v4.3.5 — Prompt optimization, pipeline improvements, config tuning | v4.3.5 Prompt 优化、流水线改进、配置调整
What's New
🔧 Prompt & Pipeline Optimization
Redesigned the chapter writing workflow to eliminate common quality issues at the source. Task 1 oracle now verifies outline completeness before proceeding. Task 3 chapter agent follows a structured subsection model (claim → evidence → analysis → counterpoint) and avoids redundant "In summary" filler. Per-section depth is now enforced naturally via tighter prompt constraints rather than post-hoc review.
🐛 Serial Progress Guard
Added a checkpoint file (serial_progress.txt) to the macOS serial mode workflow. When the main agent resumes after user interruption, it reads the file to determine the correct chapter to continue from — preventing accidental parallel dispatch or restart.
📏 Depth Balance Check
New depth-balance script command detects chapters with significantly fewer lines than average. Integrated into the QA pipeline as an informational check — flags thin chapters without blocking assembly.
📐 Character Count Alignment
Updated profiles.json with realistic character targets across all three modes (quick 15K / standard 25K / deep 45K). Chapter agents now write once without iterative trimming — a major source of latency. The stripped-char metric is maintained as the universal cross-language unit.
🧹 Word Count Display Fix
Fixed a bug where the final summary table displayed an incorrect word count value due to the agent extracting the wrong JSON field. Explicitly routed to checks.word_count.count from the qa-report JSON output.
🇨🇳 中文说明
更新内容
🔧 Prompt 与流水线优化
从源头消除常见的报告质量问题。Task 1 oracle 现先生成大纲后自动检查覆盖面完整性。Task 3 章节 agent 遵循结构化子节模型（结论→数据→因果→反方），不再写"综上所述"式重述。各节深度通过更精准的 prompt 约束保障，而非事后 Review。
🐛 串行进度文件保护
macOS 串行模式新增 checkpoints 文件。主 agent 中断恢复后读取进度文件确定从哪章继续，消除意外并行派发或重头开始的问题。
📏 章节深度均衡检查
新增 depth-balance 脚本命令，自动检测行数显著低于平均值的章节。集成到 QA 流水线中，告警但不阻断。
📐 字数目标合理化
更新 profiles.json 三档字数目标（quick 15K / standard 25K / deep 45K）。章节 agent 一次性写完，不再反复修剪——大幅降低耗时。去空格字符数作为全局统一标尺保持不变。
🧹 字数显示 bug 修复
修复最终汇报表中字数显示错误的问题——agent 原先误读了 JSON 的顶层字段。现已改为从 checks.word_count.count 读取。