獬豸(Xiezhi)小说投稿评审 MCP Server
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| XIEZHI_CACHE_DIR | No | Directory used to override the default cache location. Optional. | |
| XIEZHI_CACHE_ROOT | No | Alternative cache root override alongside XIEZHI_CACHE_DIR. Optional. | |
| XIEZHI_CONFIG_DIR | No | Directory containing an alternative language/rule configuration set: rules.yaml, platforms.yaml, benchmarks.yaml, and prompts.md. Optional. | |
| NOVEL_REVIEW_LLM_KEY | No | API key for the primary LLM gateway. Optional; used together with NOVEL_REVIEW_LLM_URL and NOVEL_REVIEW_LLM_MODEL. | |
| NOVEL_REVIEW_LLM_URL | No | OpenAI Chat Completions-compatible endpoint URL for the primary LLM gateway. Optional; without it the server runs in offline stub mode and returns '未定' rather than pretending to pass. | |
| NOVEL_REVIEW_LLM2_KEY | No | API key for the second independent LLM backend. Optional; used together with NOVEL_REVIEW_LLM2_URL and NOVEL_REVIEW_LLM2_MODEL. | |
| NOVEL_REVIEW_LLM2_URL | No | OpenAI Chat Completions-compatible endpoint URL for a second, independent LLM backend used for dual review. Optional. | |
| NOVEL_REVIEW_LLM_MODEL | No | Model name for the primary LLM gateway. Optional; used together with NOVEL_REVIEW_LLM_URL and NOVEL_REVIEW_LLM_KEY. | |
| NOVEL_REVIEW_LLM2_MODEL | No | Model name for the second independent LLM backend. Optional; used together with NOVEL_REVIEW_LLM2_URL and NOVEL_REVIEW_LLM2_KEY. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| check_redlineC | L1 形态红线 + L2 关键词打分,毫秒级;致命项直接判禁止发布。 |
| review_chapterC | 单章三审流水线:一审合规定位(硬拒短路)→二审内容质量→三审终审漂移→红队挑刺→对标头部。 |
| review_batchC | 批次评审:影响半径增量 + 摘要金字塔跨章反漂移 + 复审铁律触发。chapters=[{id,text}]。 |
| check_consistencyC | 跨章一致性:事实契约硬校验 + 摘要层 LLM 冲突发现。 |
| red_teamB | 红队挑刺:专找逻辑漏洞/动机断裂/硬巧合/读者会弃书的点。 |
| benchmarkC | 对标头部:按结构密度维度打分,输出与头部基线的 gap。 |
| platform_fitC | 平台适配度:按目标平台收稿偏好打分并列出差距。 |
| check_openingC | 黄金开篇:前 300 字主角/冲突/悬念 + 800 字金手指。 |
| check_chapter_endC | 章末钩子判定与强度。 |
| impact_planB | 查看影响半径:哪些章全文重评、哪些只走摘要、哪些复用缓存。零成本,不调 LLM。 |
| check_self_consistencyA | 单章内部自洽弱信号:时序倒退/数字互斥/称谓变体。零成本,只报 suspect 与原文证据,不判 fatal。 |
| check_pacingB | 节奏检查:爽点/冲突/钩子密度 + 节奏荒漠检测。机器层为主,LLM 仅跨章探针。 |
| extract_factsC | 从正文抽硬数据生成候选事实契约(fact_check 兼容)。零 LLM,候选需复核,含跨章同单位多值 suspect。 |
| foreshadow_registerC | 登记伏笔到持久化台账(去重);chapter_index 用于超期计算。 |
| foreshadow_closeC | 回收伏笔:按文本匹配把台账 open 条目标记为 closed。 |
| foreshadow_listC | 查看伏笔台账:status=open|closed|all。 |
| foreshadow_overdueB | 超期未收伏笔提醒;无章序信息时不臆断,返回空。 |
| next_chapter_briefB | 评审回流:把上轮评审结论+未收伏笔+人设锚转成下一章写作约束块(含可直接喂给写稿 Agent 的 prompt)。 |
| cache_statsC | 缓存命中统计与清空能力(clear=false 只看)。 |
| cache_clearA | 清空评审缓存(规则/提示词升级后建议调用)。 |
| blind_review_packageB | P2c 盲评输入包:只含正文+提示词,显式剔除既有评审字段,供独立评审员产出 judgements。 |
| blind_validate_judgementsB | 校验外部盲评 JSON 的结构与评分口径;不补分、不自动换算。 |
| blind_reviewC | P2c:用独立盲评 judgements 回注完整三审流水线,附校验与命中审计;前后清缓存防污染。 |
| blind_compareB | 事后对照原评审与盲评的分数/判定/否决差异;只呈现,不自动改写判定。 |
| list_profilesB | 列出可用平台档案、规则版本与提示词维度。 |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 25 tools
The set has many clearly dimension-specific probes, but several tools overlap: review_chapter already embeds check_redline, red_team, benchmark, and platform_fit, so an agent must decide between the aggregate pipeline and component calls. cache_stats also advertises a clear capability that duplicates cache_clear. Descriptions help but do not fully eliminate these boundary questions.
All names are snake_case and largely follow grouped verb_noun or prefixed conventions (check_*, review_*, foreshadow_*, cache_*, blind_*). A few standalone names are noun phrases (red_team, benchmark, platform_fit, impact_plan) rather than verb-led, and verb choice varies, but the pattern remains readable and casing is consistent.
25 tools is at the heavy end; the domain is genuinely complex, covering redlines, single/batch review, consistency, pacing, foreshadowing, blind review, and caching, so most tools earn a place. Still, the aggregate-vs-component duplication and granular probes make the surface larger than an agent can comfortably navigate, fitting the borderline-heavy range.
Coverage is unusually thorough for novel submission review: compliance, quality, pacing, consistency, foreshadowing lifecycle, platform fit, blind review, and next-chapter brief all have tools. Minor gaps remain: there is no explicit fact-contract review/approve/update operation after extract_facts, and no history/retrieval tool for prior review outputs beyond cache statistics. These are workable gaps.