| 1 | """Tool-batch execution and recovery bookkeeping for the typed runtime.""" |
| 2 | |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | import re |
| 6 | import shlex |
| 7 | from collections.abc import Awaitable, Callable |
| 8 | from dataclasses import dataclass, field |
| 9 | from pathlib import Path |
| 10 | from typing import Any |
| 11 | |
| 12 | from ..llm.base import Message, ToolCall |
| 13 | from .compaction import infer_preferred_next_step, summarize_confirmed_facts |
| 14 | from .context import RuntimeContext |
| 15 | from .dod import ( |
| 16 | DefinitionOfDone, |
| 17 | DefinitionOfDoneStore, |
| 18 | all_planned_artifact_outputs_exist, |
| 19 | all_planned_artifacts_exist, |
| 20 | begin_new_verification_attempt, |
| 21 | collect_planned_artifact_targets, |
| 22 | derive_verification_commands, |
| 23 | ensure_active_verification_attempt, |
| 24 | infer_next_output_file, |
| 25 | is_state_mutating_tool_call, |
| 26 | planned_artifact_target_satisfied, |
| 27 | record_successful_tool_call, |
| 28 | synthesize_todo_items, |
| 29 | ) |
| 30 | from .events import AgentEvent, TurnSummary |
| 31 | from .evidence_provenance import EvidenceProvenance, EvidenceProvenanceStatus |
| 32 | from .executor import ToolExecutionState, ToolExecutor |
| 33 | from .logging import get_runtime_logger |
| 34 | from .path_display import display_runtime_path |
| 35 | from .policy_timeline import append_verification_timeline_entry |
| 36 | from .recovery import RecoveryContext, detect_missing_mutation_payload |
| 37 | from .repair_focus import ( |
| 38 | extract_active_repair_context, |
| 39 | path_within_allowed_roots, |
| 40 | recent_repair_mutation_context_failed, |
| 41 | ) |
| 42 | from .safeguard_services import extract_shell_text_rewrite_target |
| 43 | from .tool_batch_checks import ToolBatchConfidenceGate, ToolBatchVerificationGate |
| 44 | from .tool_batch_recovery import ToolBatchRecoveryController |
| 45 | from .verification_observations import ( |
| 46 | VerificationObservation, |
| 47 | VerificationObservationStatus, |
| 48 | ) |
| 49 | from .workflow import ( |
| 50 | advance_todos_from_tool_call, |
| 51 | effective_pending_todo_items, |
| 52 | infer_pending_todo_output_target, |
| 53 | preferred_pending_todo_item, |
| 54 | reconcile_aggregate_completion_steps, |
| 55 | sync_todos_to_definition_of_done, |
| 56 | todo_describes_aggregate_mutation, |
| 57 | todo_describes_broad_setup_step, |
| 58 | ) |
| 59 | |
| 60 | EventSink = Callable[[AgentEvent], Awaitable[None]] |
| 61 | ConfirmationHandler = ( |
| 62 | Callable[[str, str, str, dict[str, Any] | None], Awaitable[bool]] | None |
| 63 | ) |
| 64 | UserQuestionHandler = Callable[[str, list[str] | None], Awaitable[str]] | None |
| 65 | |
| 66 | _VERIFY_ITEM = "Collect verification evidence" |
| 67 | _TODO_NUDGE_EXCLUDED_ITEMS = { |
| 68 | "Complete the requested work", |
| 69 | _VERIFY_ITEM, |
| 70 | } |
| 71 | _MUTATION_TODO_HINTS = ( |
| 72 | "create", |
| 73 | "creating", |
| 74 | "develop", |
| 75 | "developing", |
| 76 | "populate", |
| 77 | "populating", |
| 78 | "build", |
| 79 | "building", |
| 80 | "update", |
| 81 | "updating", |
| 82 | "edit", |
| 83 | "editing", |
| 84 | "write", |
| 85 | "writing", |
| 86 | "fix", |
| 87 | "fixing", |
| 88 | "modify", |
| 89 | "modifying", |
| 90 | "change", |
| 91 | "changing", |
| 92 | "patch", |
| 93 | "patching", |
| 94 | "replace", |
| 95 | "replacing", |
| 96 | "correct", |
| 97 | "correcting", |
| 98 | "rewrite", |
| 99 | "rewriting", |
| 100 | ) |
| 101 | _CONSISTENCY_REVIEW_HINTS = ( |
| 102 | "consistent", |
| 103 | "consistently", |
| 104 | "formatted", |
| 105 | "link", |
| 106 | "linked", |
| 107 | "navigation", |
| 108 | "work properly", |
| 109 | "all files", |
| 110 | "every file", |
| 111 | "ensure", |
| 112 | ) |
| 113 | _BOOKKEEPING_NOTE_TOOL_NAMES = { |
| 114 | "notepad_write_working", |
| 115 | "notepad_append", |
| 116 | "notepad_write_priority", |
| 117 | "notepad_write_manual", |
| 118 | } |
| 119 | _SUMMARY_ARTIFACT_NAMES = { |
| 120 | "index.html", |
| 121 | "index.htm", |
| 122 | "readme", |
| 123 | "readme.md", |
| 124 | "readme.rst", |
| 125 | "readme.txt", |
| 126 | } |
| 127 | _OBSERVATION_TOOLS = frozenset({"read", "glob", "grep", "bash"}) |
| 128 | _READ_ONLY_BASH_PREFIXES = frozenset( |
| 129 | {"ls", "pwd", "find", "stat", "cat", "head", "tail", "rg", "grep"} |
| 130 | ) |
| 131 | _MUTATING_BASH_FRAGMENTS = ( |
| 132 | " >", |
| 133 | ">>", |
| 134 | "| tee", |
| 135 | "touch ", |
| 136 | "mkdir ", |
| 137 | "rm ", |
| 138 | "mv ", |
| 139 | "cp ", |
| 140 | "sed -i", |
| 141 | "perl -pi", |
| 142 | "git add", |
| 143 | "git commit", |
| 144 | "git apply", |
| 145 | ) |
| 146 | |
| 147 | |
| 148 | def _verification_handoff_instruction() -> str: |
| 149 | return ( |
| 150 | "Finish with a final response now so Loader can run verification automatically. " |
| 151 | "Do not run more ad hoc audit commands unless you know a specific mismatch to repair." |
| 152 | ) |
| 153 | |
| 154 | |
| 155 | @dataclass |
| 156 | class ToolBatchResult: |
| 157 | """Outcome of running one assistant-proposed tool batch.""" |
| 158 | |
| 159 | actions_taken: list[str] = field(default_factory=list) |
| 160 | consecutive_errors: int = 0 |
| 161 | halted: bool = False |
| 162 | continue_after_batch: bool = False |
| 163 | final_response: str = "" |
| 164 | |
| 165 | |
| 166 | class ToolBatchRunner: |
| 167 | """Owns tool-batch execution, recovery, and post-tool bookkeeping.""" |
| 168 | |
| 169 | def __init__( |
| 170 | self, |
| 171 | context: RuntimeContext, |
| 172 | dod_store: DefinitionOfDoneStore, |
| 173 | *, |
| 174 | confidence_gate: ToolBatchConfidenceGate | None = None, |
| 175 | recovery_controller: ToolBatchRecoveryController | None = None, |
| 176 | verification_gate: ToolBatchVerificationGate | None = None, |
| 177 | ) -> None: |
| 178 | self.context = context |
| 179 | self.dod_store = dod_store |
| 180 | self.confidence_gate = confidence_gate or ToolBatchConfidenceGate(context) |
| 181 | self.recovery_controller = recovery_controller or ToolBatchRecoveryController(context) |
| 182 | self.verification_gate = verification_gate or ToolBatchVerificationGate(context) |
| 183 | |
| 184 | async def execute_batch( |
| 185 | self, |
| 186 | *, |
| 187 | tool_calls: list[ToolCall], |
| 188 | tool_source: str, |
| 189 | pending_tool_calls_seen: set[str], |
| 190 | emit: EventSink, |
| 191 | summary: TurnSummary, |
| 192 | dod: DefinitionOfDone, |
| 193 | executor: ToolExecutor, |
| 194 | on_confirmation: ConfirmationHandler, |
| 195 | on_user_question: UserQuestionHandler, |
| 196 | emit_confirmation, |
| 197 | consecutive_errors: int, |
| 198 | ) -> ToolBatchResult: |
| 199 | """Run one assistant tool batch through the shared executor seam.""" |
| 200 | |
| 201 | result = ToolBatchResult(consecutive_errors=consecutive_errors) |
| 202 | |
| 203 | # Pre-populate planned items for the entire batch so the todo |
| 204 | # widget shows what's coming, not just what's done. |
| 205 | planned_labels = _batch_planned_labels(tool_calls) |
| 206 | completed_labels: list[str] = [] |
| 207 | |
| 208 | async def _emit_batch_todos() -> None: |
| 209 | """Emit a todo update combining DoD state with batch progress.""" |
| 210 | items = synthesize_todo_items(dod) |
| 211 | for label in planned_labels: |
| 212 | if label in completed_labels: |
| 213 | continue |
| 214 | # Don't duplicate items already in DoD |
| 215 | if any(item["content"] == label for item in items): |
| 216 | continue |
| 217 | items.append({"content": label, "status": "in_progress", "active_form": label}) |
| 218 | if items: |
| 219 | await emit(AgentEvent(type="todo_update", todo_items=items)) |
| 220 | |
| 221 | await _emit_batch_todos() |
| 222 | |
| 223 | for tool_call in tool_calls: |
| 224 | cfg = self.context.config.reasoning |
| 225 | |
| 226 | if cfg.confidence_scoring: |
| 227 | should_skip = await self.confidence_gate.should_skip( |
| 228 | tool_call=tool_call, |
| 229 | emit=emit, |
| 230 | ) |
| 231 | if should_skip: |
| 232 | continue |
| 233 | |
| 234 | if await self._maybe_preempt_post_build_user_question( |
| 235 | tool_call=tool_call, |
| 236 | dod=dod, |
| 237 | emit=emit, |
| 238 | summary=summary, |
| 239 | ): |
| 240 | result.continue_after_batch = True |
| 241 | return result |
| 242 | |
| 243 | if tool_call.id not in pending_tool_calls_seen: |
| 244 | await emit( |
| 245 | AgentEvent( |
| 246 | type="tool_call", |
| 247 | tool_name=tool_call.name, |
| 248 | tool_call_id=tool_call.id, |
| 249 | tool_args=tool_call.arguments, |
| 250 | phase="assistant", |
| 251 | ) |
| 252 | ) |
| 253 | |
| 254 | result.actions_taken.append( |
| 255 | f"{tool_call.name}: {str(tool_call.arguments)[:100]}" |
| 256 | ) |
| 257 | |
| 258 | outcome = await executor.execute_tool_call( |
| 259 | tool_call, |
| 260 | on_confirmation=on_confirmation, |
| 261 | on_user_question=on_user_question, |
| 262 | emit_confirmation=emit_confirmation, |
| 263 | source=tool_source, |
| 264 | ) |
| 265 | executed_tool_call = outcome.tool_call |
| 266 | if ( |
| 267 | outcome.rollback_action is not None |
| 268 | and self.context.config.reasoning.show_rollback_plan |
| 269 | ): |
| 270 | await emit( |
| 271 | AgentEvent( |
| 272 | type="rollback", |
| 273 | content=( |
| 274 | f"Rollback tracked: {outcome.rollback_action.description}" |
| 275 | ), |
| 276 | rollback_action=outcome.rollback_action, |
| 277 | ) |
| 278 | ) |
| 279 | |
| 280 | if ( |
| 281 | outcome.state == ToolExecutionState.EXECUTED |
| 282 | and outcome.is_error |
| 283 | and self.context.config.auto_recover |
| 284 | ): |
| 285 | recovery_result = await self.recovery_controller.build_follow_up( |
| 286 | tool_call=executed_tool_call, |
| 287 | outcome=outcome, |
| 288 | emit=emit, |
| 289 | ) |
| 290 | if recovery_result is not None: |
| 291 | summary.tool_result_messages.append(recovery_result) |
| 292 | self.context.session.append(recovery_result) |
| 293 | continue |
| 294 | |
| 295 | if outcome.state == ToolExecutionState.EXECUTED and not outcome.is_error: |
| 296 | loop_response = await self._record_successful_execution( |
| 297 | tool_call=executed_tool_call, |
| 298 | outcome=outcome, |
| 299 | dod=dod, |
| 300 | emit=emit, |
| 301 | summary=summary, |
| 302 | ) |
| 303 | # Mark this tool's label as completed and emit live progress |
| 304 | label = _tool_call_label(executed_tool_call) |
| 305 | if label: |
| 306 | completed_labels.append(label) |
| 307 | await _emit_batch_todos() |
| 308 | if loop_response is not None: |
| 309 | result.halted = True |
| 310 | result.final_response = loop_response |
| 311 | return result |
| 312 | |
| 313 | if outcome.is_error: |
| 314 | if _is_recoverable_guidance_block(outcome.event_content): |
| 315 | result.consecutive_errors = 0 |
| 316 | else: |
| 317 | result.consecutive_errors += 1 |
| 318 | else: |
| 319 | result.consecutive_errors = 0 |
| 320 | |
| 321 | await emit( |
| 322 | AgentEvent( |
| 323 | type="tool_result", |
| 324 | content=outcome.event_content, |
| 325 | tool_name=executed_tool_call.name, |
| 326 | tool_call_id=outcome.tool_call.id, |
| 327 | tool_metadata=( |
| 328 | outcome.registry_result.metadata |
| 329 | if outcome.registry_result is not None |
| 330 | else None |
| 331 | ), |
| 332 | is_error=outcome.is_error, |
| 333 | phase="assistant", |
| 334 | ) |
| 335 | ) |
| 336 | |
| 337 | # Always append tool results to the session so the model sees |
| 338 | # its own output. The verification gate may inject a correction |
| 339 | # prompt, but the original result must still be in context — |
| 340 | # otherwise the model operates blind and loops. |
| 341 | self.context.session.append(outcome.message) |
| 342 | summary.tool_result_messages.append(outcome.message) |
| 343 | if outcome.state == ToolExecutionState.DUPLICATE: |
| 344 | self._queue_duplicate_mutation_nudge(tool_call, dod=dod) |
| 345 | self._queue_duplicate_observation_nudge(tool_call, dod=dod) |
| 346 | elif outcome.state == ToolExecutionState.BLOCKED: |
| 347 | self._queue_blocked_invalid_mutation_nudge( |
| 348 | tool_call, |
| 349 | outcome.event_content, |
| 350 | dod=dod, |
| 351 | ) |
| 352 | self._queue_blocked_html_declared_file_creation_nudge( |
| 353 | tool_call, |
| 354 | outcome.event_content, |
| 355 | dod=dod, |
| 356 | ) |
| 357 | self._queue_blocked_html_declared_target_nudge( |
| 358 | tool_call, |
| 359 | outcome.event_content, |
| 360 | ) |
| 361 | self._queue_blocked_html_missing_target_nudge( |
| 362 | tool_call, |
| 363 | outcome.event_content, |
| 364 | dod=dod, |
| 365 | ) |
| 366 | self._queue_blocked_html_asset_nudge( |
| 367 | tool_call, |
| 368 | outcome.event_content, |
| 369 | ) |
| 370 | self._queue_blocked_active_repair_nudge(outcome.event_content) |
| 371 | self._queue_blocked_active_repair_mutation_nudge(outcome.event_content) |
| 372 | self._queue_blocked_completed_artifact_scope_nudge( |
| 373 | outcome.event_content, |
| 374 | dod=dod, |
| 375 | ) |
| 376 | self._queue_blocked_late_reference_drift_nudge( |
| 377 | outcome.event_content, |
| 378 | dod=dod, |
| 379 | ) |
| 380 | self._queue_blocked_shell_rewrite_nudge(tool_call) |
| 381 | self._queue_blocked_html_edit_nudge( |
| 382 | tool_call, |
| 383 | outcome.event_content, |
| 384 | dod=dod, |
| 385 | ) |
| 386 | else: |
| 387 | self._queue_active_repair_success_handoff_nudge(tool_call) |
| 388 | self._queue_post_mutation_self_audit_nudge(tool_call, dod=dod) |
| 389 | self._queue_completed_artifact_observation_handoff_nudge( |
| 390 | tool_call, |
| 391 | dod=dod, |
| 392 | ) |
| 393 | completed_todowrite_response = ( |
| 394 | self._completed_todowrite_verification_response( |
| 395 | tool_call=executed_tool_call, |
| 396 | dod=dod, |
| 397 | single_tool_batch=len(tool_calls) == 1, |
| 398 | ) |
| 399 | ) |
| 400 | if completed_todowrite_response is not None: |
| 401 | summary.final_response = completed_todowrite_response |
| 402 | await emit( |
| 403 | AgentEvent( |
| 404 | type="response", |
| 405 | content=completed_todowrite_response, |
| 406 | ) |
| 407 | ) |
| 408 | result.halted = True |
| 409 | result.final_response = completed_todowrite_response |
| 410 | return result |
| 411 | if self._should_preempt_for_verification_handoff( |
| 412 | tool_call=executed_tool_call, |
| 413 | dod=dod, |
| 414 | ): |
| 415 | result.continue_after_batch = True |
| 416 | return result |
| 417 | |
| 418 | should_continue = await self.verification_gate.should_continue( |
| 419 | tool_call=tool_call, |
| 420 | outcome=outcome, |
| 421 | emit=emit, |
| 422 | ) |
| 423 | |
| 424 | rlog = get_runtime_logger() |
| 425 | rlog.tool_exec( |
| 426 | name=tool_call.name, |
| 427 | state=outcome.state.value, |
| 428 | is_error=outcome.is_error, |
| 429 | result_preview=outcome.event_content, |
| 430 | appended_to_session=True, |
| 431 | ) |
| 432 | if should_continue: |
| 433 | rlog.verification_gate(tool_call.name, should_continue=True) |
| 434 | continue |
| 435 | |
| 436 | if result.consecutive_errors >= 3: |
| 437 | final_response = ( |
| 438 | "I ran into some issues. " |
| 439 | "Let me know if you'd like me to try a different approach." |
| 440 | ) |
| 441 | summary.final_response = final_response |
| 442 | summary.failures.append("three consecutive tool errors") |
| 443 | await emit(AgentEvent(type="response", content=final_response)) |
| 444 | result.halted = True |
| 445 | result.final_response = final_response |
| 446 | |
| 447 | return result |
| 448 | |
| 449 | async def _maybe_preempt_post_build_user_question( |
| 450 | self, |
| 451 | *, |
| 452 | tool_call: ToolCall, |
| 453 | dod: DefinitionOfDone, |
| 454 | emit: EventSink, |
| 455 | summary: TurnSummary, |
| 456 | ) -> bool: |
| 457 | if tool_call.name != "AskUserQuestion": |
| 458 | return False |
| 459 | |
| 460 | pending_items = effective_pending_todo_items(dod) |
| 461 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 462 | dod, |
| 463 | project_root=self.context.project_root, |
| 464 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 465 | supplement_existing=True, |
| 466 | ) |
| 467 | if pending_items: |
| 468 | if not all( |
| 469 | item == _VERIFY_ITEM or _todo_is_consistency_review_step(item) |
| 470 | for item in pending_items |
| 471 | ): |
| 472 | return False |
| 473 | elif not verification_commands: |
| 474 | return False |
| 475 | guidance = ( |
| 476 | "The remaining work is review/verification of the generated files. " |
| 477 | "Do not ask the user for more clarification about the reference pattern now. " |
| 478 | "Use the generated files as the source of truth and repair any concrete failures " |
| 479 | f"in those files. {_verification_handoff_instruction()}" |
| 480 | ) |
| 481 | if verification_commands: |
| 482 | self.context.workflow_mode = "verify" |
| 483 | guidance += " Do not run more ad hoc audit commands." |
| 484 | |
| 485 | self.context.queue_steering_message(guidance) |
| 486 | message = Message.tool_result_message( |
| 487 | tool_call_id=tool_call.id, |
| 488 | display_content=f"[Skipped - stale post-build user question] {guidance}", |
| 489 | result_content=f"[Skipped - stale post-build user question] {guidance}", |
| 490 | is_error=False, |
| 491 | ) |
| 492 | await emit( |
| 493 | AgentEvent( |
| 494 | type="tool_result", |
| 495 | content=message.content, |
| 496 | tool_name=tool_call.name, |
| 497 | tool_call_id=tool_call.id, |
| 498 | is_error=False, |
| 499 | phase="assistant", |
| 500 | ) |
| 501 | ) |
| 502 | self.context.session.append(message) |
| 503 | summary.tool_result_messages.append(message) |
| 504 | return True |
| 505 | |
| 506 | def _should_preempt_for_verification_handoff( |
| 507 | self, |
| 508 | *, |
| 509 | tool_call: ToolCall, |
| 510 | dod: DefinitionOfDone, |
| 511 | ) -> bool: |
| 512 | """Yield back to the main loop once post-build work has clearly transitioned to verify.""" |
| 513 | |
| 514 | if dod.status in {"fixing", "done"}: |
| 515 | return False |
| 516 | if not all_planned_artifact_outputs_exist(dod, project_root=self.context.project_root): |
| 517 | return False |
| 518 | if tool_call.name in _OBSERVATION_TOOLS: |
| 519 | return True |
| 520 | |
| 521 | if self.context.workflow_mode != "verify": |
| 522 | return False |
| 523 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 524 | dod, |
| 525 | project_root=self.context.project_root, |
| 526 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 527 | supplement_existing=True, |
| 528 | ) |
| 529 | if not verification_commands: |
| 530 | return False |
| 531 | return tool_call.name in ({"TodoWrite"} | _BOOKKEEPING_NOTE_TOOL_NAMES) |
| 532 | |
| 533 | def _completed_todowrite_verification_response( |
| 534 | self, |
| 535 | *, |
| 536 | tool_call: ToolCall, |
| 537 | dod: DefinitionOfDone, |
| 538 | single_tool_batch: bool, |
| 539 | ) -> str | None: |
| 540 | if tool_call.name != "TodoWrite" or not single_tool_batch: |
| 541 | return None |
| 542 | repair = extract_active_repair_context(self.context.session.messages) |
| 543 | if repair is not None and _repair_context_is_html_quality(repair): |
| 544 | return None |
| 545 | if not all_planned_artifact_outputs_exist( |
| 546 | dod, |
| 547 | project_root=self.context.project_root, |
| 548 | ): |
| 549 | return None |
| 550 | |
| 551 | pending_items = [ |
| 552 | item |
| 553 | for item in effective_pending_todo_items( |
| 554 | dod, |
| 555 | project_root=self.context.project_root, |
| 556 | ) |
| 557 | if item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 558 | ] |
| 559 | if pending_items: |
| 560 | return None |
| 561 | |
| 562 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 563 | dod, |
| 564 | project_root=self.context.project_root, |
| 565 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 566 | supplement_existing=True, |
| 567 | ) |
| 568 | if not verification_commands: |
| 569 | return None |
| 570 | |
| 571 | self.context.set_workflow_mode("verify") |
| 572 | return ( |
| 573 | "Todo tracking is complete; running Loader verification on the generated " |
| 574 | "files now." |
| 575 | ) |
| 576 | |
| 577 | def _queue_duplicate_observation_nudge( |
| 578 | self, |
| 579 | tool_call: ToolCall, |
| 580 | *, |
| 581 | dod: DefinitionOfDone, |
| 582 | ) -> None: |
| 583 | """Queue a concrete next-step nudge after duplicate observational actions.""" |
| 584 | |
| 585 | if tool_call.name not in {"read", "glob", "grep", "bash"}: |
| 586 | return |
| 587 | |
| 588 | current_task = getattr(self.context.session, "current_task", None) |
| 589 | missing_artifact = _next_missing_planned_artifact( |
| 590 | dod, |
| 591 | project_root=self.context.project_root, |
| 592 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 593 | ) |
| 594 | next_pending = preferred_pending_todo_item( |
| 595 | dod, |
| 596 | project_root=self.context.project_root, |
| 597 | missing_artifact=missing_artifact, |
| 598 | ) |
| 599 | confirmed_facts = summarize_confirmed_facts( |
| 600 | self.context.session.messages, |
| 601 | max_items=2, |
| 602 | ) |
| 603 | edit_mismatch_target = _recent_edit_string_mismatch_target( |
| 604 | self.context.recovery_context, |
| 605 | ) |
| 606 | if edit_mismatch_target and _tool_call_targets_path(tool_call, edit_mismatch_target): |
| 607 | self.context.queue_steering_message( |
| 608 | "Reuse the earlier observation instead of repeating it. " |
| 609 | f"The last edit on `{edit_mismatch_target}` failed because `old_string` " |
| 610 | "did not exactly match the current file. Use the already-read contents " |
| 611 | "as the source of truth and send one concrete mutation now: `edit` with " |
| 612 | "an exact `old_string` copied from that file, `patch`, or `write` with " |
| 613 | "the complete replacement content if the change rewrites most of the file. " |
| 614 | "Do not read the same file again first." |
| 615 | ) |
| 616 | return |
| 617 | if _should_prioritize_missing_artifact( |
| 618 | dod=dod, |
| 619 | next_pending=next_pending, |
| 620 | missing_artifact=missing_artifact, |
| 621 | project_root=self.context.project_root, |
| 622 | ): |
| 623 | prefix = "Reuse the earlier observation instead of repeating it. " |
| 624 | if confirmed_facts: |
| 625 | prefix += f"Confirmed facts: {confirmed_facts}. " |
| 626 | self.context.queue_steering_message( |
| 627 | prefix |
| 628 | + "A declared output artifact is still missing." |
| 629 | + _missing_artifact_resume_suffix( |
| 630 | missing_artifact, |
| 631 | project_root=self.context.project_root, |
| 632 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 633 | ) |
| 634 | + " Do not switch into review or consistency-check mode until the missing artifact exists." |
| 635 | ) |
| 636 | return |
| 637 | if next_pending: |
| 638 | mutation_suffix = "" |
| 639 | if _todo_is_mutation_step(next_pending): |
| 640 | mutation_suffix = _pending_item_resume_suffix( |
| 641 | dod, |
| 642 | next_pending=next_pending, |
| 643 | missing_artifact=missing_artifact, |
| 644 | project_root=self.context.project_root, |
| 645 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 646 | ) |
| 647 | if not mutation_suffix: |
| 648 | mutation_suffix = ( |
| 649 | " You already have enough evidence for that step, so stop gathering " |
| 650 | "more reference material and perform the change now." |
| 651 | ) |
| 652 | if confirmed_facts: |
| 653 | self.context.queue_steering_message( |
| 654 | "Reuse the earlier observation instead of repeating it. " |
| 655 | f"Confirmed facts: {confirmed_facts}. " |
| 656 | f"Continue with the next pending item: `{next_pending}`. " |
| 657 | "Only gather more evidence if a specific fact required for that step is still unknown." |
| 658 | + mutation_suffix |
| 659 | ) |
| 660 | else: |
| 661 | self.context.queue_steering_message( |
| 662 | "Reuse the earlier observation instead of repeating it. " |
| 663 | f"Continue with the next pending item: `{next_pending}`. " |
| 664 | "Only gather more evidence if a specific fact required for that step is still unknown." |
| 665 | + mutation_suffix |
| 666 | ) |
| 667 | return |
| 668 | |
| 669 | if missing_artifact is not None: |
| 670 | self.context.queue_steering_message( |
| 671 | "Reuse the earlier observation instead of repeating it. " |
| 672 | + _missing_artifact_resume_suffix( |
| 673 | missing_artifact, |
| 674 | project_root=self.context.project_root, |
| 675 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 676 | ).strip() |
| 677 | ) |
| 678 | return |
| 679 | |
| 680 | if all_planned_artifact_outputs_exist(dod, project_root=self.context.project_root): |
| 681 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 682 | dod, |
| 683 | project_root=self.context.project_root, |
| 684 | task_statement=current_task, |
| 685 | supplement_existing=True, |
| 686 | ) |
| 687 | verification_suffix = ( |
| 688 | _verification_handoff_instruction() |
| 689 | if verification_commands |
| 690 | else "Finish the current review using the files already on disk." |
| 691 | ) |
| 692 | self.context.queue_steering_message( |
| 693 | "Reuse the earlier observation instead of repeating it. " |
| 694 | "All explicitly planned artifacts already exist on disk. " |
| 695 | "Use the current task artifacts as the source of truth and do not reopen " |
| 696 | "reference materials unless one specific gap is still unknown. " |
| 697 | "If anything is still wrong, repair the current files instead of expanding the artifact set. " |
| 698 | + verification_suffix |
| 699 | ) |
| 700 | return |
| 701 | |
| 702 | preferred_next_step = infer_preferred_next_step( |
| 703 | self.context.session.messages, |
| 704 | current_task=current_task, |
| 705 | ) |
| 706 | if preferred_next_step and confirmed_facts: |
| 707 | self.context.queue_steering_message( |
| 708 | "Reuse the earlier observation instead of repeating it. " |
| 709 | f"Confirmed facts: {confirmed_facts}. " |
| 710 | f"{preferred_next_step} " |
| 711 | "Only gather more evidence if a specific filename, href, or title is still unknown." |
| 712 | ) |
| 713 | return |
| 714 | |
| 715 | if preferred_next_step: |
| 716 | self.context.queue_steering_message( |
| 717 | "Reuse the earlier observation instead of repeating it. " |
| 718 | f"{preferred_next_step} " |
| 719 | "Only gather more evidence if a specific filename, href, or title is still unknown." |
| 720 | ) |
| 721 | return |
| 722 | |
| 723 | target_path = str( |
| 724 | tool_call.arguments.get("file_path") |
| 725 | or tool_call.arguments.get("path") |
| 726 | or "" |
| 727 | ).strip() |
| 728 | if target_path: |
| 729 | self.context.queue_steering_message( |
| 730 | "Reuse the earlier observation instead of repeating it. " |
| 731 | f"Use the current contents of `{target_path}` and take a different next step. " |
| 732 | "Only gather more evidence if a specific filename, href, or title is still unknown." |
| 733 | ) |
| 734 | return |
| 735 | |
| 736 | self.context.queue_steering_message( |
| 737 | "Reuse the earlier observation instead of repeating it. " |
| 738 | "Choose a different next step that makes progress." |
| 739 | ) |
| 740 | |
| 741 | def _queue_duplicate_mutation_nudge( |
| 742 | self, |
| 743 | tool_call: ToolCall, |
| 744 | *, |
| 745 | dod: DefinitionOfDone, |
| 746 | ) -> None: |
| 747 | """After a duplicate mutation, restate concrete repair deltas.""" |
| 748 | |
| 749 | if tool_call.name not in {"write", "edit", "patch"}: |
| 750 | return |
| 751 | |
| 752 | target = str( |
| 753 | tool_call.arguments.get("file_path") |
| 754 | or tool_call.arguments.get("path") |
| 755 | or "" |
| 756 | ).strip() |
| 757 | repair = extract_active_repair_context(self.context.session.messages) |
| 758 | if repair is not None: |
| 759 | repair_preview = _active_repair_focus_preview(repair.repair_lines) |
| 760 | target_label = f"`{target}`" if target else "that file" |
| 761 | self.context.queue_steering_message( |
| 762 | f"That {tool_call.name} was skipped because it would not change {target_label}. " |
| 763 | "Do not submit the same content again. " |
| 764 | f"Verification still requires these concrete repair deltas: {repair_preview} " |
| 765 | "Use the current generated file as the source of truth and make one real edit, " |
| 766 | "patch, or write that expands or changes the flagged artifact." |
| 767 | ) |
| 768 | return |
| 769 | |
| 770 | if all_planned_artifact_outputs_exist(dod, project_root=self.context.project_root): |
| 771 | target_label = f"`{target}`" if target else "the target file" |
| 772 | self.context.queue_steering_message( |
| 773 | f"That {tool_call.name} was skipped because it would not change {target_label}. " |
| 774 | "All explicitly planned artifacts already exist, so do not rewrite the same content. " |
| 775 | "If verification identified a mismatch, make a different concrete mutation that fixes it; " |
| 776 | "otherwise finish so Loader can verify the files already on disk." |
| 777 | ) |
| 778 | |
| 779 | def _queue_post_mutation_self_audit_nudge( |
| 780 | self, |
| 781 | tool_call: ToolCall, |
| 782 | *, |
| 783 | dod: DefinitionOfDone, |
| 784 | ) -> None: |
| 785 | """Steer out of rereading the file that was just written when the next output is known.""" |
| 786 | |
| 787 | if tool_call.name != "read": |
| 788 | return |
| 789 | |
| 790 | file_path = str(tool_call.arguments.get("file_path", "")).strip() |
| 791 | if not file_path: |
| 792 | return |
| 793 | |
| 794 | missing_artifact = _next_missing_planned_artifact( |
| 795 | dod, |
| 796 | project_root=self.context.project_root, |
| 797 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 798 | ) |
| 799 | if missing_artifact is None: |
| 800 | return |
| 801 | |
| 802 | read_target = Path(file_path).expanduser().resolve(strict=False) |
| 803 | last_touched = _last_touched_file_path(dod) |
| 804 | if last_touched is None or read_target != last_touched: |
| 805 | return |
| 806 | |
| 807 | self.context.queue_steering_message( |
| 808 | f"You already have the current contents of `{read_target.name}` from the successful write. " |
| 809 | "A declared output artifact is still missing." |
| 810 | + _missing_artifact_resume_suffix( |
| 811 | missing_artifact, |
| 812 | project_root=self.context.project_root, |
| 813 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 814 | ) |
| 815 | + " Do not spend another turn rereading the file you just wrote or on TodoWrite alone." |
| 816 | ) |
| 817 | |
| 818 | def _queue_completed_artifact_observation_handoff_nudge( |
| 819 | self, |
| 820 | tool_call: ToolCall, |
| 821 | *, |
| 822 | dod: DefinitionOfDone, |
| 823 | ) -> None: |
| 824 | """Turn successful post-build audit reads into verify/finalize handoffs.""" |
| 825 | |
| 826 | if tool_call.name not in _OBSERVATION_TOOLS: |
| 827 | return |
| 828 | if dod.status in {"fixing", "done"}: |
| 829 | return |
| 830 | if extract_active_repair_context(self.context.session.messages) is not None: |
| 831 | return |
| 832 | if not all_planned_artifact_outputs_exist(dod, project_root=self.context.project_root): |
| 833 | return |
| 834 | |
| 835 | observed_paths = _extract_observation_paths(tool_call) |
| 836 | if not observed_paths: |
| 837 | return |
| 838 | |
| 839 | planned_roots = _planned_output_roots( |
| 840 | dod, |
| 841 | project_root=self.context.project_root, |
| 842 | ) |
| 843 | if not planned_roots: |
| 844 | return |
| 845 | if not all(path_within_allowed_roots(path, planned_roots) for path in observed_paths): |
| 846 | return |
| 847 | |
| 848 | next_pending = preferred_pending_todo_item( |
| 849 | dod, |
| 850 | project_root=self.context.project_root, |
| 851 | ) |
| 852 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 853 | dod, |
| 854 | project_root=self.context.project_root, |
| 855 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 856 | supplement_existing=True, |
| 857 | ) |
| 858 | roots_preview = ", ".join(f"`{root}`" for root in planned_roots[:2]) |
| 859 | if len(planned_roots) > 2: |
| 860 | roots_preview += ", ..." |
| 861 | |
| 862 | if next_pending and _todo_is_consistency_review_step(next_pending): |
| 863 | verification_suffix = ( |
| 864 | " If no specific mismatch remains, finish with a final response so Loader can verify." |
| 865 | if verification_commands |
| 866 | else " If no specific mismatch remains, finish the task now." |
| 867 | ) |
| 868 | self.context.queue_ephemeral_steering_message( |
| 869 | "All explicitly planned artifacts already exist. " |
| 870 | f"Continue with `{next_pending}` using the generated files under {roots_preview} " |
| 871 | "as the source of truth, but do not keep broad-rereading the output set. " |
| 872 | "If you already know a concrete mismatch, fix it directly." |
| 873 | + verification_suffix |
| 874 | ) |
| 875 | return |
| 876 | |
| 877 | if verification_commands: |
| 878 | self.context.set_workflow_mode("verify") |
| 879 | self.context.queue_steering_message( |
| 880 | "All explicitly planned artifacts already exist. " |
| 881 | f"Use the generated files under {roots_preview} as the source of truth and stop broad rereads. " |
| 882 | "If you already know a concrete mismatch, fix it directly. " |
| 883 | f"{_verification_handoff_instruction()} Do not reopen reference materials " |
| 884 | "or keep auditing the same files." |
| 885 | ) |
| 886 | return |
| 887 | |
| 888 | verification_suffix = ( |
| 889 | _verification_handoff_instruction() |
| 890 | if verification_commands |
| 891 | else "Finish the task using the files already on disk." |
| 892 | ) |
| 893 | self.context.queue_ephemeral_steering_message( |
| 894 | "All explicitly planned artifacts already exist. " |
| 895 | f"Use the generated files under {roots_preview} as the source of truth and stop broad rereads. " |
| 896 | "If you already know a concrete mismatch, fix it directly. " |
| 897 | + verification_suffix |
| 898 | ) |
| 899 | |
| 900 | def _queue_blocked_shell_rewrite_nudge(self, tool_call: ToolCall) -> None: |
| 901 | """Steer the model back to file tools after a blocked shell text rewrite.""" |
| 902 | |
| 903 | if tool_call.name != "bash": |
| 904 | return |
| 905 | |
| 906 | target = extract_shell_text_rewrite_target( |
| 907 | str(tool_call.arguments.get("command", "")) |
| 908 | ) |
| 909 | if target is None: |
| 910 | return |
| 911 | |
| 912 | current_task = getattr(self.context.session, "current_task", None) |
| 913 | confirmed_facts = summarize_confirmed_facts( |
| 914 | self.context.session.messages, |
| 915 | max_items=2, |
| 916 | ) |
| 917 | preferred_next_step = infer_preferred_next_step( |
| 918 | self.context.session.messages, |
| 919 | current_task=current_task, |
| 920 | ) |
| 921 | |
| 922 | if preferred_next_step and confirmed_facts: |
| 923 | self.context.queue_steering_message( |
| 924 | "Use Loader's file tools for this text edit instead of a shell rewrite. " |
| 925 | f"Confirmed facts: {confirmed_facts}. " |
| 926 | f"{preferred_next_step} " |
| 927 | f"Target `{target}` with edit/patch/write rather than `bash`." |
| 928 | ) |
| 929 | return |
| 930 | |
| 931 | self.context.queue_steering_message( |
| 932 | "Use Loader's file tools for this text edit instead of a shell rewrite. " |
| 933 | f"Apply the change to `{target}` with edit/patch/write." |
| 934 | ) |
| 935 | |
| 936 | def _queue_blocked_active_repair_nudge(self, event_content: str) -> None: |
| 937 | """Reinforce active repair focus after an out-of-scope blocked observation.""" |
| 938 | |
| 939 | if "[Blocked - active repair scope:" not in event_content: |
| 940 | return |
| 941 | |
| 942 | repair = extract_active_repair_context(self.context.session.messages) |
| 943 | if repair is None: |
| 944 | return |
| 945 | |
| 946 | if repair.allowed_paths: |
| 947 | allowed_preview = ", ".join(f"`{path}`" for path in repair.allowed_paths[:3]) |
| 948 | if len(repair.allowed_paths) > 3: |
| 949 | allowed_preview += ", ..." |
| 950 | self.context.queue_steering_message( |
| 951 | "Verification already identified the active repair target. " |
| 952 | f"Stay on the concrete repair files {allowed_preview} " |
| 953 | f"and repair `{repair.artifact_path}` directly. " |
| 954 | "Do not reopen unrelated reference materials while this repair target is unresolved." |
| 955 | ) |
| 956 | return |
| 957 | |
| 958 | roots_preview = ", ".join(f"`{root}`" for root in repair.allowed_roots[:2]) |
| 959 | if len(repair.allowed_roots) > 2: |
| 960 | roots_preview += ", ..." |
| 961 | self.context.queue_steering_message( |
| 962 | "Verification already identified the active repair target. " |
| 963 | f"Stay within the current artifact set under {roots_preview} " |
| 964 | f"and repair `{repair.artifact_path}` directly. " |
| 965 | "Do not reopen unrelated reference materials while this repair target is unresolved." |
| 966 | ) |
| 967 | |
| 968 | def _queue_blocked_active_repair_mutation_nudge(self, event_content: str) -> None: |
| 969 | """Keep repair-phase mutations pinned to the named repair files.""" |
| 970 | |
| 971 | if "[Blocked - active repair mutation scope:" not in event_content: |
| 972 | return |
| 973 | |
| 974 | repair = extract_active_repair_context(self.context.session.messages) |
| 975 | if repair is None or not repair.allowed_paths: |
| 976 | return |
| 977 | |
| 978 | allowed_preview = ", ".join(f"`{path}`" for path in repair.allowed_paths[:3]) |
| 979 | if len(repair.allowed_paths) > 3: |
| 980 | allowed_preview += ", ..." |
| 981 | self.context.queue_steering_message( |
| 982 | "Verification already identified the concrete repair files. " |
| 983 | f"Keep mutations pinned to {allowed_preview} " |
| 984 | f"and repair `{repair.artifact_path}` before widening the change set." |
| 985 | ) |
| 986 | |
| 987 | def _queue_active_repair_success_handoff_nudge(self, tool_call: ToolCall) -> None: |
| 988 | """After a repair mutation, hand back to finalization instead of ad hoc edits.""" |
| 989 | |
| 990 | if tool_call.name not in {"write", "edit", "patch"}: |
| 991 | return |
| 992 | raw_path = str(tool_call.arguments.get("file_path", "")).strip() |
| 993 | if not raw_path: |
| 994 | return |
| 995 | |
| 996 | repair = extract_active_repair_context(self.context.session.messages) |
| 997 | if repair is None or not repair.allowed_paths: |
| 998 | return |
| 999 | |
| 1000 | try: |
| 1001 | changed_path = str(Path(raw_path).expanduser().resolve(strict=False)) |
| 1002 | except (OSError, RuntimeError, ValueError): |
| 1003 | changed_path = str(Path(raw_path).expanduser()) |
| 1004 | allowed_paths = { |
| 1005 | str(Path(path).expanduser().resolve(strict=False)) |
| 1006 | for path in repair.allowed_paths |
| 1007 | } |
| 1008 | if changed_path not in allowed_paths: |
| 1009 | return |
| 1010 | |
| 1011 | if _repair_context_is_html_quality(repair): |
| 1012 | next_target = _next_quality_repair_path( |
| 1013 | repair, |
| 1014 | changed_path=changed_path, |
| 1015 | ) |
| 1016 | if next_target: |
| 1017 | repair_issue = _quality_repair_issue_for_target(repair, next_target) |
| 1018 | issue_line = ( |
| 1019 | f"- {repair_issue}\n" |
| 1020 | if repair_issue |
| 1021 | else f"- Improve `{next_target}` until it satisfies the active content-quality verifier.\n" |
| 1022 | ) |
| 1023 | self.context.queue_steering_message( |
| 1024 | "The active HTML content-quality repair target was updated. " |
| 1025 | "If the current file now comfortably clears its stated threshold, " |
| 1026 | f"continue directly with the next listed quality target `{next_target}` " |
| 1027 | "using one substantial write/edit/patch anchored to current content. " |
| 1028 | "If it still looks thin, expand this same file further now. " |
| 1029 | "Do not rerun verification, reopen unrelated references, or summarize " |
| 1030 | "completion after only one quality target.\n\n" |
| 1031 | "Repair focus:\n" |
| 1032 | f"{issue_line}" |
| 1033 | f"- Immediate next step: edit `{next_target}`.\n" |
| 1034 | "- Continue with one concrete `edit`, `patch`, or `write` call that " |
| 1035 | "actually changes the current generated file." |
| 1036 | ) |
| 1037 | return |
| 1038 | |
| 1039 | if changed_path == str(Path(repair.artifact_path).expanduser().resolve(strict=False)): |
| 1040 | self.context.queue_steering_message( |
| 1041 | "The active verification repair target was updated. " |
| 1042 | "Do not keep auditing or retarget nearby links by guesswork. " |
| 1043 | "Finish with a final response now so Loader can re-run verification." |
| 1044 | ) |
| 1045 | return |
| 1046 | |
| 1047 | self.context.queue_steering_message( |
| 1048 | "The support file for the active verification repair now exists. " |
| 1049 | f"Do not retarget `{repair.artifact_path}` to a different missing path by guesswork. " |
| 1050 | "Finish with a final response now so Loader can re-run verification." |
| 1051 | ) |
| 1052 | |
| 1053 | def _queue_blocked_late_reference_drift_nudge( |
| 1054 | self, |
| 1055 | event_content: str, |
| 1056 | *, |
| 1057 | dod: DefinitionOfDone, |
| 1058 | ) -> None: |
| 1059 | """Reinforce missing-artifact progress after late-stage reference drift is blocked.""" |
| 1060 | |
| 1061 | if "[Blocked - late reference drift:" not in event_content: |
| 1062 | return |
| 1063 | |
| 1064 | missing_artifact = _next_missing_planned_artifact( |
| 1065 | dod, |
| 1066 | project_root=self.context.project_root, |
| 1067 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1068 | ) |
| 1069 | if missing_artifact is None: |
| 1070 | return |
| 1071 | |
| 1072 | planned_roots: list[str] = [] |
| 1073 | seen_roots: set[str] = set() |
| 1074 | for target, expect_directory in collect_planned_artifact_targets( |
| 1075 | dod, |
| 1076 | project_root=self.context.project_root, |
| 1077 | ): |
| 1078 | root = str(target if expect_directory else target.parent) |
| 1079 | if root in seen_roots: |
| 1080 | continue |
| 1081 | seen_roots.add(root) |
| 1082 | planned_roots.append(root) |
| 1083 | |
| 1084 | roots_preview = ", ".join(f"`{root}`" for root in planned_roots[:2]) |
| 1085 | if len(planned_roots) > 2: |
| 1086 | roots_preview += ", ..." |
| 1087 | self.context.queue_steering_message( |
| 1088 | "Late-stage reference rereads are no longer helping. " |
| 1089 | "One explicitly planned artifact is still missing." |
| 1090 | + _missing_artifact_resume_suffix( |
| 1091 | missing_artifact, |
| 1092 | project_root=self.context.project_root, |
| 1093 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1094 | ) |
| 1095 | + f" Stay within the current output roots under {roots_preview}" |
| 1096 | + " and finish that artifact before reopening older reference materials." |
| 1097 | ) |
| 1098 | |
| 1099 | def _queue_blocked_completed_artifact_scope_nudge( |
| 1100 | self, |
| 1101 | event_content: str, |
| 1102 | *, |
| 1103 | dod: DefinitionOfDone, |
| 1104 | ) -> None: |
| 1105 | """Keep post-build review anchored to the generated artifact set.""" |
| 1106 | |
| 1107 | blocked_completed_scope = ( |
| 1108 | "[Blocked - completed artifact set scope:" in event_content |
| 1109 | ) |
| 1110 | blocked_post_build_audit = "[Blocked - post-build audit loop:" in event_content |
| 1111 | if not blocked_completed_scope and not blocked_post_build_audit: |
| 1112 | return |
| 1113 | |
| 1114 | planned_roots: list[str] = [] |
| 1115 | seen_roots: set[str] = set() |
| 1116 | for target, expect_directory in collect_planned_artifact_targets( |
| 1117 | dod, |
| 1118 | project_root=self.context.project_root, |
| 1119 | ): |
| 1120 | root = str(target if expect_directory else target.parent) |
| 1121 | if root in seen_roots: |
| 1122 | continue |
| 1123 | seen_roots.add(root) |
| 1124 | planned_roots.append(root) |
| 1125 | |
| 1126 | next_pending = preferred_pending_todo_item( |
| 1127 | dod, |
| 1128 | project_root=self.context.project_root, |
| 1129 | ) |
| 1130 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 1131 | dod, |
| 1132 | project_root=self.context.project_root, |
| 1133 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 1134 | supplement_existing=True, |
| 1135 | ) |
| 1136 | if verification_commands: |
| 1137 | self.context.set_workflow_mode("verify") |
| 1138 | roots_preview = ", ".join(f"`{root}`" for root in planned_roots[:2]) |
| 1139 | if len(planned_roots) > 2: |
| 1140 | roots_preview += ", ..." |
| 1141 | if next_pending and _todo_is_consistency_review_step(next_pending): |
| 1142 | self.context.queue_steering_message( |
| 1143 | "All explicitly planned artifacts already exist. " |
| 1144 | f"Stay within the current output roots under {roots_preview} and continue " |
| 1145 | f"with `{next_pending}` using the generated files as the source of truth. " |
| 1146 | "Do not reopen earlier reference materials." |
| 1147 | + ( |
| 1148 | " Finish with a final response so Loader can verify those generated files." |
| 1149 | if verification_commands |
| 1150 | else "" |
| 1151 | ) |
| 1152 | ) |
| 1153 | return |
| 1154 | |
| 1155 | self.context.queue_steering_message( |
| 1156 | "All explicitly planned artifacts already exist. " |
| 1157 | f"Stay within the current output roots under {roots_preview} " |
| 1158 | "and finish with a final response so Loader can verify the generated files. " |
| 1159 | "Do not reopen earlier reference materials." |
| 1160 | ) |
| 1161 | |
| 1162 | def _queue_blocked_html_edit_nudge( |
| 1163 | self, |
| 1164 | tool_call: ToolCall, |
| 1165 | event_content: str, |
| 1166 | *, |
| 1167 | dod: DefinitionOfDone, |
| 1168 | ) -> None: |
| 1169 | """Keep blocked edit feedback generic; avoid task-class-specific steering.""" |
| 1170 | |
| 1171 | if tool_call.name != "edit": |
| 1172 | return |
| 1173 | if "old_string and new_string are identical - no change would occur" not in event_content: |
| 1174 | return |
| 1175 | |
| 1176 | repair = extract_active_repair_context(self.context.session.messages) |
| 1177 | if repair is None: |
| 1178 | return |
| 1179 | |
| 1180 | target = ( |
| 1181 | str(tool_call.arguments.get("file_path") or "").strip() or repair.artifact_path |
| 1182 | ) |
| 1183 | if not target: |
| 1184 | return |
| 1185 | |
| 1186 | if _repair_context_is_html_quality(repair): |
| 1187 | repair_issue = _quality_repair_issue_for_target(repair, target) |
| 1188 | issue_line = ( |
| 1189 | f"- {repair_issue}\n" |
| 1190 | if repair_issue |
| 1191 | else f"- Improve `{target}` until it satisfies the active content-quality verifier.\n" |
| 1192 | ) |
| 1193 | self.context.queue_steering_message( |
| 1194 | "That edit would make no on-disk change. " |
| 1195 | f"`{target}` already matches the change you attempted, but the active " |
| 1196 | "content-quality repair is not complete until a verifier-targeted " |
| 1197 | "mutation actually changes the file. Do not mark the task complete, " |
| 1198 | "do not use TodoWrite as a substitute for repair, and do not finish yet.\n\n" |
| 1199 | "Repair focus:\n" |
| 1200 | f"{issue_line}" |
| 1201 | f"- Immediate next step: edit `{target}`.\n" |
| 1202 | "- Submit one `edit`, `patch`, or `write` call that adds substantial " |
| 1203 | "new content or structure to the current generated file." |
| 1204 | ) |
| 1205 | return |
| 1206 | |
| 1207 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 1208 | dod, |
| 1209 | project_root=self.context.project_root, |
| 1210 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 1211 | supplement_existing=True, |
| 1212 | ) |
| 1213 | if all_planned_artifacts_exist(dod, project_root=self.context.project_root): |
| 1214 | verification_suffix = ( |
| 1215 | " " + _verification_handoff_instruction() |
| 1216 | if verification_commands |
| 1217 | else " If no concrete mismatch remains, stop editing and finish from the files already on disk." |
| 1218 | ) |
| 1219 | self.context.queue_steering_message( |
| 1220 | "That edit would make no on-disk change. " |
| 1221 | f"`{target}` already matches the change you attempted. " |
| 1222 | "All explicitly planned artifacts already exist." |
| 1223 | + verification_suffix |
| 1224 | ) |
| 1225 | return |
| 1226 | |
| 1227 | self.context.queue_steering_message( |
| 1228 | "That edit would make no on-disk change. " |
| 1229 | f"Stay on `{target}` and use the current file contents as the source of truth. " |
| 1230 | "Read the exact current text you need to change, then submit one `edit`, `patch`, " |
| 1231 | "or `write` call that actually changes the file. " |
| 1232 | "If a narrow single-line edit keeps bouncing, replace the surrounding block in one " |
| 1233 | "mutation instead of retrying the same no-op edit. " |
| 1234 | "Do not reopen unrelated reference materials while this concrete repair target is unresolved." |
| 1235 | ) |
| 1236 | |
| 1237 | def _queue_blocked_html_declared_target_nudge( |
| 1238 | self, |
| 1239 | tool_call: ToolCall, |
| 1240 | event_content: str, |
| 1241 | ) -> None: |
| 1242 | """Steer blocked HTML graph edits back to the root-declared local targets.""" |
| 1243 | |
| 1244 | if tool_call.name not in {"write", "edit", "patch"}: |
| 1245 | return |
| 1246 | if "HTML page introduces new local targets outside the current declared artifact set" not in event_content: |
| 1247 | return |
| 1248 | |
| 1249 | target = str( |
| 1250 | tool_call.arguments.get("file_path") |
| 1251 | or tool_call.arguments.get("path") |
| 1252 | or "" |
| 1253 | ).strip() |
| 1254 | if not target: |
| 1255 | return |
| 1256 | |
| 1257 | closest_targets = _extract_blocked_html_target_list( |
| 1258 | event_content, |
| 1259 | "Closest declared local targets include:", |
| 1260 | ) |
| 1261 | declared_targets = _extract_blocked_html_target_list( |
| 1262 | event_content, |
| 1263 | "Already-declared local targets include:", |
| 1264 | ) |
| 1265 | allowed_hrefs = _extract_blocked_html_target_list( |
| 1266 | event_content, |
| 1267 | "Allowed hrefs from this file include:", |
| 1268 | ) |
| 1269 | |
| 1270 | guidance = ( |
| 1271 | "That HTML mutation introduced sibling targets outside the current declared local-link set. " |
| 1272 | f"Stay on `{target}`." |
| 1273 | ) |
| 1274 | if allowed_hrefs: |
| 1275 | guidance += ( |
| 1276 | " Remove the invented hrefs and use only these exact href values " |
| 1277 | "from this file, or omit navigation links entirely: " |
| 1278 | + ", ".join(f"`{candidate}`" for candidate in allowed_hrefs[:6]) |
| 1279 | + "." |
| 1280 | ) |
| 1281 | elif closest_targets: |
| 1282 | guidance += ( |
| 1283 | " Remove the invented hrefs or replace them with the closest declared target(s): " |
| 1284 | + ", ".join(f"`{candidate}`" for candidate in closest_targets[:3]) |
| 1285 | + "." |
| 1286 | ) |
| 1287 | elif declared_targets: |
| 1288 | guidance += ( |
| 1289 | " Remove the invented hrefs or keep local links within the declared target set, for example: " |
| 1290 | + ", ".join(f"`{candidate}`" for candidate in declared_targets[:3]) |
| 1291 | + "." |
| 1292 | ) |
| 1293 | guidance += ( |
| 1294 | " Resend one concrete mutation for that same file now instead of rereading the reference guide." |
| 1295 | ) |
| 1296 | self.context.queue_steering_message(guidance) |
| 1297 | |
| 1298 | def _queue_blocked_html_declared_file_creation_nudge( |
| 1299 | self, |
| 1300 | tool_call: ToolCall, |
| 1301 | event_content: str, |
| 1302 | *, |
| 1303 | dod: DefinitionOfDone, |
| 1304 | ) -> None: |
| 1305 | """Steer blocked undeclared HTML file creation back through the root guide.""" |
| 1306 | |
| 1307 | if tool_call.name not in {"write", "edit", "patch"}: |
| 1308 | return |
| 1309 | if "HTML file creation falls outside the current declared artifact set" not in event_content: |
| 1310 | return |
| 1311 | |
| 1312 | target = str( |
| 1313 | tool_call.arguments.get("file_path") |
| 1314 | or tool_call.arguments.get("path") |
| 1315 | or "" |
| 1316 | ).strip() |
| 1317 | if not target: |
| 1318 | return |
| 1319 | |
| 1320 | target_path = Path(target).expanduser().resolve(strict=False) |
| 1321 | root_index = (target_path.parent.parent / "index.html").resolve(strict=False) |
| 1322 | relative_target = None |
| 1323 | try: |
| 1324 | relative_target = target_path.relative_to(root_index.parent).as_posix() |
| 1325 | except ValueError: |
| 1326 | relative_target = target_path.name |
| 1327 | closest_targets = _extract_blocked_html_target_list( |
| 1328 | event_content, |
| 1329 | "Closest declared local targets include:", |
| 1330 | ) |
| 1331 | declared_targets = _extract_blocked_html_target_list( |
| 1332 | event_content, |
| 1333 | "Already-declared local targets include:", |
| 1334 | ) |
| 1335 | |
| 1336 | if all_planned_artifact_outputs_exist(dod, project_root=self.context.project_root): |
| 1337 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 1338 | dod, |
| 1339 | project_root=self.context.project_root, |
| 1340 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 1341 | supplement_existing=True, |
| 1342 | ) |
| 1343 | verification_suffix = ( |
| 1344 | " " + _verification_handoff_instruction() |
| 1345 | if verification_commands |
| 1346 | else " Finish the task using the files already on disk." |
| 1347 | ) |
| 1348 | self.context.queue_steering_message( |
| 1349 | "All explicitly planned artifacts already exist on disk. " |
| 1350 | f"Do not expand the output set with `{relative_target}`. " |
| 1351 | "Use the current generated files as the source of truth and repair or verify them instead." |
| 1352 | + verification_suffix |
| 1353 | ) |
| 1354 | return |
| 1355 | |
| 1356 | guidance = ( |
| 1357 | "That new HTML file is outside the current root-declared artifact set. " |
| 1358 | f"Do not create `{relative_target}`." |
| 1359 | ) |
| 1360 | if closest_targets: |
| 1361 | guidance += ( |
| 1362 | " Stay within the declared set and continue with the closest declared target instead: " |
| 1363 | + ", ".join(f"`{candidate}`" for candidate in closest_targets[:3]) |
| 1364 | + "." |
| 1365 | ) |
| 1366 | else: |
| 1367 | guidance += ( |
| 1368 | f" Before creating `{relative_target}`, update `{root_index}` so the guide root " |
| 1369 | "explicitly links to that page, then retry the file creation." |
| 1370 | ) |
| 1371 | if declared_targets: |
| 1372 | guidance += ( |
| 1373 | " Already-declared local targets include: " |
| 1374 | + ", ".join(f"`{candidate}`" for candidate in declared_targets[:3]) |
| 1375 | + "." |
| 1376 | ) |
| 1377 | guidance += " Stay on the active guide files; do not reopen the earlier reference guide first." |
| 1378 | self.context.queue_steering_message(guidance) |
| 1379 | |
| 1380 | def _queue_blocked_html_missing_target_nudge( |
| 1381 | self, |
| 1382 | tool_call: ToolCall, |
| 1383 | event_content: str, |
| 1384 | *, |
| 1385 | dod: DefinitionOfDone, |
| 1386 | ) -> None: |
| 1387 | """Turn post-build missing-link expansions into verify/repair handoffs.""" |
| 1388 | |
| 1389 | if tool_call.name not in {"write", "edit", "patch"}: |
| 1390 | return |
| 1391 | if "Edited HTML links point to files that do not exist" not in event_content: |
| 1392 | return |
| 1393 | if not all_planned_artifact_outputs_exist(dod, project_root=self.context.project_root): |
| 1394 | return |
| 1395 | |
| 1396 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 1397 | dod, |
| 1398 | project_root=self.context.project_root, |
| 1399 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 1400 | supplement_existing=True, |
| 1401 | ) |
| 1402 | verification_suffix = ( |
| 1403 | " " + _verification_handoff_instruction() |
| 1404 | if verification_commands |
| 1405 | else " Finish the task using the files already on disk." |
| 1406 | ) |
| 1407 | target = str( |
| 1408 | tool_call.arguments.get("file_path") |
| 1409 | or tool_call.arguments.get("path") |
| 1410 | or "" |
| 1411 | ).strip() |
| 1412 | target_suffix = f" Stay on `{target}`." if target else "" |
| 1413 | self.context.queue_steering_message( |
| 1414 | "All explicitly planned artifacts already exist on disk. " |
| 1415 | + target_suffix |
| 1416 | + " " |
| 1417 | "Do not introduce new local-link targets beyond the current output set. " |
| 1418 | "Repair the existing generated files instead of expanding the guide. " |
| 1419 | "Replace broken hrefs with existing local targets or remove the broken link." |
| 1420 | + verification_suffix |
| 1421 | ) |
| 1422 | |
| 1423 | def _queue_blocked_html_asset_nudge( |
| 1424 | self, |
| 1425 | tool_call: ToolCall, |
| 1426 | event_content: str, |
| 1427 | ) -> None: |
| 1428 | """Make missing local asset feedback sticky for the immediate retry.""" |
| 1429 | |
| 1430 | if tool_call.name not in {"write", "edit", "patch"}: |
| 1431 | return |
| 1432 | if "HTML local asset references do not exist" not in event_content: |
| 1433 | return |
| 1434 | |
| 1435 | target = str( |
| 1436 | tool_call.arguments.get("file_path") |
| 1437 | or tool_call.arguments.get("path") |
| 1438 | or "" |
| 1439 | ).strip() |
| 1440 | if not target: |
| 1441 | return |
| 1442 | |
| 1443 | missing_assets = _extract_blocked_html_target_list( |
| 1444 | event_content, |
| 1445 | "Missing local asset href(s):", |
| 1446 | ) |
| 1447 | asset_preview = "" |
| 1448 | if missing_assets: |
| 1449 | asset_preview = ( |
| 1450 | " Remove or replace " |
| 1451 | + ", ".join(f"`{asset}`" for asset in missing_assets[:3]) |
| 1452 | + "." |
| 1453 | ) |
| 1454 | |
| 1455 | repeat_count = _count_recent_blocked_html_asset_events( |
| 1456 | self.context.session.messages, |
| 1457 | missing_assets, |
| 1458 | ) |
| 1459 | if repeat_count >= 2 and missing_assets: |
| 1460 | missing_preview = ", ".join(f"`{asset}`" for asset in missing_assets[:3]) |
| 1461 | self.context.queue_steering_message( |
| 1462 | f"The same HTML mutation for `{target}` has now been blocked " |
| 1463 | f"{repeat_count} times because local asset href(s) {missing_preview} " |
| 1464 | "do not exist. Do not resend another `write`/`edit`/`patch` for " |
| 1465 | f"`{target}` while it still contains those hrefs. Recommended next " |
| 1466 | "action: retry the same file with the entire stylesheet `<link>` " |
| 1467 | "line removed and inline any necessary styling. Alternative: create " |
| 1468 | "the referenced asset file first, then link to it. Do not claim " |
| 1469 | "completion until this blocked file write succeeds." |
| 1470 | ) |
| 1471 | return |
| 1472 | |
| 1473 | self.context.queue_steering_message( |
| 1474 | f"The last HTML mutation for `{target}` was blocked, so that file was " |
| 1475 | "not created or updated. Retry the same file with one concrete " |
| 1476 | "`write`, `edit`, or `patch` call that omits missing local asset hrefs." |
| 1477 | f"{asset_preview} Inline necessary styling/content, or create the referenced " |
| 1478 | "asset before linking it. Do not resend the same `<link>` tag and do " |
| 1479 | "not claim completion until the blocked file write succeeds." |
| 1480 | ) |
| 1481 | |
| 1482 | def _queue_blocked_invalid_mutation_nudge( |
| 1483 | self, |
| 1484 | tool_call: ToolCall, |
| 1485 | event_content: str, |
| 1486 | *, |
| 1487 | dod: DefinitionOfDone, |
| 1488 | ) -> None: |
| 1489 | """Recover blocked mutations that omitted a real target path or text payload.""" |
| 1490 | |
| 1491 | fix = detect_missing_mutation_payload( |
| 1492 | tool_call.name, |
| 1493 | tool_call.arguments, |
| 1494 | event_content, |
| 1495 | ) |
| 1496 | if fix is None: |
| 1497 | return |
| 1498 | |
| 1499 | self._record_blocked_invalid_mutation_attempt(tool_call, event_content) |
| 1500 | |
| 1501 | messages = list(getattr(self.context.session, "messages", []) or []) |
| 1502 | missing_artifact = _next_missing_planned_artifact( |
| 1503 | dod, |
| 1504 | project_root=self.context.project_root, |
| 1505 | messages=messages, |
| 1506 | ) |
| 1507 | next_pending = preferred_pending_todo_item( |
| 1508 | dod, |
| 1509 | project_root=self.context.project_root, |
| 1510 | missing_artifact=missing_artifact, |
| 1511 | ) |
| 1512 | missing_artifact = _prefer_missing_artifact_for_pending_item( |
| 1513 | dod, |
| 1514 | missing_artifact=missing_artifact, |
| 1515 | next_pending=next_pending, |
| 1516 | project_root=self.context.project_root, |
| 1517 | ) |
| 1518 | resume_target = _preferred_resume_target_path( |
| 1519 | dod, |
| 1520 | next_pending=next_pending, |
| 1521 | missing_artifact=missing_artifact, |
| 1522 | project_root=self.context.project_root, |
| 1523 | messages=messages, |
| 1524 | ) |
| 1525 | resume_suffix = _pending_item_resume_suffix( |
| 1526 | dod, |
| 1527 | next_pending=next_pending, |
| 1528 | missing_artifact=missing_artifact, |
| 1529 | project_root=self.context.project_root, |
| 1530 | messages=messages, |
| 1531 | ) |
| 1532 | target_label = f"`{resume_target.name or str(resume_target)}`" if resume_target else "" |
| 1533 | |
| 1534 | if fix.get("kind") == "missing_target": |
| 1535 | prefix = f"That `{tool_call.name}` call did not provide a valid `file_path`." |
| 1536 | if target_label: |
| 1537 | prefix += f" Stay on {target_label}." |
| 1538 | self.context.queue_steering_message( |
| 1539 | prefix |
| 1540 | + resume_suffix |
| 1541 | + " Resend one concrete " |
| 1542 | + _invalid_mutation_call_shape(tool_call.name) |
| 1543 | + " now instead of another working note, reread, or empty response." |
| 1544 | ) |
| 1545 | return |
| 1546 | |
| 1547 | invalid_fields = ", ".join(f"`{field}`" for field in fix["invalid_fields"]) |
| 1548 | prefix = f"That `{tool_call.name}` call omitted the real text payload." |
| 1549 | if invalid_fields: |
| 1550 | prefix += f" {invalid_fields} are summary fields, not valid mutation inputs." |
| 1551 | if target_label: |
| 1552 | prefix += f" Stay on {target_label}." |
| 1553 | self.context.queue_steering_message( |
| 1554 | prefix |
| 1555 | + resume_suffix |
| 1556 | + " Resend one concrete " |
| 1557 | + _invalid_mutation_call_shape(tool_call.name) |
| 1558 | + " now instead of rereading more files." |
| 1559 | ) |
| 1560 | |
| 1561 | def _record_blocked_invalid_mutation_attempt( |
| 1562 | self, |
| 1563 | tool_call: ToolCall, |
| 1564 | error: str, |
| 1565 | ) -> None: |
| 1566 | """Seed recovery state from blocked malformed mutations for later retry guidance.""" |
| 1567 | |
| 1568 | recovery_context = self.context.recovery_context |
| 1569 | if recovery_context is None or not recovery_context.is_related_failure( |
| 1570 | tool_call.name, |
| 1571 | tool_call.arguments, |
| 1572 | error, |
| 1573 | ): |
| 1574 | recovery_context = RecoveryContext( |
| 1575 | original_tool=tool_call.name, |
| 1576 | original_args=tool_call.arguments, |
| 1577 | max_retries=self.context.config.max_recovery_attempts, |
| 1578 | ) |
| 1579 | self.context.recovery_context = recovery_context |
| 1580 | |
| 1581 | if not recovery_context.is_similar_attempt( |
| 1582 | tool_call.name, |
| 1583 | tool_call.arguments, |
| 1584 | ): |
| 1585 | recovery_context.add_attempt( |
| 1586 | tool_call.name, |
| 1587 | tool_call.arguments, |
| 1588 | error, |
| 1589 | ) |
| 1590 | |
| 1591 | async def _record_successful_execution( |
| 1592 | self, |
| 1593 | *, |
| 1594 | tool_call: ToolCall, |
| 1595 | outcome, |
| 1596 | dod: DefinitionOfDone, |
| 1597 | emit: EventSink, |
| 1598 | summary: TurnSummary, |
| 1599 | ) -> str | None: |
| 1600 | """Update DoD bookkeeping after a successful tool execution.""" |
| 1601 | |
| 1602 | is_mutating = is_state_mutating_tool_call(tool_call) |
| 1603 | previously_verified = dod.last_verification_result == "passed" |
| 1604 | record_successful_tool_call(dod, tool_call) |
| 1605 | if tool_call.name == "TodoWrite" and outcome.registry_result is not None: |
| 1606 | repair = extract_active_repair_context(self.context.session.messages) |
| 1607 | if repair is None or not _repair_context_is_html_quality(repair): |
| 1608 | new_todos = outcome.registry_result.metadata.get("new_todos", []) |
| 1609 | if isinstance(new_todos, list): |
| 1610 | sync_todos_to_definition_of_done( |
| 1611 | dod, |
| 1612 | new_todos, |
| 1613 | project_root=self.context.project_root, |
| 1614 | ) |
| 1615 | self._refresh_todowrite_outcome_summary(outcome=outcome, dod=dod) |
| 1616 | self._queue_todowrite_resume_nudge(dod=dod) |
| 1617 | else: |
| 1618 | pending_before = list(dod.pending_items) |
| 1619 | if advance_todos_from_tool_call(dod, tool_call): |
| 1620 | reconcile_aggregate_completion_steps( |
| 1621 | dod, |
| 1622 | project_root=self.context.project_root, |
| 1623 | ) |
| 1624 | self._queue_next_pending_todo_nudge( |
| 1625 | tool_call=tool_call, |
| 1626 | pending_before=pending_before, |
| 1627 | dod=dod, |
| 1628 | ) |
| 1629 | self._queue_bookkeeping_resume_nudge( |
| 1630 | tool_call=tool_call, |
| 1631 | dod=dod, |
| 1632 | ) |
| 1633 | self._queue_missing_artifact_progress_nudge( |
| 1634 | tool_call=tool_call, |
| 1635 | dod=dod, |
| 1636 | ) |
| 1637 | self._queue_planned_artifact_handoff_nudge( |
| 1638 | tool_call=tool_call, |
| 1639 | dod=dod, |
| 1640 | ) |
| 1641 | if previously_verified and is_mutating: |
| 1642 | _mark_verification_stale( |
| 1643 | context=self.context, |
| 1644 | summary=summary, |
| 1645 | dod=dod, |
| 1646 | tool_call=tool_call, |
| 1647 | ) |
| 1648 | elif is_mutating and _should_plan_verification_for_tool_call( |
| 1649 | dod, |
| 1650 | tool_call=tool_call, |
| 1651 | project_root=self.context.project_root, |
| 1652 | ): |
| 1653 | _mark_verification_planned( |
| 1654 | context=self.context, |
| 1655 | summary=summary, |
| 1656 | dod=dod, |
| 1657 | tool_call=tool_call, |
| 1658 | ) |
| 1659 | self.dod_store.save(dod) |
| 1660 | recovery_context = self.context.recovery_context |
| 1661 | if recovery_context is not None: |
| 1662 | recovery_context.note_success(tool_call.name, tool_call.arguments) |
| 1663 | if recovery_context.should_clear_after_success( |
| 1664 | tool_call.name, |
| 1665 | tool_call.arguments, |
| 1666 | ): |
| 1667 | self.context.recovery_context = None |
| 1668 | return None |
| 1669 | |
| 1670 | def _refresh_todowrite_outcome_summary( |
| 1671 | self, |
| 1672 | *, |
| 1673 | outcome, |
| 1674 | dod: DefinitionOfDone, |
| 1675 | ) -> None: |
| 1676 | """Rewrite TodoWrite summaries from reconciled DoD state instead of raw snapshots.""" |
| 1677 | |
| 1678 | pending = [ |
| 1679 | item |
| 1680 | for item in effective_pending_todo_items( |
| 1681 | dod, |
| 1682 | project_root=self.context.project_root, |
| 1683 | ) |
| 1684 | if item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 1685 | ] |
| 1686 | completed = [ |
| 1687 | item |
| 1688 | for item in dod.completed_items |
| 1689 | if item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 1690 | ] |
| 1691 | summary_parts = [ |
| 1692 | "updated todo list", |
| 1693 | f"{len(completed)} completed", |
| 1694 | f"{len(pending)} pending", |
| 1695 | ] |
| 1696 | if pending: |
| 1697 | summary_parts.append(f"next pending: {pending[0]}") |
| 1698 | if not pending and ( |
| 1699 | dod.verification_commands |
| 1700 | or all_planned_artifact_outputs_exist( |
| 1701 | dod, |
| 1702 | project_root=self.context.project_root, |
| 1703 | ) |
| 1704 | ): |
| 1705 | summary_parts.append("final response should be provided next for Loader verification") |
| 1706 | |
| 1707 | result = "; ".join(summary_parts) |
| 1708 | content = f"Observation [TodoWrite]: Result: {result}" |
| 1709 | outcome.event_content = content |
| 1710 | outcome.result_output = content |
| 1711 | outcome.message.content = content |
| 1712 | if outcome.message.tool_results: |
| 1713 | outcome.message.tool_results[0].content = content |
| 1714 | |
| 1715 | def _queue_next_pending_todo_nudge( |
| 1716 | self, |
| 1717 | *, |
| 1718 | tool_call: ToolCall, |
| 1719 | pending_before: list[str], |
| 1720 | dod: DefinitionOfDone, |
| 1721 | ) -> None: |
| 1722 | if is_state_mutating_tool_call(tool_call): |
| 1723 | return |
| 1724 | if tool_call.name not in {"read", "glob", "grep", "bash"}: |
| 1725 | return |
| 1726 | if tool_call.name == "bash": |
| 1727 | command = str(tool_call.arguments.get("command", "")).lower() |
| 1728 | if not any( |
| 1729 | token in command |
| 1730 | for token in ( |
| 1731 | "ls ", |
| 1732 | " ls", |
| 1733 | "find ", |
| 1734 | "grep ", |
| 1735 | "rg ", |
| 1736 | "cat ", |
| 1737 | "sed ", |
| 1738 | "head ", |
| 1739 | "tail ", |
| 1740 | ) |
| 1741 | ): |
| 1742 | return |
| 1743 | |
| 1744 | completed_label = next( |
| 1745 | ( |
| 1746 | item |
| 1747 | for item in pending_before |
| 1748 | if item not in dod.pending_items |
| 1749 | and item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 1750 | ), |
| 1751 | None, |
| 1752 | ) |
| 1753 | missing_artifact = _next_missing_planned_artifact( |
| 1754 | dod, |
| 1755 | project_root=self.context.project_root, |
| 1756 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1757 | ) |
| 1758 | next_pending = preferred_pending_todo_item( |
| 1759 | dod, |
| 1760 | project_root=self.context.project_root, |
| 1761 | missing_artifact=missing_artifact, |
| 1762 | ) |
| 1763 | missing_artifact = _prefer_missing_artifact_for_pending_item( |
| 1764 | dod, |
| 1765 | missing_artifact=missing_artifact, |
| 1766 | next_pending=next_pending, |
| 1767 | project_root=self.context.project_root, |
| 1768 | ) |
| 1769 | has_substantive_file_artifact_progress = ( |
| 1770 | _has_confirmed_substantive_file_artifact_progress( |
| 1771 | dod, |
| 1772 | project_root=self.context.project_root, |
| 1773 | ) |
| 1774 | ) |
| 1775 | if not completed_label or not next_pending or next_pending == completed_label: |
| 1776 | return |
| 1777 | if _should_prioritize_missing_artifact( |
| 1778 | dod=dod, |
| 1779 | next_pending=next_pending, |
| 1780 | missing_artifact=missing_artifact, |
| 1781 | project_root=self.context.project_root, |
| 1782 | ): |
| 1783 | if not has_substantive_file_artifact_progress: |
| 1784 | compact_handoff = _compact_missing_artifact_handoff( |
| 1785 | missing_artifact, |
| 1786 | project_root=self.context.project_root, |
| 1787 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1788 | encourage_initial_version=True, |
| 1789 | ) |
| 1790 | if compact_handoff: |
| 1791 | self.context.queue_steering_message( |
| 1792 | f"Confirmed progress: `{completed_label}` is now satisfied by the successful " |
| 1793 | f"`{tool_call.name}` result. {compact_handoff}" |
| 1794 | " Do not reread reference material or spend the next turn on bookkeeping." |
| 1795 | ) |
| 1796 | return |
| 1797 | self.context.queue_steering_message( |
| 1798 | f"Confirmed progress: `{completed_label}` is now satisfied by the successful " |
| 1799 | f"`{tool_call.name}` result. One declared output artifact is still missing." |
| 1800 | + _missing_artifact_resume_suffix( |
| 1801 | missing_artifact, |
| 1802 | project_root=self.context.project_root, |
| 1803 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1804 | ) |
| 1805 | + " Do not switch into review or consistency-check mode until the missing artifact exists." |
| 1806 | ) |
| 1807 | return |
| 1808 | |
| 1809 | mutation_suffix = "" |
| 1810 | if _todo_is_mutation_step(next_pending): |
| 1811 | mutation_suffix = _pending_item_resume_suffix( |
| 1812 | dod, |
| 1813 | next_pending=next_pending, |
| 1814 | missing_artifact=missing_artifact, |
| 1815 | project_root=self.context.project_root, |
| 1816 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1817 | ) |
| 1818 | if not mutation_suffix: |
| 1819 | mutation_suffix = ( |
| 1820 | " You already have enough evidence for that step, so stop gathering " |
| 1821 | "more reference material and perform the change now." |
| 1822 | ) |
| 1823 | |
| 1824 | self.context.queue_steering_message( |
| 1825 | f"Confirmed progress: `{completed_label}` is now satisfied by the successful " |
| 1826 | f"`{tool_call.name}` result. Continue with the next pending item: " |
| 1827 | f"`{next_pending}` instead of rereading the same evidence.{mutation_suffix}" |
| 1828 | ) |
| 1829 | |
| 1830 | def _queue_planned_artifact_handoff_nudge( |
| 1831 | self, |
| 1832 | *, |
| 1833 | tool_call: ToolCall, |
| 1834 | dod: DefinitionOfDone, |
| 1835 | ) -> None: |
| 1836 | if not is_state_mutating_tool_call(tool_call): |
| 1837 | return |
| 1838 | if not all_planned_artifact_outputs_exist(dod, project_root=self.context.project_root): |
| 1839 | return |
| 1840 | repair = extract_active_repair_context(self.context.session.messages) |
| 1841 | if repair is not None and _repair_context_is_html_quality(repair): |
| 1842 | return |
| 1843 | |
| 1844 | next_pending = preferred_pending_todo_item( |
| 1845 | dod, |
| 1846 | project_root=self.context.project_root, |
| 1847 | ) |
| 1848 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 1849 | dod, |
| 1850 | project_root=self.context.project_root, |
| 1851 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 1852 | supplement_existing=True, |
| 1853 | ) |
| 1854 | |
| 1855 | if next_pending and _todo_is_consistency_review_step(next_pending): |
| 1856 | verification_suffix = ( |
| 1857 | " Finish with a final response once no specific mismatch remains so Loader can verify." |
| 1858 | if verification_commands |
| 1859 | else " Avoid another full reread unless one specific inconsistency is still unknown." |
| 1860 | ) |
| 1861 | self.context.queue_steering_message( |
| 1862 | "All explicitly planned artifacts now exist on disk. " |
| 1863 | f"Continue with the next pending item: `{next_pending}`. " |
| 1864 | "Use the files already on disk as the source of truth instead of restarting " |
| 1865 | "discovery or inventing alternate filenames." |
| 1866 | + verification_suffix |
| 1867 | ) |
| 1868 | return |
| 1869 | |
| 1870 | if verification_commands: |
| 1871 | self.context.queue_steering_message( |
| 1872 | "All explicitly planned artifacts now exist on disk. " |
| 1873 | "Do not expand the artifact set or restart discovery unless a specific gap is " |
| 1874 | "still known. Finish with a final response now so Loader can verify the files " |
| 1875 | "that already exist." |
| 1876 | ) |
| 1877 | |
| 1878 | def _queue_missing_artifact_progress_nudge( |
| 1879 | self, |
| 1880 | *, |
| 1881 | tool_call: ToolCall, |
| 1882 | dod: DefinitionOfDone, |
| 1883 | ) -> None: |
| 1884 | if not is_state_mutating_tool_call(tool_call): |
| 1885 | return |
| 1886 | missing_artifact = _next_missing_planned_artifact( |
| 1887 | dod, |
| 1888 | project_root=self.context.project_root, |
| 1889 | ) |
| 1890 | if missing_artifact is None: |
| 1891 | return |
| 1892 | next_pending = preferred_pending_todo_item( |
| 1893 | dod, |
| 1894 | project_root=self.context.project_root, |
| 1895 | missing_artifact=missing_artifact, |
| 1896 | ) |
| 1897 | missing_artifact = _prefer_missing_artifact_for_pending_item( |
| 1898 | dod, |
| 1899 | missing_artifact=missing_artifact, |
| 1900 | next_pending=next_pending, |
| 1901 | project_root=self.context.project_root, |
| 1902 | ) |
| 1903 | |
| 1904 | current_label = _current_mutation_label(tool_call) |
| 1905 | has_substantive_file_artifact_progress = ( |
| 1906 | _has_confirmed_substantive_file_artifact_progress( |
| 1907 | dod, |
| 1908 | project_root=self.context.project_root, |
| 1909 | ) |
| 1910 | ) |
| 1911 | resume_target = _preferred_resume_target_path( |
| 1912 | dod, |
| 1913 | next_pending=next_pending, |
| 1914 | missing_artifact=missing_artifact, |
| 1915 | project_root=self.context.project_root, |
| 1916 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1917 | ) |
| 1918 | resume_suffix = _pending_item_resume_suffix( |
| 1919 | dod, |
| 1920 | next_pending=next_pending, |
| 1921 | missing_artifact=missing_artifact, |
| 1922 | project_root=self.context.project_root, |
| 1923 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1924 | ) |
| 1925 | if ( |
| 1926 | not has_substantive_file_artifact_progress |
| 1927 | and _is_pure_directory_creation_tool_call(tool_call) |
| 1928 | ): |
| 1929 | if ( |
| 1930 | next_pending |
| 1931 | and _todo_is_mutation_step(next_pending) |
| 1932 | and resume_target is not None |
| 1933 | and resume_target.suffix |
| 1934 | ): |
| 1935 | compact_resume = _compact_missing_artifact_handoff( |
| 1936 | (resume_target, False), |
| 1937 | project_root=self.context.project_root, |
| 1938 | messages=list(getattr(self.context.session, "messages", []) or []), |
| 1939 | encourage_initial_version=True, |
| 1940 | ) |
| 1941 | if compact_resume: |
| 1942 | self.context.queue_steering_message( |
| 1943 | "Directory setup is complete. " |
| 1944 | + compact_resume |
| 1945 | + " Do not reread older reference files before that mutation." |
| 1946 | ) |
| 1947 | return |
| 1948 | self.context.queue_steering_message( |
| 1949 | f"Directory setup is complete. Continue with the next pending item: `{next_pending}`." |
| 1950 | + resume_suffix |
| 1951 | + " Do not reread older reference files before that mutation." |
| 1952 | ) |
| 1953 | return |
| 1954 | use_persistent_handoff = _should_use_persistent_missing_artifact_handoff( |
| 1955 | dod, |
| 1956 | project_root=self.context.project_root, |
| 1957 | ) |
| 1958 | session_messages = list(getattr(self.context.session, "messages", []) or []) |
| 1959 | if ( |
| 1960 | use_persistent_handoff |
| 1961 | and _recent_recovery_prompt(session_messages) |
| 1962 | and not _should_preserve_first_child_handoff_after_recovery( |
| 1963 | tool_call=tool_call, |
| 1964 | resume_target=resume_target, |
| 1965 | dod=dod, |
| 1966 | project_root=self.context.project_root, |
| 1967 | ) |
| 1968 | ): |
| 1969 | use_persistent_handoff = False |
| 1970 | queue_message = ( |
| 1971 | self.context.queue_steering_message |
| 1972 | if use_persistent_handoff |
| 1973 | else self.context.queue_ephemeral_steering_message |
| 1974 | ) |
| 1975 | if resume_target is not None and resume_target.suffix: |
| 1976 | compact_resume = _compact_missing_artifact_handoff( |
| 1977 | (resume_target, False), |
| 1978 | project_root=self.context.project_root, |
| 1979 | messages=session_messages, |
| 1980 | encourage_initial_version=False, |
| 1981 | ) |
| 1982 | if compact_resume: |
| 1983 | queue_message( |
| 1984 | f"Confirmed progress: {current_label} is now recorded. " |
| 1985 | + compact_resume |
| 1986 | + " Do not reread reference material or spend the next turn on bookkeeping." |
| 1987 | ) |
| 1988 | return |
| 1989 | todo_refresh = _todo_refresh_guidance( |
| 1990 | dod, |
| 1991 | project_root=self.context.project_root, |
| 1992 | ) |
| 1993 | if not has_substantive_file_artifact_progress: |
| 1994 | compact_handoff = _compact_missing_artifact_handoff( |
| 1995 | missing_artifact, |
| 1996 | project_root=self.context.project_root, |
| 1997 | messages=session_messages, |
| 1998 | encourage_initial_version=False, |
| 1999 | ) |
| 2000 | if compact_handoff: |
| 2001 | queue_message( |
| 2002 | f"Confirmed progress: {current_label} is now recorded. " |
| 2003 | + compact_handoff |
| 2004 | + " Do not reread reference material or spend the next turn on bookkeeping." |
| 2005 | ) |
| 2006 | return |
| 2007 | if _late_stage_missing_artifact_build( |
| 2008 | dod, |
| 2009 | project_root=self.context.project_root, |
| 2010 | ): |
| 2011 | queue_message( |
| 2012 | f"Confirmed progress: {current_label} is now recorded." |
| 2013 | + resume_suffix |
| 2014 | + " No TodoWrite, no verification, no rereads until that artifact exists." |
| 2015 | ) |
| 2016 | return |
| 2017 | queue_message( |
| 2018 | f"Confirmed progress: {current_label} is now recorded." |
| 2019 | " One declared output artifact is still missing." |
| 2020 | + resume_suffix |
| 2021 | + todo_refresh |
| 2022 | + " Do not move to verification, final confirmation, or TodoWrite-only " |
| 2023 | "bookkeeping until that artifact exists." |
| 2024 | + " Do not spend another turn on working notes or rediscovery alone." |
| 2025 | ) |
| 2026 | |
| 2027 | def _queue_todowrite_resume_nudge( |
| 2028 | self, |
| 2029 | *, |
| 2030 | dod: DefinitionOfDone, |
| 2031 | ) -> None: |
| 2032 | repair = extract_active_repair_context(self.context.session.messages) |
| 2033 | if repair is not None and _repair_context_is_html_quality(repair): |
| 2034 | target = repair.artifact_path or ( |
| 2035 | repair.allowed_paths[0] if repair.allowed_paths else "" |
| 2036 | ) |
| 2037 | if not target: |
| 2038 | return |
| 2039 | repair_issue = _quality_repair_issue_for_target(repair, target) |
| 2040 | issue_line = ( |
| 2041 | f"- {repair_issue}\n" |
| 2042 | if repair_issue |
| 2043 | else f"- Improve `{target}` until it satisfies the active content-quality verifier.\n" |
| 2044 | ) |
| 2045 | force_write = recent_repair_mutation_context_failed( |
| 2046 | self.context.session.messages, |
| 2047 | target, |
| 2048 | ) |
| 2049 | if force_write: |
| 2050 | immediate_step = ( |
| 2051 | f"- Immediate next step: rewrite `{target}` with one `write` call.\n" |
| 2052 | "- Recent `edit`/`patch` attempts for this file failed against stale " |
| 2053 | "or malformed context. Use `write(file_path=..., content=...)` with " |
| 2054 | "a complete valid HTML document, and do not call `read`, `edit`, " |
| 2055 | "`patch`, or TodoWrite again first." |
| 2056 | ) |
| 2057 | else: |
| 2058 | immediate_step = ( |
| 2059 | f"- Immediate next step: edit `{target}`.\n" |
| 2060 | "- Continue with one concrete `edit`, `patch`, or `write` call that " |
| 2061 | "actually changes the current generated file." |
| 2062 | ) |
| 2063 | self.context.set_workflow_mode("execute") |
| 2064 | self.context.queue_steering_message( |
| 2065 | "Todo tracking is updated, but verification still has an active " |
| 2066 | "HTML content-quality repair. TodoWrite cannot satisfy that verifier " |
| 2067 | "or close the repair by itself. Do not mark the task complete and do " |
| 2068 | "not finish yet.\n\n" |
| 2069 | "Repair focus:\n" |
| 2070 | f"{issue_line}" |
| 2071 | f"{immediate_step}" |
| 2072 | ) |
| 2073 | return |
| 2074 | |
| 2075 | session_messages = list(getattr(self.context.session, "messages", []) or []) |
| 2076 | missing_artifact = _next_missing_planned_artifact( |
| 2077 | dod, |
| 2078 | project_root=self.context.project_root, |
| 2079 | messages=session_messages, |
| 2080 | ) |
| 2081 | next_pending = preferred_pending_todo_item( |
| 2082 | dod, |
| 2083 | project_root=self.context.project_root, |
| 2084 | missing_artifact=missing_artifact, |
| 2085 | ) |
| 2086 | missing_artifact = _prefer_missing_artifact_for_pending_item( |
| 2087 | dod, |
| 2088 | missing_artifact=missing_artifact, |
| 2089 | next_pending=next_pending, |
| 2090 | project_root=self.context.project_root, |
| 2091 | ) |
| 2092 | outputs_exist = all_planned_artifact_outputs_exist( |
| 2093 | dod, |
| 2094 | project_root=self.context.project_root, |
| 2095 | ) |
| 2096 | if missing_artifact is None: |
| 2097 | if next_pending and _todo_is_mutation_step(next_pending) and not outputs_exist: |
| 2098 | pending_target = infer_pending_todo_output_target( |
| 2099 | dod, |
| 2100 | next_pending, |
| 2101 | project_root=self.context.project_root, |
| 2102 | ) |
| 2103 | if pending_target is not None: |
| 2104 | concrete_message = ( |
| 2105 | "Todo tracking is updated. Continue with the next pending item: " |
| 2106 | f"`{next_pending}`. Resume by creating `{pending_target.name}` now. " |
| 2107 | f"Prefer one `write` call for `{pending_target}` instead of more rereads. " |
| 2108 | ) |
| 2109 | if not pending_target.parent.exists(): |
| 2110 | concrete_message += ( |
| 2111 | "The `write` tool can create that file's parent directories " |
| 2112 | "automatically, so do the write in one step instead of stopping " |
| 2113 | "for a separate mkdir. " |
| 2114 | ) |
| 2115 | concrete_message += ( |
| 2116 | "Use the current output files as the source of truth, and do not " |
| 2117 | "reopen reference materials unless one specific fact required for " |
| 2118 | "that step is still unknown. Make your next response the concrete " |
| 2119 | "mutation tool call itself, not another bookkeeping-only turn. " |
| 2120 | "Perform the mutation now instead of spending another turn on " |
| 2121 | "planning, rereads, or verification." |
| 2122 | ) |
| 2123 | self.context.queue_steering_message(concrete_message) |
| 2124 | return |
| 2125 | self.context.queue_steering_message( |
| 2126 | "Todo tracking is updated. Continue with the next pending item: " |
| 2127 | f"`{next_pending}`. Use the current output files as the source of " |
| 2128 | "truth, and do not reopen reference materials unless one specific " |
| 2129 | "fact required for that step is still unknown. Perform the mutation " |
| 2130 | "now instead of spending another turn on planning, rereads, or " |
| 2131 | "verification." |
| 2132 | ) |
| 2133 | return |
| 2134 | |
| 2135 | if ( |
| 2136 | next_pending |
| 2137 | and _todo_is_consistency_review_step(next_pending) |
| 2138 | and not outputs_exist |
| 2139 | ): |
| 2140 | self.context.queue_ephemeral_steering_message( |
| 2141 | "Todo tracking is updated. Continue with the next pending item: " |
| 2142 | f"`{next_pending}`. Use the current output files as the source of " |
| 2143 | "truth, and do not reopen reference materials unless one specific " |
| 2144 | "mismatch is still unknown." |
| 2145 | ) |
| 2146 | return |
| 2147 | |
| 2148 | if not outputs_exist: |
| 2149 | return |
| 2150 | |
| 2151 | verification_commands = dod.verification_commands or derive_verification_commands( |
| 2152 | dod, |
| 2153 | project_root=self.context.project_root, |
| 2154 | task_statement=getattr(self.context.session, "current_task", "") or "", |
| 2155 | supplement_existing=True, |
| 2156 | ) |
| 2157 | if next_pending and _todo_is_consistency_review_step(next_pending): |
| 2158 | verification_suffix = ( |
| 2159 | " Finish with a final response once no specific mismatch remains so Loader can verify." |
| 2160 | if verification_commands |
| 2161 | else " Finish the targeted consistency pass without reopening reference materials." |
| 2162 | ) |
| 2163 | self.context.queue_ephemeral_steering_message( |
| 2164 | "Todo tracking is updated. All explicitly planned artifacts now exist on disk. " |
| 2165 | f"Continue with the next pending item: `{next_pending}`. " |
| 2166 | "Use the current output files as the source of truth, and do not restart " |
| 2167 | "early discovery or reopen reference materials." |
| 2168 | + verification_suffix |
| 2169 | ) |
| 2170 | return |
| 2171 | |
| 2172 | if verification_commands: |
| 2173 | self.context.set_workflow_mode("verify") |
| 2174 | self.context.queue_steering_message( |
| 2175 | "Todo tracking is updated. All explicitly planned artifacts now exist on disk. " |
| 2176 | "Finish with a final response now so Loader can run verification automatically. " |
| 2177 | "Use the current output files as the source of truth, and do not restart discovery, " |
| 2178 | "reopen reference materials, run more ad hoc audit commands, or spend another turn " |
| 2179 | "on TodoWrite alone." |
| 2180 | ) |
| 2181 | return |
| 2182 | |
| 2183 | verification_suffix = ( |
| 2184 | " " + _verification_handoff_instruction() |
| 2185 | if verification_commands |
| 2186 | else " Finish the task using the files already on disk." |
| 2187 | ) |
| 2188 | self.context.queue_ephemeral_steering_message( |
| 2189 | "Todo tracking is updated. All explicitly planned artifacts now exist on disk. " |
| 2190 | "Do not restart discovery, reopen reference materials, or spend another turn " |
| 2191 | "on TodoWrite alone. Repair or verify the current files instead of expanding the artifact set." |
| 2192 | + verification_suffix |
| 2193 | ) |
| 2194 | return |
| 2195 | |
| 2196 | todo_refresh = _todo_refresh_guidance( |
| 2197 | dod, |
| 2198 | project_root=self.context.project_root, |
| 2199 | ) |
| 2200 | resume_target = _preferred_resume_target_path( |
| 2201 | dod, |
| 2202 | next_pending=next_pending, |
| 2203 | missing_artifact=missing_artifact, |
| 2204 | project_root=self.context.project_root, |
| 2205 | messages=session_messages, |
| 2206 | ) |
| 2207 | pending_target = _preferred_pending_target_path( |
| 2208 | dod, |
| 2209 | next_pending=next_pending, |
| 2210 | project_root=self.context.project_root, |
| 2211 | ) |
| 2212 | next_pending_suffix = _pending_item_handoff_prefix( |
| 2213 | next_pending, |
| 2214 | pending_target=pending_target, |
| 2215 | resume_target=resume_target, |
| 2216 | ) |
| 2217 | if resume_target is not None and resume_target.suffix: |
| 2218 | compact_resume = _compact_missing_artifact_handoff( |
| 2219 | (resume_target, False), |
| 2220 | project_root=self.context.project_root, |
| 2221 | messages=session_messages, |
| 2222 | encourage_initial_version=False, |
| 2223 | ) |
| 2224 | if compact_resume: |
| 2225 | self.context.queue_steering_message( |
| 2226 | "Todo tracking is updated. " |
| 2227 | + compact_resume |
| 2228 | + " Do not spend the next turn on TodoWrite alone, bookkeeping notes, " |
| 2229 | "verification, or final confirmation until that artifact exists." |
| 2230 | ) |
| 2231 | return |
| 2232 | self.context.queue_steering_message( |
| 2233 | "Todo tracking is updated. A declared output artifact is still missing." |
| 2234 | + next_pending_suffix |
| 2235 | + _missing_artifact_resume_suffix( |
| 2236 | missing_artifact, |
| 2237 | project_root=self.context.project_root, |
| 2238 | messages=session_messages, |
| 2239 | ) |
| 2240 | + todo_refresh |
| 2241 | + " Do not spend the next turn on TodoWrite alone, bookkeeping notes, " |
| 2242 | "verification, or final confirmation until that artifact exists." |
| 2243 | ) |
| 2244 | |
| 2245 | def _queue_bookkeeping_resume_nudge( |
| 2246 | self, |
| 2247 | *, |
| 2248 | tool_call: ToolCall, |
| 2249 | dod: DefinitionOfDone, |
| 2250 | ) -> None: |
| 2251 | if tool_call.name not in _BOOKKEEPING_NOTE_TOOL_NAMES: |
| 2252 | return |
| 2253 | |
| 2254 | session_messages = list(getattr(self.context.session, "messages", []) or []) |
| 2255 | missing_artifact = _next_missing_planned_artifact( |
| 2256 | dod, |
| 2257 | project_root=self.context.project_root, |
| 2258 | messages=session_messages, |
| 2259 | ) |
| 2260 | if missing_artifact is None: |
| 2261 | return |
| 2262 | |
| 2263 | next_pending = preferred_pending_todo_item( |
| 2264 | dod, |
| 2265 | project_root=self.context.project_root, |
| 2266 | missing_artifact=missing_artifact, |
| 2267 | ) |
| 2268 | missing_artifact = _prefer_missing_artifact_for_pending_item( |
| 2269 | dod, |
| 2270 | missing_artifact=missing_artifact, |
| 2271 | next_pending=next_pending, |
| 2272 | project_root=self.context.project_root, |
| 2273 | ) |
| 2274 | todo_refresh = _todo_refresh_guidance( |
| 2275 | dod, |
| 2276 | project_root=self.context.project_root, |
| 2277 | ) |
| 2278 | if ( |
| 2279 | next_pending |
| 2280 | and not _todo_is_mutation_step(next_pending) |
| 2281 | and not _todo_is_consistency_review_step(next_pending) |
| 2282 | and not _should_prioritize_missing_artifact( |
| 2283 | dod=dod, |
| 2284 | next_pending=next_pending, |
| 2285 | missing_artifact=( |
| 2286 | missing_artifact |
| 2287 | if _has_confirmed_artifact_progress( |
| 2288 | dod, |
| 2289 | project_root=self.context.project_root, |
| 2290 | ) |
| 2291 | else None |
| 2292 | ), |
| 2293 | project_root=self.context.project_root, |
| 2294 | ) |
| 2295 | ): |
| 2296 | self.context.queue_ephemeral_steering_message( |
| 2297 | "Bookkeeping note is recorded. Continue with the next pending item: " |
| 2298 | f"`{next_pending}`. Make your next response one concrete evidence-gathering " |
| 2299 | "tool call that advances that step, not another bookkeeping-only turn." |
| 2300 | + todo_refresh |
| 2301 | + " Do not jump ahead to later artifact creation, verification, or final " |
| 2302 | "confirmation until that step is satisfied." |
| 2303 | ) |
| 2304 | return |
| 2305 | |
| 2306 | self.context.queue_ephemeral_steering_message( |
| 2307 | "Bookkeeping note is recorded. A declared output artifact is still missing." |
| 2308 | + _missing_artifact_resume_suffix( |
| 2309 | missing_artifact, |
| 2310 | project_root=self.context.project_root, |
| 2311 | messages=session_messages, |
| 2312 | ) |
| 2313 | + todo_refresh |
| 2314 | + " Do not spend the next turn on additional notes, rediscovery, " |
| 2315 | "verification, or final confirmation until that artifact exists." |
| 2316 | ) |
| 2317 | |
| 2318 | |
| 2319 | def _todo_is_consistency_review_step(item: str) -> bool: |
| 2320 | text = item.lower() |
| 2321 | return any(hint in text for hint in _CONSISTENCY_REVIEW_HINTS) |
| 2322 | |
| 2323 | |
| 2324 | def _planned_output_roots( |
| 2325 | dod: DefinitionOfDone, |
| 2326 | *, |
| 2327 | project_root: Path, |
| 2328 | ) -> tuple[str, ...]: |
| 2329 | planned_roots: list[str] = [] |
| 2330 | seen_roots: set[str] = set() |
| 2331 | for target, expect_directory in collect_planned_artifact_targets( |
| 2332 | dod, |
| 2333 | project_root=project_root, |
| 2334 | ): |
| 2335 | root = str(target if expect_directory else target.parent) |
| 2336 | if root in seen_roots: |
| 2337 | continue |
| 2338 | seen_roots.add(root) |
| 2339 | planned_roots.append(root) |
| 2340 | return tuple(planned_roots) |
| 2341 | |
| 2342 | |
| 2343 | def _extract_observation_paths(tool_call: ToolCall) -> list[str]: |
| 2344 | arguments = tool_call.arguments |
| 2345 | if tool_call.name == "read": |
| 2346 | file_path = str(arguments.get("file_path", "")).strip() |
| 2347 | return [file_path] if file_path else [] |
| 2348 | |
| 2349 | if tool_call.name in {"glob", "grep"}: |
| 2350 | candidates: list[str] = [] |
| 2351 | search_path = str(arguments.get("path", "")).strip() |
| 2352 | if search_path: |
| 2353 | anchored_path = _derive_search_anchor( |
| 2354 | search_path, |
| 2355 | str(arguments.get("pattern", "")).strip(), |
| 2356 | ) |
| 2357 | candidates.append(anchored_path or search_path) |
| 2358 | pattern = str(arguments.get("pattern", "")).strip() |
| 2359 | if not search_path and pattern.startswith(("/", "~")): |
| 2360 | candidates.append(str(Path(pattern).expanduser().parent)) |
| 2361 | return candidates |
| 2362 | |
| 2363 | command = str(arguments.get("command", "")).strip() |
| 2364 | if not _is_read_only_bash(command): |
| 2365 | return [] |
| 2366 | return _extract_bash_paths(command) |
| 2367 | |
| 2368 | |
| 2369 | def _derive_search_anchor(search_path: str, pattern: str) -> str: |
| 2370 | base = str(Path(search_path).expanduser()) |
| 2371 | normalized_pattern = pattern.strip() |
| 2372 | if not normalized_pattern: |
| 2373 | return base |
| 2374 | if normalized_pattern.startswith(("~", "/")): |
| 2375 | pattern_path = Path(normalized_pattern).expanduser() |
| 2376 | try: |
| 2377 | return str(pattern_path.parent.resolve(strict=False)) |
| 2378 | except Exception: |
| 2379 | return str(pattern_path.parent) |
| 2380 | if "/" in normalized_pattern: |
| 2381 | prefix = normalized_pattern.rsplit("/", 1)[0].strip() |
| 2382 | if prefix and prefix not in {".", ".."}: |
| 2383 | joined = Path(base).joinpath(prefix).expanduser() |
| 2384 | try: |
| 2385 | return str(joined.resolve(strict=False)) |
| 2386 | except Exception: |
| 2387 | return str(joined) |
| 2388 | return base |
| 2389 | |
| 2390 | |
| 2391 | def _is_read_only_bash(command: str) -> bool: |
| 2392 | normalized = " ".join(command.split()) |
| 2393 | if not normalized: |
| 2394 | return False |
| 2395 | if extract_shell_text_rewrite_target(normalized) is not None: |
| 2396 | return False |
| 2397 | if any(fragment in normalized for fragment in _MUTATING_BASH_FRAGMENTS): |
| 2398 | return False |
| 2399 | try: |
| 2400 | argv = shlex.split(normalized) |
| 2401 | except ValueError: |
| 2402 | return False |
| 2403 | if not argv: |
| 2404 | return False |
| 2405 | return argv[0] in _READ_ONLY_BASH_PREFIXES |
| 2406 | |
| 2407 | |
| 2408 | def _extract_bash_paths(command: str) -> list[str]: |
| 2409 | try: |
| 2410 | argv = shlex.split(command) |
| 2411 | except ValueError: |
| 2412 | return [] |
| 2413 | if not argv: |
| 2414 | return [] |
| 2415 | |
| 2416 | command_name = argv[0] |
| 2417 | if command_name == "pwd": |
| 2418 | return [str(Path.cwd())] |
| 2419 | |
| 2420 | paths: list[str] = [] |
| 2421 | for arg in argv[1:]: |
| 2422 | if arg.startswith("-"): |
| 2423 | continue |
| 2424 | if command_name in {"ls", "stat", "cat", "head", "tail"}: |
| 2425 | paths.append(arg) |
| 2426 | continue |
| 2427 | if command_name in {"find", "rg", "grep"}: |
| 2428 | paths.append(str(Path.cwd()) if arg in {".", "./"} else arg) |
| 2429 | break |
| 2430 | return paths |
| 2431 | |
| 2432 | |
| 2433 | def _should_prioritize_missing_artifact( |
| 2434 | *, |
| 2435 | dod: DefinitionOfDone, |
| 2436 | next_pending: str | None, |
| 2437 | missing_artifact: tuple[Path, bool] | None, |
| 2438 | project_root: Path, |
| 2439 | ) -> bool: |
| 2440 | if missing_artifact is None: |
| 2441 | return False |
| 2442 | if not next_pending: |
| 2443 | return True |
| 2444 | if _pending_todo_conflicts_with_missing_artifact( |
| 2445 | dod, |
| 2446 | item=next_pending, |
| 2447 | missing_artifact=missing_artifact, |
| 2448 | project_root=project_root, |
| 2449 | ): |
| 2450 | return True |
| 2451 | if _todo_is_consistency_review_step(next_pending): |
| 2452 | return True |
| 2453 | return not _todo_is_mutation_step(next_pending) |
| 2454 | |
| 2455 | |
| 2456 | def _pending_todo_conflicts_with_missing_artifact( |
| 2457 | dod: DefinitionOfDone, |
| 2458 | *, |
| 2459 | item: str, |
| 2460 | missing_artifact: tuple[Path, bool], |
| 2461 | project_root: Path, |
| 2462 | ) -> bool: |
| 2463 | text = item.strip().lower() |
| 2464 | if not text or item in _TODO_NUDGE_EXCLUDED_ITEMS: |
| 2465 | return False |
| 2466 | |
| 2467 | target, expect_directory = missing_artifact |
| 2468 | inferred_target = infer_pending_todo_output_target( |
| 2469 | dod, |
| 2470 | item, |
| 2471 | project_root=project_root, |
| 2472 | ) |
| 2473 | if inferred_target is None: |
| 2474 | return not expect_directory and _todo_is_mutation_step(item) |
| 2475 | |
| 2476 | inferred_target = inferred_target.resolve(strict=False) |
| 2477 | target = target.resolve(strict=False) |
| 2478 | if expect_directory: |
| 2479 | return target != inferred_target and target not in inferred_target.parents |
| 2480 | return inferred_target != target |
| 2481 | |
| 2482 | |
| 2483 | def _next_missing_planned_artifact( |
| 2484 | dod: DefinitionOfDone, |
| 2485 | *, |
| 2486 | project_root: Path, |
| 2487 | messages: list[Any] | None = None, |
| 2488 | ) -> tuple[Path, bool] | None: |
| 2489 | for target, expect_directory in collect_planned_artifact_targets( |
| 2490 | dod, |
| 2491 | project_root=project_root, |
| 2492 | max_paths=12, |
| 2493 | ): |
| 2494 | if not planned_artifact_target_satisfied( |
| 2495 | dod, |
| 2496 | target=target, |
| 2497 | expect_directory=expect_directory, |
| 2498 | project_root=project_root, |
| 2499 | ): |
| 2500 | return target, expect_directory |
| 2501 | if not _pending_mutation_work_requires_inferred_children( |
| 2502 | dod, |
| 2503 | project_root=project_root, |
| 2504 | ): |
| 2505 | return None |
| 2506 | for target, expect_directory in collect_planned_artifact_targets( |
| 2507 | dod, |
| 2508 | project_root=project_root, |
| 2509 | max_paths=12, |
| 2510 | ): |
| 2511 | if not expect_directory or not target.is_dir(): |
| 2512 | continue |
| 2513 | next_output_file, _ = infer_next_output_file( |
| 2514 | target=target, |
| 2515 | project_root=project_root, |
| 2516 | messages=list(messages or []), |
| 2517 | ) |
| 2518 | if next_output_file is not None and not next_output_file.exists(): |
| 2519 | return next_output_file, False |
| 2520 | return None |
| 2521 | |
| 2522 | |
| 2523 | def _pending_mutation_work_requires_inferred_children( |
| 2524 | dod: DefinitionOfDone, |
| 2525 | *, |
| 2526 | project_root: Path, |
| 2527 | ) -> bool: |
| 2528 | actionable_pending = [ |
| 2529 | item |
| 2530 | for item in effective_pending_todo_items(dod, project_root=project_root) |
| 2531 | if item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 2532 | ] |
| 2533 | return any(_todo_is_mutation_step(item) for item in actionable_pending) |
| 2534 | |
| 2535 | |
| 2536 | def _prefer_missing_artifact_for_pending_item( |
| 2537 | dod: DefinitionOfDone, |
| 2538 | *, |
| 2539 | missing_artifact: tuple[Path, bool] | None, |
| 2540 | next_pending: str | None, |
| 2541 | project_root: Path, |
| 2542 | ) -> tuple[Path, bool] | None: |
| 2543 | if missing_artifact is None or not next_pending: |
| 2544 | return missing_artifact |
| 2545 | |
| 2546 | inferred_target = infer_pending_todo_output_target( |
| 2547 | dod, |
| 2548 | next_pending, |
| 2549 | project_root=project_root, |
| 2550 | ) |
| 2551 | if inferred_target is None or inferred_target.exists(): |
| 2552 | return missing_artifact |
| 2553 | |
| 2554 | normalized_target = inferred_target.expanduser().resolve(strict=False) |
| 2555 | for planned_target, expect_directory in collect_planned_artifact_targets( |
| 2556 | dod, |
| 2557 | project_root=project_root, |
| 2558 | max_paths=12, |
| 2559 | ): |
| 2560 | normalized_planned = planned_target.expanduser().resolve(strict=False) |
| 2561 | if expect_directory and normalized_planned == normalized_target: |
| 2562 | return normalized_target, True |
| 2563 | if expect_directory: |
| 2564 | try: |
| 2565 | normalized_target.relative_to(normalized_planned) |
| 2566 | except ValueError: |
| 2567 | continue |
| 2568 | return normalized_target, False |
| 2569 | if normalized_planned == normalized_target: |
| 2570 | return normalized_target, False |
| 2571 | return missing_artifact |
| 2572 | |
| 2573 | |
| 2574 | def _late_stage_missing_artifact_build( |
| 2575 | dod: DefinitionOfDone, |
| 2576 | *, |
| 2577 | project_root: Path, |
| 2578 | ) -> bool: |
| 2579 | completed = 0 |
| 2580 | missing = 0 |
| 2581 | for target, expect_directory in collect_planned_artifact_targets( |
| 2582 | dod, |
| 2583 | project_root=project_root, |
| 2584 | max_paths=12, |
| 2585 | ): |
| 2586 | if planned_artifact_target_satisfied( |
| 2587 | dod, |
| 2588 | target=target, |
| 2589 | expect_directory=expect_directory, |
| 2590 | project_root=project_root, |
| 2591 | ): |
| 2592 | completed += 1 |
| 2593 | else: |
| 2594 | missing += 1 |
| 2595 | return completed >= 7 and missing > 0 |
| 2596 | |
| 2597 | |
| 2598 | def _has_confirmed_artifact_progress( |
| 2599 | dod: DefinitionOfDone, |
| 2600 | *, |
| 2601 | project_root: Path, |
| 2602 | ) -> bool: |
| 2603 | for target, expect_directory in collect_planned_artifact_targets( |
| 2604 | dod, |
| 2605 | project_root=project_root, |
| 2606 | max_paths=12, |
| 2607 | ): |
| 2608 | if planned_artifact_target_satisfied( |
| 2609 | dod, |
| 2610 | target=target, |
| 2611 | expect_directory=expect_directory, |
| 2612 | project_root=project_root, |
| 2613 | ): |
| 2614 | return True |
| 2615 | return bool(dod.touched_files) |
| 2616 | |
| 2617 | |
| 2618 | def _has_confirmed_file_artifact_progress( |
| 2619 | dod: DefinitionOfDone, |
| 2620 | *, |
| 2621 | project_root: Path, |
| 2622 | ) -> bool: |
| 2623 | return _confirmed_file_artifact_count(dod, project_root=project_root) > 0 |
| 2624 | |
| 2625 | |
| 2626 | def _has_confirmed_substantive_file_artifact_progress( |
| 2627 | dod: DefinitionOfDone, |
| 2628 | *, |
| 2629 | project_root: Path, |
| 2630 | ) -> bool: |
| 2631 | return _confirmed_substantive_file_artifact_count( |
| 2632 | dod, |
| 2633 | project_root=project_root, |
| 2634 | ) > 0 |
| 2635 | |
| 2636 | |
| 2637 | def _last_touched_file_path(dod: DefinitionOfDone) -> Path | None: |
| 2638 | for raw_path in reversed(dod.touched_files): |
| 2639 | path_text = str(raw_path or "").strip() |
| 2640 | if not path_text: |
| 2641 | continue |
| 2642 | candidate = Path(path_text).expanduser().resolve(strict=False) |
| 2643 | if candidate.suffix: |
| 2644 | return candidate |
| 2645 | return None |
| 2646 | |
| 2647 | |
| 2648 | def _confirmed_file_artifact_count( |
| 2649 | dod: DefinitionOfDone, |
| 2650 | *, |
| 2651 | project_root: Path, |
| 2652 | ) -> int: |
| 2653 | count = 0 |
| 2654 | for target, expect_directory in collect_planned_artifact_targets( |
| 2655 | dod, |
| 2656 | project_root=project_root, |
| 2657 | max_paths=12, |
| 2658 | ): |
| 2659 | if expect_directory: |
| 2660 | continue |
| 2661 | if planned_artifact_target_satisfied( |
| 2662 | dod, |
| 2663 | target=target, |
| 2664 | expect_directory=False, |
| 2665 | project_root=project_root, |
| 2666 | ): |
| 2667 | count += 1 |
| 2668 | if count: |
| 2669 | return count |
| 2670 | return sum( |
| 2671 | 1 |
| 2672 | for path in dod.touched_files |
| 2673 | if str(path).strip() |
| 2674 | and Path(path).expanduser().resolve(strict=False).suffix |
| 2675 | ) |
| 2676 | |
| 2677 | |
| 2678 | def _confirmed_substantive_file_artifact_count( |
| 2679 | dod: DefinitionOfDone, |
| 2680 | *, |
| 2681 | project_root: Path, |
| 2682 | ) -> int: |
| 2683 | count = 0 |
| 2684 | for target, expect_directory in collect_planned_artifact_targets( |
| 2685 | dod, |
| 2686 | project_root=project_root, |
| 2687 | max_paths=12, |
| 2688 | ): |
| 2689 | if expect_directory or _is_summary_artifact_path(target): |
| 2690 | continue |
| 2691 | if planned_artifact_target_satisfied( |
| 2692 | dod, |
| 2693 | target=target, |
| 2694 | expect_directory=False, |
| 2695 | project_root=project_root, |
| 2696 | ): |
| 2697 | count += 1 |
| 2698 | if count: |
| 2699 | return count |
| 2700 | return sum( |
| 2701 | 1 |
| 2702 | for path in dod.touched_files |
| 2703 | if str(path).strip() |
| 2704 | and Path(path).expanduser().resolve(strict=False).suffix |
| 2705 | and not _is_summary_artifact_path(path) |
| 2706 | ) |
| 2707 | |
| 2708 | |
| 2709 | def _should_use_persistent_missing_artifact_handoff( |
| 2710 | dod: DefinitionOfDone, |
| 2711 | *, |
| 2712 | project_root: Path, |
| 2713 | ) -> bool: |
| 2714 | return _confirmed_substantive_file_artifact_count( |
| 2715 | dod, |
| 2716 | project_root=project_root, |
| 2717 | ) == 0 |
| 2718 | |
| 2719 | |
| 2720 | def _should_preserve_first_child_handoff_after_recovery( |
| 2721 | *, |
| 2722 | tool_call: ToolCall, |
| 2723 | resume_target: Path | None, |
| 2724 | dod: DefinitionOfDone, |
| 2725 | project_root: Path, |
| 2726 | ) -> bool: |
| 2727 | if resume_target is None or not resume_target.suffix or _is_summary_artifact_path(resume_target): |
| 2728 | return False |
| 2729 | raw_target = str(tool_call.arguments.get("file_path", "")).strip() |
| 2730 | if not raw_target: |
| 2731 | return False |
| 2732 | written_target = Path(raw_target).expanduser().resolve(strict=False) |
| 2733 | if not _is_summary_artifact_path(written_target): |
| 2734 | return False |
| 2735 | return _confirmed_substantive_file_artifact_count( |
| 2736 | dod, |
| 2737 | project_root=project_root, |
| 2738 | ) == 0 |
| 2739 | |
| 2740 | |
| 2741 | def _is_summary_artifact_path(path: str | Path) -> bool: |
| 2742 | return Path(path).name.lower() in _SUMMARY_ARTIFACT_NAMES |
| 2743 | |
| 2744 | |
| 2745 | def _next_missing_planned_file_within_directory( |
| 2746 | dod: DefinitionOfDone, |
| 2747 | *, |
| 2748 | target: Path, |
| 2749 | project_root: Path, |
| 2750 | ) -> Path | None: |
| 2751 | normalized_target = target.expanduser().resolve(strict=False) |
| 2752 | if normalized_target.suffix: |
| 2753 | return None |
| 2754 | |
| 2755 | for planned_target, expect_directory in collect_planned_artifact_targets( |
| 2756 | dod, |
| 2757 | project_root=project_root, |
| 2758 | max_paths=12, |
| 2759 | ): |
| 2760 | if expect_directory: |
| 2761 | continue |
| 2762 | normalized_planned = planned_target.expanduser().resolve(strict=False) |
| 2763 | try: |
| 2764 | normalized_planned.relative_to(normalized_target) |
| 2765 | except ValueError: |
| 2766 | continue |
| 2767 | if planned_artifact_target_satisfied( |
| 2768 | dod, |
| 2769 | target=normalized_planned, |
| 2770 | expect_directory=False, |
| 2771 | project_root=project_root, |
| 2772 | ): |
| 2773 | continue |
| 2774 | return normalized_planned |
| 2775 | return None |
| 2776 | |
| 2777 | |
| 2778 | def _missing_artifact_resume_suffix( |
| 2779 | missing_artifact: tuple[Path, bool] | None, |
| 2780 | *, |
| 2781 | project_root: Path, |
| 2782 | messages: list[Any] | None = None, |
| 2783 | ) -> str: |
| 2784 | if missing_artifact is None: |
| 2785 | return "" |
| 2786 | |
| 2787 | target, expect_directory = missing_artifact |
| 2788 | return _resume_suffix_for_target( |
| 2789 | target, |
| 2790 | expect_directory=expect_directory, |
| 2791 | project_root=project_root, |
| 2792 | messages=messages, |
| 2793 | ) |
| 2794 | |
| 2795 | |
| 2796 | def _pending_item_resume_suffix( |
| 2797 | dod: DefinitionOfDone, |
| 2798 | *, |
| 2799 | next_pending: str | None, |
| 2800 | missing_artifact: tuple[Path, bool] | None, |
| 2801 | project_root: Path, |
| 2802 | messages: list[Any] | None = None, |
| 2803 | ) -> str: |
| 2804 | if next_pending: |
| 2805 | pending_target = infer_pending_todo_output_target( |
| 2806 | dod, |
| 2807 | next_pending, |
| 2808 | project_root=project_root, |
| 2809 | ) |
| 2810 | if pending_target is not None and not pending_target.exists(): |
| 2811 | normalized_target = pending_target.expanduser().resolve(strict=False) |
| 2812 | return _resume_suffix_for_target( |
| 2813 | normalized_target, |
| 2814 | expect_directory=not bool(normalized_target.suffix), |
| 2815 | project_root=project_root, |
| 2816 | messages=messages, |
| 2817 | allow_inferred_child=False, |
| 2818 | ) |
| 2819 | if missing_artifact is not None and missing_artifact[1]: |
| 2820 | next_planned_file = _next_missing_planned_file_within_directory( |
| 2821 | dod, |
| 2822 | target=missing_artifact[0], |
| 2823 | project_root=project_root, |
| 2824 | ) |
| 2825 | if next_planned_file is not None: |
| 2826 | parent_label = missing_artifact[0].name or str(missing_artifact[0]) |
| 2827 | return ( |
| 2828 | f" Resume by creating `{next_planned_file.name}` now." |
| 2829 | f" It is the next missing declared output under `{parent_label}/`." |
| 2830 | f" Prefer one `write` call for `{next_planned_file}` instead of more rereads." |
| 2831 | " Make your next response the concrete mutation tool call itself, not another" |
| 2832 | " bookkeeping-only turn." |
| 2833 | ) |
| 2834 | return _missing_artifact_resume_suffix( |
| 2835 | missing_artifact, |
| 2836 | project_root=project_root, |
| 2837 | messages=messages, |
| 2838 | ) |
| 2839 | |
| 2840 | |
| 2841 | def _pending_item_handoff_prefix( |
| 2842 | next_pending: str | None, |
| 2843 | *, |
| 2844 | pending_target: Path | None, |
| 2845 | resume_target: Path | None, |
| 2846 | ) -> str: |
| 2847 | if not next_pending: |
| 2848 | return "" |
| 2849 | if ( |
| 2850 | pending_target is None |
| 2851 | and resume_target is not None |
| 2852 | and resume_target.suffix |
| 2853 | and todo_describes_aggregate_mutation(next_pending) |
| 2854 | and not todo_describes_broad_setup_step(next_pending) |
| 2855 | ): |
| 2856 | return f" Continue with the next concrete output: `{resume_target.name}`." |
| 2857 | return f" Continue with the next pending item: `{next_pending}`." |
| 2858 | |
| 2859 | |
| 2860 | def _preferred_pending_target_path( |
| 2861 | dod: DefinitionOfDone, |
| 2862 | *, |
| 2863 | next_pending: str | None, |
| 2864 | project_root: Path, |
| 2865 | ) -> Path | None: |
| 2866 | if not next_pending: |
| 2867 | return None |
| 2868 | pending_target = infer_pending_todo_output_target( |
| 2869 | dod, |
| 2870 | next_pending, |
| 2871 | project_root=project_root, |
| 2872 | ) |
| 2873 | if pending_target is None: |
| 2874 | return None |
| 2875 | return pending_target.expanduser().resolve(strict=False) |
| 2876 | |
| 2877 | |
| 2878 | def _preferred_resume_target_path( |
| 2879 | dod: DefinitionOfDone, |
| 2880 | *, |
| 2881 | next_pending: str | None, |
| 2882 | missing_artifact: tuple[Path, bool] | None, |
| 2883 | project_root: Path, |
| 2884 | messages: list[Any] | None = None, |
| 2885 | ) -> Path | None: |
| 2886 | if next_pending: |
| 2887 | pending_target = infer_pending_todo_output_target( |
| 2888 | dod, |
| 2889 | next_pending, |
| 2890 | project_root=project_root, |
| 2891 | ) |
| 2892 | if pending_target is not None and not pending_target.exists(): |
| 2893 | return pending_target.expanduser().resolve(strict=False) |
| 2894 | |
| 2895 | if missing_artifact is None: |
| 2896 | return None |
| 2897 | |
| 2898 | target, expect_directory = missing_artifact |
| 2899 | normalized_target = target.expanduser().resolve(strict=False) |
| 2900 | if not expect_directory: |
| 2901 | return normalized_target |
| 2902 | |
| 2903 | next_planned_file = _next_missing_planned_file_within_directory( |
| 2904 | dod, |
| 2905 | target=normalized_target, |
| 2906 | project_root=project_root, |
| 2907 | ) |
| 2908 | if next_planned_file is not None: |
| 2909 | return next_planned_file.expanduser().resolve(strict=False) |
| 2910 | |
| 2911 | next_output_file, _ = infer_next_output_file( |
| 2912 | target=normalized_target, |
| 2913 | project_root=project_root, |
| 2914 | messages=list(messages or []), |
| 2915 | ) |
| 2916 | if next_output_file is not None: |
| 2917 | return next_output_file.expanduser().resolve(strict=False) |
| 2918 | return normalized_target |
| 2919 | |
| 2920 | |
| 2921 | def _invalid_mutation_call_shape(tool_name: str) -> str: |
| 2922 | if tool_name == "write": |
| 2923 | return "`write(file_path=..., content=...)`" |
| 2924 | if tool_name == "edit": |
| 2925 | return "`edit(file_path=..., old_string=..., new_string=...)`" |
| 2926 | if tool_name == "patch": |
| 2927 | return "`patch(file_path=..., patch='...')` or `patch(..., hunks=[...])`" |
| 2928 | return f"`{tool_name}(...)`" |
| 2929 | |
| 2930 | |
| 2931 | def _extract_blocked_html_target_list(event_content: str, marker: str) -> list[str]: |
| 2932 | if marker not in event_content: |
| 2933 | return [] |
| 2934 | tail = event_content.split(marker, 1)[1].strip() |
| 2935 | target_text = tail.split(". ", 1)[0].strip() |
| 2936 | if not target_text: |
| 2937 | return [] |
| 2938 | return [item.strip() for item in target_text.split(",") if item.strip()] |
| 2939 | |
| 2940 | |
| 2941 | def _count_recent_blocked_html_asset_events( |
| 2942 | messages: list[Any], |
| 2943 | missing_assets: list[str], |
| 2944 | ) -> int: |
| 2945 | if not missing_assets: |
| 2946 | return 0 |
| 2947 | |
| 2948 | count = 0 |
| 2949 | for message in reversed(messages[-12:]): |
| 2950 | content = str(getattr(message, "content", "") or "") |
| 2951 | if "HTML local asset references do not exist" not in content: |
| 2952 | continue |
| 2953 | if any(asset and asset in content for asset in missing_assets): |
| 2954 | count += 1 |
| 2955 | return count |
| 2956 | |
| 2957 | |
| 2958 | def _resume_suffix_for_target( |
| 2959 | target: Path, |
| 2960 | *, |
| 2961 | expect_directory: bool, |
| 2962 | project_root: Path, |
| 2963 | messages: list[Any] | None = None, |
| 2964 | allow_inferred_child: bool = True, |
| 2965 | ) -> str: |
| 2966 | label = target.name or str(target) |
| 2967 | display_target = display_runtime_path(target) |
| 2968 | if expect_directory and not label.endswith("/"): |
| 2969 | label += "/" |
| 2970 | if expect_directory: |
| 2971 | if allow_inferred_child: |
| 2972 | next_output_file, next_output_source = infer_next_output_file( |
| 2973 | target=target, |
| 2974 | project_root=project_root, |
| 2975 | messages=list(messages or []), |
| 2976 | ) |
| 2977 | if next_output_file is not None: |
| 2978 | guidance_origin = ( |
| 2979 | f"It is the next missing declared output under `{label}`." |
| 2980 | if next_output_source == "declared" |
| 2981 | else ( |
| 2982 | "It mirrors the observed filename pattern from another " |
| 2983 | f"`{label}` directory you already inspected." |
| 2984 | ) |
| 2985 | ) |
| 2986 | guidance = ( |
| 2987 | f" Resume by creating `{next_output_file.name}` now. {guidance_origin} " |
| 2988 | f"Prefer one `write` call for " |
| 2989 | f"`{display_runtime_path(next_output_file)}` instead of more rereads." |
| 2990 | ) |
| 2991 | if not next_output_file.parent.exists(): |
| 2992 | guidance += ( |
| 2993 | " The `write` tool can create that file's parent directories automatically," |
| 2994 | " so do the write in one step instead of stopping for a separate mkdir." |
| 2995 | ) |
| 2996 | guidance += ( |
| 2997 | " Make your next response the concrete mutation tool call itself, not another" |
| 2998 | " bookkeeping-only turn." |
| 2999 | ) |
| 3000 | return guidance |
| 3001 | if target.is_dir(): |
| 3002 | return ( |
| 3003 | f" Resume by creating the next output file under `{label}` now. Prefer one " |
| 3004 | f"concrete `write` call for a file inside `{display_target}` instead of more rereads." |
| 3005 | " Make your next response the concrete mutation tool call itself, not another" |
| 3006 | " bookkeeping-only turn." |
| 3007 | ) |
| 3008 | return ( |
| 3009 | f" Resume by creating `{label}` now. Prefer one concrete directory-creation " |
| 3010 | f"step for `{display_target}` instead of more rereads." |
| 3011 | ) |
| 3012 | guidance = ( |
| 3013 | f" Resume by creating `{label}` now. Prefer one `write` call for `{display_target}` " |
| 3014 | "instead of more rereads." |
| 3015 | ) |
| 3016 | if not target.parent.exists(): |
| 3017 | guidance += ( |
| 3018 | " The `write` tool can create that file's parent directories automatically," |
| 3019 | " so do the write in one step instead of stopping for a separate mkdir." |
| 3020 | ) |
| 3021 | guidance += ( |
| 3022 | " Make your next response the concrete mutation tool call itself, not another" |
| 3023 | " bookkeeping-only turn." |
| 3024 | ) |
| 3025 | return guidance |
| 3026 | |
| 3027 | |
| 3028 | def _compact_missing_artifact_handoff( |
| 3029 | missing_artifact: tuple[Path, bool] | None, |
| 3030 | *, |
| 3031 | project_root: Path, |
| 3032 | messages: list[Any] | None = None, |
| 3033 | encourage_initial_version: bool = False, |
| 3034 | ) -> str: |
| 3035 | """Build a shorter first-mutation handoff once the next output target is known.""" |
| 3036 | |
| 3037 | if missing_artifact is None: |
| 3038 | return "" |
| 3039 | |
| 3040 | target, expect_directory = missing_artifact |
| 3041 | label = target.name or str(target) |
| 3042 | display_target = display_runtime_path(target) |
| 3043 | if expect_directory and not label.endswith("/"): |
| 3044 | label += "/" |
| 3045 | if expect_directory: |
| 3046 | next_output_file, _ = infer_next_output_file( |
| 3047 | target=target, |
| 3048 | project_root=project_root, |
| 3049 | messages=list(messages or []), |
| 3050 | ) |
| 3051 | if next_output_file is None: |
| 3052 | if target.is_dir(): |
| 3053 | return ( |
| 3054 | f"Next step: create the next output file under `{label}`. Prefer one " |
| 3055 | f"concrete `write` call inside `{display_target}` now." |
| 3056 | ) |
| 3057 | return ( |
| 3058 | f"Next step: create `{label}`. Prefer one concrete directory-creation step " |
| 3059 | f"for `{display_target}` now." |
| 3060 | ) |
| 3061 | guidance = ( |
| 3062 | f"Next step: create `{next_output_file.name}`. Prefer one " |
| 3063 | f"`write(file_path=..., content=...)` call for `{display_runtime_path(next_output_file)}` now." |
| 3064 | ) |
| 3065 | if not next_output_file.parent.exists(): |
| 3066 | guidance += ( |
| 3067 | " The `write` tool can create that file's parent directories automatically." |
| 3068 | ) |
| 3069 | if encourage_initial_version: |
| 3070 | guidance += ( |
| 3071 | " Write a compact but real initial version of that file now; you can expand " |
| 3072 | "or refine it in later edits." |
| 3073 | ) |
| 3074 | guidance += " Make your next response the concrete mutation tool call itself." |
| 3075 | return guidance |
| 3076 | |
| 3077 | guidance = ( |
| 3078 | f"Next step: create `{label}`. Prefer one " |
| 3079 | f"`write(file_path=..., content=...)` call for `{display_target}` now." |
| 3080 | ) |
| 3081 | if not target.parent.exists(): |
| 3082 | guidance += ( |
| 3083 | " The `write` tool can create that file's parent directories automatically." |
| 3084 | ) |
| 3085 | if encourage_initial_version: |
| 3086 | guidance += ( |
| 3087 | " Write a compact but real initial version of that file now; you can expand " |
| 3088 | "or refine it in later edits." |
| 3089 | ) |
| 3090 | guidance += " Make your next response the concrete mutation tool call itself." |
| 3091 | return guidance |
| 3092 | |
| 3093 | |
| 3094 | def _todo_refresh_guidance( |
| 3095 | dod: DefinitionOfDone, |
| 3096 | *, |
| 3097 | project_root: Path | None = None, |
| 3098 | ) -> str: |
| 3099 | non_special_pending = [ |
| 3100 | item |
| 3101 | for item in effective_pending_todo_items(dod, project_root=project_root) |
| 3102 | if item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 3103 | ] |
| 3104 | non_special_completed = [ |
| 3105 | item for item in dod.completed_items if item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 3106 | ] |
| 3107 | if len(dod.touched_files) < 2 and (len(non_special_pending) + len(non_special_completed)) < 3: |
| 3108 | return "" |
| 3109 | return ( |
| 3110 | " If the tracked steps no longer match the confirmed progress, refresh `TodoWrite` " |
| 3111 | "in the same response as the next concrete step instead of spending a full turn on " |
| 3112 | "bookkeeping alone." |
| 3113 | ) |
| 3114 | |
| 3115 | |
| 3116 | def _mark_verification_stale( |
| 3117 | *, |
| 3118 | context: RuntimeContext, |
| 3119 | summary: TurnSummary, |
| 3120 | dod: DefinitionOfDone, |
| 3121 | tool_call: ToolCall, |
| 3122 | ) -> None: |
| 3123 | detail = _stale_verification_detail(tool_call) |
| 3124 | stale_attempt = ensure_active_verification_attempt(dod) |
| 3125 | next_attempt = begin_new_verification_attempt( |
| 3126 | dod, |
| 3127 | supersedes_attempt_id=stale_attempt.attempt_id, |
| 3128 | ) |
| 3129 | append_verification_timeline_entry( |
| 3130 | context, |
| 3131 | summary, |
| 3132 | reason_code="verification_stale", |
| 3133 | reason_summary="previous verification became stale after new mutating work", |
| 3134 | evidence_summary=[f"fresh verification required after {detail}"], |
| 3135 | evidence_provenance=_stale_verification_provenance(dod, detail=detail), |
| 3136 | verification_observations=_stale_verification_observations( |
| 3137 | dod, |
| 3138 | detail=detail, |
| 3139 | stale_attempt_id=stale_attempt.attempt_id, |
| 3140 | stale_attempt_number=stale_attempt.attempt_number, |
| 3141 | superseded_by_attempt_id=next_attempt.attempt_id, |
| 3142 | ), |
| 3143 | ) |
| 3144 | dod.last_verification_result = VerificationObservationStatus.STALE.value |
| 3145 | dod.evidence = [] |
| 3146 | while _VERIFY_ITEM in dod.completed_items: |
| 3147 | dod.completed_items.remove(_VERIFY_ITEM) |
| 3148 | if _VERIFY_ITEM not in dod.pending_items: |
| 3149 | dod.pending_items.append(_VERIFY_ITEM) |
| 3150 | |
| 3151 | |
| 3152 | def _todo_is_mutation_step(label: str) -> bool: |
| 3153 | lowered = label.lower() |
| 3154 | return any(token in lowered for token in _MUTATION_TODO_HINTS) |
| 3155 | |
| 3156 | |
| 3157 | def _should_plan_verification_for_tool_call( |
| 3158 | dod: DefinitionOfDone, |
| 3159 | *, |
| 3160 | tool_call: ToolCall, |
| 3161 | project_root: Path, |
| 3162 | ) -> bool: |
| 3163 | actionable_pending = [ |
| 3164 | item |
| 3165 | for item in effective_pending_todo_items( |
| 3166 | dod, |
| 3167 | project_root=project_root, |
| 3168 | ) |
| 3169 | if item not in _TODO_NUDGE_EXCLUDED_ITEMS |
| 3170 | ] |
| 3171 | if any( |
| 3172 | _todo_is_mutation_step(item) or _todo_is_consistency_review_step(item) |
| 3173 | for item in actionable_pending |
| 3174 | ): |
| 3175 | return False |
| 3176 | if tool_call.name in {"write", "edit", "patch"}: |
| 3177 | return True |
| 3178 | if tool_call.name != "bash": |
| 3179 | return False |
| 3180 | if any( |
| 3181 | Path(path).expanduser().resolve(strict=False).suffix |
| 3182 | for path in dod.touched_files |
| 3183 | if str(path).strip() |
| 3184 | ): |
| 3185 | return True |
| 3186 | return any( |
| 3187 | not expect_directory |
| 3188 | and planned_artifact_target_satisfied( |
| 3189 | dod, |
| 3190 | target=target, |
| 3191 | expect_directory=False, |
| 3192 | project_root=project_root, |
| 3193 | ) |
| 3194 | for target, expect_directory in collect_planned_artifact_targets( |
| 3195 | dod, |
| 3196 | project_root=project_root, |
| 3197 | max_paths=12, |
| 3198 | ) |
| 3199 | ) |
| 3200 | |
| 3201 | |
| 3202 | def _mark_verification_planned( |
| 3203 | *, |
| 3204 | context: RuntimeContext, |
| 3205 | summary: TurnSummary, |
| 3206 | dod: DefinitionOfDone, |
| 3207 | tool_call: ToolCall, |
| 3208 | ) -> None: |
| 3209 | if dod.last_verification_result in { |
| 3210 | VerificationObservationStatus.PLANNED.value, |
| 3211 | VerificationObservationStatus.PENDING.value, |
| 3212 | VerificationObservationStatus.STALE.value, |
| 3213 | }: |
| 3214 | return |
| 3215 | if not dod.verification_commands: |
| 3216 | dod.verification_commands = derive_verification_commands( |
| 3217 | dod, |
| 3218 | project_root=context.project_root, |
| 3219 | task_statement=dod.task_statement, |
| 3220 | ) |
| 3221 | commands = [command for command in dod.verification_commands if command] |
| 3222 | if not commands: |
| 3223 | return |
| 3224 | |
| 3225 | attempt = begin_new_verification_attempt(dod) |
| 3226 | detail = _stale_verification_detail(tool_call) |
| 3227 | append_verification_timeline_entry( |
| 3228 | context, |
| 3229 | summary, |
| 3230 | reason_code="verification_planned", |
| 3231 | reason_summary="verification is planned after new mutating work", |
| 3232 | evidence_summary=[f"verification planned for `{command}`" for command in commands[:2]], |
| 3233 | evidence_provenance=[ |
| 3234 | EvidenceProvenance( |
| 3235 | category="verification", |
| 3236 | source="dod.verification_commands", |
| 3237 | summary=f"verification planned for `{command}`", |
| 3238 | status=EvidenceProvenanceStatus.MISSING.value, |
| 3239 | subject=command, |
| 3240 | detail=detail, |
| 3241 | ) |
| 3242 | for command in commands |
| 3243 | ], |
| 3244 | verification_observations=[ |
| 3245 | VerificationObservation( |
| 3246 | status=VerificationObservationStatus.PLANNED.value, |
| 3247 | summary=f"verification planned for `{command}`", |
| 3248 | command=command, |
| 3249 | kind="runtime", |
| 3250 | detail=detail, |
| 3251 | attempt_id=attempt.attempt_id, |
| 3252 | attempt_number=attempt.attempt_number, |
| 3253 | ) |
| 3254 | for command in commands |
| 3255 | ], |
| 3256 | ) |
| 3257 | dod.last_verification_result = VerificationObservationStatus.PLANNED.value |
| 3258 | while _VERIFY_ITEM in dod.completed_items: |
| 3259 | dod.completed_items.remove(_VERIFY_ITEM) |
| 3260 | if _VERIFY_ITEM not in dod.pending_items: |
| 3261 | dod.pending_items.append(_VERIFY_ITEM) |
| 3262 | |
| 3263 | |
| 3264 | def _stale_verification_observations( |
| 3265 | dod: DefinitionOfDone, |
| 3266 | *, |
| 3267 | detail: str, |
| 3268 | stale_attempt_id: str, |
| 3269 | stale_attempt_number: int, |
| 3270 | superseded_by_attempt_id: str, |
| 3271 | ) -> list[VerificationObservation]: |
| 3272 | return [ |
| 3273 | VerificationObservation( |
| 3274 | status=VerificationObservationStatus.STALE.value, |
| 3275 | summary=f"verification became stale for `{command}` after new mutating work", |
| 3276 | command=command, |
| 3277 | kind="runtime", |
| 3278 | detail=detail, |
| 3279 | attempt_id=stale_attempt_id, |
| 3280 | attempt_number=stale_attempt_number, |
| 3281 | supersedes_attempt_id=superseded_by_attempt_id, |
| 3282 | ) |
| 3283 | for command in _stale_verification_commands(dod) |
| 3284 | ] |
| 3285 | |
| 3286 | |
| 3287 | def _stale_verification_provenance( |
| 3288 | dod: DefinitionOfDone, |
| 3289 | *, |
| 3290 | detail: str, |
| 3291 | ) -> list[EvidenceProvenance]: |
| 3292 | return [ |
| 3293 | EvidenceProvenance( |
| 3294 | category="verification", |
| 3295 | source="tool_execution", |
| 3296 | summary=f"fresh verification required for `{command}` after new mutating work", |
| 3297 | status=EvidenceProvenanceStatus.MISSING.value, |
| 3298 | subject=command, |
| 3299 | detail=detail, |
| 3300 | ) |
| 3301 | for command in _stale_verification_commands(dod) |
| 3302 | ] |
| 3303 | |
| 3304 | |
| 3305 | def _stale_verification_commands(dod: DefinitionOfDone) -> list[str]: |
| 3306 | commands = [command for command in dod.verification_commands if command] |
| 3307 | if commands: |
| 3308 | return commands |
| 3309 | observed = [evidence.command for evidence in dod.evidence if evidence.command] |
| 3310 | if observed: |
| 3311 | return observed |
| 3312 | return ["verification"] |
| 3313 | |
| 3314 | |
| 3315 | def _stale_verification_detail(tool_call: ToolCall) -> str: |
| 3316 | if tool_call.name in {"write", "edit", "patch"}: |
| 3317 | file_path = str(tool_call.arguments.get("file_path", "")).strip() |
| 3318 | if file_path: |
| 3319 | return f"{tool_call.name} changed {file_path}" |
| 3320 | if tool_call.name == "bash": |
| 3321 | command = str(tool_call.arguments.get("command", "")).strip() |
| 3322 | if command: |
| 3323 | return f"bash ran `{command}`" |
| 3324 | return f"{tool_call.name} changed the workspace" |
| 3325 | |
| 3326 | |
| 3327 | def _current_mutation_label(tool_call: ToolCall) -> str: |
| 3328 | if tool_call.name in {"write", "edit", "patch"}: |
| 3329 | file_path = str(tool_call.arguments.get("file_path", "")).strip() |
| 3330 | if file_path: |
| 3331 | return f"`{Path(file_path).name or file_path}`" |
| 3332 | if tool_call.name == "bash": |
| 3333 | command = str(tool_call.arguments.get("command", "")).strip() |
| 3334 | if command: |
| 3335 | return f"`{command}`" |
| 3336 | return f"the successful `{tool_call.name}` result" |
| 3337 | |
| 3338 | |
| 3339 | def _is_pure_directory_creation_tool_call(tool_call: ToolCall) -> bool: |
| 3340 | if tool_call.name != "bash": |
| 3341 | return False |
| 3342 | command = str(tool_call.arguments.get("command", "")).strip() |
| 3343 | if not command or any( |
| 3344 | operator in command for operator in ("&&", "||", ";", "|", "$(", ">", "<") |
| 3345 | ): |
| 3346 | return False |
| 3347 | try: |
| 3348 | parts = shlex.split(command) |
| 3349 | except ValueError: |
| 3350 | return False |
| 3351 | return bool(parts) and parts[0] == "mkdir" |
| 3352 | |
| 3353 | |
| 3354 | def _recent_recovery_prompt(messages: list[Any]) -> bool: |
| 3355 | for message in reversed(messages[-4:]): |
| 3356 | role = getattr(message, "role", None) |
| 3357 | if getattr(role, "value", role) != "user": |
| 3358 | continue |
| 3359 | content = getattr(message, "content", "") |
| 3360 | if not isinstance(content, str): |
| 3361 | continue |
| 3362 | if content.startswith("[EMPTY ASSISTANT RESPONSE]"): |
| 3363 | return True |
| 3364 | if content.startswith("[CONTINUE CURRENT STEP]"): |
| 3365 | return True |
| 3366 | return False |
| 3367 | |
| 3368 | |
| 3369 | def _is_recoverable_guidance_block(event_content: str) -> bool: |
| 3370 | """Return whether a blocked observation should steer without tripping fatal error limits.""" |
| 3371 | |
| 3372 | normalized = str(event_content or "") |
| 3373 | return ( |
| 3374 | "[Blocked - completed artifact set scope:" in normalized |
| 3375 | or "[Blocked - post-build audit loop:" in normalized |
| 3376 | ) |
| 3377 | |
| 3378 | |
| 3379 | def _recent_edit_string_mismatch_target(recovery_context: RecoveryContext | None) -> str: |
| 3380 | """Return the active edit target when recovery is from an old_string miss.""" |
| 3381 | |
| 3382 | if recovery_context is None: |
| 3383 | return "" |
| 3384 | for attempt in reversed(recovery_context.attempts): |
| 3385 | if attempt.tool_name != "edit": |
| 3386 | continue |
| 3387 | if "old_string not found" not in str(attempt.error or "").lower(): |
| 3388 | continue |
| 3389 | target = str( |
| 3390 | attempt.arguments.get("file_path") |
| 3391 | or attempt.arguments.get("path") |
| 3392 | or "" |
| 3393 | ).strip() |
| 3394 | if target: |
| 3395 | return target |
| 3396 | return "" |
| 3397 | |
| 3398 | |
| 3399 | def _repair_context_is_html_quality(repair: Any) -> bool: |
| 3400 | """Return whether the active repair context is for generated HTML quality.""" |
| 3401 | |
| 3402 | return any( |
| 3403 | _repair_line_is_html_quality(line) |
| 3404 | for line in getattr(repair, "repair_lines", ()) or () |
| 3405 | ) |
| 3406 | |
| 3407 | |
| 3408 | def _repair_line_is_html_quality(line: str) -> bool: |
| 3409 | lowered = str(line or "").lower() |
| 3410 | return ( |
| 3411 | "thin content" in lowered |
| 3412 | or "insufficient structured content" in lowered |
| 3413 | or "content-quality" in lowered |
| 3414 | or "content quality" in lowered |
| 3415 | ) |
| 3416 | |
| 3417 | |
| 3418 | def _next_quality_repair_path(repair: Any, *, changed_path: str) -> str: |
| 3419 | """Return the next concrete repair file after a successful quality mutation.""" |
| 3420 | |
| 3421 | try: |
| 3422 | normalized_changed = str(Path(changed_path).expanduser().resolve(strict=False)) |
| 3423 | except (OSError, RuntimeError, ValueError): |
| 3424 | normalized_changed = str(Path(changed_path).expanduser()) |
| 3425 | |
| 3426 | normalized_paths: list[str] = [] |
| 3427 | for raw_path in getattr(repair, "allowed_paths", ()) or (): |
| 3428 | try: |
| 3429 | normalized = str(Path(raw_path).expanduser().resolve(strict=False)) |
| 3430 | except (OSError, RuntimeError, ValueError): |
| 3431 | normalized = str(Path(raw_path).expanduser()) |
| 3432 | if normalized and normalized not in normalized_paths: |
| 3433 | normalized_paths.append(normalized) |
| 3434 | |
| 3435 | if normalized_changed in normalized_paths: |
| 3436 | index = normalized_paths.index(normalized_changed) |
| 3437 | if index + 1 < len(normalized_paths): |
| 3438 | return normalized_paths[index + 1] |
| 3439 | for normalized in normalized_paths: |
| 3440 | if normalized != normalized_changed: |
| 3441 | return normalized |
| 3442 | return "" |
| 3443 | |
| 3444 | |
| 3445 | def _tool_call_targets_path(tool_call: ToolCall, target: str) -> bool: |
| 3446 | if not target: |
| 3447 | return False |
| 3448 | candidate = str( |
| 3449 | tool_call.arguments.get("file_path") |
| 3450 | or tool_call.arguments.get("path") |
| 3451 | or "" |
| 3452 | ).strip() |
| 3453 | if not candidate and tool_call.name == "bash": |
| 3454 | paths = _extract_bash_paths( |
| 3455 | str(tool_call.arguments.get("command") or "").strip(), |
| 3456 | ) |
| 3457 | candidate = paths[0] if paths else "" |
| 3458 | if not candidate: |
| 3459 | return False |
| 3460 | try: |
| 3461 | return Path(candidate).expanduser().resolve(strict=False) == Path( |
| 3462 | target |
| 3463 | ).expanduser().resolve(strict=False) |
| 3464 | except (OSError, RuntimeError, ValueError): |
| 3465 | return candidate == target |
| 3466 | |
| 3467 | |
| 3468 | def _active_repair_focus_preview(repair_lines: list[str], *, max_lines: int = 4) -> str: |
| 3469 | """Compact repair-focus bullets for steering after no-op mutations.""" |
| 3470 | |
| 3471 | preview: list[str] = [] |
| 3472 | for raw_line in repair_lines: |
| 3473 | line = str(raw_line or "").strip() |
| 3474 | if not line.startswith("- "): |
| 3475 | continue |
| 3476 | if line.startswith("- Immediate next step:"): |
| 3477 | continue |
| 3478 | preview.append(line[2:].strip()) |
| 3479 | if len(preview) >= max_lines: |
| 3480 | break |
| 3481 | if not preview: |
| 3482 | return "the active verifier repair focus" |
| 3483 | return "; ".join(preview) |
| 3484 | |
| 3485 | |
| 3486 | def _quality_repair_issue_for_target(repair: Any, target: str) -> str: |
| 3487 | """Return the most relevant content-quality repair line for a target path.""" |
| 3488 | |
| 3489 | target_text = str(target or "").strip() |
| 3490 | try: |
| 3491 | normalized_target = str(Path(target_text).expanduser().resolve(strict=False)) |
| 3492 | except (OSError, RuntimeError, ValueError): |
| 3493 | normalized_target = str(Path(target_text).expanduser()) if target_text else "" |
| 3494 | |
| 3495 | first_quality_line = "" |
| 3496 | for raw_line in getattr(repair, "repair_lines", ()) or (): |
| 3497 | line = str(raw_line or "").strip() |
| 3498 | if not line: |
| 3499 | continue |
| 3500 | clean_line = line[2:].strip() if line.startswith("- ") else line |
| 3501 | if not _repair_line_is_html_quality(clean_line): |
| 3502 | continue |
| 3503 | if not first_quality_line: |
| 3504 | first_quality_line = clean_line |
| 3505 | if target_text and target_text in clean_line: |
| 3506 | return clean_line |
| 3507 | if normalized_target and normalized_target in clean_line: |
| 3508 | return clean_line |
| 3509 | for candidate in re.findall(r"`([^`]+)`", clean_line): |
| 3510 | try: |
| 3511 | normalized_candidate = str( |
| 3512 | Path(candidate).expanduser().resolve(strict=False) |
| 3513 | ) |
| 3514 | except (OSError, RuntimeError, ValueError): |
| 3515 | normalized_candidate = str(Path(candidate).expanduser()) |
| 3516 | if normalized_target and normalized_candidate == normalized_target: |
| 3517 | return clean_line |
| 3518 | |
| 3519 | return first_quality_line |
| 3520 | |
| 3521 | |
| 3522 | def _tool_call_label(tool_call: ToolCall) -> str: |
| 3523 | """Human-readable label for one tool call.""" |
| 3524 | name = tool_call.name |
| 3525 | if name in ("write", "edit", "patch"): |
| 3526 | path = str(tool_call.arguments.get("file_path", "")).strip() |
| 3527 | if path: |
| 3528 | short = Path(path).name |
| 3529 | verb = "Write" if name == "write" else "Edit" |
| 3530 | return f"{verb} {short}" |
| 3531 | if name == "bash": |
| 3532 | cmd = str(tool_call.arguments.get("command", "")).strip() |
| 3533 | if cmd: |
| 3534 | return f"Run {cmd[:40]}" |
| 3535 | if name == "read": |
| 3536 | path = str(tool_call.arguments.get("file_path", "")).strip() |
| 3537 | if path: |
| 3538 | return f"Read {Path(path).name}" |
| 3539 | if name == "glob": |
| 3540 | pattern = str(tool_call.arguments.get("pattern", "")).strip() |
| 3541 | if pattern: |
| 3542 | return f"Search {pattern[:30]}" |
| 3543 | return "" |
| 3544 | |
| 3545 | |
| 3546 | def _batch_planned_labels(tool_calls: list[ToolCall]) -> list[str]: |
| 3547 | """Build labels for all tool calls in a batch (for upfront planning display).""" |
| 3548 | labels = [] |
| 3549 | for tc in tool_calls: |
| 3550 | label = _tool_call_label(tc) |
| 3551 | if label and label not in labels: |
| 3552 | labels.append(label) |
| 3553 | return labels |