novelcode-dsh
DiscoveredDSH-native long-form fiction author studio with Narrative V2 memory, simulation, approvals, and a read-only Web workbench
NovelCode DSH
NovelCode DSH is an independent DeepSeek Harness product for long-form fiction analysis, planning, state maintenance, style control, segmented drafting, review, and immutable manuscript assembly.
Developer preview. The current release is source-deployed and pinned to DeepSeek Harness
0.1.0-rc.8. It requires a separate Novel Workbench checkout for the Narrative V2 JSONL service; that checkout is never vendored here.
The product replaces the generic Harness layer only. It does not rewrite the existing Narrative V2/GBrain domain core. During the first migration phase, a workspace-scoped Node provider starts the existing Bun JSONL service and exposes five stable model-facing tools through DSH:
narrative_projectnarrative_sourcenarrative_memorynarrative_plannarrative_writing
Status
Phase 0-5 product foundation:
- frozen 35-action V2 contract;
- workspace-scoped JSONL sidecar provider;
- DSH
ctx.novelservice and five tool surfaces; - high-risk author approval policy;
- dedicated
apps/noveldshproduct CLI and workspace ownership gate; - log-only novel session event vocabulary;
- pure temporal/epistemic memory contracts and deterministic state projection;
- pure branch sandbox with causal, state, POV-knowledge, stale-canon, and candidate-ranking checks;
- Memory OS adapters that enter the existing Brain proposal/review lifecycle;
- viable branch to provenance/Blueprint candidate compilation;
- approved revision to evidence-bound state candidate compilation;
- four packaged DSH fiction workflow Skills;
- trusted confirmed-Brain snapshot and live baseline hash;
- versioned branch/Blueprint and approved-revision/state Host envelopes on the existing five tools;
- per-workspace mutation serialization;
- dedicated rc.8 Web client package with loopback-only Workspace RPC;
- seven-surface read-only workbench for project, sources, temporal memory, plot, style, chapter runs, and immutable revisions;
- timeline, character-knowledge matrix, deterministic state projection, and final-ledger inspection without manuscript bodies;
- synthetic contract tests with no private novel content.
Interactive branch/envelope forms, approval impact diffs, Narrative V2-native
temporal schema expansion, multi-agent role composition, and long-form recovery
qualification follow the reviewed implementation plan under
docs/design-architecture/dsh-novel-author-studio/.
Primary documents:
- Product and architecture design
- Deployment and usage guide
- Phase 0/1 implementation plan
- Open-source reuse analysis
- Installation and verification guide
- Phase 2 memory and simulation status
- Phase 3 Blueprint and state loop status
- Phase 4 Host orchestration status
- Phase 5 Web workbench foundation status
Quick start
git clone https://github.com/WOO-woo-Waf/novelcode-dsh.git
cd novelcode-dsh
cp .env.example .env
pnpm install --frozen-lockfile
pnpm verify
pnpm add --global @deepseek-ai/dsh@0.1.0-rc.8
export NOVELCODE_DSH_CORE_ROOT=/absolute/path/to/novel-workbench
export DSH_BUNDLED_SKILL_DIR="$PWD/packages/dsh-novel-author/skills"
dsh plugin --profile web add "$PWD/packages/dsh-novel-author" "$PWD/packages/web"
pnpm noveldsh workspace init /path/to/new-novel
pnpm noveldsh workspace doctor /path/to/new-novel
cd /path/to/new-novel
dsh web --no-open
Open http://127.0.0.1:3080, configure the model provider in DSH, create a
session whose workspace is the novel directory, and select the NovelCode novel
director preset after installing it as described in the
deployment and usage guide.
The marker created by workspace init only binds an empty directory to this
product; it is not novel memory or canon. Existing Novel Workbench V2 projects
with .narrative/v2/project.json are accepted without a marker.
The source checkout configured by NOVELCODE_DSH_CORE_ROOT is read-only. Novel
workspaces are selected by each DSH session's immutable cwd; the provider starts
one sidecar per active workspace and terminates it when the plugin is disposed.
Deployment boundary
- Supported now: a local workstation, or a remote host accessed through an SSH tunnel, with DSH bound to loopback.
- Run only one writing Host against a novel workspace. The current queue is process-local and does not coordinate multiple DSH processes.
- Do not expose the Web port directly to the public Internet. DSH's current Web carrier is a loopback development surface, not an authentication gateway.
- Model credentials remain in DSH settings or private environment files. Never commit them to this repository or a novel workspace.
License
MIT. Third-party projects listed in the reference manifest are not vendored and retain their own licenses.
Comments
Loading…
Similar plugins
by qinpeizhan77
Local-first fiction workspace that links outlines, chapters, characters, timelines, writing rules, and continuity checks to the conversation.
★ 9
↓ 438/wk
MIT
TypeScript
Aug 30, 2026
dsh plugin --profile web add deepseek-harness-novel-studioby zenstory-ai
A DSH plugin for novel writing and short-drama production, powered by Oh Story and Drama Skills.
★ 489
↓ 898/wk
MIT
Python
Oct 10, 2026
dsh plugin --profile web add @oh-story/dshby x2802490130-prog
Writing engine for DeepSeek Harness: long-form web-novel orchestration with a separate DeepSeek key, lore management, semantic retrieval, and a corpus library.
★ 8
↓ 361/wk
MIT
JavaScript
Aug 25, 2026
dsh plugin --profile agent add dsh-tool-writingby peterwangze
DSH (DeepSeek Harness) 自动化小说写作发布流水线插件:claude-writing-workflow 迁移版 agent 预设 + 小说工作台(可视化/实时渲染/章节编辑)+ 多平台发布配置与数据驱动优化闭环
★ 7
JavaScript
Sep 16, 2026
dsh plugin --profile web add dsh-novel-writingby Tkingxiao
DeepSeek Harness 的「单核写作」插件:面向量化小模型做了充分的工具瘦身与输出加固,适合在本机用本地模型跑长篇小说流水线。
★ 11
↓ 672/wk
MIT
JavaScript
Oct 3, 2026
dsh plugin --profile web add dsh-novel-soloby shuaweng
写书鲸:面向 DeepSeek Harness 的原生 AI 小说创作工作台插件
★ 4
MIT
JavaScript
Sep 7, 2026
dsh plugin --profile web add @xieshujing/dsh-plugin