Repository navigation
Releases
Releases · Zn-Dk/dsh-zhipu-toolkit
dsh-zhipu-toolkit v0.3.0
适配 DSH 0.1.5-rc.1(破坏性变更)
依赖 DSH ≥ 0.1.5-rc.1;settings namespace 与既有配置完全兼容,~/.dsh/settings.yaml 无需改动。
修复 / Fixed
- BigModel 模型调用即崩:rc.1 的
ResolvedPiAiProviderProfile新增必需字段modelErrors(llm-pi-aiadapter 无守卫消费profile.modelErrors.get(model)),本插件构造 profile 时补modelErrors: new Map()。修复前症状:Models 页能列出 BigModel,一选模型调用就Cannot read properties of undefined (reading 'get')。 - 设置卡片迁移到官方 Remote 架构:不再经自建
connection.rpc通道(/zhipu-toolkit-settings)通信,改走 api-gateway 暴露的ctx.remote.settings/ctx.remote.credentials。自建通道在 rc.1 下每个调用都会落到静态资源 fallback 返回 HTTP 405。 - 读取形状修正:
settings.describe()的 value 是{ namespaces: [...] }包装对象、credentials.describe()的 value 是按引用键控的 map(均非裸数组),已按官方dsh-llm-stepfun同款写法修正并写进测试断言。
变更 / Changed
- 移除本机用量统计面板(
usage-stats会话日志聚合):remote 架构下无自定义聚合通道,官方 console 才是权威来源;保留两个官方 console 深链 + 静态计费规则说明。 - 删除
settings-rpc.ts(RPC 桥)与usage-stats.ts及相应单测,净删约 1800 行。Host 半的settings.installSection、双端点路由、实时模型发现、本地 key 模式均无改动。 - 新增
engines.dsh: >=0.1.5-rc.1声明。
验证 / Verification
- host + client
tsc全绿;i18n / esm 自检 OK;vitest 54 passed(9 个既有失败源于 pi-ai 0.82→0.85 上游目录漂移的断言过时,与本版无关)。 - 真机 GUI 验收通过:设置卡片加载、字段显示、mutate 保存、BigModel 模型调用。
dsh-zhipu-toolkit v0.2.1
dsh-zhipu-toolkit v0.2.1
新增 / Added
- 本机用量统计面板:聚合本机 DSH session 日志(多帧 zstd 解码、按文件增量缓存),按模型汇总 token 并按官方系数折算积分;三档窗口切换(5 小时滚动近似 / 今天 / 本周),KMB 缩写、积分降序、近似系数徽章。数据只在本机聚合,绝无外发。
- 设置卡片迁移至「插件」配置 tab:与官方插件卡片同位(settings.plugin.item slot),折叠卡外壳对齐官方视觉;侧栏不再有独立入口。
- glm-5.3-flash 图像输入支持(官方证实原生多模态)。
- 扫描遥测:结果尾部显示「已扫描 N 个会话文件,耗时 X 秒」。
优化 / Changed
- 扫描状态可见化(双语提示可离开页面)+ Host 后台延迟预扫(60s,不与启动抢 CPU)+ 60s 结果缓存。
- 用量排版卡片化(tabular-nums、近似系数徽章);大文件保护(>10MB 跳过)。
npm: npm install dsh-zhipu-toolkit@^0.2.1 · npm 页面
Added (EN)
- Local BigModel usage statistics panel (multi-frame zstd session-log aggregation, per-file incremental cache, official credit factors, 5h/today/week window switcher, KMB tokens). Data stays local.
- Settings card migrated into the official Plugins tab (settings.plugin.item slot, collapsible shell matching host chrome).
- glm-5.3-flash image input support.
- Scan telemetry line ("N session files, X.X s").
dsh-zhipu-toolkit v0.2.0
dsh-zhipu-toolkit v0.2.0
新增 / Added
- 本机 BigModel 路由用量统计:聚合本机 DSH session 日志(多帧 zstd 解码、增量缓存热扫 ~36ms),按模型汇总请求数与 token,按官方系数折算积分(GLM-5.3 6.9/24、Flash 2.3/8 每万 token,其余标注近似);附最近 5 小时窗口。数据只在本机聚合,绝无外发。
- 「用量与额度」设置区块:官方 console 深链按钮 + 静态计费规则说明 + 权威数据免责说明。
- glm-5.3-flash 图像输入支持:官方证实原生多模态,目录条目 input=['text','image'](视频/文件待 pi-ai 类型支持)。
- 只读 usage-stats RPC 端点。
优化 / Changed
- 用量扫描状态可见化(双语提示可离开页面)+ Host 后台预扫 + 按文件增量缓存(重扫 ~36ms)
- 用量数字卡片化排版(tabular-nums 等宽数字、近似系数徽章)
- 大文件保护(>10MB 跳过告警,扫描上限 1500 文件,并发 3)
npm: npm install dsh-zhipu-toolkit@^0.2.0 · npm 页面
Added (EN)
- Local BigModel route usage statistics (multi-frame zstd session-log aggregation, incremental cache ~36ms hot rescan, per-model tokens/requests with official credit factors, 5h window). Data stays local — never sent anywhere.
- "Usage & Quota" settings block with official console deep links and billing rules.
- glm-5.3-flash image input support (officially confirmed native multimodal).
- Read-only usage-stats RPC endpoint.
- Scan visibility, background prescan, card-based layout with tabular numerals and approx-factor badges.
dsh-zhipu-toolkit v0.1.0
dsh-zhipu-toolkit v0.1.0
English | 简体中文
English
First public release. A DeepSeek Harness plugin that registers the Zhipu BigModel GLM family as a model-provider route, covering both BigModel endpoints with one served at a time.
Added
- Dual-endpoint GLM provider route — serve the Coding Plan endpoint (
zai, subscription billing) or the ordinary OpenAPI endpoint (zhipu, per-token billing); both expose identical model lists and parameter behavior, so the choice is purely the billing channel (endpoints: coding | paas, defaultcoding). - Self-maintained GLM catalog — the full GLM chat family (GLM-4.5 → GLM-5.3-Flash) with verified context windows and output caps, overlaid on pi-ai's builtin catalog (map-carrying-descriptor-wins merge).
- Verified GLM-5.3 thinking semantics —
off/minimalremoved from the picker;low/medium/highmap onto thehighwire tier,xhigh/maxontomax; never wiresthinking: disabled(API error 1210). - Live model discovery —
/modelspolled at startup and after every settings/credential change; directly-verified models survive endpoint lag. - Local API key on the settings card — optional local key mode stored in the dsh credentials store (
ZHIPU_TOOLKIT_API_KEY, never in settings.yaml), masked display with edit-in-place / show / clear, env-reference fallback. - Web settings card —
settings.sectioncard over a whitelisted RPC bridge; endpoint choice, credential refs, display name, default reasoning tier, local API key, base-URL overrides; hot-reloadable, bilingual (zh-CN / en).
Install
dsh plugin --profile web add dsh-zhipu-toolkitRestart dsh web after installing.
- npm: https://www.npmjs.com/package/dsh-zhipu-toolkit
- Tarball direct link: https://github.com/Zn-Dk/dsh-zhipu-toolkit/raw/v0.1.0/dsh-zhipu-toolkit-0.1.0.tgz
简体中文
首个公开发布版本。一个 DeepSeek Harness 插件,把智谱 BigModel GLM 全家族注册为模型 provider 路由,双端点一次服务其一。
新增
- 双端点 GLM provider 路由 —— 服务 Coding Plan 端点(
zai,订阅计费)或普通 OpenAPI 端点(zhipu,按 token 计费);两端点模型列表与参数行为完全一致,区别仅在计费渠道(endpoints: coding | paas,默认coding)。 - 自维护 GLM 模型目录 —— 覆盖 GLM 全对话家族(GLM-4.5 → GLM-5.3-Flash),上下文窗口与输出上限均经实测验证,以 map-carrying-descriptor-wins 合并语义叠加在 pi-ai 内置目录之上。
- 已验证 GLM-5.3 思考语义 —— picker 移除
off/minimal;low/medium/high映射到线上high档,xhigh/max解锁max;绝不发送thinking: disabled(API 报错 1210)。 - 实时模型发现 —— 启动时与每次设置/凭证变更后轮询
/models;直连验证过的模型在端点滞后时保留。 - 设置卡片本地 API key —— 可选本地 key 模式,存入 dsh credentials store(
ZHIPU_TOOLKIT_API_KEY,绝不进 settings.yaml),掩码显示 + 原位编辑 / 显示 / 清除,env 引用兜底。 - Web 设置卡片 —— 走白名单 RPC 桥的
settings.section卡片:端点选择、凭证引用、显示名、默认推理档位、本地 API key、base-URL 覆盖;热更新,中英双语。
安装
dsh plugin --profile web add dsh-zhipu-toolkit安装后需重启 dsh web。