novel-pipeline:可恢复流水线

SkillDev tools

Chains all features into a resumable pipeline: (co-creation →) writing → review → rewrite → delivery, executed stage by stage with checkpoint resume. Triggers: "一条龙写完整本书" (write a whole book end to end), "从头跑完整个流程" (run the entire flow from scratch), "中断后接着跑" (resume after interruption). Use when yo

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the novel-pipeline:可恢复流水线 skill

What this skill tells your AI

The instructions your AI receives, as published by xiaoyangy/novel-studio in skills/novel-pipeline/SKILL.md and read by ahel’s review.

按阶段顺序跑完整流程,状态存 output/novel/meta/pipeline.json已完成的阶段重跑时自动跳过, 从断点继续。流水线只做阶段编排;每个阶段复用对应子命令逻辑,阶段内部还有更细的恢复 (write 走 checkpoint、review/rewrite 按章号),两层恢复叠加。

阶段

cocreate(可选,多轮澄清)→ write(创作)→ review(评审)→ rewrite(重写)→ deliver(交付沉淀:推进台账 + RAG 事实入库 + 交付快照)

默认序列:write,review,rewrite,deliver(假设已有创作指令)。

前置条件

  • 已完成首次配置(见 skills/README.md),LLM 可用(先 novel-check)。
  • cocreate 阶段时需要交互终端。

执行

# 标准:给创作指令,跑写作→评审→重写→交付
novel-studio --pipeline --prompt "写一本东方玄幻长篇,主角从边陲小城起步"

# 从文件读创作指令
novel-studio --pipeline --prompt-file ./requirement.md

# 先共创澄清再一条龙(首阶段交互,之后无人值守)
novel-studio --pipeline --stages cocreate,write,review,rewrite,deliver

# 自定义阶段子集
novel-studio --pipeline --prompt "..." --stages write,deliver

# 单阶段评审 / 重写也走 pipeline,可指定章号范围
novel-studio --pipeline --stages review --from 3 --to 8 --budget 5m
novel-studio --pipeline --stages rewrite --from 3 --to 8 --role coordinator

# 中断后再次运行同一命令即从断点继续;要从头重跑用 --restart
novel-studio --pipeline --prompt "..."
novel-studio --pipeline --prompt "..." --restart

参数

参数默认说明
--prompt <text> / --prompt-file <path>创作指令(write 阶段用;- 表示 stdin)
--stages a,b,cwrite,review,rewrite,deliver阶段子集与顺序
--write-to <n>0write 阶段写到指定章节后暂停;0 表示写到全书完结
--from <n> / --to <n>0review / rewrite 阶段章号范围
--budget <dur>阶段默认review / rewrite 阶段每章 LLM 调用预算
--role <name>writerrewrite 阶段模型角色
--restart清空已保存状态,从头重跑

恢复语义

  • 每个阶段成功返回后还要通过证据校验,才写入 pipeline.jsoncompletedevidence;失败或证据不足的阶段不标记完成,下次重跑从它继续
  • write 阶段:本书已完结则跳过;已有进度则恢复创作;全新项目用创作指令起新书
  • 已标记完成的阶段重跑时会先复核证据;若产物被删或 checkpoint 缺失,会清掉完成标记并重跑该阶段
  • 证据口径:writeprogress.phase=complete、章节文件和 commit_chapter checkpoint;reviewreviews/*.mdmeta/review-summary.mdrewrite 看章节文件和 .pre-rewrite.md 备份;delivermeta/delivery_log.jsonl / meta/delivery_log.md
  • novel-studio --diag 会只读报告 meta/pipeline.json 的证据漂移;真正清标记和重跑只发生在再次执行 --pipeline
  • 改了 --stages 列表会重置进度(保留已捕获的创作指令)
  • 状态文件用临时文件 + rename 原子写入,写一半崩溃不损坏
  • --headless --prompt--review-existing--rewrite-existing 只保留为兼容别名,新任务应直接调用 --pipeline

长期资产

流水线过程中会持续读写 meta/writing_assets.jsonbook_world.jsonmeta/resource_ledger.jsonmeta/rag/*。这些文件是可编辑、可复用的创作资产;不要把写法、世界观或资源状态只写在 一次性 prompt 里。

Signals

GitHub stars
109
Forks
18
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
novel-pipeline
Source
github.com/xiaoyangy/novel-studio