""" Fanfic Writer v2.0 - Writing Loop (Phase 6) Core writing pipeline with QC, Attempt cycle, and FORCED handling """ import json import re from pathlib import Path from typing import Dict, Any, Optional, List, Tuple, Callable from dataclasses import dataclass from enum import Enum from .config_manager import ConfigManager from .state_manager import StateManager, Evidence from .prompt_assembly import PromptBuilder from .price_table import PriceTableManager from .atomic_io import atomic_write_text, atomic_write_json, atomic_append_jsonl from .utils import get_timestamp_iso, generate_event_id, sanitize_chapter_filename class QCStatus(Enum): INIT = "INIT" PASS = "PASS" WARNING = "WARNING" REVISE = "REVISE" FORCED = "FORCED" @dataclass class QCResult: """Quality Check result""" score: int status: QCStatus dimensions: Dict[str, int] pros: List[str] cons: List[str] rewrite_plan: str def to_dict(self) -> Dict[str, Any]: return { 'score': self.score, 'status': self.status.value, 'dimensions': self.dimensions, 'pros': self.pros, 'cons': self.cons, 'rewrite_plan': self.rewrite_plan } class WritingLoop: """ Phase 6: Chapter-by-chapter writing with quality control Flow: 6.1 Sanitizer -> 6.2 Outline Confirm -> 6.3 Draft -> 6.4 QC -> 6.5 Content Confirm -> 6.6 State Commit """ def __init__( self, run_dir: Path, model_callable: Callable, config_manager: Optional[ConfigManager] = None, state_manager: Optional[StateManager] = None, prompt_builder: Optional[PromptBuilder] = None ): self.run_dir = Path(run_dir) self.model_callable = model_callable self.config = (config_manager or ConfigManager(run_dir)).load() self.state = state_manager or StateManager(run_dir) self.prompt_builder = prompt_builder self.price_mgr = PriceTableManager(run_dir) # QC config qc_config = self.config.get('qc', {}) self.pass_threshold = qc_config.get('pass_threshold', 85) self.warning_threshold = qc_config.get('warning_threshold', 75) self.max_attempts = self.config.get('generation', {}).get('max_attempts', 3) self.mode = self.config.get('generation', {}).get('mode', 'manual') # Ending state check self.target_words = self.config.get('book', {}).get('target_word_count', 100000) self.max_chapters = 200 # Hard limit def _get_forced_streak(self) -> int: """Get current forced_streak value""" state_path = self.run_dir / "4-state" / "4-writing-state.json" try: with open(state_path, 'r', encoding='utf-8') as f: state = json.load(f) return state.get('forced_streak', 0) except: return 0 # ======================================================================== # 6.1 Sanitizer # ======================================================================== def sanitizer(self, chapter_num: int) -> Dict[str, Any]: """ 6.1: State Sanitizer Reads state panels and extracts invariants for chapter generation """ print(f"[6.1] Sanitizer for Chapter {chapter_num}") # Get invariants invariants = self.state.get_invariants(chapter_num) # Check for open backpatch issues backpatch_path = self.run_dir / "4-state" / "backpatch.jsonl" open_issues = [] if backpatch_path.exists(): with open(backpatch_path, 'r', encoding='utf-8') as f: for line in f: try: issue = json.loads(line.strip()) if issue.get('status') == 'open': open_issues.append(issue) except: pass # Build sanitized context context_parts = ["## 不变量 (Invariants) - 必须延续"] for char_name, fields in invariants.get('characters', {}).items(): context_parts.append(f"\n### {char_name}") for field, value in fields.items(): context_parts.append(f"- {field}: {value}") if open_issues: context_parts.append("\n## Open Backpatch Issues") for issue in open_issues[:3]: # Top 3 context_parts.append(f"- 第{issue.get('chapter')}章: {issue.get('issue', '')}") sanitized_context = "\n".join(context_parts) # Log sanitizer output sanitizer_record = { 'timestamp': get_timestamp_iso(), 'chapter': chapter_num, 'invariants_enforced': list(invariants.get('characters', {}).keys()), 'soft_retcons_applied': [], 'reason': '提取不变量用于第{chapter_num}章生成', 'sanitized_context': sanitized_context[:500] # Truncated } atomic_append_jsonl( self.run_dir / "4-state" / "sanitizer_output.jsonl", sanitizer_record ) print(f"[6.1] Complete: {len(invariants.get('characters', {}))} characters") return { 'invariants': invariants, 'sanitized_context': sanitized_context, 'open_backpatch_issues': open_issues } # ======================================================================== # 6.2 Outline Generation & Confirm # ======================================================================== def generate_chapter_outline( self, chapter_num: int, previous_content: str = "" ) -> str: """6.2: Generate detailed chapter outline""" print(f"[6.2] Generate outline for Chapter {chapter_num}") # Load chapter summary from plan plan_path = self.run_dir / "2-planning" / "2-chapter-plan.json" chapter_summary = "(暂无概要)" chapter_title = f"第{chapter_num}章" if plan_path.exists(): with open(plan_path, 'r', encoding='utf-8') as f: plan = json.load(f) for ch in plan.get('chapters', []): if ch.get('chapter_number') == chapter_num: chapter_summary = ch.get('summary', chapter_summary) chapter_title = ch.get('title', chapter_title) break target_words = self.config['book']['chapter_target_words'] # Build prompt if self.prompt_builder: prompt, log_path = self.prompt_builder.build_chapter_outline_prompt( run_id=self.config['run_id'], chapter_num=chapter_num, chapter_title=chapter_title, chapter_summary=chapter_summary, previous_content=previous_content, target_words=target_words ) else: # Simple fallback prompt = f"生成第{chapter_num}章详细大纲" # Call model (placeholder) outline = self.model_callable(prompt) # Save to drafts outline_path = self.run_dir / "drafts" / "outlines" / f"Ch{chapter_num:03d}_outline_attempt1.md" atomic_write_text(outline_path, outline) print(f"[6.2] Complete: outline saved") return outline # ======================================================================== # 6.3 Draft Generation # ======================================================================== def generate_draft( self, chapter_num: int, outline: str, previous_content: str = "", attempt: int = 1 ) -> str: """6.3: Generate chapter draft""" print(f"[6.3] Generate draft for Chapter {chapter_num} (Attempt {attempt})") # Split outline into segments if needed # For now, generate as one piece target_words = self.config['book']['chapter_target_words'] # Build prompt if self.prompt_builder: prompt, log_path = self.prompt_builder.build_chapter_draft_prompt( run_id=self.config['run_id'], chapter_num=chapter_num, chapter_title=f"第{chapter_num}章", detailed_outline=outline, previous_content=previous_content, segment_summary="本章全部内容", segment_words=target_words, is_first_segment=True, event_id=generate_event_id(self.config['run_id'], '6.3', chapter_num) ) else: prompt = f"根据大纲生成第{chapter_num}章正文" # Call model draft = self.model_callable(prompt) # Save to drafts draft_path = self.run_dir / "drafts" / "chapters" / f"Ch{chapter_num:03d}_draft_attempt{attempt}.md" atomic_write_text(draft_path, draft) print(f"[6.3] Complete: {len(draft)} chars") return draft # ======================================================================== # 6.4 Quality Check # ======================================================================== def qc_evaluate( self, chapter_num: int, draft: str, outline: str, previous_content: str = "" ) -> QCResult: """ 6.4: Quality Check with multi-perspective review """ print(f"[6.4] QC for Chapter {chapter_num}") # In real implementation, would call 3 critic models # For now, placeholder with simple heuristic # Simple QC: check word count, basic format word_count = len(draft) target = self.config['book']['chapter_target_words'] # Calculate base score score = 80 # Start neutral # Word count check (+/- 10% is fine) if abs(word_count - target) / target > 0.1: score -= 5 # Check for outline elements if outline[:100] in draft or any(line[:30] in draft for line in outline.split('\n')[:5]): score -= 10 # Outline leaked into draft # Multi-perspective simulation perspectives = ['editor', 'logic', 'continuity'] scores = [] for perspective in perspectives: # Would call actual critic model here perspective_score = score + (5 if perspective == 'editor' else 0) scores.append(perspective_score) final_score = int(sum(scores) / len(scores)) # Determine status if final_score >= self.pass_threshold: status = QCStatus.PASS elif final_score >= self.warning_threshold: status = QCStatus.WARNING else: status = QCStatus.REVISE # Build result result = QCResult( score=final_score, status=status, dimensions={ 'outline_adherence': final_score - 5, 'main_plot': final_score, 'character': final_score - 2, 'logic': final_score - 3, 'continuity': final_score - 1, 'pacing': final_score + 2, 'style': final_score }, pros=[ "章节结构完整", "情节推进自然" if final_score > 75 else "情节有待加强", "人物行为符合设定" if final_score > 80 else "人物刻画需改进" ], cons=[] if status == QCStatus.PASS else [ "建议加强细节描写", "部分对话可更口语化" ], rewrite_plan="" if status == QCStatus.PASS else "根据cons进行针对性修改" ) # Save QC result qc_path = self.run_dir / "drafts" / "qc" / f"Ch{chapter_num:03d}_qc_attempt1.md" atomic_write_json(qc_path, result.to_dict()) print(f"[6.4] Complete: score={final_score}, status={status.value}") return result # ======================================================================== # Attempt Cycle # ======================================================================== def attempt_cycle( self, chapter_num: int, outline: str, previous_content: str = "" ) -> Tuple[str, QCResult, int]: """ Run Attempt 1-3 cycle Returns: (final_draft, qc_result, attempt_number) """ attempt = 1 best_draft = "" best_result = None while attempt <= self.max_attempts: print(f"\n[Attempt {attempt}/{self.max_attempts}]") # Generate draft draft = self.generate_draft( chapter_num, outline, previous_content, attempt ) # QC result = self.qc_evaluate(chapter_num, draft, outline, previous_content) # Track best if best_result is None or result.score > best_result.score: best_draft = draft best_result = result # Check if passed if result.status in [QCStatus.PASS, QCStatus.WARNING]: print(f"[Attempt {attempt}] Passed with score {result.score}") return draft, result, attempt # Continue to next attempt if attempt < self.max_attempts: print(f"[Attempt {attempt}] Failed ({result.status.value}), retrying...") # In real implementation, would apply rewrite_plan for targeted refinement attempt += 1 # All attempts exhausted -> FORCED print(f"[FORCED] All {self.max_attempts} attempts failed, best score: {best_result.score}") best_result.status = QCStatus.FORCED return best_draft, best_result, self.max_attempts # ======================================================================== # 6.5 Save Chapter # ======================================================================== def save_chapter( self, chapter_num: int, draft: str, qc_result: QCResult, chapter_title: str = "" ) -> Path: """6.5: Save chapter to chapters/ directory""" print(f"[6.5] Save Chapter {chapter_num}") if not chapter_title: chapter_title = f"第{chapter_num}章" # Determine filename based on QC status is_forced = qc_result.status == QCStatus.FORCED filename = sanitize_chapter_filename(chapter_num, chapter_title, is_forced) chapter_path = self.run_dir / "chapters" / filename # Add metadata header metadata = f"""[Chapter Metadata] Number: {chapter_num} Title: {chapter_title} Word Count: {len(draft)} QC Score: {qc_result.score} QC Status: {qc_result.status.value} Saved: {get_timestamp_iso()} [End Metadata] --- """ full_content = metadata + draft atomic_write_text(chapter_path, full_content) print(f"[6.5] Complete: {chapter_path}") return chapter_path # ======================================================================== # 6.6 State Commit # ======================================================================== def state_commit( self, chapter_num: int, draft: str, qc_result: QCResult, state_changes: Optional[Dict[str, Any]] = None ): """6.6: Commit state changes after chapter completion""" print(f"[6.6] State Commit for Chapter {chapter_num}") # Update writing state writing_state_path = self.run_dir / "4-state" / "4-writing-state.json" with open(writing_state_path, 'r', encoding='utf-8') as f: writing_state = json.load(f) writing_state['current_chapter'] = chapter_num writing_state['completed_chapters'].append(chapter_num) writing_state['qc_score'] = qc_result.score writing_state['qc_status'] = qc_result.status.value writing_state['updated_at'] = get_timestamp_iso() # Handle forced_streak if qc_result.status == QCStatus.FORCED: writing_state['forced_streak'] = writing_state.get('forced_streak', 0) + 1 writing_state['flags']['prev_chapter_forced'] = True # Add to backpatch backpatch_record = { 'id': generate_event_id(self.config['run_id'], 'BP', chapter_num), 'chapter': chapter_num, 'issue': f"QC score {qc_result.score} below threshold", 'severity': 'high' if qc_result.score < 70 else 'medium', 'evidence': f"QC result: {qc_result.cons}", 'status': 'open', 'created_at': get_timestamp_iso(), 'closed_at': None, 'fix_strategy': None, 'qc_after_fix': None } atomic_append_jsonl( self.run_dir / "4-state" / "backpatch.jsonl", backpatch_record ) else: writing_state['forced_streak'] = 0 writing_state['flags']['prev_chapter_forced'] = False # Check forced_streak threshold if writing_state['forced_streak'] >= 2: writing_state['flags']['is_paused'] = True print("[ALERT] forced_streak >= 2, pausing for manual review") atomic_write_json(writing_state_path, writing_state) # Commit state changes if provided if state_changes: self.state.commit_chapter_state(chapter_num, state_changes) print(f"[6.6] Complete: forced_streak={writing_state['forced_streak']}") # ======================================================================== # Write Single Chapter # ======================================================================== def write_chapter( self, chapter_num: int, previous_content: str = "", auto_continue: bool = False ) -> Dict[str, Any]: """ Write a single chapter through full 6.x pipeline Returns: Result dict with paths and status """ print(f"\n{'='*50}") print(f"Writing Chapter {chapter_num}") print(f"{'='*50}\n") # 6.1 Sanitizer sanitizer_result = self.sanitizer(chapter_num) # 6.2 Generate Outline outline = self.generate_chapter_outline(chapter_num, previous_content) # In manual mode, would wait for user confirmation here if self.mode == 'manual' and not auto_continue: print("[Manual Mode] Waiting for outline confirmation...") # Would pause here in real implementation # 6.3-6.4 Attempt Cycle draft, qc_result, attempt_num = self.attempt_cycle( chapter_num, outline, previous_content ) # 6.5 Save Chapter chapter_path = self.save_chapter(chapter_num, draft, qc_result) # 6.6 State Commit self.state_commit(chapter_num, draft, qc_result) return { 'chapter_num': chapter_num, 'chapter_path': chapter_path, 'qc_score': qc_result.score, 'qc_status': qc_result.status.value, 'attempt': attempt_num, 'forced_streak': self._get_forced_streak() } # ============================================================================ # Module Test # ============================================================================ if __name__ == "__main__": import tempfile print("=== Writing Loop Test (Phase 6) ===\n") # Mock model callable def mock_model(prompt: str) -> str: return "这是生成的内容...\n（模拟模型输出）\n字数填充：" + "内容" * 100 with tempfile.TemporaryDirectory() as tmpdir: # Setup from .phase_runner import PhaseRunner, WorkspaceManager workspace = WorkspaceManager(Path(tmpdir) / "novels") runner = PhaseRunner(workspace) # Run phases 1-5 results = runner.run_all( book_title="测试小说", genre="都市异能", target_words=50000, mode="auto" ) run_dir = results['run_dir'] # Create writing loop loop = WritingLoop( run_dir=run_dir, model_callable=mock_model ) # Write chapter 1 result = loop.write_chapter(1) print(f"\n[Test] Chapter 1 result:") print(f" Status: {result['qc_status']}") print(f" Score: {result['qc_score']}") print(f" Attempt: {result['attempt']}") print(f" Path: {result['chapter_path'].exists()}") print("\n=== All tests completed ===")