dsh-novel-solo
A single-author novel-writing plugin for DeepSeek Harness: a subagent concurrency setting plus a complete novel-creation preset (persona). It is tuned for quantized small models — the tool catalog is slimmed and output behavior hardened — so you can run a full-length novel pipeline locally.
Features
- GUI setting: a "Subagent count (1-12)" card on the plugin page (0.1.7: left rail → Plugins; 0.1.6: Settings → Plugins), with built-in zh/en i18n.
- Full creation preset: ships the
novel-solo preset with a self-driven persona that follows a fixed pipeline — 叙事方法 (narrative method) → 核心世界观 (core worldview) → 名词索引 (noun index) → 大纲 (outline) → 章节目录 (chapter list) → 人物档案 (character files) → chapter-by-chapter writing → per-chapter review → final whole-book review & assembly.
- Quantized-safe protocol: plain CJK + common punctuation only, no JSON, no escapes, avoids fragile tokens; minimal tool calls, one at a time.
- Review loop: every chapter is reviewed (green/yellow/red) across 7 dimensions (setting / character / catalog / narrative / text rules / AI-cliché / plot logic); each report is written to a markdown file; when all chapters pass, a whole-book review runs and the book is assembled.
- No AI clichés: judges AI-flavor across six dimensions (sentence templates / stock vocabulary / emotion telling / structural sameness / diluted information / character distortion), distinguishing narrator-level repetition from character-level consistency; shared by writing and review.
- Slimmed catalog: hard-disables whole tool rows (shell / jobs / skills / goals / web) inside the preset to shrink the tool schema.
- Versioned writing: each chapter is
第N章-章节名字-vX.md; full-chapter rewrites first copy to v(X+1) and never overwrite older drafts.
Install
dsh plugin --profile web add "dsh-novel-solo"
dsh web
Then open the plugin page — Plugins in the left rail on 0.1.7, Settings → Plugins on 0.1.6 — and the dsh-novel-solo "Subagent count" card appears.
On hosts older than 0.1.7 the plugin additionally deploys the preset idempotently from template/ to <dshHome>/.agent-presets/novel-solo/, because that was the only preset contract there (an existing target is skipped — your edited preset is never overwritten). On 0.1.7 nothing is deployed: the preset arrives as a declarative row in the profile patch.
Host compatibility
The plugin supports DSH 0.1.6-alpha.1 / 0.1.6-alpha.2 and, from 0.1.7-alpha.1 on, every later host release and picks its route per host. Both halves probe the host's own version through ctx.profileContext.installAnchor — the YAML disabled: gate in cordis.patch.yml and the same check in lib/index.js — and fall back to the 0.1.6 route whenever the version cannot be read. Everything from 0.1.7 on takes the same route, 0.2.0 included.
| 0.1.6 | 0.1.7 and later |
|---|
| Preset delivery | deploy template/ into <dshHome>/.agent-presets/novel-solo/ | declarative @deepseek-ai/dsh-agent-preset row in the patch |
Where N is saved | dsh-novel-solo settings namespace the plugin registers | the plugin's own Config (count is volatile), projected into the profile patch |
| Card slot | settings.plugin.item via settingsScope | plugins.item via configForms |
Both routes are exercised against a real host; the 0.1.7 one on 0.1.7-rc.1 (card renders on the Plugins page, saves land in cordis.patch.yml, preset shows up in the picker). The 0.1.6 route is covered by the same version gate offline, since no 0.1.6 host binary is installed next to the 0.1.7 one. The hosts actually installed here now are 0.2.0-rc.1, 0.2.1-alpha.3 and 0.2.1-rc.2, and both plugins were installed and booted side by side on each: the node half runs (the .dsh-novel-solo-data mirror is still written), --dump-config shows both declaration rows and the novel-solo roster id, and nothing gets force-disabled. The card itself was not re-clicked in a 0.2.1 browser, but every seam it rides on was checked against the host source at 0.2.1-rc.2 — the plugins.item slot, configForms.get / whileServed, settings.configure / update, the client manifest's id/url/rev, and bind / register on @deepseek-ai/dsh-client-locale — all identical to 0.1.7-rc.1; the same seam list was re-read at 0.2.1-rc.3 and none of it moved (the plugin-manager page only gained an additive openList() reflection method for the new entry-customization plugin). 0.1.7-rc.2, 0.2.0-rc.2 and 0.2.1-alpha.1 remain source-checked only, not run.
Four 0.2.x changes this package has caught up with: @deepseek-ai/dsh-workflow-worker-thread was folded into the PTC engine, so that row is now @deepseek-ai/dsh-workflow-ptc (still disabled: true); tool-subagent dropped backgroundMode and enableRunInBackground (they no longer error, they simply do nothing, so they were removed); and since 0.2.x a browser only loads a client graph whose release matches the host's, so after a host upgrade the page must be reloaded before the card re-mounts. The fourth is how the persona is interpolated: since 0.2.1 the host resolves {{…}} strictly, and an unregistered name throws unknown prompt variable ... in section "deployment:persona-prefix". The persona's {{cwd}} was never a host variable — the whole host registers only provider and model — so the preset row failed to load and no message could be sent at all (issue #2). It no longer depends on template variables: the working directory comes from the required Current working directory entry in the runtime context (neither complete nor includeRuntimeContext suppresses a required contribution), and an absolute path is one working_directory call away; npm run check now also lints the persona so an unregistered variable fails the build instead of the preset.
On 0.2.2-alpha.2 one thing moved that this preset does touch: @deepseek-ai/dsh-tool-subagent-control/list-agents no longer exists as a subpath export — the list_agents tool was folded into that package's main entry, and factory standard deleted its own tool-subagent-list-agents row in the same release. The row is therefore probed rather than declared: disabled resolves the subpath and switches the row off when it cannot be found, which keeps the tool on 0.2.1 and avoids a duplicate registration on 0.2.2. A disabled row is filtered before the Loader resolves any package name, so a host without the subpath loses nothing else. Two 0.2.2 facts that do not reach this package: tool-subagent renamed modelSelectionSettings to modelSelection (this preset never set that key), and {{…}} interpolation itself is unchanged — the persona no longer references any template variable, so the registered-variable set is not something this preset depends on. template/agent.cordis.yml is the single source for the whole row list: cordis.patch.yml's generated region is produced from it by node scripts/sync-preset-patch.mjs, so the probe belongs in the template and the patch is regenerated, never hand-edited.
Whether the plugin loads at all is decided before its code runs: since 0.1.7-rc.1 the host evaluates every @deepseek-ai/dsh* range in peerDependencies against the running release and disables the row if one of them doesn't match, and since 0.2.1-alpha.4 it consults engines.dsh first — once that field exists it decides the verdict alone and peerDependencies no longer participates (each rule was tested against the host code of 0.2.1-alpha.3 and 0.2.1-rc.2 respectively). So both lists are kept: they name 0.1.6-alpha.1 through 0.2.2-alpha.2, then end in an open arm — >=0.1.7-alpha.1, with no upper bound — because nothing in this plugin is bound to a particular host build: its route is chosen by its own version gate at load time, so an unchecked release should still get a working plugin rather than lose one to a check that fails closed. The versions named before the open arm are the ones that were actually compared. The umbrella does not promise behavior across a contract change: the declarative preset row is the part a future host would break first, and the fix for that belongs in the patch file, not in the range — and likewise the open arm only promises that the plugin loads, not that this preset grows whatever factory standard adds next; the time-context, tool-schedule and plugin-market/tools rows the factory added since 0.1.7 are still deliberately absent, and adopting them is a cordis.patch.yml change. Reinstalling on a host that previously refused the plugin needs a dsh web restart, since the decision is made while the profile is composed.
Since 0.1.5, the persona uses the @deepseek-ai/dsh-persona config key prefix (the text key of 0.1.2-rc.1 was removed), so hosts older than that need the persona row renamed by hand. Each home keeps its own deployed copy; homes provisioned by older hosts keep their text: copy and still work when you switch back.
How the subagent count takes effect
DSH's agent/request waterfall only lets a plugin rewrite LLM routing/config — it cannot inject or rewrite system/messages — so "GUI → model prompt" dynamic injection cannot go through the request waterfall. This plugin wires the value through files instead:
- The card saves
N into the host's settings store (see the table above). That is the source of truth.
- On load, the node half mirrors it to
<dshHome>/.dsh-novel-solo-data/agent-count.json.
- The persona row resolves
!!js while the preset loads: it reads template/persona.md and replaces that file's 并发上限 N=<number> anchor with the mirrored number. The persona then decides: N=1 the main agent does everything itself; N>1 writing/review tasks are delegated to subagents while the main agent only dispatches and silently waits.
Note: the mirror and the persona are both resolved while the host loads its plugins, so a GUI change reaches the prompt only after a host restart. Volatile settings give the node side no change callback, so there is nothing to push the new value out live.
Preset at a glance
The persona body lives in template/persona.md; template/agent.cordis.yml (mirrored into cordis.patch.yml) carries the tool rows around it. Content:
| Section | Content |
|---|
| Division of labor | concurrency anchor N=1 (default), with N=1 / N>1 execution paths |
| Quantized-safe rules | top-priority constraints on output and tool calls |
| Six-doc standard structures | per-field templates for 叙事方法 / 核心世界观 / 名词索引 / 大纲 / 章节目录 / 人物档案 |
| Writing & review rules | chapter rules, AI-cliché lists, A–G review dimensions + green/yellow/red, md review reports, whole-book review & assembly |
| Tool usage standards | per-tool rules for read/write/edit/glob/grep/subagent, etc. |
| Project & discipline | the project dir is the 项目名 folder under the working directory; paths come from the runtime context's Current working directory or the working_directory tool — never from a template variable — one thing at a time |
Tool slimming
The preset hard-disables these rows via disabled: true: tool-bash, tool-jobs, skill-filesystem, tool-skill, tool-goal, plan-mode, subagent_codex, subagent_claude_code, workflow-ptc, tool-workflow, tool-ralph, tool-todo, tool-web. It keeps tool-fs (read/write/edit, with read_image switched off), tool-fs-search (glob/grep), str-replace-editor (view/create/str_replace/insert), tool-pwsh (filesystem management only: rename/delete files & folders), subagent/subagent_fork, list_agents, etc.
0.2.1's standard preset grew four more model-facing toolchains, and this preset screens every one of them out: tool-schedule (cron tasks), command-goal (the /goal command), tool-present (deliverable presentation), and the tools entry of plugin-manager / plugin-market. They are spelled out as disabled: true rows — a preset's plugins is a complete composition, so an unlisted row is already absent, but the host itself disables rather than deletes on the web layer because an omitted row quietly reappears the day someone reorders the composition; and disabled rows are filtered before name resolution, so an older host without those packages is never killed by them. Two prompt-only rows are simply left out: agent-instructions (AGENTS.md injection) and time-context (clock context) would be dead weight under complete: true.
One thing the preset cannot govern: the working_directory tool is a host-plane row every session receives, so the persona treats it as the source of truth for paths instead of fighting it.
It also ships a preset-scoped vendored plugin (active only for sessions mounting the novel-solo preset):
template/plugins/llm-tool-choice-pin/index.mjs, exported as dsh-novel-solo/plugins — pins toolChoice to auto for the llama provider so retained tools like edit stay callable (avoids per-request tool decisions under small models).
File structure
lib/index.js node half: host version gate, count mirror, settings seam (0.1.7 volatile `Config` form / 0.1.6 namespace), preset deploy on legacy hosts
lib/client.js browser half: the "Subagent count" card, mounted on `plugins.item` (0.1.7) and `settings.plugin.item` (0.1.6)
cordis.patch.yml installed into the web profile: the plugin row plus the declarative novel-solo preset row (disabled below 0.1.7)
cordis.yml development overlay: card only, no preset rows
template/ the novel-solo preset (agent.cordis.yml + persona.md + preset.yml + vendored plugin), shipped with the package
scripts/sync-preset-patch.mjs copies template/agent.cordis.yml into the patch's generated block; `--check` verifies it and rejects a persona that references an unregistered `{{variable}}`
package.json dsh.bundle / dsh.client metadata so the plugin is recognizable by the plugin market/manifest
npm run check syntax-checks both halves, re-runs the patch/template consistency check, and rejects any {{variable}} in template/persona.md that the host does not register (only provider / model exist; one wrong name is a load-time failure, not a cosmetic gap in the UI).
Environment variables
| Variable | Purpose | Default |
|---|
DSH_HOME | dsh home directory | ~/.dsh |
DSH_NOVEL_SKIP_DEPLOY | 1 skips preset deployment (legacy hosts only) | none |
DSH_NOVEL_REDEPLOY | 1 forcibly overwrites an existing preset (use with care) | none |
Test environment
This plugin was tested end-to-end with a local model:
- Runtime: llama.cpp (
llama-b10615-bin-win-cuda-13.3-x64)
- Model:
Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-IQ4_XS.gguf
- Context:
ctx=65536, reasoning on
- Hardware: laptop RTX 4060 8GB + 32GB RAM + AMD 7840H CPU
- Measured on a ~10k-char novel: 22 token/s, 98% cache hit; 78.8k input / 36.9k output / 2.3M cache; 22 turns / 63 steps, 1h03m42s total
Key server flags (llama-server):
llama-server.exe -m <model.gguf> --no-mmproj --load-mode none --n-cpu-moe 30 -c 65536 -ngl 999 -t 12 -b 1024 -ub 512 -ctk q8_0 -ctv q8_0 -fa on --fit off --no-warmup --poll 0 --temp 0.85 --top-k 20 --top-p 0.95 --min-p 0.05 --repeat-penalty 1.35 --presence-penalty 0.2 --frequency-penalty 0.2 --dry-multiplier 0.8 --dry-base 1.75 --jinja --reasoning on --reasoning-effort medium --reasoning-budget 2048 --reasoning-format deepseek --reasoning-preserve --cont-batching -np 1 --alias "qwen3.6-novel-nsfw-reason" --port 8090 --host 127.0.0.1 --ui --keep -1 --cache-ram 4096 --ctx-checkpoints 64
References
This plugin's design is inspired by sailoumili/novel-writer.
License
MIT License
Copyright (c) 2026 Tkingxiao
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
