shiyan688/dsh-novel-craft

dsh-novel-craft is a native writing plugin for the DeepSeek Harness (DSH) agent platform, installed via the single command dsh plugin --profile web add dsh-novel-craft and mounted as two halves: a host half exposing 23 HTTP routes under /novel-craft/api/ (state, marks, evidence, distill, rules, profile, discover, project, workspace, chapter, setup, annotation, pack, revise, revise-apply, tension, check, stage, newchapter, rule-evidence) that bound to 127.0.0.1 only, and a browser half delivered as a seven-tab workbench (抽卡工作台) that DSH serves on demand with no-cache so a page refresh picks up code changes. Its core capability is "taste calibration by clicking": the reader steps through candidate paragraphs with j/k/G/B/Space/n/p, marks good or bad with a left 3px color bar and soft background, then triggers a single one-shot distillation call capped at 40 paragraphs × 240 chars / 20KB total, where the system prompt forces the model to output only 喜欢： / 避免： sections and forbids copying source text. Each distilled rule must be individually adopted to enter the profile, and every rule carries a 🔍 evidence button pointing back to the original marked spans.

Practical specifics of "open a new chapter": the user supplies a chapter goal (number, objective, must-happen, forbidden), requests N direction labels (default 10) that propose scene-level decisions rather than adjective swaps, edits or deselects them, then the system writes each surviving direction to disk immediately as 第13章-A-保守精修.txt (visible "第 3/10 篇", mid-run abort safe, single-file failure isolated) and the approved paragraphs can be merged into a final draft that flags 〔此处需过渡〕 gaps instead of bridging them. Revision only touches annotated paragraphs (each ≤800 chars, batch ≤20), assembles ±80 chars of context, and a literal byte-diff guard rejects writes when any non-annotated paragraph drifts or the paragraph count changes, with originals auto-backed up before write-back. The 情节 tab runs purely local string statistics — tension curve (manual 1–5 vs estimated dashed with per-point rationale), outline drift, unresolved foreshadowing ledger, word-count imbalance, event density, score decline — never feeding the novel body into the model.

Practical specifics of dependencies, limits, lifecycle: the writing pack sent into the model is 9000 chars across 10 sections (chapter task, author preference profile, previous chapter tail ≤900 chars, recap from summaries, characters, props/buffs ledger, plot nodes, pending foreshadowing, anti-AI-tone list, writing requirements) trimmed in a fixed priority order with a budget table appended and an explicit "do not load into context" footer that lists evidence excerpts, annotation files, full-book merges, and so on; the three buttons that may carry body text (提炼规律, 补齐来源, 按批注微调) each have hard limits printed in the README. The host half is loaded at DSH startup so the user must restart DSH after install; peer dependencies are declared as *, verified against DSH 0.1.0-rc.7, 0.1.2-rc.1, 0.1.5-rc.1 (latest), 0.1.5-rc.2 (next), 0.1.6-alpha.1 via node scripts/check-dsh-compat.mjs next (latest/alpha/specific), which spins up a temporary DSH install with isolated DSH_HOME and asserts host routes + client bundle delivery then cleans up. DSH ≥0.1.5 requires Node 22 because Node 20 makes dsh web exit silently (no log, no port), and the compat script auto-searches for Node 22 and reports that as an environment failure rather than a plugin regression. Tests run via npm test (7 files, 775 assertions, no host-path dependency, fixtures built in temp dirs) and the package is MIT-licensed.