@serviceme/devtools-shared 0.4.5 → 0.4.7

package/dist/index.mjs

@@ -1,3 +1,19 @@
+// src/ai/protocol.ts
+function protocolForBaseUrl(baseUrl) {
+  try {
+    const url = new URL(baseUrl);
+    return /(^|\/)anthropic(\/|$)/i.test(url.pathname) ? "anthropic" : "openai";
+  } catch {
+    return "openai";
+  }
+}
+function effectiveAdapterType(configuredType, baseUrl) {
+  if (configuredType === "zhipu" && protocolForBaseUrl(baseUrl) === "anthropic") {
+    return "anthropic-compatible";
+  }
+  return configuredType;
+}
+
 // src/ai/providers.base-url.ts
 var PROVIDER_BASE_URL_PRESETS = {
   "openai-compatible": [],
@@ -12,8 +28,48 @@ 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 / 智谱 GLM 6 endpoint paths × 2 hosts. The 4
+    // "credential channels" the GLM-for-copilot reference
+    // distinguishes (region × apiMode — each with its own API
+    // key) collapse to a 6-row baseUrl dropdown here because we
+    // keep one API key per provider, not one per channel. The
+    // user picks the host + path that matches the API key
+    // they actually have; the curated `MODEL_METADATA` prices
+    // are host-based (CNY vs USD via `currencyForBaseUrl`).
+    //
+    // Source: https://bigmodel.cn/pricing (CN platform, CNY) +
+    // https://z.ai/pricing (international, USD). The 6 paths
+    // map to:
+    //   - `/api/paas/v4`          → 标准 API (Standard)
+    //   - `/api/coding/paas/v4`   → Coding Plan (订阅套餐)
+    //   - `/api/anthropic`        → Anthropic 兼容协议
+    //
+    // Earlier single-entry dropdown omitted the Coding Plan
+    // path and the international Z.ai host entirely — users on
+    // the Coding Plan subscription were 404'ing because they
+    // pasted `open.bigmodel.cn/api/paas/v4` into a Coding Plan
+    // key, and Z.ai users had no preset to pick.
+    // ── 国内 (open.bigmodel.cn — CNY) ───────────────────────
+    { label: "\u56FD\u5185 \xB7 \u6807\u51C6 API", baseUrl: "https://open.bigmodel.cn/api/paas/v4" },
+    { label: "\u56FD\u5185 \xB7 Coding Plan", baseUrl: "https://open.bigmodel.cn/api/coding/paas/v4" },
+    // 重要：Anthropic 协议端点 **复用** Coding Plan 的 key（与
+    // `/api/coding/paas/v4` 共用同一凭证；不是 Standard API key，
+    // 也没有独立的 "Anthropic API key"）。
+    // Reference: `docs/references/GLM-for-copilot-main/src/i18n.ts:578-579`
+    //   "Coding Plan and Standard API credentials are independent.
+    //    OpenAI and Anthropic endpoints in the same region share
+    //    the Coding Plan key."
+    // label 故意重复 "Coding Plan" 两次，让用户从下拉里一眼看出：
+    //   (a) 这个端点**只能配 Coding Plan key**；
+    //   (b) 这是 Coding Plan 的**协议变体**，不是 Standard API 的。
+    {
+      label: "\u56FD\u5185 \xB7 Coding Plan \xB7 Anthropic \u534F\u8BAE",
+      baseUrl: "https://open.bigmodel.cn/api/anthropic"
+    },
+    // ── 国际 (api.z.ai — USD) ───────────────────────────────
+    { label: "\u56FD\u9645 \xB7 \u6807\u51C6 API", baseUrl: "https://api.z.ai/api/paas/v4" },
+    { label: "\u56FD\u9645 \xB7 Coding Plan", baseUrl: "https://api.z.ai/api/coding/paas/v4" },
+    { label: "\u56FD\u9645 \xB7 Coding Plan \xB7 Anthropic \u534F\u8BAE", baseUrl: "https://api.z.ai/api/anthropic" }
   ],
   stepfun: [{ label: "\u5B98\u65B9", baseUrl: "https://api.stepfun.com/v1" }],
   siliconflow: [
@@ -60,7 +116,7 @@ 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,
@@ -93,8 +149,17 @@ var MODEL_METADATA = {
     detail: "M2.7 high-speed: same quality, faster (~100 TPS)",
     imageInput: false,
     toolCalling: true,
+    // Per https://minimax-ai.chat/pricing (2026-07 verified):
+    // M2.7-highspeed is 2× M2.7 base on input/output, but
+    // identical on cache hit (same model + same infra, just a
+    // serving-side TPS bump). The CNY values mirror USD at the
+    // project's 1:7 CNY-per-USD convention.
+    // Previously the CNY input/output were the same as base
+    // (¥2.1 / ¥8.4) while USD was already 2× — that left the
+    // USD/CNY ratio at 3.5× instead of 7× and silently
+    // under-reported CNY cost for users on the China platform.
     pricingUSD: { input: 0.6, output: 2.4, cacheRead: 0.06 },
-    pricingCNY: { input: 2.1, output: 8.4, cacheRead: 0.42 },
+    pricingCNY: { input: 4.2, output: 16.8, cacheRead: 0.42 },
     priceCategory: "low",
     // Inherits M2.7's context window.
     maxInputTokens: 131072,
@@ -104,8 +169,20 @@ var MODEL_METADATA = {
     detail: "Fast, general-purpose model",
     imageInput: true,
     toolCalling: true,
-    pricingUSD: { input: 0.14, output: 0.28, cacheRead: 28e-4 },
-    pricingCNY: { input: 1, output: 2, cacheRead: 0.02 },
+    // Per https://api-docs.deepseek.com/quick_start/pricing/ and
+    // https://new.qq.com/rain/a/20260813A0DROS00 (2026-08-13
+    // announcement, effective 2026-08-17 00:00 Beijing): peak/
+    // off-peak tiered pricing. Pinned the OFF-PEAK rate since
+    // peak hours (01:00–04:00 + 06:00–10:00 UTC = 09:00–12:00 +
+    // 14:00–18:00 Beijing) cover only 8 of 24 hours — most chat
+    // sessions land off-peak. Peak is exactly 2× off-peak per
+    // the official page.
+    //   Off-peak: $0.007 cache hit / $0.22 input / $0.66 output
+    //             ¥0.05 cache hit / ¥1.5 input / ¥4.5 output
+    // The pre-2026-08-17 rate was 1/3 of the current off-peak;
+    // 8/17 调价 raised cache hit 6× and output 2.25×.
+    pricingUSD: { input: 0.22, output: 0.66, cacheRead: 7e-3 },
+    pricingCNY: { input: 1.5, output: 4.5, cacheRead: 0.05 },
     priceCategory: "low",
     // Official docs (api-docs.deepseek.com/quick_start/pricing, fetched
     // 2026-07-27): "THINKING MODE: Supports both non-thinking and
@@ -121,8 +198,20 @@ var MODEL_METADATA = {
     detail: "Most capable reasoning model",
     imageInput: true,
     toolCalling: true,
-    pricingUSD: { input: 0.435, output: 0.87, cacheRead: 3625e-6 },
-    pricingCNY: { input: 2.1, output: 4.2, cacheRead: 0.025 },
+    // Per https://api-docs.deepseek.com/quick_start/pricing/ and
+    // https://new.qq.com/rain/a/20260813A0DROS00 (2026-08-13
+    // announcement, effective 2026-08-17 00:00 Beijing): peak/
+    // off-peak tiered pricing. Pinned the OFF-PEAK rate (peak
+    // hours are 8/24; most chat sessions land off-peak; peak is
+    // exactly 2× off-peak per the official page).
+    //   Off-peak: $0.022 cache hit / $0.66 input / $1.98 output
+    //             ¥0.15 cache hit / ¥4.5 input / ¥13.5 output
+    // The 8/17 调价 raised cache hit 6× (¥0.025 → ¥0.15),
+    // input 1.5×, and output 2.25×. The pre-08-17 USD values
+    // (0.435/0.87/0.003625) and CNY values (2.1/4.2/0.025) did
+    // not correspond to any DeepSeek-published rate; corrected.
+    pricingUSD: { input: 0.66, output: 1.98, cacheRead: 0.022 },
+    pricingCNY: { input: 4.5, output: 13.5, cacheRead: 0.15 },
     priceCategory: "low",
     thinkingSchema: "thinkingEnabled",
     maxInputTokens: 655360,
@@ -250,56 +339,133 @@ var MODEL_METADATA = {
   // GLM-5's explicit "Agentic 长程规划与执行" description. Corrected
   // from false (inconsistent with the rest of the GLM-5 family).
   "glm-5.2": {
-    detail: "GLM-5.2 \u2014 1M \u4E0A\u4E0B\u6587\uFF0C\u6700\u5927\u8F93\u51FA 128K",
+    detail: "GLM-5.2 \u2014 1M \u4E0A\u4E0B\u6587\uFF0C\u6700\u5927\u8F93\u51FA 128K\uFF08\u5355\u6863 pricing\uFF09",
     imageInput: false,
     toolCalling: true,
-    // ¥8 input / ¥28 output / ¥2 cache hit per 1M tokens (输入长度 32K+ 档)
+    // Single rate (no input-length tier split) per
+    // bigmodel.cn/pricing 2026-08-18:
+    //   ¥8 input / ¥28 output / ¥2 cache hit per 1M tokens
+    // The previous entry's comment said "输入长度 32K+ 档" — that
+    // was wrong: GLM-5.2 has no tier split on the official page.
     pricingUSD: { input: 1.12, output: 3.92, cacheRead: 0.28 },
     pricingCNY: { input: 8, output: 28, cacheRead: 2 },
     priceCategory: "high",
     maxInputTokens: 1e6,
     maxOutputTokens: 128e3
   },
-  "glm-4.7": {
-    detail: "200K \u4E0A\u4E0B\u6587\uFF0C\u5DE5\u5177\u8C03\u7528",
+  // 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\uFF08[0, 32K) tier \u955C\u50CF GLM-5.1\uFF09",
     imageInput: false,
     toolCalling: true,
-    // cc-switch 标价：$0.6 input / $2.2 output
+    // Mirrors GLM-5.1 [0, 32K) tier per bigmodel.cn/pricing 2026-08-18.
+    // TileRT is a serving-side optim; the per-token rate is the same
+    // architecture as the base model.
     pricingUSD: { input: 0.6, output: 2.2, cacheRead: 0.11 },
-    pricingCNY: { input: 4.3, output: 15.7, cacheRead: 0.79 },
+    pricingCNY: { input: 6, output: 24, cacheRead: 1.3 },
+    priceCategory: "medium",
+    maxInputTokens: 2e5,
+    maxOutputTokens: 128e3
+  },
+  // GLM-4.7-Flash (2026-01-19) — 200K context, fully-free tier on
+  // bigmodel.cn/pricing (fetched 2026-08-18). Lightweight + high-
+  // frequency optimised; coding / writing / translation / reasoning
+  // at "best-in-class-for-its-size" per the Zhipu release notes.
+  // Distinct from `glm-4.7-flashx` (the latter is the 快速版 with
+  // paid pricing).
+  "glm-4.7-flash": {
+    detail: "GLM-4.7 Flash \u2014 \u5B8C\u5168\u514D\u8D39\uFF08200K \u4E0A\u4E0B\u6587\uFF09",
+    imageInput: false,
+    toolCalling: true,
+    // Free tier — input / output / cache hit all 0 (bigmodel.cn
+    // 2026-08-18 lists "免费" for every column). USD mirrors CNY
+    // rather than inventing a rate.
+    pricingUSD: { input: 0, output: 0, cacheRead: 0 },
+    pricingCNY: { input: 0, output: 0, cacheRead: 0 },
+    priceCategory: "low",
+    maxInputTokens: 2e5,
+    maxOutputTokens: 128e3
+  },
+  "glm-4.7": {
+    detail: "GLM-4.7 \u2014 200K \u4E0A\u4E0B\u6587\uFF0C\u5DE5\u5177\u8C03\u7528\uFF083-tier pricing\uFF09",
+    imageInput: false,
+    toolCalling: true,
+    // Pinned the LOWEST tier per bigmodel.cn/pricing 2026-08-18:
+    //   [0, 32K) input × [0, 0.2K) output — ¥2 / ¥8 / ¥0.4 cache hit
+    //   [0, 32K) input × [0.2K+) output — ¥3 / ¥14 / ¥0.6 cache hit
+    //   [32K, 200K) input — ¥4 / ¥16 / ¥0.8 cache hit
+    // Most real prompts are < 32K input and < 0.2K output, so the
+    // lowest tier is the most representative per-request price.
+    // Re-pick from a higher tier if the picker adds a length slider.
+    pricingUSD: { input: 0.28, output: 1.12, cacheRead: 0.056 },
+    pricingCNY: { input: 2, output: 8, cacheRead: 0.4 },
     priceCategory: "medium",
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
   },
   "glm-5.1": {
-    detail: "GLM-5.1 \u2014 200K \u4E0A\u4E0B\u6587\uFF0C\u6700\u5927\u8F93\u51FA 128K",
+    detail: "GLM-5.1 \u2014 200K \u4E0A\u4E0B\u6587\uFF0C\u6700\u5927\u8F93\u51FA 128K\uFF082-tier pricing\uFF09",
     imageInput: false,
     toolCalling: true,
-    // ¥8 input / ¥28 output / ¥2 cache hit per 1M tokens (输入长度 32K+ 档)
-    pricingUSD: { input: 1.12, output: 3.92, cacheRead: 0.28 },
-    pricingCNY: { input: 8, output: 28, cacheRead: 2 },
+    // Pinned the LOWER tier per bigmodel.cn/pricing 2026-08-18:
+    //   [0, 32K)  — ¥6 input / ¥24 output / ¥1.3 cache hit
+    //   [32K+)     — ¥8 input / ¥28 output / ¥2   cache hit
+    // The previous entry had the higher tier; switched to the lower
+    // tier so a typical < 32K prompt shows the more accurate price.
+    pricingUSD: { input: 0.84, output: 3.36, cacheRead: 0.182 },
+    pricingCNY: { input: 6, output: 24, cacheRead: 1.3 },
     priceCategory: "high",
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
   },
   "glm-5": {
-    detail: "GLM-5 \u2014 200K \u4E0A\u4E0B\u6587\uFF0CAgentic \u5DE5\u5177\u8C03\u7528\uFF0C\u6700\u5927\u8F93\u51FA 128K",
+    detail: "GLM-5 \u2014 200K \u4E0A\u4E0B\u6587\uFF0CAgentic \u5DE5\u5177\u8C03\u7528\uFF0C\u6700\u5927\u8F93\u51FA 128K\uFF082-tier pricing\uFF09",
     imageInput: false,
     toolCalling: true,
-    // ¥6 input / ¥22 output / ¥1.5 cache hit per 1M tokens (输入长度 32K+ 档)
-    pricingUSD: { input: 0.84, output: 3.08, cacheRead: 0.21 },
-    pricingCNY: { input: 6, output: 22, cacheRead: 1.5 },
+    // Pinned the LOWER tier per bigmodel.cn/pricing 2026-08-18:
+    //   [0, 32K)  — ¥4 input / ¥18 output / ¥1   cache hit
+    //   [32K+)     — ¥6 input / ¥22 output / ¥1.5 cache hit
+    pricingUSD: { input: 0.56, output: 2.52, cacheRead: 0.14 },
+    pricingCNY: { input: 4, output: 18, cacheRead: 1 },
     priceCategory: "high",
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
   },
   "glm-5-turbo": {
-    detail: "GLM-5 Turbo \u2014 200K \u4E0A\u4E0B\u6587\uFF0C\u6700\u5927\u8F93\u51FA 128K",
+    detail: "GLM-5 Turbo \u2014 200K \u4E0A\u4E0B\u6587\uFF0C\u6700\u5927\u8F93\u51FA 128K\uFF082-tier pricing\uFF09",
     imageInput: false,
     toolCalling: true,
-    // ¥7 input / ¥26 output / ¥1.8 cache hit per 1M tokens (输入长度 32K+ 档)
-    pricingUSD: { input: 0.98, output: 3.64, cacheRead: 0.252 },
-    pricingCNY: { input: 7, output: 26, cacheRead: 1.8 },
+    // Pinned the LOWER tier per bigmodel.cn/pricing 2026-08-18:
+    //   [0, 32K)  — ¥5 input / ¥22 output / ¥1.2 cache hit
+    //   [32K+)     — ¥7 input / ¥26 output / ¥1.8 cache hit
+    pricingUSD: { input: 0.7, output: 3.08, cacheRead: 0.168 },
+    pricingCNY: { input: 5, output: 22, cacheRead: 1.2 },
     priceCategory: "medium",
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
@@ -325,8 +491,19 @@ var MODEL_METADATA = {
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
   },
-  "glm-4.5-air": {
-    detail: "GLM-4.5 Air \u2014 \u5DE5\u5177\u8C03\u7528",
+  // `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 },
@@ -335,14 +512,35 @@ var MODEL_METADATA = {
     maxInputTokens: 128e3,
     maxOutputTokens: 96e3
   },
+  "glm-4.5-air": {
+    detail: "GLM-4.5 Air \u2014 \u5DE5\u5177\u8C03\u7528\uFF083-tier pricing\uFF09",
+    imageInput: false,
+    toolCalling: true,
+    // Pinned the LOWEST tier per bigmodel.cn/pricing 2026-08-18:
+    //   [0, 32K) × [0, 0.2K) output — ¥0.8 / ¥2   / ¥0.16 cache hit
+    //   [0, 32K) × [0.2K+)    output — ¥0.8 / ¥6   / ¥0.16 cache hit
+    //   [32K, 128K)              — ¥1.2 / ¥8   / ¥0.24 cache hit
+    // All cache-hit rates are 4× lower than input — the
+    // explicit-cache-discount half of BYOM-depth #1.
+    pricingUSD: { input: 0.112, output: 0.28, cacheRead: 0.0224 },
+    pricingCNY: { input: 0.8, output: 2, cacheRead: 0.16 },
+    priceCategory: "low",
+    maxInputTokens: 128e3,
+    maxOutputTokens: 96e3
+  },
   "glm-4.5-airx": {
-    detail: "GLM-4.5 AirX \u2014 \u5FEB\u901F\u7248",
+    detail: "GLM-4.5 AirX \u2014 \u5FEB\u901F\u7248\uFF08\xA510/M \u5355\u6863\uFF09",
     imageInput: false,
     toolCalling: false,
-    pricingUSD: { input: 0, output: 0, cacheRead: null },
-    pricingCNY: { input: 0, output: 0, cacheRead: null },
+    // ¥10 / M tokens (single rate, input == output) per
+    // bigmodel.cn/pricing 2026-08-18 — listed under the "模型推理
+    // → Language Models" sub-tab, NOT the flagship text section.
+    // 8K context window per the same sub-tab; 96K max output is a
+    // best-guess from sibling Air-tier models.
+    pricingUSD: { input: 1.4, output: 1.4, cacheRead: null },
+    pricingCNY: { input: 10, output: 10, cacheRead: null },
     priceCategory: "low",
-    maxInputTokens: 128e3,
+    maxInputTokens: 8192,
     maxOutputTokens: 96e3
   },
   "glm-4-long": {
@@ -379,24 +577,29 @@ var MODEL_METADATA = {
     maxOutputTokens: 4e3
   },
   "glm-4.5v": {
-    detail: "GLM-4.5V \u89C6\u89C9\u63A8\u7406\u6A21\u578B \u2014 \u56FE\u50CF/\u89C6\u9891/\u6587\u6863/GUI",
+    detail: "GLM-4.5V \u89C6\u89C9\u63A8\u7406\u6A21\u578B \u2014 \u56FE\u50CF/\u89C6\u9891/\u6587\u6863/GUI\uFF082-tier pricing\uFF09",
     imageInput: true,
     toolCalling: true,
-    pricingUSD: { input: 0, output: 0, cacheRead: null },
-    pricingCNY: { input: 0, output: 0, cacheRead: null },
+    // Pinned the LOWER tier per bigmodel.cn/pricing 2026-08-18:
+    //   [0, 32K)  — ¥2 input / ¥6 output / ¥0.4 cache hit
+    //   [32, 64K) — ¥4 input / ¥12 output / ¥0.8 cache hit
+    pricingUSD: { input: 0.28, output: 0.84, cacheRead: 0.056 },
+    pricingCNY: { input: 2, output: 6, cacheRead: 0.4 },
     priceCategory: "medium",
-    maxInputTokens: 128e3,
+    maxInputTokens: 64e3,
     maxOutputTokens: 8192
   },
   "glm-5v-turbo": {
-    detail: "GLM-5V Turbo \u2014 \u591A\u6A21\u6001 Coding \u6A21\u578B",
+    detail: "GLM-5V Turbo \u2014 \u591A\u6A21\u6001 Coding \u6A21\u578B\uFF082-tier pricing\uFF09",
     imageInput: true,
     toolCalling: true,
-    pricingUSD: { input: 0, output: 0, cacheRead: null },
-    pricingCNY: { input: 0, output: 0, cacheRead: null },
+    // Pinned the LOWER tier per bigmodel.cn/pricing 2026-08-18:
+    //   [0, 32K)  — ¥5 input / ¥22 output / ¥1.2 cache hit
+    //   [32K+)     — ¥7 input / ¥26 output / ¥1.8 cache hit
+    // Vendor-published context: 200K / 128K max output.
+    pricingUSD: { input: 0.7, output: 3.08, cacheRead: 0.168 },
+    pricingCNY: { input: 5, output: 22, cacheRead: 1.2 },
     priceCategory: "medium",
-    // Official model overview: 200K context / 128K max output
-    // (previously mis-set to 128K/8_192 — corrected 2026-07-27).
     maxInputTokens: 2e5,
     maxOutputTokens: 128e3
   },
@@ -424,8 +627,15 @@ var MODEL_METADATA = {
     // Official model page lists "🛠️ 工具调用: 可靠的工具调用能力，支持多步
     // 任务分解与计划执行" as a core capability — was mis-set to false.
     toolCalling: true,
-    // ¥1.35 input / ¥8.1 output / ¥0.27 cache hit per 1M tokens
-    pricingUSD: { input: 0.189, output: 1.134, cacheRead: 0.038 },
+    // Per https://platform.stepfun.com/docs/zh/pricing/details
+    // (2026-08-18 fetched): ¥1.35 input / ¥8.1 output /
+    // ¥0.27 cache hit per 1M tokens, USD = $0.20 / $1.15 /
+    // $0.04 (StepFun is USD-billed at the same rate as CNY/7
+    // with small rounding per the official pricing page).
+    // USD values previously 0.189/1.134/0.038 — slightly off
+    // from the official page (rounding error from dividing CNY
+    // by hand), corrected.
+    pricingUSD: { input: 0.2, output: 1.15, cacheRead: 0.04 },
     pricingCNY: { input: 1.35, output: 8.1, cacheRead: 0.27 },
     priceCategory: "medium",
     thinkingSchema: "reasoningEffort",
@@ -441,8 +651,13 @@ var MODEL_METADATA = {
     // Official model page lists "🛠️ 工具调用: 可靠的 tools / tool_choice
     // 调用能力" as a core capability — was mis-set to false.
     toolCalling: true,
-    // ¥0.7 input / ¥2.1 output / ¥0.14 cache hit per 1M tokens
-    pricingUSD: { input: 0.098, output: 0.294, cacheRead: 0.02 },
+    // Per https://platform.stepfun.com/docs/zh/pricing/details
+    // (2026-08-18 fetched): ¥0.7 input / ¥2.1 output /
+    // ¥0.14 cache hit per 1M tokens, USD = $0.10 / $0.30 /
+    // $0.02. USD values previously 0.098/0.294/0.02 — slightly
+    // off from the official page (rounding error), corrected
+    // to the exact published values.
+    pricingUSD: { input: 0.1, output: 0.3, cacheRead: 0.02 },
     pricingCNY: { input: 0.7, output: 2.1, cacheRead: 0.14 },
     priceCategory: "low",
     thinkingSchema: "reasoningEffort",
@@ -453,9 +668,13 @@ var MODEL_METADATA = {
     detail: "Step 1o Turbo Vision \u2014 \u89C6\u89C9\u6A21\u578B",
     imageInput: true,
     toolCalling: false,
-    // ¥2.5 input / ¥8 output per 1M tokens
-    pricingUSD: { input: 0.35, output: 1.12, cacheRead: null },
-    pricingCNY: { input: 2.5, output: 8, cacheRead: null },
+    // Per https://platform.stepfun.com/docs/zh/pricing/details
+    // (2026-08-18 fetched): ¥2.5 cache miss / ¥0.5 cache hit /
+    // ¥8 output per 1M tokens. USD = $0.357 / $0.071 / $1.143
+    // (CNY/7 with rounding). Cache hit was previously
+    // undocumented in the curated entry — added.
+    pricingUSD: { input: 0.357, output: 1.143, cacheRead: 0.071 },
+    pricingCNY: { input: 2.5, output: 8, cacheRead: 0.5 },
     priceCategory: "low",
     // Official model overview: 32K context window.
     maxInputTokens: 32768,
@@ -486,18 +705,112 @@ 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,
+    // Per https://minimax-ai.chat/pricing (M2.5 legacy line):
+    // ¥2.1 input / ¥8.4 output / ¥0.21 cache hit per 1M tokens;
+    // USD = $0.30 / $1.20 / $0.03 (cloudprice.net 2026-08-13).
+    // Cache hit IS the published rate — the previous comment
+    // "按官方 10% cache 命中率回填" was wrong (it implied we
+    // were estimating, when actually the cache rate is
+    // documented at ¥0.21 / $0.03 per 1M tokens).
+    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];
 }
 function currencyForBaseUrl(baseUrl) {
   try {
     const hostname = new URL(baseUrl).hostname.toLowerCase();
-    if (hostname === "api.minimaxi.com" || hostname === "api.minimaxi.cn" || hostname === "api.deepseek.com" || hostname === "api.moonshot.cn" || hostname === "open.bigmodel.cn") {
+    if (hostname === "api.minimaxi.com" || hostname === "api.minimaxi.cn" || hostname === "api.deepseek.com" || hostname === "api.moonshot.cn" || hostname === "open.bigmodel.cn" || // Zhipu legacy v3 host. Per the GLM-for-copilot reference
+    // (`docs/references/GLM-for-copilot-main/src/endpoint.ts:4`)
+    // this was retired to `bigmodel.cn` but is still
+    // resolvable for accounts that haven't migrated — we
+    // don't surface it in the baseUrl dropdown, but a user
+    // may paste it from a saved settings.json, so the
+    // currency has to match (CNY, same as the new host).
+    hostname === "dev.bigmodel.cn") {
       return "CNY";
     }
-    if (hostname === "api.minimax.io" || hostname === "api.minimaxi.io" || hostname === "api.siliconflow.cn" || hostname === "api.siliconflow.com" || hostname === "api.stepfun.com" || hostname === "openrouter.ai" || hostname === "api.novita.ai") {
+    if (hostname === "api.minimax.io" || hostname === "api.minimaxi.io" || hostname === "api.siliconflow.cn" || hostname === "api.siliconflow.com" || hostname === "api.stepfun.com" || hostname === "openrouter.ai" || hostname === "api.novita.ai" || // Z.ai / Zhipu international. Billed in USD per the
+    // official `bigmodel.cn/pricing` page (the CNY-billed
+    // list is the China-domiciled `open.bigmodel.cn` only;
+    // the international `api.z.ai` is USD regardless of
+    // which apiMode / protocol path the user picked). The
+    // GLM-for-copilot reference uses the same split
+    // (`docs/references/GLM-for-copilot-main/src/endpoint.ts:160-173`).
+    // Without this explicit entry, `api.z.ai` would still
+    // resolve to USD via the catch-all below — adding it
+    // here makes the intent grep-able and pins the host
+    // list against accidental removal.
+    hostname === "api.z.ai") {
       return "USD";
     }
   } catch {
@@ -555,7 +868,12 @@ 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 +922,81 @@ 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 +1011,23 @@ 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 +1037,21 @@ 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) {
@@ -1200,6 +1595,7 @@ export {
   checkGitHubOrgMembership,
   createConsoleLogger,
   currencyForBaseUrl,
+  effectiveAdapterType,
   fetchGitHubUser,
   getBuiltinProviderPreset,
   getGitHubOrgMembership,
@@ -1212,6 +1608,7 @@ export {
   normalizeErrorForLog,
   normalizeGitUrl,
   parsePayload,
+  protocolForBaseUrl,
   resolvePrimaryEmail,
   safeJson
 };