dsh-novel-solo

Running a novel pipeline on a local quantized small model is deceptively hard. Small models wobble on tool calls, emit JSON and escape tokens nobody asked for, invoke tools that should not exist in the schema, and produce prose dripping with AI clichés. Most agent frameworks assume a large model with dozens of tools and a generous context window. dsh-novel-solo is purpose-built for the opposite scenario: it strips the tool catalog to the bare minimum, locks output to plain CJK text and common punctuation, and bakes the entire workflow into a single preset so a laptop GPU can finish a full-length novel without a cloud API in sight.

The plugin ships two things you actually touch. First, a "Subagent count (1-12)" selector under Settings → General, with built-in zh/en i18n. Second, the novel-solo creation preset, which defines a fixed pipeline: narrative method → core worldview → noun index → outline → chapter list → character files → chapter writing → per-chapter review → whole-book review and assembly. Every chapter is graded green / yellow / red across seven dimensions (setting, character, catalog, narrative, text rules, AI cliché, plot logic), and the report is written to a markdown file. A shared six-dimension AI-cliché judge catches sentence templates, stock vocabulary, emotion-telling, structural sameness, diluted information, and character distortion during both writing and review. Versioning is overwrite-safe: a full-chapter rewrite copies to v(X+1) and the old draft is always kept.

Who is this for? Solo authors and hobbyist writers who want to finish a book on a laptop GPU (tested with RTX 4060 8 GB + llama.cpp + Qwen3-35B at IQ4_XS) at zero API cost. Also developers experimenting with local small-model agent pipelines who need a hardened, deterministic workflow rather than trial-and-error tool-schema tuning. No cloud, no large model, no manual tool-list debugging—install, pick a concurrency number, and start writing.