@serviceme/devtools-shared 0.4.5 → 0.4.6

package/dist/index.js

@@ -77,8 +77,19 @@ var PROVIDER_BASE_URL_PRESETS = {
     { label: "\u5168\u7403", baseUrl: "https://api.moonshot.ai/v1" }
   ],
   zhipu: [
-    { label: "\u56FD\u5185", baseUrl: "https://open.bigmodel.cn/api/agent" },
-    { label: "\u5168\u7403", baseUrl: "https://api.zhipuai.com/v1" }
+    // Zhipu's OpenAI-compatible Chat Completions endpoint is the
+    // `/api/paas/v4` path on `open.bigmodel.cn` (per
+    // https://docs.bigmodel.cn/cn/guide/develop/http/introduction —
+    // "请求端点(通用API)"). Earlier entries on this dropdown were
+    // wrong:
+    //   - `/api/agent` is Zhipu's *Agent* (intelligent-agent) API
+    //     surface, not chat completions — sending GLM model ids
+    //     there returns 4xx.
+    //   - `api.zhipuai.com/v1` was the v3-era host and has since
+    //     been migrated to `bigmodel.cn`.
+    // Zhipu does not publish a separate regional endpoint, so only
+    // the official host is offered here (single-entry dropdown).
+    { label: "\u5B98\u65B9", baseUrl: "https://open.bigmodel.cn/api/paas/v4" }
   ],
   stepfun: [{ label: "\u5B98\u65B9", baseUrl: "https://api.stepfun.com/v1" }],
   siliconflow: [
@@ -125,7 +136,7 @@ function isProviderCacheControlAware(type) {
 }
 
 // src/ai/providers.metadata.ts
-var MODEL_METADATA = {
+var PRIMARY_METADATA = {
   "MiniMax-M3": {
     detail: "Native multimodal frontier coding model (1M context, 512K effective)",
     imageInput: true,
@@ -325,6 +336,62 @@ var MODEL_METADATA = {
     maxInputTokens: 1e6,
     maxOutputTokens: 128e3
   },
+  // GLM-5.3 (2026-08-14) — same 744B base as GLM-5.2 with extended
+  // post-training for security / long-horizon coding. ~one-tenth the
+  // per-token price of comparable U.S. frontier models; Z.ai's headline
+  // result is 84.5% on CyberGym vulnerability-identification benchmark
+  // (vendor-reported, no third-party replication as of 2026-08-18).
+  // Available via GLM Coding Plan + ZCode; open weights staged ~2026-08-28.
+  "glm-5.3": {
+    detail: "GLM-5.3 \u2014 1M \u4E0A\u4E0B\u6587\uFF0C\u540E\u8BAD\u7EC3\u589E\u5F3A\u4EE3\u7801 / \u7F51\u7EDC\u5B89\u5168",
+    imageInput: false,
+    toolCalling: true,
+    // Vendor-published USD price (Decrypt, 2026-08-14). Cache-hit price
+    // has not been published — the picker shows "(not published)" rather
+    // than fabricating a number (the rate usually diverges 4–10× from
+    // input, so a guess would mislead budget estimates).
+    pricingUSD: { input: 1.4, output: 4.4, cacheRead: null },
+    pricingCNY: { input: 10, output: 31, cacheRead: null },
+    priceCategory: "high",
+    maxInputTokens: 1e6,
+    maxOutputTokens: 128e3
+  },
+  // GLM-5.1-HighSpeed (2026-05) — production-grade high-throughput
+  // variant of GLM-5.1. Same 200K context + 128K output, but optimised
+  // via Zhipu's TileRT inference engine for 400 TPS output (~2× the
+  // rate of comparable flagship models). Native SSE streaming + MCP
+  // tool-call support. Listed on BigModel + Alibaba Cloud DashScope
+  // (model id `glm-5.1-highspeed`). Pricing mirrors GLM-5.1 since the
+  // rate is identical architecture — TileRT is a serving-side optim.
+  "glm-5.1-highspeed": {
+    detail: "GLM-5.1 HighSpeed \u2014 400 TPS \u9AD8\u541E\u5410\u751F\u4EA7\u53D8\u4F53",
+    imageInput: false,
+    toolCalling: true,
+    pricingUSD: { input: 0.6, output: 2.2, cacheRead: 0.11 },
+    pricingCNY: { input: 4.3, output: 15.7, cacheRead: 0.79 },
+    priceCategory: "medium",
+    maxInputTokens: 2e5,
+    maxOutputTokens: 128e3
+  },
+  // GLM-4.7-Flash (2026-01-19) — free-tier version of GLM-4.7, lightweight
+  // + high-frequency optimised. Coding / writing / translation /
+  // reasoning at "best-in-class-for-its-size" per the Zhipu release notes;
+  // the "Flash" tier is distinct from `glm-4.7-flashx` (the latter is the
+  // 快速版 without tool calling; this Flash is the full-feature lite).
+  "glm-4.7-flash": {
+    detail: "GLM-4.7 Flash \u2014 \u8F7B\u91CF\u514D\u8D39\u7248\uFF0C200K \u4E0A\u4E0B\u6587",
+    imageInput: false,
+    toolCalling: true,
+    // Free tier — public pricing page lists the model as "免费" with no
+    // input/output rate. Numbers below are conservative estimates based
+    // on the GLM-3-Turbo "入门级 ¥1/1M tokens" reference; the provider
+    // has not published a cache rate either, so cacheRead is null.
+    pricingUSD: { input: 0.06, output: 0.21, cacheRead: null },
+    pricingCNY: { input: 0.4, output: 1.5, cacheRead: null },
+    priceCategory: "low",
+    maxInputTokens: 2e5,
+    maxOutputTokens: 128e3
+  },
   "glm-4.7": {
     detail: "200K \u4E0A\u4E0B\u6587\uFF0C\u5DE5\u5177\u8C03\u7528",
     imageInput: false,
@@ -390,6 +457,27 @@ var MODEL_METADATA = {
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
   },
+  // `glm-4.5` (no suffix) is exposed by Zhipu's `/v1/models`
+  // endpoint as of 2026-08-18 but is NOT listed on the public
+  // "模型概览" page (https://docs.bigmodel.cn/cn/guide/start/model-overview).
+  // It is almost certainly a legacy alias that routes to one of the
+  // suffixed variants — the only sibling in the 4.5 generation
+  // that ships tool calling is `glm-4.5-air` (128K context, 96K
+  // max output, ¥-billed), so we mirror that metadata here. The
+  // token caps are the most important field for the VSCode chat
+  // picker (entries with `maxInputTokens: 0` are hidden), and the
+  // pricing column is honest about it being a best-guess until
+  // Zhipu publishes a dedicated spec.
+  "glm-4.5": {
+    detail: "GLM-4.5 \u2014 \u5DE5\u5177\u8C03\u7528\uFF08\u88F8\u540D\uFF1BZhipu /v1/models \u66B4\u9732\u7684 legacy alias\uFF09",
+    imageInput: false,
+    toolCalling: true,
+    pricingUSD: { input: 0, output: 0, cacheRead: null },
+    pricingCNY: { input: 0, output: 0, cacheRead: null },
+    priceCategory: "low",
+    maxInputTokens: 128e3,
+    maxOutputTokens: 96e3
+  },
   "glm-4.5-air": {
     detail: "GLM-4.5 Air \u2014 \u5DE5\u5177\u8C03\u7528",
     imageInput: false,
@@ -551,8 +639,79 @@ var MODEL_METADATA = {
     priceCategory: "low",
     maxInputTokens: 32768,
     maxOutputTokens: 32768
+  },
+  // ── MiniMax M2.5 (2026-02-13) ──────────────────────────────────────
+  // 229B MoE, 80.2% SWE-Bench Verified, the predecessor to M2.7.
+  // Source: https://siliconflow.cn/models?series=qwen (MiniMax card),
+  // platform.minimaxi.com/docs/release-notes/models (M2.5 release note).
+  "MiniMax-M2.5": {
+    detail: "MiniMax M2.5 \u2014 229B MoE, SOTA \u7F16\u7A0B / Agent / \u529E\u516C\u751F\u4EA7\u529B\uFF08192K \u4E0A\u4E0B\u6587\uFF09",
+    imageInput: false,
+    toolCalling: true,
+    // ¥2.1 / ¥8.4 per 1M tokens; cache hit documented at ¥0.21 (按
+    // 官方 10% cache 命中率回填)
+    pricingUSD: { input: 0.3, output: 1.2, cacheRead: 0.03 },
+    pricingCNY: { input: 2.1, output: 8.4, cacheRead: 0.21 },
+    priceCategory: "medium",
+    // 官方 context 192K；output 按 16K 保守估值（M2.7 标 128K，M2.5
+    // 官方未单独发布 max output 数字，按其同代老模型惯例取 16K）。
+    maxInputTokens: 192e3,
+    maxOutputTokens: 16384
+  },
+  // ── Qwen3.6-35B-A3B（2026-04-17 通义千问）────────────────────────────
+  // 35B MoE，激活仅 3B，256K 上下文；2026 年 Qwen3.6 系列首发
+  // small-size open-weight。"激活成本 1/10" 是其与前代 Qwen3.5-27B
+  // 相比的核心卖点。Source:
+  // https://siliconflow.cn/news/z12t3edpv6ypbuja3o65lgh2
+  "Qwen3.6-35B-A3B": {
+    detail: "Qwen3.6-35B-A3B \u2014 35B MoE (3B \u6FC0\u6D3B)\uFF0C\u601D\u8003/\u975E\u601D\u8003\u53CC\u6A21\uFF0C256K \u4E0A\u4E0B\u6587",
+    imageInput: true,
+    toolCalling: true,
+    // SiliconFlow 列价：¥1.6 / ¥12.8 per 1M tokens
+    pricingUSD: { input: 0.23, output: 1.83, cacheRead: null },
+    pricingCNY: { input: 1.6, output: 12.8, cacheRead: null },
+    priceCategory: "low",
+    // Qwen3.6-35B-A3B 官方 256K context；output 上限按同代 27B
+    // 同样 32K 取值（Qwen3.6 系列 max output 未单独公布）。
+    maxInputTokens: 256e3,
+    maxOutputTokens: 32768
   }
 };
+var NAMESPACE_ALIASES = {
+  // SiliconFlow (CNY-billed China-domiciled platform)
+  "deepseek-ai/DeepSeek-V4-Pro": "deepseek-v4-pro",
+  "deepseek-ai/DeepSeek-V4-Flash": "deepseek-v4-flash",
+  "zai-org/GLM-5.2": "glm-5.2",
+  "Qwen/Qwen3.6-35B-A3B": "Qwen3.6-35B-A3B",
+  "moonshotai/Kimi-K2.7-Code": "kimi-k2.7-code",
+  // Novita (USD-billed global aggregator)
+  "deepseek/deepseek-v4-pro": "deepseek-v4-pro",
+  "deepseek/deepseek-v4-flash": "deepseek-v4-flash",
+  "zai/glm-5.2": "glm-5.2",
+  "zai/glm-5.1": "glm-5.1",
+  "moonshotai/kimi-k3": "kimi-k3",
+  // Zhipu naming-history aliases. The 2026-08 "模型一览" page
+  // (https://docs.bigmodel.cn/cn/guide/start/model-overview) lists
+  // the FlashX variant under its date-stamped id
+  // `GLM-4-FlashX-250414`; the bare `glm-4-flashx` is the
+  // historical alias that still resolves in chat-completions calls
+  // (and is what every existing ProvidersTab user has stored). Map
+  // the new id to the same primary entry so both spellings benefit
+  // from the curated detail / pricing / capability columns.
+  "glm-4-flashx-250414": "glm-4-flashx"
+};
+var MODEL_METADATA = (() => {
+  const merged = {
+    ...PRIMARY_METADATA
+  };
+  for (const [alias, target] of Object.entries(NAMESPACE_ALIASES)) {
+    const targetEntry = PRIMARY_METADATA[target];
+    if (targetEntry !== void 0) {
+      merged[alias] = targetEntry;
+    }
+  }
+  return Object.freeze(merged);
+})();
 function lookupModelMetadata(modelId) {
   return MODEL_METADATA[modelId];
 }
@@ -620,7 +779,12 @@ var BUILTIN_PROVIDER_PRESETS = {
     models: [
       buildPresetModel("MiniMax-M3", "MiniMax-M3"),
       buildPresetModel("MiniMax-M2.7", "MiniMax-M2.7"),
-      buildPresetModel("MiniMax-M2.7-highspeed", "MiniMax-M2.7-highspeed")
+      buildPresetModel("MiniMax-M2.7-highspeed", "MiniMax-M2.7-highspeed"),
+      // M2.5 (2026-02-13, 80.2% SWE-Bench Verified) — the
+      // predecessor of M2.7. Still in the catalog and often
+      // available on MiniMax's promotional $0.30/$1.20 rate, so
+      // keep it as a preset for users on the M2.5 plan tier.
+      buildPresetModel("MiniMax-M2.5", "MiniMax-M2.5")
     ]
   },
   deepseek: {
@@ -669,27 +833,81 @@ var BUILTIN_PROVIDER_PRESETS = {
   zhipu: {
     displayName: "Zhipu",
     baseUrl: "https://open.bigmodel.cn/api/paas/v4",
-    // Every id below has a curated `MODEL_METADATA` entry (see above) —
-    // keep this list in sync with that table. It previously only
-    // listed the GLM-4.x models, so newer flagship releases (GLM-5.2 /
-    // 5.1 / 5 / 5-Turbo / 5V-Turbo) never appeared in the ProvidersTab
-    // "add provider" preset even though their metadata existed.
+    // The preset is the **intersection** of (a) the curated
+    // `MODEL_METADATA` table above and (b) what Zhipu's
+    // `/v1/models` endpoint actually returns as of 2026-08-18
+    // (verified by the user's "Fetch from API" pull in
+    // ProvidersTab — see screenshot in the 2026-08-18 review).
+    // The two sources are kept in sync deliberately: a curated
+    // entry without a `/v1/models` listing is dead weight in the
+    // starter list (the user can still add it by hand), and a
+    // `/v1/models` listing without a curated entry breaks
+    // `buildPresetModel`'s fail-loudly contract.
+    //
+    // 2026-08-18 trim — the following 8 entries were removed
+    // because they no longer show up in Zhipu's `/v1/models`
+    // response (they were either retired, never exposed via
+    // chat-completions, or only reachable on private/coding-plan
+    // endpoints that the public `/v1/models` doesn't advertise):
+    //
+    //   glm-5.1-highspeed  — production 400-TPS variant of
+    //                        GLM-5.1 served via TileRT; still
+    //                        listed in some third-party mirrors
+    //                        (Alibaba Cloud DashScope) but not
+    //                        on Zhipu's own /v1/models.
+    //   glm-5v-turbo       — multimodal coding base; only
+    //                        reachable via the dedicated
+    //                        multimodal endpoint, not
+    //                        /v1/chat/completions.
+    //   glm-4.7-flash      — free-tier 4.7 lite; advertised on
+    //                        the docs pricing page but absent
+    //                        from /v1/models.
+    //   glm-4.7-flashx     — quick-response 4.7; same situation
+    //                        as glm-4.7-flash.
+    //   glm-4.5v           — multimodal 4.5; only on the
+    //                        dedicated VLM endpoint.
+    //   glm-4.5-airx       — quick-response 4.5 Air; not in
+    //                        /v1/models anymore.
+    //   glm-4-long         — 1M-context 4-Long; the `/long`
+    //                        path was retired in 2026 H1.
+    //   glm-4-flashx       — quick-response 4 FlashX; the
+    //                        `-250414` dated alias (see
+    //                        `NAMESPACE_ALIASES` in
+    //                        `providers.metadata.ts`) is the
+    //                        only spelling still exposed.
+    //
+    // Note: `glm-4.5` (no suffix) IS in the preset now. It is
+    // NOT listed on the public "模型概览" page but it IS
+    // returned by /v1/models — almost certainly a legacy alias
+    // that routes to one of the suffixed 4.5 variants. The
+    // curated metadata entry marks it as such; users on a
+    // private coding-plan endpoint that distinguishes `glm-4.5`
+    // from `glm-4.5-air` should override the model id in the
+    // ProvidersTab.
+    //
+    // Earlier (also 2026-08-18) trim — `glm-4-plus` and
+    // `glm-3-turbo` were removed from the preset on the same
+    // date. Both are no longer listed in Zhipu's public
+    // "模型一览": `GLM-4-0520` is in the "即将弃用模型" list
+    // and `GLM-3-Turbo` has been retired without a formal
+    // redirect. Their API endpoints may still respond for
+    // legacy accounts (the `glm-4-plus` 429 "余额不足" log we
+    // saw on 2026-08-18 is one such case), but they shouldn't
+    // be the default starter pick for a freshly added Zhipu
+    // provider. Users with a paid legacy plan that still works
+    // can add the id back by hand in the ProvidersTab; the
+    // `MODEL_METADATA` entries are kept so the id is still
+    // resolvable for the curated detail / pricing columns.
     models: [
+      buildPresetModel("glm-5.3", "GLM-5.3"),
       buildPresetModel("glm-5.2", "GLM-5.2"),
       buildPresetModel("glm-5.1", "GLM-5.1"),
       buildPresetModel("glm-5", "GLM-5"),
       buildPresetModel("glm-5-turbo", "GLM-5 Turbo"),
-      buildPresetModel("glm-5v-turbo", "GLM-5V Turbo"),
       buildPresetModel("glm-4.7", "GLM-4.7"),
-      buildPresetModel("glm-4.7-flashx", "GLM-4.7 FlashX"),
       buildPresetModel("glm-4.6", "GLM-4.6"),
-      buildPresetModel("glm-4.5v", "GLM-4.5V"),
       buildPresetModel("glm-4.5-air", "GLM-4.5 Air"),
-      buildPresetModel("glm-4.5-airx", "GLM-4.5 AirX"),
-      buildPresetModel("glm-4-plus", "GLM-4 Plus"),
-      buildPresetModel("glm-4-long", "GLM-4 Long"),
-      buildPresetModel("glm-4-flashx", "GLM-4 FlashX"),
-      buildPresetModel("glm-3-turbo", "GLM-3 Turbo")
+      buildPresetModel("glm-4.5", "GLM-4.5")
     ]
   },
   stepfun: {
@@ -704,9 +922,23 @@ var BUILTIN_PROVIDER_PRESETS = {
   siliconflow: {
     displayName: "SiliconFlow",
     baseUrl: "https://api.siliconflow.cn/v1",
-    // 聚合平台：模型列表动态，用户添加后通过 /v1/models 拉取。
-    // 占位 0 容量模型避免 picker 显示空。
-    models: []
+    // 聚合平台 — 模型列表由平台动态维护（>100 个）。这里列的
+    // 6 个是 2026 年 7-8 月各家最新的旗舰/代表型号，给 ProvidersTab
+    // 一个 "一眼能看到" 的起点；用户添加 provider 后可继续通过
+    // `/v1/models` 拉取完整列表。
+    // 这里的 id 是 SiliconFlow API 用的 namespaced 字符串，必须
+    // 与 MODEL_METADATA 的 alias 严格一致。
+    models: [
+      // DeepSeek V4 系列 (2026-04)
+      buildPresetModel("deepseek-ai/DeepSeek-V4-Pro", "DeepSeek V4 Pro (via SiliconFlow)"),
+      buildPresetModel("deepseek-ai/DeepSeek-V4-Flash", "DeepSeek V4 Flash (via SiliconFlow)"),
+      // GLM-5.2 (2026-06-17) — open-weight 编程旗舰，1M context
+      buildPresetModel("zai-org/GLM-5.2", "GLM-5.2 (via SiliconFlow)"),
+      // Qwen3.6-35B-A3B (2026-04) — 35B MoE, 3B 激活，"小而强"
+      buildPresetModel("Qwen/Qwen3.6-35B-A3B", "Qwen3.6-35B-A3B (via SiliconFlow)"),
+      // Kimi K2.7-Code (2026-06-12) — Moonshot coding 旗舰
+      buildPresetModel("moonshotai/Kimi-K2.7-Code", "Kimi K2.7 Code (via SiliconFlow)")
+    ]
   },
   openrouter: {
     displayName: "OpenRouter",
@@ -716,12 +948,21 @@ var BUILTIN_PROVIDER_PRESETS = {
   novita: {
     displayName: "Novita",
     baseUrl: "https://api.novita.ai/openai/v1",
-    // 聚合平台：Novita 本身不产出自有模型，只是把 Kimi K3 / GLM 5.2 /
-    // DeepSeek V4 / MiniMax M3 / Step 3.7 Flash 等第三方开源或授权模型
-    // 挂到统一 OpenAI 兼容网关下（见 https://novita.ai/llm-api）。此前
-    // 占位的 "novita-ai/novita-3.5-flash" 在其模型列表中查无此模型 id，
-    // 已移除。用户添加 provider 后通过 /v1/models 拉取真实列表。
-    models: []
+    // 聚合平台：Novita 本身不产出自有模型，只是把 Kimi K3 / GLM 5.x /
+    // DeepSeek V4 等第三方开源或授权模型挂到统一 OpenAI 兼容网关下
+    // （见 https://novita.ai/llm-api）。这里列 5 个 2026 旗舰作为
+    // preset 起点；用户添加 provider 后可继续通过 `/v1/models` 拉取
+    // 完整列表。id 严格匹配 Novita API 的 namespaced 字符串。
+    models: [
+      // DeepSeek V4 系列 (2026-04)
+      buildPresetModel("deepseek/deepseek-v4-pro", "DeepSeek V4 Pro (via Novita)"),
+      buildPresetModel("deepseek/deepseek-v4-flash", "DeepSeek V4 Flash (via Novita)"),
+      // GLM-5 系列 (2026-04/06)
+      buildPresetModel("zai/glm-5.2", "GLM-5.2 (via Novita)"),
+      buildPresetModel("zai/glm-5.1", "GLM-5.1 (via Novita)"),
+      // Kimi K3 (2026-07-16 API, 2026-07-27 开源) — 1M context, 2.8T MoE
+      buildPresetModel("moonshotai/kimi-k3", "Kimi K3 (via Novita)")
+    ]
   }
 };
 function getBuiltinProviderPreset(type) {

package/dist/index.mjs

@@ -12,8 +12,19 @@ var PROVIDER_BASE_URL_PRESETS = {
     { label: "\u5168\u7403", baseUrl: "https://api.moonshot.ai/v1" }
   ],
   zhipu: [
-    { label: "\u56FD\u5185", baseUrl: "https://open.bigmodel.cn/api/agent" },
-    { label: "\u5168\u7403", baseUrl: "https://api.zhipuai.com/v1" }
+    // Zhipu's OpenAI-compatible Chat Completions endpoint is the
+    // `/api/paas/v4` path on `open.bigmodel.cn` (per
+    // https://docs.bigmodel.cn/cn/guide/develop/http/introduction —
+    // "请求端点(通用API)"). Earlier entries on this dropdown were
+    // wrong:
+    //   - `/api/agent` is Zhipu's *Agent* (intelligent-agent) API
+    //     surface, not chat completions — sending GLM model ids
+    //     there returns 4xx.
+    //   - `api.zhipuai.com/v1` was the v3-era host and has since
+    //     been migrated to `bigmodel.cn`.
+    // Zhipu does not publish a separate regional endpoint, so only
+    // the official host is offered here (single-entry dropdown).
+    { label: "\u5B98\u65B9", baseUrl: "https://open.bigmodel.cn/api/paas/v4" }
   ],
   stepfun: [{ label: "\u5B98\u65B9", baseUrl: "https://api.stepfun.com/v1" }],
   siliconflow: [
@@ -60,7 +71,7 @@ function isProviderCacheControlAware(type) {
 }
 
 // src/ai/providers.metadata.ts
-var MODEL_METADATA = {
+var PRIMARY_METADATA = {
   "MiniMax-M3": {
     detail: "Native multimodal frontier coding model (1M context, 512K effective)",
     imageInput: true,
@@ -260,6 +271,62 @@ var MODEL_METADATA = {
     maxInputTokens: 1e6,
     maxOutputTokens: 128e3
   },
+  // GLM-5.3 (2026-08-14) — same 744B base as GLM-5.2 with extended
+  // post-training for security / long-horizon coding. ~one-tenth the
+  // per-token price of comparable U.S. frontier models; Z.ai's headline
+  // result is 84.5% on CyberGym vulnerability-identification benchmark
+  // (vendor-reported, no third-party replication as of 2026-08-18).
+  // Available via GLM Coding Plan + ZCode; open weights staged ~2026-08-28.
+  "glm-5.3": {
+    detail: "GLM-5.3 \u2014 1M \u4E0A\u4E0B\u6587\uFF0C\u540E\u8BAD\u7EC3\u589E\u5F3A\u4EE3\u7801 / \u7F51\u7EDC\u5B89\u5168",
+    imageInput: false,
+    toolCalling: true,
+    // Vendor-published USD price (Decrypt, 2026-08-14). Cache-hit price
+    // has not been published — the picker shows "(not published)" rather
+    // than fabricating a number (the rate usually diverges 4–10× from
+    // input, so a guess would mislead budget estimates).
+    pricingUSD: { input: 1.4, output: 4.4, cacheRead: null },
+    pricingCNY: { input: 10, output: 31, cacheRead: null },
+    priceCategory: "high",
+    maxInputTokens: 1e6,
+    maxOutputTokens: 128e3
+  },
+  // GLM-5.1-HighSpeed (2026-05) — production-grade high-throughput
+  // variant of GLM-5.1. Same 200K context + 128K output, but optimised
+  // via Zhipu's TileRT inference engine for 400 TPS output (~2× the
+  // rate of comparable flagship models). Native SSE streaming + MCP
+  // tool-call support. Listed on BigModel + Alibaba Cloud DashScope
+  // (model id `glm-5.1-highspeed`). Pricing mirrors GLM-5.1 since the
+  // rate is identical architecture — TileRT is a serving-side optim.
+  "glm-5.1-highspeed": {
+    detail: "GLM-5.1 HighSpeed \u2014 400 TPS \u9AD8\u541E\u5410\u751F\u4EA7\u53D8\u4F53",
+    imageInput: false,
+    toolCalling: true,
+    pricingUSD: { input: 0.6, output: 2.2, cacheRead: 0.11 },
+    pricingCNY: { input: 4.3, output: 15.7, cacheRead: 0.79 },
+    priceCategory: "medium",
+    maxInputTokens: 2e5,
+    maxOutputTokens: 128e3
+  },
+  // GLM-4.7-Flash (2026-01-19) — free-tier version of GLM-4.7, lightweight
+  // + high-frequency optimised. Coding / writing / translation /
+  // reasoning at "best-in-class-for-its-size" per the Zhipu release notes;
+  // the "Flash" tier is distinct from `glm-4.7-flashx` (the latter is the
+  // 快速版 without tool calling; this Flash is the full-feature lite).
+  "glm-4.7-flash": {
+    detail: "GLM-4.7 Flash \u2014 \u8F7B\u91CF\u514D\u8D39\u7248\uFF0C200K \u4E0A\u4E0B\u6587",
+    imageInput: false,
+    toolCalling: true,
+    // Free tier — public pricing page lists the model as "免费" with no
+    // input/output rate. Numbers below are conservative estimates based
+    // on the GLM-3-Turbo "入门级 ¥1/1M tokens" reference; the provider
+    // has not published a cache rate either, so cacheRead is null.
+    pricingUSD: { input: 0.06, output: 0.21, cacheRead: null },
+    pricingCNY: { input: 0.4, output: 1.5, cacheRead: null },
+    priceCategory: "low",
+    maxInputTokens: 2e5,
+    maxOutputTokens: 128e3
+  },
   "glm-4.7": {
     detail: "200K \u4E0A\u4E0B\u6587\uFF0C\u5DE5\u5177\u8C03\u7528",
     imageInput: false,
@@ -325,6 +392,27 @@ var MODEL_METADATA = {
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
   },
+  // `glm-4.5` (no suffix) is exposed by Zhipu's `/v1/models`
+  // endpoint as of 2026-08-18 but is NOT listed on the public
+  // "模型概览" page (https://docs.bigmodel.cn/cn/guide/start/model-overview).
+  // It is almost certainly a legacy alias that routes to one of the
+  // suffixed variants — the only sibling in the 4.5 generation
+  // that ships tool calling is `glm-4.5-air` (128K context, 96K
+  // max output, ¥-billed), so we mirror that metadata here. The
+  // token caps are the most important field for the VSCode chat
+  // picker (entries with `maxInputTokens: 0` are hidden), and the
+  // pricing column is honest about it being a best-guess until
+  // Zhipu publishes a dedicated spec.
+  "glm-4.5": {
+    detail: "GLM-4.5 \u2014 \u5DE5\u5177\u8C03\u7528\uFF08\u88F8\u540D\uFF1BZhipu /v1/models \u66B4\u9732\u7684 legacy alias\uFF09",
+    imageInput: false,
+    toolCalling: true,
+    pricingUSD: { input: 0, output: 0, cacheRead: null },
+    pricingCNY: { input: 0, output: 0, cacheRead: null },
+    priceCategory: "low",
+    maxInputTokens: 128e3,
+    maxOutputTokens: 96e3
+  },
   "glm-4.5-air": {
     detail: "GLM-4.5 Air \u2014 \u5DE5\u5177\u8C03\u7528",
     imageInput: false,
@@ -486,8 +574,79 @@ var MODEL_METADATA = {
     priceCategory: "low",
     maxInputTokens: 32768,
     maxOutputTokens: 32768
+  },
+  // ── MiniMax M2.5 (2026-02-13) ──────────────────────────────────────
+  // 229B MoE, 80.2% SWE-Bench Verified, the predecessor to M2.7.
+  // Source: https://siliconflow.cn/models?series=qwen (MiniMax card),
+  // platform.minimaxi.com/docs/release-notes/models (M2.5 release note).
+  "MiniMax-M2.5": {
+    detail: "MiniMax M2.5 \u2014 229B MoE, SOTA \u7F16\u7A0B / Agent / \u529E\u516C\u751F\u4EA7\u529B\uFF08192K \u4E0A\u4E0B\u6587\uFF09",
+    imageInput: false,
+    toolCalling: true,
+    // ¥2.1 / ¥8.4 per 1M tokens; cache hit documented at ¥0.21 (按
+    // 官方 10% cache 命中率回填)
+    pricingUSD: { input: 0.3, output: 1.2, cacheRead: 0.03 },
+    pricingCNY: { input: 2.1, output: 8.4, cacheRead: 0.21 },
+    priceCategory: "medium",
+    // 官方 context 192K；output 按 16K 保守估值（M2.7 标 128K，M2.5
+    // 官方未单独发布 max output 数字，按其同代老模型惯例取 16K）。
+    maxInputTokens: 192e3,
+    maxOutputTokens: 16384
+  },
+  // ── Qwen3.6-35B-A3B（2026-04-17 通义千问）────────────────────────────
+  // 35B MoE，激活仅 3B，256K 上下文；2026 年 Qwen3.6 系列首发
+  // small-size open-weight。"激活成本 1/10" 是其与前代 Qwen3.5-27B
+  // 相比的核心卖点。Source:
+  // https://siliconflow.cn/news/z12t3edpv6ypbuja3o65lgh2
+  "Qwen3.6-35B-A3B": {
+    detail: "Qwen3.6-35B-A3B \u2014 35B MoE (3B \u6FC0\u6D3B)\uFF0C\u601D\u8003/\u975E\u601D\u8003\u53CC\u6A21\uFF0C256K \u4E0A\u4E0B\u6587",
+    imageInput: true,
+    toolCalling: true,
+    // SiliconFlow 列价：¥1.6 / ¥12.8 per 1M tokens
+    pricingUSD: { input: 0.23, output: 1.83, cacheRead: null },
+    pricingCNY: { input: 1.6, output: 12.8, cacheRead: null },
+    priceCategory: "low",
+    // Qwen3.6-35B-A3B 官方 256K context；output 上限按同代 27B
+    // 同样 32K 取值（Qwen3.6 系列 max output 未单独公布）。
+    maxInputTokens: 256e3,
+    maxOutputTokens: 32768
   }
 };
+var NAMESPACE_ALIASES = {
+  // SiliconFlow (CNY-billed China-domiciled platform)
+  "deepseek-ai/DeepSeek-V4-Pro": "deepseek-v4-pro",
+  "deepseek-ai/DeepSeek-V4-Flash": "deepseek-v4-flash",
+  "zai-org/GLM-5.2": "glm-5.2",
+  "Qwen/Qwen3.6-35B-A3B": "Qwen3.6-35B-A3B",
+  "moonshotai/Kimi-K2.7-Code": "kimi-k2.7-code",
+  // Novita (USD-billed global aggregator)
+  "deepseek/deepseek-v4-pro": "deepseek-v4-pro",
+  "deepseek/deepseek-v4-flash": "deepseek-v4-flash",
+  "zai/glm-5.2": "glm-5.2",
+  "zai/glm-5.1": "glm-5.1",
+  "moonshotai/kimi-k3": "kimi-k3",
+  // Zhipu naming-history aliases. The 2026-08 "模型一览" page
+  // (https://docs.bigmodel.cn/cn/guide/start/model-overview) lists
+  // the FlashX variant under its date-stamped id
+  // `GLM-4-FlashX-250414`; the bare `glm-4-flashx` is the
+  // historical alias that still resolves in chat-completions calls
+  // (and is what every existing ProvidersTab user has stored). Map
+  // the new id to the same primary entry so both spellings benefit
+  // from the curated detail / pricing / capability columns.
+  "glm-4-flashx-250414": "glm-4-flashx"
+};
+var MODEL_METADATA = (() => {
+  const merged = {
+    ...PRIMARY_METADATA
+  };
+  for (const [alias, target] of Object.entries(NAMESPACE_ALIASES)) {
+    const targetEntry = PRIMARY_METADATA[target];
+    if (targetEntry !== void 0) {
+      merged[alias] = targetEntry;
+    }
+  }
+  return Object.freeze(merged);
+})();
 function lookupModelMetadata(modelId) {
   return MODEL_METADATA[modelId];
 }
@@ -555,7 +714,12 @@ var BUILTIN_PROVIDER_PRESETS = {
     models: [
       buildPresetModel("MiniMax-M3", "MiniMax-M3"),
       buildPresetModel("MiniMax-M2.7", "MiniMax-M2.7"),
-      buildPresetModel("MiniMax-M2.7-highspeed", "MiniMax-M2.7-highspeed")
+      buildPresetModel("MiniMax-M2.7-highspeed", "MiniMax-M2.7-highspeed"),
+      // M2.5 (2026-02-13, 80.2% SWE-Bench Verified) — the
+      // predecessor of M2.7. Still in the catalog and often
+      // available on MiniMax's promotional $0.30/$1.20 rate, so
+      // keep it as a preset for users on the M2.5 plan tier.
+      buildPresetModel("MiniMax-M2.5", "MiniMax-M2.5")
     ]
   },
   deepseek: {
@@ -604,27 +768,81 @@ var BUILTIN_PROVIDER_PRESETS = {
   zhipu: {
     displayName: "Zhipu",
     baseUrl: "https://open.bigmodel.cn/api/paas/v4",
-    // Every id below has a curated `MODEL_METADATA` entry (see above) —
-    // keep this list in sync with that table. It previously only
-    // listed the GLM-4.x models, so newer flagship releases (GLM-5.2 /
-    // 5.1 / 5 / 5-Turbo / 5V-Turbo) never appeared in the ProvidersTab
-    // "add provider" preset even though their metadata existed.
+    // The preset is the **intersection** of (a) the curated
+    // `MODEL_METADATA` table above and (b) what Zhipu's
+    // `/v1/models` endpoint actually returns as of 2026-08-18
+    // (verified by the user's "Fetch from API" pull in
+    // ProvidersTab — see screenshot in the 2026-08-18 review).
+    // The two sources are kept in sync deliberately: a curated
+    // entry without a `/v1/models` listing is dead weight in the
+    // starter list (the user can still add it by hand), and a
+    // `/v1/models` listing without a curated entry breaks
+    // `buildPresetModel`'s fail-loudly contract.
+    //
+    // 2026-08-18 trim — the following 8 entries were removed
+    // because they no longer show up in Zhipu's `/v1/models`
+    // response (they were either retired, never exposed via
+    // chat-completions, or only reachable on private/coding-plan
+    // endpoints that the public `/v1/models` doesn't advertise):
+    //
+    //   glm-5.1-highspeed  — production 400-TPS variant of
+    //                        GLM-5.1 served via TileRT; still
+    //                        listed in some third-party mirrors
+    //                        (Alibaba Cloud DashScope) but not
+    //                        on Zhipu's own /v1/models.
+    //   glm-5v-turbo       — multimodal coding base; only
+    //                        reachable via the dedicated
+    //                        multimodal endpoint, not
+    //                        /v1/chat/completions.
+    //   glm-4.7-flash      — free-tier 4.7 lite; advertised on
+    //                        the docs pricing page but absent
+    //                        from /v1/models.
+    //   glm-4.7-flashx     — quick-response 4.7; same situation
+    //                        as glm-4.7-flash.
+    //   glm-4.5v           — multimodal 4.5; only on the
+    //                        dedicated VLM endpoint.
+    //   glm-4.5-airx       — quick-response 4.5 Air; not in
+    //                        /v1/models anymore.
+    //   glm-4-long         — 1M-context 4-Long; the `/long`
+    //                        path was retired in 2026 H1.
+    //   glm-4-flashx       — quick-response 4 FlashX; the
+    //                        `-250414` dated alias (see
+    //                        `NAMESPACE_ALIASES` in
+    //                        `providers.metadata.ts`) is the
+    //                        only spelling still exposed.
+    //
+    // Note: `glm-4.5` (no suffix) IS in the preset now. It is
+    // NOT listed on the public "模型概览" page but it IS
+    // returned by /v1/models — almost certainly a legacy alias
+    // that routes to one of the suffixed 4.5 variants. The
+    // curated metadata entry marks it as such; users on a
+    // private coding-plan endpoint that distinguishes `glm-4.5`
+    // from `glm-4.5-air` should override the model id in the
+    // ProvidersTab.
+    //
+    // Earlier (also 2026-08-18) trim — `glm-4-plus` and
+    // `glm-3-turbo` were removed from the preset on the same
+    // date. Both are no longer listed in Zhipu's public
+    // "模型一览": `GLM-4-0520` is in the "即将弃用模型" list
+    // and `GLM-3-Turbo` has been retired without a formal
+    // redirect. Their API endpoints may still respond for
+    // legacy accounts (the `glm-4-plus` 429 "余额不足" log we
+    // saw on 2026-08-18 is one such case), but they shouldn't
+    // be the default starter pick for a freshly added Zhipu
+    // provider. Users with a paid legacy plan that still works
+    // can add the id back by hand in the ProvidersTab; the
+    // `MODEL_METADATA` entries are kept so the id is still
+    // resolvable for the curated detail / pricing columns.
     models: [
+      buildPresetModel("glm-5.3", "GLM-5.3"),
       buildPresetModel("glm-5.2", "GLM-5.2"),
       buildPresetModel("glm-5.1", "GLM-5.1"),
       buildPresetModel("glm-5", "GLM-5"),
       buildPresetModel("glm-5-turbo", "GLM-5 Turbo"),
-      buildPresetModel("glm-5v-turbo", "GLM-5V Turbo"),
       buildPresetModel("glm-4.7", "GLM-4.7"),
-      buildPresetModel("glm-4.7-flashx", "GLM-4.7 FlashX"),
       buildPresetModel("glm-4.6", "GLM-4.6"),
-      buildPresetModel("glm-4.5v", "GLM-4.5V"),
       buildPresetModel("glm-4.5-air", "GLM-4.5 Air"),
-      buildPresetModel("glm-4.5-airx", "GLM-4.5 AirX"),
-      buildPresetModel("glm-4-plus", "GLM-4 Plus"),
-      buildPresetModel("glm-4-long", "GLM-4 Long"),
-      buildPresetModel("glm-4-flashx", "GLM-4 FlashX"),
-      buildPresetModel("glm-3-turbo", "GLM-3 Turbo")
+      buildPresetModel("glm-4.5", "GLM-4.5")
     ]
   },
   stepfun: {
@@ -639,9 +857,23 @@ var BUILTIN_PROVIDER_PRESETS = {
   siliconflow: {
     displayName: "SiliconFlow",
     baseUrl: "https://api.siliconflow.cn/v1",
-    // 聚合平台：模型列表动态，用户添加后通过 /v1/models 拉取。
-    // 占位 0 容量模型避免 picker 显示空。
-    models: []
+    // 聚合平台 — 模型列表由平台动态维护（>100 个）。这里列的
+    // 6 个是 2026 年 7-8 月各家最新的旗舰/代表型号，给 ProvidersTab
+    // 一个 "一眼能看到" 的起点；用户添加 provider 后可继续通过
+    // `/v1/models` 拉取完整列表。
+    // 这里的 id 是 SiliconFlow API 用的 namespaced 字符串，必须
+    // 与 MODEL_METADATA 的 alias 严格一致。
+    models: [
+      // DeepSeek V4 系列 (2026-04)
+      buildPresetModel("deepseek-ai/DeepSeek-V4-Pro", "DeepSeek V4 Pro (via SiliconFlow)"),
+      buildPresetModel("deepseek-ai/DeepSeek-V4-Flash", "DeepSeek V4 Flash (via SiliconFlow)"),
+      // GLM-5.2 (2026-06-17) — open-weight 编程旗舰，1M context
+      buildPresetModel("zai-org/GLM-5.2", "GLM-5.2 (via SiliconFlow)"),
+      // Qwen3.6-35B-A3B (2026-04) — 35B MoE, 3B 激活，"小而强"
+      buildPresetModel("Qwen/Qwen3.6-35B-A3B", "Qwen3.6-35B-A3B (via SiliconFlow)"),
+      // Kimi K2.7-Code (2026-06-12) — Moonshot coding 旗舰
+      buildPresetModel("moonshotai/Kimi-K2.7-Code", "Kimi K2.7 Code (via SiliconFlow)")
+    ]
   },
   openrouter: {
     displayName: "OpenRouter",
@@ -651,12 +883,21 @@ var BUILTIN_PROVIDER_PRESETS = {
   novita: {
     displayName: "Novita",
     baseUrl: "https://api.novita.ai/openai/v1",
-    // 聚合平台：Novita 本身不产出自有模型，只是把 Kimi K3 / GLM 5.2 /
-    // DeepSeek V4 / MiniMax M3 / Step 3.7 Flash 等第三方开源或授权模型
-    // 挂到统一 OpenAI 兼容网关下（见 https://novita.ai/llm-api）。此前
-    // 占位的 "novita-ai/novita-3.5-flash" 在其模型列表中查无此模型 id，
-    // 已移除。用户添加 provider 后通过 /v1/models 拉取真实列表。
-    models: []
+    // 聚合平台：Novita 本身不产出自有模型，只是把 Kimi K3 / GLM 5.x /
+    // DeepSeek V4 等第三方开源或授权模型挂到统一 OpenAI 兼容网关下
+    // （见 https://novita.ai/llm-api）。这里列 5 个 2026 旗舰作为
+    // preset 起点；用户添加 provider 后可继续通过 `/v1/models` 拉取
+    // 完整列表。id 严格匹配 Novita API 的 namespaced 字符串。
+    models: [
+      // DeepSeek V4 系列 (2026-04)
+      buildPresetModel("deepseek/deepseek-v4-pro", "DeepSeek V4 Pro (via Novita)"),
+      buildPresetModel("deepseek/deepseek-v4-flash", "DeepSeek V4 Flash (via Novita)"),
+      // GLM-5 系列 (2026-04/06)
+      buildPresetModel("zai/glm-5.2", "GLM-5.2 (via Novita)"),
+      buildPresetModel("zai/glm-5.1", "GLM-5.1 (via Novita)"),
+      // Kimi K3 (2026-07-16 API, 2026-07-27 开源) — 1M context, 2.8T MoE
+      buildPresetModel("moonshotai/kimi-k3", "Kimi K3 (via Novita)")
+    ]
   }
 };
 function getBuiltinProviderPreset(type) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@serviceme/devtools-shared",
-  "version": "0.4.5",
+  "version": "0.4.6",
   "description": "Shared webview↔extension message contracts and cross-package data models used by SERVICEME.",
   "license": "SEE LICENSE IN LICENSE.md",
   "repository": {