index_path = guide_root / "index.html" |
| 975 | chapter_one = chapters / "01-introduction.html" |
| 976 | |
| 977 | implementation_plan = temp_dir / "implementation.md" |
| 978 | implementation_plan.write_text( |
| 979 | "\n".join( |
| 980 | [ |
| 981 | "# Implementation Plan", |
| 982 | "", |
| 983 | "## File Changes", |
| 984 | f"- `{guide_root}/`", |
| 985 | f"- `{chapters}/`", |
| 986 | f"- `{index_path}`", |
| 987 | f"- `{chapter_one}`", |
| 988 | "", |
| 989 | ] |
| 990 | ) |
| 991 | ) |
| 992 | |
| 993 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 994 | dod.implementation_plan = str(implementation_plan) |
| 995 | dod.completed_items.extend( |
| 996 | [ |
| 997 | "First, examine the existing Fortran guide structure to understand the format and depth", |
| 998 | "Create the new nginx guide directory structure", |
| 999 | ] |
| 1000 | ) |
| 1001 | dod.pending_items.append("Develop the main index.html file with proper structure") |
| 1002 | |
| 1003 | decision = repairer.handle_empty_response( |
| 1004 | task="Create a multi-file nginx guide.", |
| 1005 | original_task=None, |
| 1006 | empty_retry_count=4, |
| 1007 | max_empty_retries=4, |
| 1008 | dod=dod, |
| 1009 | ) |
| 1010 | |
| 1011 | assert decision.should_continue is True |
| 1012 | assert decision.retry_message is not None |
| 1013 | assert "Next missing planned artifact: `index.html`" in decision.retry_message |
| 1014 | assert ( |
| 1015 | "Resume with this exact next step: continue `Develop the main index.html file with proper structure` " |
| 1016 | "by creating `index.html`." |
| 1017 | in decision.retry_message |
| 1018 | ) |
| 1019 | assert "Next missing planned artifact: `chapters/`" not in decision.retry_message |
| 1020 | assert "Remaining planned artifacts:" not in decision.retry_message |
| 1021 | assert ( |
| 1022 | "Next observed output pattern under `chapters/`: `01-introduction.html`" |
| 1023 | not in decision.retry_message |
| 1024 | ) |
| 1025 | |
| 1026 | |
| 1027 | def test_empty_response_retry_uses_concrete_file_language_for_aggregate_chapter_step( |
| 1028 | temp_dir: Path, |
| 1029 | ) -> None: |
| 1030 | context = build_context( |
| 1031 | temp_dir=temp_dir, |
| 1032 | use_react=False, |
| 1033 | ) |
| 1034 | repairer = ResponseRepairer(context) |
| 1035 | |
| 1036 | guide_root = temp_dir / "guides" / "nginx" |
| 1037 | chapters = guide_root / "chapters" |
| 1038 | chapters.mkdir(parents=True) |
| 1039 | index_path = guide_root / "index.html" |
| 1040 | index_path.write_text( |
| 1041 | "\n".join( |
| 1042 | [ |
| 1043 | "<html>", |
| 1044 | '<a href="chapters/01-introduction.html">Chapter 1: Introduction to Nginx</a>', |
| 1045 | '<a href="chapters/02-installation.html">Chapter 2: Installation and Setup</a>', |
| 1046 | "</html>", |
| 1047 | ] |
| 1048 | ) |
| 1049 | + "\n" |
| 1050 | ) |
| 1051 | |
| 1052 | implementation_plan = temp_dir / "implementation.md" |
| 1053 | implementation_plan.write_text( |
| 1054 | "\n".join( |
| 1055 | [ |
| 1056 | "# Implementation Plan", |
| 1057 | "", |
| 1058 | "## File Changes", |
| 1059 | f"- `{guide_root}/`", |
| 1060 | f"- `{chapters}/`", |
| 1061 | f"- `{index_path}`", |
| 1062 | "", |
| 1063 | ] |
| 1064 | ) |
| 1065 | ) |
| 1066 | |
| 1067 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1068 | dod.implementation_plan = str(implementation_plan) |
| 1069 | dod.touched_files.append(str(index_path)) |
| 1070 | dod.completed_items.append("Develop the main index.html file with proper structure") |
| 1071 | dod.pending_items.append("Create chapter files with content and structure") |
| 1072 | |
| 1073 | decision = repairer.handle_empty_response( |
| 1074 | task="Create a multi-file nginx guide.", |
| 1075 | original_task=None, |
| 1076 | empty_retry_count=3, |
| 1077 | max_empty_retries=4, |
| 1078 | dod=dod, |
| 1079 | ) |
| 1080 | |
| 1081 | assert decision.should_continue is True |
| 1082 | assert decision.retry_message is not None |
| 1083 | assert "Next missing planned artifact:" not in decision.retry_message |
| 1084 | assert ( |
| 1085 | "Continue `Create chapter files with content and structure` by creating `01-introduction.html`." |
| 1086 | in decision.retry_message |
| 1087 | ) |
| 1088 | assert ( |
| 1089 | 'Emit this tool shape now: `write(file_path="' |
| 1090 | in decision.retry_message |
| 1091 | ) |
| 1092 | assert ( |
| 1093 | "Write a compact but real initial version of this file now, then refine or expand it in later edits." |
| 1094 | in decision.retry_message |
| 1095 | ) |
| 1096 | assert "No narration, no TodoWrite, no rereads, and no empty response" in decision.retry_message |
| 1097 | assert "Follow the same full-payload one-file-at-a-time write pattern" not in decision.retry_message |
| 1098 | assert "Remaining planned artifacts:" not in decision.retry_message |
| 1099 | assert "Next pending item:" not in decision.retry_message |
| 1100 | |
| 1101 | |
| 1102 | def test_empty_response_retry_keeps_concrete_second_chapter_for_aggregate_chapter_step( |
| 1103 | temp_dir: Path, |
| 1104 | ) -> None: |
| 1105 | context = build_context( |
| 1106 | temp_dir=temp_dir, |
| 1107 | use_react=False, |
| 1108 | ) |
| 1109 | repairer = ResponseRepairer(context) |
| 1110 | |
| 1111 | guide_root = temp_dir / "guides" / "nginx" |
| 1112 | chapters = guide_root / "chapters" |
| 1113 | chapters.mkdir(parents=True) |
| 1114 | index_path = guide_root / "index.html" |
| 1115 | chapter_one = chapters / "01-introduction.html" |
| 1116 | chapter_two = chapters / "02-installation.html" |
| 1117 | index_path.write_text( |
| 1118 | "\n".join( |
| 1119 | [ |
| 1120 | "<html>", |
| 1121 | '<a href="chapters/01-introduction.html">Chapter 1: Introduction to Nginx</a>', |
| 1122 | '<a href="chapters/02-installation.html">Chapter 2: Installation and Setup</a>', |
| 1123 | "</html>", |
| 1124 | ] |
| 1125 | ) |
| 1126 | + "\n" |
| 1127 | ) |
| 1128 | chapter_one.write_text("<h1>Introduction</h1>\n") |
| 1129 | |
| 1130 | implementation_plan = temp_dir / "implementation.md" |
| 1131 | implementation_plan.write_text( |
| 1132 | "\n".join( |
| 1133 | [ |
| 1134 | "# Implementation Plan", |
| 1135 | "", |
| 1136 | "## File Changes", |
| 1137 | f"- `{guide_root}/`", |
| 1138 | f"- `{chapters}/`", |
| 1139 | f"- `{index_path}`", |
| 1140 | "", |
| 1141 | ] |
| 1142 | ) |
| 1143 | ) |
| 1144 | |
| 1145 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1146 | dod.implementation_plan = str(implementation_plan) |
| 1147 | dod.touched_files.extend([str(index_path), str(chapter_one)]) |
| 1148 | dod.completed_items.extend( |
| 1149 | [ |
| 1150 | "Develop the main index.html file with proper structure", |
| 1151 | "Create first chapter file (01-introduction.html)", |
| 1152 | ] |
| 1153 | ) |
| 1154 | dod.pending_items.append("Create chapter files following the established pattern") |
| 1155 | |
| 1156 | decision = repairer.handle_empty_response( |
| 1157 | task="Create a multi-file nginx guide.", |
| 1158 | original_task=None, |
| 1159 | empty_retry_count=1, |
| 1160 | max_empty_retries=2, |
| 1161 | dod=dod, |
| 1162 | ) |
| 1163 | |
| 1164 | assert decision.should_continue is True |
| 1165 | assert decision.retry_message is not None |
| 1166 | assert "Next pending item:" not in decision.retry_message |
| 1167 | assert ( |
| 1168 | "Resume with this exact next step: continue `Create chapter files following the established pattern` " |
| 1169 | "by creating `02-installation.html`." |
| 1170 | in decision.retry_message |
| 1171 | ) |
| 1172 | assert ( |
| 1173 | "It is the next concrete output needed to continue `Create chapter files following the established pattern`." |
| 1174 | in decision.retry_message |
| 1175 | ) |
| 1176 | assert f"`{display_runtime_path(chapter_two)}`" in decision.retry_message |
| 1177 | assert "Follow the same full-payload one-file-at-a-time write pattern" in decision.retry_message |
| 1178 | |
| 1179 | |
| 1180 | def test_empty_response_retry_reuses_known_reference_structure_for_first_substantive_file( |
| 1181 | temp_dir: Path, |
| 1182 | ) -> None: |
| 1183 | context = build_context( |
| 1184 | temp_dir=temp_dir, |
| 1185 | use_react=False, |
| 1186 | ) |
| 1187 | repairer = ResponseRepairer(context) |
| 1188 | |
| 1189 | guide_root = temp_dir / "guides" / "nginx" |
| 1190 | chapters = guide_root / "chapters" |
| 1191 | chapters.mkdir(parents=True) |
| 1192 | index_path = guide_root / "index.html" |
| 1193 | reference_chapter = temp_dir / "guides" / "fortran" / "chapters" / "01-introduction.html" |
| 1194 | reference_chapter.parent.mkdir(parents=True) |
| 1195 | reference_chapter.write_text("<h1>Chapter 1: Introduction to Fortran</h1>\n") |
| 1196 | index_path.write_text( |
| 1197 | "\n".join( |
| 1198 | [ |
| 1199 | "<html>", |
| 1200 | '<a href="chapters/01-introduction.html">Chapter 1: Introduction to Nginx</a>', |
| 1201 | "</html>", |
| 1202 | ] |
| 1203 | ) |
| 1204 | + "\n" |
| 1205 | ) |
| 1206 | context.session.append( |
| 1207 | Message( |
| 1208 | role=Role.ASSISTANT, |
| 1209 | content="", |
| 1210 | tool_calls=[ |
| 1211 | ToolCall( |
| 1212 | id="call_ref", |
| 1213 | name="read", |
| 1214 | arguments={"file_path": str(reference_chapter)}, |
| 1215 | ) |
| 1216 | ], |
| 1217 | ) |
| 1218 | ) |
| 1219 | |
| 1220 | implementation_plan = temp_dir / "implementation.md" |
| 1221 | implementation_plan.write_text( |
| 1222 | "\n".join( |
| 1223 | [ |
| 1224 | "# Implementation Plan", |
| 1225 | "", |
| 1226 | "## File Changes", |
| 1227 | f"- `{guide_root}/`", |
| 1228 | f"- `{chapters}/`", |
| 1229 | f"- `{index_path}`", |
| 1230 | "", |
| 1231 | ] |
| 1232 | ) |
| 1233 | ) |
| 1234 | |
| 1235 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1236 | dod.implementation_plan = str(implementation_plan) |
| 1237 | dod.touched_files.append(str(index_path)) |
| 1238 | dod.completed_items.append("Develop the main index.html file with proper structure") |
| 1239 | dod.pending_items.append("Create chapter files following the established pattern") |
| 1240 | |
| 1241 | decision = repairer.handle_empty_response( |
| 1242 | task="Create a multi-file nginx guide.", |
| 1243 | original_task=None, |
| 1244 | empty_retry_count=1, |
| 1245 | max_empty_retries=2, |
| 1246 | dod=dod, |
| 1247 | ) |
| 1248 | |
| 1249 | assert decision.should_continue is True |
| 1250 | assert decision.retry_message is not None |
| 1251 | assert ( |
| 1252 | f"You already read `{display_runtime_path(reference_chapter)}`; reuse its overall structure " |
| 1253 | "as the starting pattern for this new file, then adapt the content to the current target." |
| 1254 | in decision.retry_message |
| 1255 | ) |
| 1256 | |
| 1257 | |
| 1258 | def test_compact_first_substantive_retry_reuses_known_reference_structure( |
| 1259 | temp_dir: Path, |
| 1260 | ) -> None: |
| 1261 | context = build_context( |
| 1262 | temp_dir=temp_dir, |
| 1263 | use_react=False, |
| 1264 | ) |
| 1265 | repairer = ResponseRepairer(context) |
| 1266 | |
| 1267 | guide_root = temp_dir / "guides" / "nginx" |
| 1268 | chapters = guide_root / "chapters" |
| 1269 | chapters.mkdir(parents=True) |
| 1270 | index_path = guide_root / "index.html" |
| 1271 | reference_chapter = temp_dir / "guides" / "fortran" / "chapters" / "01-introduction.html" |
| 1272 | reference_chapter.parent.mkdir(parents=True) |
| 1273 | reference_chapter.write_text("<h1>Chapter 1: Introduction to Fortran</h1>\n") |
| 1274 | index_path.write_text( |
| 1275 | "\n".join( |
| 1276 | [ |
| 1277 | "<html>", |
| 1278 | '<a href="chapters/01-introduction.html">Chapter 1: Introduction to Nginx</a>', |
| 1279 | "</html>", |
| 1280 | ] |
| 1281 | ) |
| 1282 | + "\n" |
| 1283 | ) |
| 1284 | context.session.append( |
| 1285 | Message( |
| 1286 | role=Role.ASSISTANT, |
| 1287 | content="", |
| 1288 | tool_calls=[ |
| 1289 | ToolCall( |
| 1290 | id="call_ref", |
| 1291 | name="read", |
| 1292 | arguments={"file_path": str(reference_chapter)}, |
| 1293 | ) |
| 1294 | ], |
| 1295 | ) |
| 1296 | ) |
| 1297 | |
| 1298 | implementation_plan = temp_dir / "implementation.md" |
| 1299 | implementation_plan.write_text( |
| 1300 | "\n".join( |
| 1301 | [ |
| 1302 | "# Implementation Plan", |
| 1303 | "", |
| 1304 | "## File Changes", |
| 1305 | f"- `{guide_root}/`", |
| 1306 | f"- `{chapters}/`", |
| 1307 | f"- `{index_path}`", |
| 1308 | "", |
| 1309 | ] |
| 1310 | ) |
| 1311 | ) |
| 1312 | |
| 1313 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1314 | dod.implementation_plan = str(implementation_plan) |
| 1315 | dod.touched_files.append(str(index_path)) |
| 1316 | dod.completed_items.append("Develop the main index.html file with proper structure") |
| 1317 | dod.pending_items.append("Create chapter files following the established pattern") |
| 1318 | |
| 1319 | decision = repairer.handle_empty_response( |
| 1320 | task="Create a multi-file nginx guide.", |
| 1321 | original_task=None, |
| 1322 | empty_retry_count=3, |
| 1323 | max_empty_retries=4, |
| 1324 | dod=dod, |
| 1325 | ) |
| 1326 | |
| 1327 | assert decision.should_continue is True |
| 1328 | assert decision.retry_message is not None |
| 1329 | assert ( |
| 1330 | f"You already read `{display_runtime_path(reference_chapter)}`; reuse its overall structure " |
| 1331 | "as the starting pattern for this new file, then adapt the content to the current target." |
| 1332 | in decision.retry_message |
| 1333 | ) |
| 1334 | |
| 1335 | |
| 1336 | def test_empty_response_retry_prefers_output_index_over_reference_index_with_same_name( |
| 1337 | temp_dir: Path, |
| 1338 | ) -> None: |
| 1339 | context = build_context( |
| 1340 | temp_dir=temp_dir, |
| 1341 | use_react=False, |
| 1342 | ) |
| 1343 | repairer = ResponseRepairer(context) |
| 1344 | |
| 1345 | nginx_root = temp_dir / "Loader" / "guides" / "nginx" |
| 1346 | fortran_root = temp_dir / "Loader" / "guides" / "fortran" |
| 1347 | nginx_root.mkdir(parents=True) |
| 1348 | fortran_root.mkdir(parents=True) |
| 1349 | reference_index = fortran_root / "index.html" |
| 1350 | reference_index.write_text("<html>fortran</html>\n") |
| 1351 | output_index = nginx_root / "index.html" |
| 1352 | |
| 1353 | implementation_plan = temp_dir / "implementation.md" |
| 1354 | implementation_plan.write_text( |
| 1355 | "\n".join( |
| 1356 | [ |
| 1357 | "# Implementation Plan", |
| 1358 | "", |
| 1359 | "## File Changes", |
| 1360 | f"- `{output_index}`", |
| 1361 | f"- `{nginx_root / 'chapters'}/`", |
| 1362 | f"- `{reference_index}`", |
| 1363 | "", |
| 1364 | ] |
| 1365 | ) |
| 1366 | ) |
| 1367 | |
| 1368 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1369 | dod.implementation_plan = str(implementation_plan) |
| 1370 | dod.touched_files.append(str(reference_index)) |
| 1371 | dod.completed_items.append( |
| 1372 | "First, examine the existing Fortran guide structure and content" |
| 1373 | ) |
| 1374 | dod.pending_items.append("Develop the nginx index.html file") |
| 1375 | |
| 1376 | decision = repairer.handle_empty_response( |
| 1377 | task="Create a multi-file nginx guide.", |
| 1378 | original_task=None, |
| 1379 | empty_retry_count=2, |
| 1380 | max_empty_retries=2, |
| 1381 | dod=dod, |
| 1382 | ) |
| 1383 | |
| 1384 | assert decision.should_continue is True |
| 1385 | assert decision.retry_message is not None |
| 1386 | assert ( |
| 1387 | "Prefer one `write(content=...)` call for " |
| 1388 | f"`{display_runtime_path(output_index)}` before more research." |
| 1389 | in decision.retry_message |
| 1390 | ) |
| 1391 | assert str(reference_index) not in decision.retry_message |
| 1392 | |
| 1393 | |
| 1394 | def test_empty_response_retry_points_at_declared_child_file_within_incomplete_output_directory( |
| 1395 | temp_dir: Path, |
| 1396 | ) -> None: |
| 1397 | context = build_context( |
| 1398 | temp_dir=temp_dir, |
| 1399 | use_react=False, |
| 1400 | ) |
| 1401 | repairer = ResponseRepairer(context) |
| 1402 | |
| 1403 | guide_root = temp_dir / "guides" / "nginx" |
| 1404 | chapters = guide_root / "chapters" |
| 1405 | chapters.mkdir(parents=True) |
| 1406 | index_path = guide_root / "index.html" |
| 1407 | index_path.write_text( |
| 1408 | "\n".join( |
| 1409 | [ |
| 1410 | "<html>", |
| 1411 | '<a href="chapters/introduction.html">Introduction</a>', |
| 1412 | '<a href="chapters/installation.html">Installation</a>', |
| 1413 | "</html>", |
| 1414 | ] |
| 1415 | ) |
| 1416 | + "\n" |
| 1417 | ) |
| 1418 | |
| 1419 | implementation_plan = temp_dir / "implementation.md" |
| 1420 | implementation_plan.write_text( |
| 1421 | "\n".join( |
| 1422 | [ |
| 1423 | "# Implementation Plan", |
| 1424 | "", |
| 1425 | "## File Changes", |
| 1426 | f"- `{guide_root}/`", |
| 1427 | f"- `{chapters}/`", |
| 1428 | f"- `{index_path}`", |
| 1429 | f"- `{chapters / '02-installation.html'}`", |
| 1430 | "", |
| 1431 | ] |
| 1432 | ) |
| 1433 | ) |
| 1434 | |
| 1435 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1436 | dod.implementation_plan = str(implementation_plan) |
| 1437 | dod.touched_files.append(str(index_path)) |
| 1438 | dod.pending_items.append("Write the introduction chapter") |
| 1439 | |
| 1440 | decision = repairer.handle_empty_response( |
| 1441 | task="Create a multi-file nginx guide.", |
| 1442 | original_task=None, |
| 1443 | empty_retry_count=1, |
| 1444 | max_empty_retries=2, |
| 1445 | dod=dod, |
| 1446 | ) |
| 1447 | |
| 1448 | assert decision.should_continue is True |
| 1449 | assert decision.retry_message is not None |
| 1450 | assert "Next missing planned artifact: `introduction.html`" in decision.retry_message |
| 1451 | assert ( |
| 1452 | "Resume with this exact next step: continue `Write the introduction chapter` " |
| 1453 | "by creating `introduction.html`." |
| 1454 | in decision.retry_message |
| 1455 | ) |
| 1456 | assert "Next declared output under `chapters/`" not in decision.retry_message |
| 1457 | assert ( |
| 1458 | f"Prefer one `write(content=...)` call for `{(chapters / 'introduction.html').resolve(strict=False)}` " |
| 1459 | "before more research." |
| 1460 | in decision.retry_message |
| 1461 | ) |
| 1462 | |
| 1463 | |
| 1464 | def test_empty_response_retry_infers_concrete_file_from_pending_todo_after_broad_artifacts_exist( |
| 1465 | temp_dir: Path, |
| 1466 | ) -> None: |
| 1467 | context = build_context( |
| 1468 | temp_dir=temp_dir, |
| 1469 | use_react=False, |
| 1470 | ) |
| 1471 | repairer = ResponseRepairer(context) |
| 1472 | |
| 1473 | guide_root = temp_dir / "guides" / "nginx" |
| 1474 | chapters = guide_root / "chapters" |
| 1475 | chapters.mkdir(parents=True) |
| 1476 | index_path = guide_root / "index.html" |
| 1477 | chapter_one = chapters / "01-introduction.html" |
| 1478 | index_path.write_text("<html></html>\n") |
| 1479 | chapter_one.write_text("<html></html>\n") |
| 1480 | |
| 1481 | implementation_plan = temp_dir / "implementation.md" |
| 1482 | implementation_plan.write_text( |
| 1483 | "\n".join( |
| 1484 | [ |
| 1485 | "# Implementation Plan", |
| 1486 | "", |
| 1487 | "## File Changes", |
| 1488 | f"- `{guide_root}/`", |
| 1489 | f"- `{chapters}/`", |
| 1490 | f"- `{index_path}`", |
| 1491 | f"- `{chapters / '02-installation.html'}`", |
| 1492 | "", |
| 1493 | ] |
| 1494 | ) |
| 1495 | ) |
| 1496 | |
| 1497 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1498 | dod.implementation_plan = str(implementation_plan) |
| 1499 | dod.touched_files.extend([str(index_path), str(chapter_one)]) |
| 1500 | dod.completed_items.extend( |
| 1501 | [ |
| 1502 | "Create index.html for nginx guide", |
| 1503 | "Create first chapter file (01-introduction.html)", |
| 1504 | ] |
| 1505 | ) |
| 1506 | dod.pending_items.append("Create second chapter file (02-installation.html)") |
| 1507 | |
| 1508 | decision = repairer.handle_empty_response( |
| 1509 | task="Create a multi-file nginx guide.", |
| 1510 | original_task=None, |
| 1511 | empty_retry_count=2, |
| 1512 | max_empty_retries=2, |
| 1513 | dod=dod, |
| 1514 | ) |
| 1515 | |
| 1516 | assert decision.should_continue is True |
| 1517 | assert decision.retry_message is not None |
| 1518 | assert ( |
| 1519 | "Resume with this exact next step: continue `Create second chapter file " |
| 1520 | "(02-installation.html)` by creating `02-installation.html`." |
| 1521 | in decision.retry_message |
| 1522 | ) |
| 1523 | assert ( |
| 1524 | "Prefer one `write(content=...)` call for " |
| 1525 | f"`{display_runtime_path(chapters / '02-installation.html')}` " |
| 1526 | "before more research." |
| 1527 | in decision.retry_message |
| 1528 | ) |
| 1529 | assert "Do not return another working note or empty response" in decision.retry_message |
| 1530 | |
| 1531 | |
| 1532 | def test_empty_response_retry_maps_title_style_todo_to_html_graph_target( |
| 1533 | temp_dir: Path, |
| 1534 | ) -> None: |
| 1535 | context = build_context( |
| 1536 | temp_dir=temp_dir, |
| 1537 | use_react=False, |
| 1538 | ) |
| 1539 | repairer = ResponseRepairer(context) |
| 1540 | |
| 1541 | guide_root = temp_dir / "guides" / "nginx" |
| 1542 | chapters = guide_root / "chapters" |
| 1543 | chapters.mkdir(parents=True) |
| 1544 | index_path = guide_root / "index.html" |
| 1545 | chapter_one = chapters / "01-introduction.html" |
| 1546 | index_path.write_text( |
| 1547 | "\n".join( |
| 1548 | [ |
| 1549 | "<html>", |
| 1550 | '<a href="chapters/01-introduction.html">Chapter 1: Introduction to NGINX Tool</a>', |
| 1551 | '<a href="chapters/02-installation.html">Chapter 2: Installation and Setup</a>', |
| 1552 | "</html>", |
| 1553 | ] |
| 1554 | ) |
| 1555 | + "\n" |
| 1556 | ) |
| 1557 | chapter_one.write_text("<html></html>\n") |
| 1558 | |
| 1559 | implementation_plan = temp_dir / "implementation.md" |
| 1560 | implementation_plan.write_text( |
| 1561 | "\n".join( |
| 1562 | [ |
| 1563 | "# Implementation Plan", |
| 1564 | "", |
| 1565 | "## File Changes", |
| 1566 | f"- `{guide_root}/`", |
| 1567 | f"- `{chapters}/`", |
| 1568 | f"- `{index_path}`", |
| 1569 | f"- `{chapters / '02-installation.html'}`", |
| 1570 | "", |
| 1571 | ] |
| 1572 | ) |
| 1573 | ) |
| 1574 | |
| 1575 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1576 | dod.implementation_plan = str(implementation_plan) |
| 1577 | dod.touched_files.extend([str(index_path), str(chapter_one)]) |
| 1578 | dod.completed_items.extend( |
| 1579 | [ |
| 1580 | "Create index.html for nginx guide", |
| 1581 | "Create Chapter 1: Introduction to NGINX Tool", |
| 1582 | ] |
| 1583 | ) |
| 1584 | dod.pending_items.append("Creating Chapter 2: Installation and Setup") |
| 1585 | |
| 1586 | decision = repairer.handle_empty_response( |
| 1587 | task="Create a multi-file nginx guide.", |
| 1588 | original_task=None, |
| 1589 | empty_retry_count=2, |
| 1590 | max_empty_retries=2, |
| 1591 | dod=dod, |
| 1592 | ) |
| 1593 | |
| 1594 | assert decision.should_continue is True |
| 1595 | assert decision.retry_message is not None |
| 1596 | assert ( |
| 1597 | "Resume with this exact next step: continue `Creating Chapter 2: Installation and Setup` " |
| 1598 | "by creating `02-installation.html`." |
| 1599 | in decision.retry_message |
| 1600 | ) |
| 1601 | assert ( |
| 1602 | "Prefer one `write(content=...)` call for " |
| 1603 | f"`{display_runtime_path(chapters / '02-installation.html')}` " |
| 1604 | "before more research." |
| 1605 | in decision.retry_message |
| 1606 | ) |
| 1607 | assert ( |
| 1608 | f'Emit this tool shape now: `write(file_path="{display_runtime_path(chapters / "02-installation.html")}", content="...")`.' |
| 1609 | in decision.retry_message |
| 1610 | ) |
| 1611 | assert ( |
| 1612 | "Use the existing outline label `Chapter 2: Installation and Setup` for that file " |
| 1613 | "so it matches the current guide structure." |
| 1614 | in decision.retry_message |
| 1615 | ) |
| 1616 | |
| 1617 | |
| 1618 | def test_empty_response_retry_reminds_model_to_resend_real_write_payload( |
| 1619 | temp_dir: Path, |
| 1620 | ) -> None: |
| 1621 | context = build_context( |
| 1622 | temp_dir=temp_dir, |
| 1623 | use_react=False, |
| 1624 | ) |
| 1625 | repairer = ResponseRepairer(context) |
| 1626 | |
| 1627 | guide_root = temp_dir / "guides" / "nginx" |
| 1628 | chapters = guide_root / "chapters" |
| 1629 | chapters.mkdir(parents=True) |
| 1630 | chapter_one = chapters / "01-introduction.html" |
| 1631 | chapter_one.write_text("<html></html>\n") |
| 1632 | |
| 1633 | implementation_plan = temp_dir / "implementation.md" |
| 1634 | implementation_plan.write_text( |
| 1635 | "\n".join( |
| 1636 | [ |
| 1637 | "# Implementation Plan", |
| 1638 | "", |
| 1639 | "## File Changes", |
| 1640 | f"- `{guide_root}/`", |
| 1641 | f"- `{chapters}/`", |
| 1642 | f"- `{guide_root / 'index.html'}`", |
| 1643 | f"- `{chapters / '01-introduction.html'}`", |
| 1644 | "", |
| 1645 | ] |
| 1646 | ) |
| 1647 | ) |
| 1648 | |
| 1649 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1650 | dod.implementation_plan = str(implementation_plan) |
| 1651 | dod.touched_files.append(str(chapter_one)) |
| 1652 | dod.completed_items.append("Create first chapter file (01-introduction.html)") |
| 1653 | dod.pending_items.append("Develop the main index.html file for the nginx guide") |
| 1654 | |
| 1655 | recovery_context = RecoveryContext( |
| 1656 | original_tool="write", |
| 1657 | original_args={ |
| 1658 | "file_path": "~/Loader/guides/nginx/index.html", |
| 1659 | "content_chars": 1354, |
| 1660 | "content_lines": 30, |
| 1661 | }, |
| 1662 | ) |
| 1663 | recovery_context.add_attempt( |
| 1664 | "write", |
| 1665 | { |
| 1666 | "file_path": "~/Loader/guides/nginx/index.html", |
| 1667 | "content_chars": 1354, |
| 1668 | "content_lines": 30, |
| 1669 | }, |
| 1670 | "WriteTool.execute() missing 1 required positional argument: 'content'", |
| 1671 | ) |
| 1672 | context.recovery_context = recovery_context |
| 1673 | |
| 1674 | decision = repairer.handle_empty_response( |
| 1675 | task="Create a multi-file nginx guide.", |
| 1676 | original_task=None, |
| 1677 | empty_retry_count=2, |
| 1678 | max_empty_retries=2, |
| 1679 | dod=dod, |
| 1680 | ) |
| 1681 | |
| 1682 | assert decision.should_continue is True |
| 1683 | assert decision.retry_message is not None |
| 1684 | assert "resend `write`" in decision.retry_message |
| 1685 | assert "content_chars" in decision.retry_message |
| 1686 | assert "index.html" in decision.retry_message |
| 1687 | |
| 1688 | |
| 1689 | def test_empty_response_retry_uses_compact_prompt_after_early_progress_with_concrete_next_file( |
| 1690 | temp_dir: Path, |
| 1691 | ) -> None: |
| 1692 | context = build_context( |
| 1693 | temp_dir=temp_dir, |
| 1694 | use_react=False, |
| 1695 | ) |
| 1696 | repairer = ResponseRepairer(context) |
| 1697 | |
| 1698 | guide_root = temp_dir / "guides" / "nginx" |
| 1699 | chapters = guide_root / "chapters" |
| 1700 | chapters.mkdir(parents=True) |
| 1701 | index_path = guide_root / "index.html" |
| 1702 | chapter_one = chapters / "01-introduction.html" |
| 1703 | index_path.write_text( |
| 1704 | "\n".join( |
| 1705 | [ |
| 1706 | "<html>", |
| 1707 | '<a href="chapters/01-introduction.html">Introduction</a>', |
| 1708 | '<a href="chapters/02-installation.html">Installation</a>', |
| 1709 | "</html>", |
| 1710 | ] |
| 1711 | ) |
| 1712 | + "\n" |
| 1713 | ) |
| 1714 | chapter_one.write_text("<html></html>\n") |
| 1715 | |
| 1716 | implementation_plan = temp_dir / "implementation.md" |
| 1717 | implementation_plan.write_text( |
| 1718 | "\n".join( |
| 1719 | [ |
| 1720 | "# Implementation Plan", |
| 1721 | "", |
| 1722 | "## File Changes", |
| 1723 | f"- `{guide_root}/`", |
| 1724 | f"- `{chapters}/`", |
| 1725 | f"- `{index_path}`", |
| 1726 | f"- `{chapters / '02-installation.html'}`", |
| 1727 | "", |
| 1728 | ] |
| 1729 | ) |
| 1730 | ) |
| 1731 | |
| 1732 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1733 | dod.implementation_plan = str(implementation_plan) |
| 1734 | dod.touched_files.extend([str(index_path), str(chapter_one)]) |
| 1735 | dod.completed_items.extend( |
| 1736 | [ |
| 1737 | "Create index.html for nginx guide", |
| 1738 | "Create first chapter file (01-introduction.html)", |
| 1739 | ] |
| 1740 | ) |
| 1741 | dod.pending_items.append("Create second chapter file (02-installation.html)") |
| 1742 | |
| 1743 | decision = repairer.handle_empty_response( |
| 1744 | task="Create a multi-file nginx guide.", |
| 1745 | original_task=None, |
| 1746 | empty_retry_count=1, |
| 1747 | max_empty_retries=2, |
| 1748 | dod=dod, |
| 1749 | ) |
| 1750 | |
| 1751 | assert decision.should_continue is True |
| 1752 | assert decision.retry_message is not None |
| 1753 | assert "Continue from the exact next step below." in decision.retry_message |
| 1754 | assert "Confirmed completed work:" not in decision.retry_message |
| 1755 | assert "Next pending item:" not in decision.retry_message |
| 1756 | assert ( |
| 1757 | "Resume with this exact next step: continue `Create second chapter file " |
| 1758 | "(02-installation.html)` by creating `02-installation.html`." |
| 1759 | in decision.retry_message |
| 1760 | ) |
| 1761 | |
| 1762 | |
| 1763 | def test_empty_response_retry_ignores_stale_setup_todo_after_files_created( |
| 1764 | temp_dir: Path, |
| 1765 | ) -> None: |
| 1766 | context = build_context( |
| 1767 | temp_dir=temp_dir, |
| 1768 | use_react=False, |
| 1769 | ) |
| 1770 | repairer = ResponseRepairer(context) |
| 1771 | |
| 1772 | guide_root = temp_dir / "guides" / "nginx" |
| 1773 | chapters = guide_root / "chapters" |
| 1774 | chapters.mkdir(parents=True) |
| 1775 | index_path = guide_root / "index.html" |
| 1776 | chapter_one = chapters / "01-introduction.html" |
| 1777 | index_path.write_text("<html></html>\n") |
| 1778 | chapter_one.write_text("<html></html>\n") |
| 1779 | |
| 1780 | implementation_plan = temp_dir / "implementation.md" |
| 1781 | implementation_plan.write_text( |
| 1782 | "\n".join( |
| 1783 | [ |
| 1784 | "# Implementation Plan", |
| 1785 | "", |
| 1786 | "## File Changes", |
| 1787 | f"- `{guide_root}/`", |
| 1788 | f"- `{chapters}/`", |
| 1789 | f"- `{index_path}`", |
| 1790 | f"- `{chapter_one}`", |
| 1791 | f"- `{chapters / '02-installation.html'}`", |
| 1792 | "", |
| 1793 | ] |
| 1794 | ) |
| 1795 | ) |
| 1796 | |
| 1797 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1798 | dod.implementation_plan = str(implementation_plan) |
| 1799 | dod.touched_files.extend([str(index_path), str(chapter_one)]) |
| 1800 | dod.completed_items.extend( |
| 1801 | [ |
| 1802 | "Develop the main index.html file for the nginx guide", |
| 1803 | "Create first chapter file (01-introduction.html)", |
| 1804 | ] |
| 1805 | ) |
| 1806 | dod.pending_items.extend( |
| 1807 | [ |
| 1808 | "Create the nginx directory structure", |
| 1809 | "Create second chapter file (02-installation.html)", |
| 1810 | ] |
| 1811 | ) |
| 1812 | |
| 1813 | decision = repairer.handle_empty_response( |
| 1814 | task="Create a multi-file nginx guide.", |
| 1815 | original_task=None, |
| 1816 | empty_retry_count=1, |
| 1817 | max_empty_retries=2, |
| 1818 | dod=dod, |
| 1819 | ) |
| 1820 | |
| 1821 | assert decision.should_continue is True |
| 1822 | assert decision.retry_message is not None |
| 1823 | assert "Create the nginx directory structure" not in decision.retry_message |
| 1824 | assert "02-installation.html" in decision.retry_message |
| 1825 | |
| 1826 | |
| 1827 | def test_empty_response_retry_fails_after_extended_late_stage_budget_is_exhausted( |
| 1828 | temp_dir: Path, |
| 1829 | ) -> None: |
| 1830 | context = build_context( |
| 1831 | temp_dir=temp_dir, |
| 1832 | use_react=False, |
| 1833 | ) |
| 1834 | repairer = ResponseRepairer(context) |
| 1835 | |
| 1836 | guide_root = temp_dir / "guides" / "nginx" |
| 1837 | chapters = guide_root / "chapters" |
| 1838 | chapters.mkdir(parents=True) |
| 1839 | index_path = guide_root / "index.html" |
| 1840 | chapter_one = chapters / "01-getting-started.html" |
| 1841 | chapter_two = chapters / "02-installation.html" |
| 1842 | chapter_three = chapters / "03-first-website.html" |
| 1843 | chapter_four = chapters / "04-configuration-basics.html" |
| 1844 | index_path.write_text("<html></html>\n") |
| 1845 | chapter_one.write_text("<h1>One</h1>\n") |
| 1846 | chapter_two.write_text("<h1>Two</h1>\n") |
| 1847 | chapter_three.write_text("<h1>Three</h1>\n") |
| 1848 | |
| 1849 | implementation_plan = temp_dir / "implementation.md" |
| 1850 | implementation_plan.write_text( |
| 1851 | "\n".join( |
| 1852 | [ |
| 1853 | "# Implementation Plan", |
| 1854 | "", |
| 1855 | "## File Changes", |
| 1856 | f"- `{guide_root}/`", |
| 1857 | f"- `{chapters}/`", |
| 1858 | f"- `{index_path}`", |
| 1859 | f"- `{chapter_one}`", |
| 1860 | f"- `{chapter_two}`", |
| 1861 | f"- `{chapter_three}`", |
| 1862 | f"- `{chapter_four}`", |
| 1863 | "", |
| 1864 | ] |
| 1865 | ) |
| 1866 | ) |
| 1867 | |
| 1868 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1869 | dod.implementation_plan = str(implementation_plan) |
| 1870 | dod.touched_files.extend( |
| 1871 | [str(index_path), str(chapter_one), str(chapter_two), str(chapter_three)] |
| 1872 | ) |
| 1873 | dod.completed_items.extend( |
| 1874 | [ |
| 1875 | "Create the directory structure for the new nginx guide", |
| 1876 | "Create the main index.html file with proper structure", |
| 1877 | ] |
| 1878 | ) |
| 1879 | dod.pending_items.append("Create each chapter file in sequence") |
| 1880 | |
| 1881 | decision = repairer.handle_empty_response( |
| 1882 | task="Create a multi-file nginx guide.", |
| 1883 | original_task=None, |
| 1884 | empty_retry_count=5, |
| 1885 | max_empty_retries=2, |
| 1886 | dod=dod, |
| 1887 | ) |
| 1888 | |
| 1889 | assert decision.should_continue is False |
| 1890 | assert decision.final_response is not None |
| 1891 | assert "retrying 4 times" in decision.final_response |
| 1892 | |
| 1893 | |
| 1894 | def test_empty_response_retry_mentions_todowrite_when_progress_has_outpaced_tracking( |
| 1895 | temp_dir: Path, |
| 1896 | ) -> None: |
| 1897 | context = build_context( |
| 1898 | temp_dir=temp_dir, |
| 1899 | use_react=False, |
| 1900 | ) |
| 1901 | repairer = ResponseRepairer(context) |
| 1902 | |
| 1903 | guide_root = temp_dir / "guides" / "nginx" |
| 1904 | chapters = guide_root / "chapters" |
| 1905 | chapters.mkdir(parents=True) |
| 1906 | implementation_plan = temp_dir / "implementation.md" |
| 1907 | implementation_plan.write_text( |
| 1908 | "\n".join( |
| 1909 | [ |
| 1910 | "# Implementation Plan", |
| 1911 | "", |
| 1912 | "## File Changes", |
| 1913 | f"- `{guide_root / 'index.html'}`", |
| 1914 | f"- `{chapters / '01-getting-started.html'}`", |
| 1915 | f"- `{chapters / '02-installation.html'}`", |
| 1916 | "", |
| 1917 | ] |
| 1918 | ) |
| 1919 | ) |
| 1920 | |
| 1921 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1922 | dod.implementation_plan = str(implementation_plan) |
| 1923 | dod.touched_files.extend( |
| 1924 | [ |
| 1925 | str(guide_root / "index.html"), |
| 1926 | str(chapters / "01-getting-started.html"), |
| 1927 | ] |
| 1928 | ) |
| 1929 | dod.completed_items.extend( |
| 1930 | [ |
| 1931 | "Create the directory structure for the new nginx guide", |
| 1932 | "Create the main index.html file with proper structure", |
| 1933 | ] |
| 1934 | ) |
| 1935 | dod.pending_items.append("Create each chapter file in sequence") |
| 1936 | |
| 1937 | decision = repairer.handle_empty_response( |
| 1938 | task="Create a multi-file nginx guide.", |
| 1939 | original_task=None, |
| 1940 | empty_retry_count=1, |
| 1941 | max_empty_retries=2, |
| 1942 | dod=dod, |
| 1943 | ) |
| 1944 | |
| 1945 | assert decision.retry_message is not None |
| 1946 | assert "Continue from the exact next step below." in decision.retry_message |
| 1947 | assert "refresh `TodoWrite` alongside the next concrete mutation" not in decision.retry_message |
| 1948 | |
| 1949 | |
| 1950 | def test_empty_response_retry_omits_stale_aggregate_completed_work_when_artifacts_missing( |
| 1951 | temp_dir: Path, |
| 1952 | ) -> None: |
| 1953 | context = build_context( |
| 1954 | temp_dir=temp_dir, |
| 1955 | use_react=False, |
| 1956 | ) |
| 1957 | repairer = ResponseRepairer(context) |
| 1958 | |
| 1959 | guide_root = temp_dir / "guides" / "nginx" |
| 1960 | chapters = guide_root / "chapters" |
| 1961 | chapters.mkdir(parents=True) |
| 1962 | index_path = guide_root / "index.html" |
| 1963 | chapter_one = chapters / "01-getting-started.html" |
| 1964 | chapter_two = chapters / "02-installation.html" |
| 1965 | chapter_three = chapters / "03-first-website.html" |
| 1966 | index_path.write_text("<html></html>\n") |
| 1967 | chapter_one.write_text("<h1>One</h1>\n") |
| 1968 | chapter_two.write_text("<h1>Two</h1>\n") |
| 1969 | |
| 1970 | implementation_plan = temp_dir / "implementation.md" |
| 1971 | implementation_plan.write_text( |
| 1972 | "\n".join( |
| 1973 | [ |
| 1974 | "# Implementation Plan", |
| 1975 | "", |
| 1976 | "## File Changes", |
| 1977 | f"- `{guide_root}/`", |
| 1978 | f"- `{chapters}/`", |
| 1979 | f"- `{index_path}`", |
| 1980 | f"- `{chapter_one}`", |
| 1981 | f"- `{chapter_two}`", |
| 1982 | f"- `{chapter_three}`", |
| 1983 | "", |
| 1984 | ] |
| 1985 | ) |
| 1986 | ) |
| 1987 | |
| 1988 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 1989 | dod.implementation_plan = str(implementation_plan) |
| 1990 | dod.touched_files.extend([str(index_path), str(chapter_one), str(chapter_two)]) |
| 1991 | dod.completed_items.extend( |
| 1992 | [ |
| 1993 | "Create the main index.html file with proper structure", |
| 1994 | "Link all chapters together properly", |
| 1995 | ] |
| 1996 | ) |
| 1997 | dod.pending_items.append("Create each chapter file in sequence") |
| 1998 | |
| 1999 | decision = repairer.handle_empty_response( |
| 2000 | task="Create a multi-file nginx guide.", |
| 2001 | original_task=None, |
| 2002 | empty_retry_count=1, |
| 2003 | max_empty_retries=2, |
| 2004 | dod=dod, |
| 2005 | ) |
| 2006 | |
| 2007 | assert decision.retry_message is not None |
| 2008 | assert "Link all chapters together properly" not in decision.retry_message |
| 2009 | assert "Continue from the exact next step below." in decision.retry_message |
| 2010 | assert "Resume with this exact next step:" in decision.retry_message |
| 2011 | |
| 2012 | |
| 2013 | def test_empty_response_retry_names_next_file_from_observed_sibling_directory( |
| 2014 | temp_dir: Path, |
| 2015 | ) -> None: |
| 2016 | context = build_context( |
| 2017 | temp_dir=temp_dir, |
| 2018 | use_react=False, |
| 2019 | ) |
| 2020 | repairer = ResponseRepairer(context) |
| 2021 | |
| 2022 | reference_chapters = temp_dir / "fortran" / "chapters" |
| 2023 | reference_chapters.mkdir(parents=True) |
| 2024 | (reference_chapters / "01-introduction.html").write_text("<h1>Introduction</h1>\n") |
| 2025 | |
| 2026 | guide_root = temp_dir / "guides" / "nginx" |
| 2027 | chapters = guide_root / "chapters" |
| 2028 | chapters.mkdir(parents=True) |
| 2029 | index_path = guide_root / "index.html" |
| 2030 | index_path.write_text("<html></html>\n") |
| 2031 | |
| 2032 | implementation_plan = temp_dir / "implementation.md" |
| 2033 | implementation_plan.write_text( |
| 2034 | "\n".join( |
| 2035 | [ |
| 2036 | "# Implementation Plan", |
| 2037 | "", |
| 2038 | "## File Changes", |
| 2039 | f"- `{guide_root}/`", |
| 2040 | f"- `{chapters}/`", |
| 2041 | f"- `{index_path}`", |
| 2042 | "", |
| 2043 | ] |
| 2044 | ) |
| 2045 | ) |
| 2046 | |
| 2047 | dod = create_definition_of_done("Create a multi-file nginx guide.") |
| 2048 | dod.implementation_plan = str(implementation_plan) |
| 2049 | dod.touched_files.append(str(index_path)) |
| 2050 | dod.pending_items.append("Write the introduction chapter") |
| 2051 | context.session.append( |
| 2052 | Message( |
| 2053 | role=Role.ASSISTANT, |
| 2054 | content="", |
| 2055 | tool_calls=[ |
| 2056 | ToolCall( |
| 2057 | id="read-ref-1", |
| 2058 | name="read", |
| 2059 | arguments={"file_path": str(reference_chapters / "01-introduction.html")}, |
| 2060 | ) |
| 2061 | ], |
| 2062 | ) |
| 2063 | ) |
| 2064 | |
| 2065 | decision = repairer.handle_empty_response( |
| 2066 | task="Create a multi-file nginx guide.", |
| 2067 | original_task=None, |
| 2068 | empty_retry_count=1, |
| 2069 | max_empty_retries=2, |
| 2070 | dod=dod, |
| 2071 | ) |
| 2072 | |
| 2073 | assert decision.should_continue is True |
| 2074 | assert decision.retry_message is not None |
| 2075 | assert "Next missing planned artifact: `01-introduction.html`" in decision.retry_message |
| 2076 | assert ( |
| 2077 | "Resume with this exact next step: continue `Write the introduction chapter` " |
| 2078 | "by creating `01-introduction.html`." |
| 2079 | in decision.retry_message |
| 2080 | ) |
| 2081 | assert "Next observed output pattern under `chapters/`" not in decision.retry_message |
| 2082 | assert ( |
| 2083 | "It mirrors the observed filename pattern from another `chapters/` directory " |
| 2084 | "you already inspected." |
| 2085 | in decision.retry_message |
| 2086 | ) |