Commit Graph
5 Commits
Author SHA1 Message Date
kaiandUser <human> 1b47b50d3c v0.6-wip: Python-based Phase 4 translation pipeline
架构变更:把 dr-translator 从 opencode agent 降级为 Python 脚本编排下的 LLM
调用。根本原因是 agent 一次性处理 19k 英文词整文,单次 output token 接近
Sonnet 4.6 上限(~32k),多次重跑都卡在同一个坑里——问题是架构本身,不是
prompt。

新架构:

scripts/lib/zenmux_client.py     HTTP 客户端,指数退避重试、token 统计
                                  JSONL 日志、secrets.env 自动加载
scripts/lib/markdown_chunker.py   按 H1/H2 切块,稳定 anchor ID(order+title
                                  sha1),支持合并/统计
scripts/prompts/translate_system.txt  英译中 prompt,用自定义 <<<TRANSLATION>>>
                                       分隔符格式(规避 Markdown-in-JSON 问题)
scripts/prompts/polish_system.txt     中文润色 prompt(留给下一步 polish.py)
scripts/translate.py              主入口:章节级切块 → 逐块翻译 → 拼接

关键设计:
- 0 依赖 LLM 遵从性:Python 控制切块/循环/重试,LLM 只做单块翻译
- 断点续传:每块翻译完立即写 phase4/zh_chunks/<order>-<anchor>.md
- 术语表累积:每块的 glossary_patch 合并回 phase4/glossary.json
- 失败隔离:单块失败不影响其他块,重跑只补缺
- 调试友好:--only N,M / --limit K / --force

实测(dual-target-rnai-pipeline-2026):
- 63 块全部成功,17 分钟,$1.70
- 33,441 中文字(符合"研究类 ≥30,000 字"硬标准)
- 310 条双语术语
- 翻译质量:接近母语咨询分析师写作

下一步:polish.py(按 H2 section 润色)、merge_chapters.py(从 phase2/drafts
合并生成 final_en.md)、重构 dr-editor-in-chief 调度脚本、更新 /dr-finalize。

Co-authored-by: User <human>
2026-04-22 10:43:43 +08:00
kai 701bc1887e v0.5.2: dr-translator chunked translation protocol
Root cause: dr-translator was trying to write entire final_zh.md in one
write call, hitting Sonnet 4-6 output token limit for long reports
(~19k English words → ~27k Chinese chars → blown past 32k token cap).

Fix: explicit chunk-and-append protocol
- Split final_en.md by H1 (# ) then H2 (## ) boundaries
- Each chunk ≤ 2,500 English words
- First chunk uses write to create final_zh.md
- Subsequent chunks use edit or read+write to append
- Per-chunk Chinese output kept under ~5,000 characters (safe margin)
- Preserves glossary.json updates across chunks
2026-04-21 23:10:10 +08:00
kai 333b7bb8d5 v0.5.1: disable apply_patch in agents prone to append-mode failures
Root cause: apply_patch finds anchor lines in read-cached file state,
but file may have been modified between read and patch, causing stalls.

Changes:
- dr-verifier: disable apply_patch AND edit; force read-then-write protocol for evidence file appends
- dr-analyst: document write-preferred protocol for sources.jsonl appends
- dr-polisher: disable apply_patch; keep edit for small string replacements
- dr-editor-in-chief / dr-translator: disable apply_patch

Recovery procedure documented in dr-verifier for write failures.
2026-04-21 14:44:03 +08:00
kai a092af4398 v0.5: deep quality refactor (P0+P1+P2)
- Split dr-chief-editor (Phase 3 read-only) vs new dr-editor-in-chief (Opus, Phase 4 lead)
- New dr-translator (en->zh) and new humanizer-cn / output-hygiene / en-zh-translation skills
- Switch to English working language (Phase 2-3), final Chinese translation (Phase 4)
- /dr-init: add report title proposals + word budget mode
- /dr-frame: bilingual framework
- /dr-finalize: new chain editor->translator->polisher->reporter
- report-template.py: widows/orphans/keepWithNext, 3-color hierarchy, confidentiality banner
- dr-reporter: mandatory citations backfill + output hygiene check
- dr-pm: batch-level context compression via manifest.batches_summary
- mckinsey-method: SCQA only for Executive Summary + chapter intros (no explicit labels)
- length-budget: 4 word-budget modes + en/zh 1:1.4 ratio
v0.5
2026-04-21 13:02:54 +08:00
kai 4a38f6bed1 snapshot before v0.5 refactor v0.4-final 2026-04-21 12:31:58 +08:00