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