@serviceme/devtools-shared 0.4.6 → 0.4.8

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": [],
@@ -12,19 +28,60 @@ var PROVIDER_BASE_URL_PRESETS = {
     { label: "\u5168\u7403", baseUrl: "https://api.moonshot.ai/v1" }
   ],
   zhipu: [
-    // 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" }
+    // 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: [
@@ -37,6 +94,8 @@ var PROVIDER_BASE_URL_PRESETS = {
     { label: "\u56FD\u5185", baseUrl: "https://api.agnes-ai.cn/v1" },
     { label: "\u5168\u7403", baseUrl: "https://apihub.agnes-ai.com/v1" }
   ],
+  // Medalsoft internal LLM gateway — single OpenAI-compatible endpoint.
+  medalsoft: [{ label: "\u5185\u90E8\u7F51\u5173", baseUrl: "https://llm.proxy.alio.wang/v1" }],
   "vscode-builtin": []
 };
 function getProviderBaseUrlPresets(type) {
@@ -64,6 +123,12 @@ var PROVIDER_CACHE_CONTROL_METADATA = {
   agnes: {
     supportsCacheControl: true
   },
+  // Medalsoft internal gateway — OpenAI-compatible pass-through; the
+  // upstream's prompt-cache behaviour is unknown from outside, so we
+  // don't declare support (no `prompt_cache_key` header is written).
+  // Flip to `supportsCacheControl: true` once the gateway is verified
+  // to honour stable prompt caching.
+  medalsoft: {},
   "vscode-builtin": {}
 };
 function isProviderCacheControlAware(type) {
@@ -104,8 +169,17 @@ var PRIMARY_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,
@@ -115,8 +189,20 @@ var PRIMARY_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
@@ -132,8 +218,20 @@ var PRIMARY_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,
@@ -176,6 +274,44 @@ var PRIMARY_METADATA = {
     maxInputTokens: 934976,
     maxOutputTokens: 65536
   },
+  "agnes-2.5-pro": {
+    detail: "Paid reasoning model \u2014 commercial stable of 2.5 Pro Alpha (advanced coding, scientific reasoning, long context, multimodal). 1M context / 65K max output",
+    imageInput: true,
+    toolCalling: true,
+    // Official pricing per https://wiki.agnes-ai.com/en/docs/agnes-25-pro.md
+    // (fetched 2026-08-19) — identical to 2.5 Pro Alpha:
+    // $0.45 input / $0.0038 cache read / $0.90 output per 1M tokens.
+    // No separate CNY tier is published, so pricingCNY mirrors USD.
+    pricingUSD: { input: 0.45, output: 0.9, cacheRead: 38e-4 },
+    pricingCNY: { input: 0.45, output: 0.9, cacheRead: 38e-4 },
+    priceCategory: "medium",
+    // Same binary Thinking on/off switch as the other Agnes models
+    // (`chat_template_kwargs.enable_thinking` on the OpenAI wire).
+    thinkingSchema: "thinkingEnabled",
+    // Same published window as 2.5 Pro Alpha: 1M context /
+    // 65536 max output — input cap mirrors the alpha entry.
+    maxInputTokens: 934976,
+    maxOutputTokens: 65536
+  },
+  "agnes-2.5-flash": {
+    detail: "GA upgrade of Agnes 2.0 Flash \u2014 stronger coding, agent workflows, tool calling, image understanding (512K context)",
+    imageInput: true,
+    toolCalling: true,
+    // Standard (post-promo) pricing per
+    // https://wiki.agnes-ai.com/en/docs/agnes-25-flash.md (fetched
+    // 2026-08-19) — same list price as 2.0 Flash ($0.03 / $0.15 per
+    // 1M tokens); the temporary $0 / $0 promo is ignored per the
+    // same policy as the 2.0 Flash entry. No separate CNY tier is
+    // published, so pricingCNY mirrors the USD numbers.
+    pricingUSD: { input: 0.03, output: 0.15, cacheRead: null },
+    pricingCNY: { input: 0.03, output: 0.15, cacheRead: null },
+    priceCategory: "low",
+    // Same binary Thinking on/off switch as 2.0 Flash
+    // (`chat_template_kwargs.enable_thinking` on the OpenAI wire).
+    thinkingSchema: "thinkingEnabled",
+    maxInputTokens: 512e3,
+    maxOutputTokens: 65536
+  },
   // ── Kimi (月之暗面 / Moonshot AI) ────────────────────────────────
   // Source: https://platform.kimi.com/docs/pricing/chat-k3,
   // chat-k27-code, chat-k26 (fetched 2026-07-27). Prices are per 1M
@@ -261,10 +397,21 @@ var PRIMARY_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)
+    // Thinking-mode dropdown + `reasoning_effort`. Zhipu official
+    // OpenAPI (fetched 2026-08-19): "仅 GLM-5.2 及其以上模型支持"
+    // reasoning_effort; for 5.2 the server maps none/minimal → give
+    // up thinking, low/medium → high, xhigh → max (we only send the
+    // native high/max so no client-side aliasing is needed).
+    thinkingSchema: "reasoningEffort",
+    supportsReasoningEffort: true,
+    // 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",
@@ -281,6 +428,14 @@ var PRIMARY_METADATA = {
     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,
+    // Thinking-mode dropdown (思考模式 停用/标准/深度) + the
+    // `reasoning_effort` field. Zhipu official OpenAPI (fetched
+    // 2026-08-19): GLM-5.3 restricts `thinking.type` to "enabled"
+    // and controls depth via `reasoning_effort` (low/high/max
+    // only). Mirrors `supportsReasoningEffort: true` in
+    // `docs/references/GLM-for-copilot-main/src/consts.ts`.
+    thinkingSchema: "reasoningEffort",
+    supportsReasoningEffort: 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
@@ -299,74 +454,105 @@ var PRIMARY_METADATA = {
   // (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",
+    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,
+    // Thinking switch (no reasoning_effort — 5.1 < 5.2 threshold).
+    thinkingSchema: "reasoningEffort",
+    // 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) — 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 (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 \u8F7B\u91CF\u514D\u8D39\u7248\uFF0C200K \u4E0A\u4E0B\u6587",
+    detail: "GLM-4.7 Flash \u2014 \u5B8C\u5168\u514D\u8D39\uFF08200K \u4E0A\u4E0B\u6587\uFF09",
     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 },
+    // Thinking switch (no reasoning_effort — 4.7 < 5.2 threshold).
+    thinkingSchema: "reasoningEffort",
+    // 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: "200K \u4E0A\u4E0B\u6587\uFF0C\u5DE5\u5177\u8C03\u7528",
+    detail: "GLM-4.7 \u2014 200K \u4E0A\u4E0B\u6587\uFF0C\u5DE5\u5177\u8C03\u7528\uFF083-tier pricing\uFF09",
     imageInput: false,
     toolCalling: true,
-    // cc-switch 标价：$0.6 input / $2.2 output
-    pricingUSD: { input: 0.6, output: 2.2, cacheRead: 0.11 },
-    pricingCNY: { input: 4.3, output: 15.7, cacheRead: 0.79 },
+    // Thinking switch (GLM-4.7 forces thinking when enabled per the
+    // official ChatThinking doc; no reasoning_effort — 4.7 < 5.2).
+    thinkingSchema: "reasoningEffort",
+    // 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 },
+    // Thinking switch (no reasoning_effort — 5.1 < 5.2 threshold).
+    thinkingSchema: "reasoningEffort",
+    // 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 },
+    // Thinking switch (no reasoning_effort — 5 < 5.2 threshold).
+    thinkingSchema: "reasoningEffort",
+    // 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 },
+    // Thinking switch (listed in the official ChatThinking doc's
+    // auto-think family; no reasoning_effort — 5-Turbo < 5.2).
+    thinkingSchema: "reasoningEffort",
+    // 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
@@ -375,6 +561,8 @@ var PRIMARY_METADATA = {
     detail: "GLM-4.7 FlashX \u2014 \u5FEB\u901F\u7248",
     imageInput: false,
     toolCalling: false,
+    // Thinking switch (4.7 family; no reasoning_effort).
+    thinkingSchema: "reasoningEffort",
     // ¥0.5 input / ¥3 output / ¥0.1 cache hit per 1M tokens
     pricingUSD: { input: 0.07, output: 0.42, cacheRead: 0.014 },
     pricingCNY: { input: 0.5, output: 3, cacheRead: 0.1 },
@@ -386,6 +574,9 @@ var PRIMARY_METADATA = {
     detail: "GLM-4.6 \u2014 \u5DE5\u5177\u8C03\u7528",
     imageInput: false,
     toolCalling: true,
+    // Thinking switch (auto-think family per the official
+    // ChatThinking doc; no reasoning_effort — 4.6 < 5.2).
+    thinkingSchema: "reasoningEffort",
     pricingUSD: { input: 0.6, output: 2.2, cacheRead: 0.11 },
     pricingCNY: { input: 4.3, output: 15.7, cacheRead: 0.79 },
     priceCategory: "medium",
@@ -407,6 +598,9 @@ var PRIMARY_METADATA = {
     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,
+    // Thinking switch (4.5 generation is the thinking-param floor
+    // per the official ChatThinking doc; no reasoning_effort).
+    thinkingSchema: "reasoningEffort",
     pricingUSD: { input: 0, output: 0, cacheRead: null },
     pricingCNY: { input: 0, output: 0, cacheRead: null },
     priceCategory: "low",
@@ -414,23 +608,38 @@ var PRIMARY_METADATA = {
     maxOutputTokens: 96e3
   },
   "glm-4.5-air": {
-    detail: "GLM-4.5 Air \u2014 \u5DE5\u5177\u8C03\u7528",
+    detail: "GLM-4.5 Air \u2014 \u5DE5\u5177\u8C03\u7528\uFF083-tier pricing\uFF09",
     imageInput: false,
     toolCalling: true,
-    pricingUSD: { input: 0, output: 0, cacheRead: null },
-    pricingCNY: { input: 0, output: 0, cacheRead: null },
+    // Thinking switch (4.5 family; no reasoning_effort).
+    thinkingSchema: "reasoningEffort",
+    // 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 },
+    // Thinking switch (4.5 family; no reasoning_effort).
+    thinkingSchema: "reasoningEffort",
+    // ¥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": {
@@ -467,24 +676,69 @@ var PRIMARY_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 },
+    // Thinking switch (official ChatThinking doc: GLM-4.5V 为强制
+    // 思考 when enabled; no reasoning_effort).
+    thinkingSchema: "reasoningEffort",
+    // 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: 64e3,
+    maxOutputTokens: 8192
+  },
+  // GLM-4.6V (2026 H1) — 视觉推理模型 (paid version, 2-tier
+  // pricing). Distinct from `glm-4.6v-flash` (免费版, 0/0/0) and
+  // `glm-4.6v-flashx` (快速版, ¥0.15/¥1.5/¥0.03). Per
+  // bigmodel.cn/pricing 2026-08-18 (the same snapshot the
+  // §7 table in `byom-glm-learnings.md` was sourced from, where
+  // this row was first documented but never landed in
+  // MODEL_METADATA until 2026-08-19):
+  //   [0, 32K)    — ¥1 input / ¥3 output / ¥0.2 cache hit
+  //   [32, 128K)  — ¥2 input / ¥6 output / ¥0.4 cache hit
+  // The lower tier is pinned per the project's "multi-tier →
+  // lowest tier" convention (most prompts < 32K input × < 0.2K
+  // output). 4.6V sits between glm-4.5v (¥2/¥6/¥0.4) and the
+  // flashx variant — the 4.6 generation is meaningfully cheaper
+  // per the 2026 mid-year price reset.
+  "glm-4.6v": {
+    detail: "GLM-4.6V \u89C6\u89C9\u63A8\u7406\u6A21\u578B \u2014 \u56FE\u50CF/\u89C6\u9891/\u6587\u6863/GUI\uFF082-tier pricing\uFF09",
+    imageInput: true,
+    toolCalling: true,
+    // Thinking switch (same auto-think family as glm-4.5v;
+    // 4.6 < 5.2 threshold, so no reasoning_effort).
+    thinkingSchema: "reasoningEffort",
+    // Pinned the LOWER tier [0, 32K) per bigmodel.cn/pricing
+    // 2026-08-18. USD mirrors the CNY/7 ratio with 2-decimal
+    // rounding (matches glm-4.5v's USD rounding style).
+    pricingUSD: { input: 0.14, output: 0.42, cacheRead: 0.028 },
+    pricingCNY: { input: 1, output: 3, cacheRead: 0.2 },
+    // Tier boundary is 128K (vs glm-4.5v's 64K), so a
+    // `medium` tier tag is still appropriate — the 4.6V is
+    // cheaper than glm-4.5v but not a "low" tier like the
+    // 0.5/3/0.1 flashx family.
     priceCategory: "medium",
     maxInputTokens: 128e3,
     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 },
+    // Thinking switch (listed in the official ChatThinking doc's
+    // auto-think family; no reasoning_effort).
+    thinkingSchema: "reasoningEffort",
+    // 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
   },
@@ -512,8 +766,15 @@ var PRIMARY_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",
@@ -529,8 +790,13 @@ var PRIMARY_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",
@@ -541,9 +807,13 @@ var PRIMARY_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,
@@ -583,8 +853,13 @@ var PRIMARY_METADATA = {
     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)
+    // 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",
@@ -635,6 +910,20 @@ var NAMESPACE_ALIASES = {
   // from the curated detail / pricing / capability columns.
   "glm-4-flashx-250414": "glm-4-flashx"
 };
+var NAMESPACE_ALIAS_FAMILY = {
+  // SiliconFlow
+  "deepseek-ai/DeepSeek-V4-Pro": "SiliconFlow (aggregator)",
+  "deepseek-ai/DeepSeek-V4-Flash": "SiliconFlow (aggregator)",
+  "zai-org/GLM-5.2": "SiliconFlow (aggregator)",
+  "Qwen/Qwen3.6-35B-A3B": "SiliconFlow (aggregator)",
+  "moonshotai/Kimi-K2.7-Code": "SiliconFlow (aggregator)",
+  // Novita
+  "deepseek/deepseek-v4-pro": "Novita (aggregator)",
+  "deepseek/deepseek-v4-flash": "Novita (aggregator)",
+  "zai/glm-5.2": "Novita (aggregator)",
+  "zai/glm-5.1": "Novita (aggregator)",
+  "moonshotai/kimi-k3": "Novita (aggregator)"
+};
 var MODEL_METADATA = (() => {
   const merged = {
     ...PRIMARY_METADATA
@@ -653,10 +942,28 @@ function lookupModelMetadata(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.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 {
@@ -665,13 +972,14 @@ function currencyForBaseUrl(baseUrl) {
 }
 
 // src/ai/providers.presets.ts
-function buildPresetModel(id, displayName) {
+function buildPresetModel(id, displayName, baseUrl) {
   const meta = MODEL_METADATA[id];
   if (!meta) {
     throw new Error(
       `buildPresetModel: no curated MODEL_METADATA entry for '${id}' \u2014 add one before referencing it from BUILTIN_PROVIDER_PRESETS.`
     );
   }
+  const currency = baseUrl ? currencyForBaseUrl(baseUrl) : "USD";
   return {
     id,
     displayName,
@@ -682,11 +990,95 @@ function buildPresetModel(id, displayName) {
       supportsImageToText: meta.imageInput,
       supportsToolCalling: meta.toolCalling
     },
-    pricing: meta.pricingUSD,
+    pricing: meta[`pricing${currency}`],
     priceCategory: meta.priceCategory,
     thinkingSchema: meta.thinkingSchema ?? "none"
   };
 }
+function unionProviderModelWithPreset(fetched, baseUrl) {
+  const meta = MODEL_METADATA[fetched.id];
+  if (!meta) {
+    return fetched;
+  }
+  const currency = baseUrl ? currencyForBaseUrl(baseUrl) : "USD";
+  const presetDisplayName = getPresetModelDisplayName(fetched.id);
+  const displayName = typeof fetched.displayName === "string" && fetched.displayName.length > 0 ? fetched.displayName : presetDisplayName;
+  return {
+    // Spread fetched first so `id` wins, then overlay the
+    // curated fields. (Curated fields ALWAYS win over fetched
+    // per the precedence above — that's why they're explicit,
+    // not relying on spread order.)
+    ...fetched,
+    displayName,
+    detail: meta.detail,
+    capabilities: {
+      supportsImageToText: meta.imageInput,
+      supportsToolCalling: meta.toolCalling
+    },
+    pricing: meta[`pricing${currency}`],
+    priceCategory: meta.priceCategory,
+    // thinkingSchema: prefer preset (curated), fall back to
+    // fetched (in case the API publishes a schema the
+    // preset doesn't know about), then "none" to match
+    // `buildPresetModel`'s explicit-none convention so the
+    // picker renders identically for preset vs fetched rows.
+    thinkingSchema: meta.thinkingSchema ?? fetched.thinkingSchema ?? "none",
+    // Token caps: preset is curated; if the preset doesn't
+    // publish a cap, keep whatever fetched supplied (the
+    // user may have entered it by hand earlier).
+    maxInputTokens: meta.maxInputTokens ?? fetched.maxInputTokens,
+    maxOutputTokens: meta.maxOutputTokens ?? fetched.maxOutputTokens
+  };
+}
+var PRESET_MODEL_FAMILIES = [
+  "GLM",
+  "DeepSeek",
+  "Kimi",
+  "StepFun",
+  "MiniMax",
+  "Agnes",
+  "Qwen",
+  "SiliconFlow (aggregator)",
+  "Novita (aggregator)",
+  "OpenRouter (aggregator)"
+];
+function vendorFamilyForId(id) {
+  const aliasFamily = NAMESPACE_ALIAS_FAMILY[id];
+  if (aliasFamily) return aliasFamily;
+  if (id.startsWith("glm-")) return "GLM";
+  if (id.startsWith("deepseek-")) return "DeepSeek";
+  if (id.startsWith("kimi-")) return "Kimi";
+  if (id.startsWith("step-")) return "StepFun";
+  if (id.startsWith("MiniMax-")) return "MiniMax";
+  if (id.startsWith("agnes-")) return "Agnes";
+  if (id.startsWith("Qwen")) return "Qwen";
+  if (id.startsWith("openrouter/")) return "OpenRouter (aggregator)";
+  return "Other";
+}
+var LISTABLE_PRESET_MODELS = (() => {
+  const aggregatorAliasKeys = new Set(Object.keys(NAMESPACE_ALIAS_FAMILY));
+  const allIds = Object.keys(MODEL_METADATA).filter((id) => {
+    if (NAMESPACE_ALIASES[id] === void 0) return true;
+    return aggregatorAliasKeys.has(id);
+  }).sort();
+  return allIds.map((id) => ({
+    id,
+    displayName: id,
+    vendorFamily: vendorFamilyForId(id)
+  }));
+})();
+function listPresetModelGroups() {
+  const byFamily = /* @__PURE__ */ new Map();
+  for (const entry of LISTABLE_PRESET_MODELS) {
+    const bucket = byFamily.get(entry.vendorFamily) ?? [];
+    bucket.push(entry);
+    byFamily.set(entry.vendorFamily, bucket);
+  }
+  return PRESET_MODEL_FAMILIES.filter((f) => byFamily.has(f)).map((family) => ({
+    family,
+    entries: byFamily.get(family) ?? []
+  }));
+}
 var BUILTIN_PROVIDER_PRESETS = {
   minimax: {
     displayName: "MiniMax",
@@ -712,14 +1104,18 @@ var BUILTIN_PROVIDER_PRESETS = {
     // (otherwise the user's added model never shows up in the
     // Copilot Chat picker).
     models: [
-      buildPresetModel("MiniMax-M3", "MiniMax-M3"),
-      buildPresetModel("MiniMax-M2.7", "MiniMax-M2.7"),
-      buildPresetModel("MiniMax-M2.7-highspeed", "MiniMax-M2.7-highspeed"),
+      buildPresetModel("MiniMax-M3", "MiniMax-M3", "https://api.minimaxi.com/anthropic"),
+      buildPresetModel("MiniMax-M2.7", "MiniMax-M2.7", "https://api.minimaxi.com/anthropic"),
+      buildPresetModel(
+        "MiniMax-M2.7-highspeed",
+        "MiniMax-M2.7-highspeed",
+        "https://api.minimaxi.com/anthropic"
+      ),
       // 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")
+      buildPresetModel("MiniMax-M2.5", "MiniMax-M2.5", "https://api.minimaxi.com/anthropic")
     ]
   },
   deepseek: {
@@ -741,8 +1137,8 @@ var BUILTIN_PROVIDER_PRESETS = {
     // the recommended default — users still on V3 can override
     // these defaults in the ProvidersTab form.
     models: [
-      buildPresetModel("deepseek-v4-flash", "DeepSeek V4 Flash"),
-      buildPresetModel("deepseek-v4-pro", "DeepSeek V4 Pro")
+      buildPresetModel("deepseek-v4-flash", "DeepSeek V4 Flash", "https://api.deepseek.com/v1"),
+      buildPresetModel("deepseek-v4-pro", "DeepSeek V4 Pro", "https://api.deepseek.com/v1")
     ]
   },
   agnes: {
@@ -750,19 +1146,34 @@ var BUILTIN_PROVIDER_PRESETS = {
     // Sapiens AI's Agnes — OpenAI-compatible endpoint per
     // https://wiki.agnes-ai.com (`POST /v1/chat/completions`).
     baseUrl: "https://apihub.agnes-ai.com/v1",
+    // Order mirrors the vendor docs index (wiki.agnes-ai.com,
+    // fetched 2026-08-19): paid reasoning models first — the
+    // commercial stable `agnes-2.5-pro` leads, then the
+    // benchmark-listed alpha — followed by the flash tier
+    // newest-first.
     models: [
-      buildPresetModel("agnes-2.5-pro-alpha", "Agnes 2.5 Pro Alpha"),
-      buildPresetModel("agnes-2.0-flash", "Agnes 2.0 Flash")
+      buildPresetModel("agnes-2.5-pro", "Agnes 2.5 Pro", "https://apihub.agnes-ai.com/v1"),
+      buildPresetModel(
+        "agnes-2.5-pro-alpha",
+        "Agnes 2.5 Pro Alpha",
+        "https://apihub.agnes-ai.com/v1"
+      ),
+      buildPresetModel("agnes-2.5-flash", "Agnes 2.5 Flash", "https://apihub.agnes-ai.com/v1"),
+      buildPresetModel("agnes-2.0-flash", "Agnes 2.0 Flash", "https://apihub.agnes-ai.com/v1")
     ]
   },
   kimi: {
     displayName: "Kimi",
     baseUrl: "https://api.moonshot.cn/v1",
     models: [
-      buildPresetModel("kimi-k3", "Kimi K3"),
-      buildPresetModel("kimi-k2.7-code", "Kimi K2.7 Code"),
-      buildPresetModel("kimi-k2.7-code-highspeed", "Kimi K2.7 Code HighSpeed"),
-      buildPresetModel("kimi-k2.6", "Kimi K2.6")
+      buildPresetModel("kimi-k3", "Kimi K3", "https://api.moonshot.cn/v1"),
+      buildPresetModel("kimi-k2.7-code", "Kimi K2.7 Code", "https://api.moonshot.cn/v1"),
+      buildPresetModel(
+        "kimi-k2.7-code-highspeed",
+        "Kimi K2.7 Code HighSpeed",
+        "https://api.moonshot.cn/v1"
+      ),
+      buildPresetModel("kimi-k2.6", "Kimi K2.6", "https://api.moonshot.cn/v1")
     ]
   },
   zhipu: {
@@ -834,24 +1245,28 @@ var BUILTIN_PROVIDER_PRESETS = {
     // `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-4.7", "GLM-4.7"),
-      buildPresetModel("glm-4.6", "GLM-4.6"),
-      buildPresetModel("glm-4.5-air", "GLM-4.5 Air"),
-      buildPresetModel("glm-4.5", "GLM-4.5")
+      buildPresetModel("glm-5.3", "GLM-5.3", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-5.2", "GLM-5.2", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-5.1", "GLM-5.1", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-5", "GLM-5", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-5-turbo", "GLM-5 Turbo", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-4.7", "GLM-4.7", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-4.6", "GLM-4.6", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-4.5-air", "GLM-4.5 Air", "https://open.bigmodel.cn/api/paas/v4"),
+      buildPresetModel("glm-4.5", "GLM-4.5", "https://open.bigmodel.cn/api/paas/v4")
     ]
   },
   stepfun: {
     displayName: "StepFun",
     baseUrl: "https://api.stepfun.com/v1",
     models: [
-      buildPresetModel("step-3.7-flash", "Step 3.7 Flash"),
-      buildPresetModel("step-3.5-flash", "Step 3.5 Flash"),
-      buildPresetModel("step-1o-turbo-vision", "Step 1o Turbo Vision")
+      buildPresetModel("step-3.7-flash", "Step 3.7 Flash", "https://api.stepfun.com/v1"),
+      buildPresetModel("step-3.5-flash", "Step 3.5 Flash", "https://api.stepfun.com/v1"),
+      buildPresetModel(
+        "step-1o-turbo-vision",
+        "Step 1o Turbo Vision",
+        "https://api.stepfun.com/v1"
+      )
     ]
   },
   siliconflow: {
@@ -865,20 +1280,42 @@ var BUILTIN_PROVIDER_PRESETS = {
     // 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)"),
+      buildPresetModel(
+        "deepseek-ai/DeepSeek-V4-Pro",
+        "DeepSeek V4 Pro (via SiliconFlow)",
+        "https://api.siliconflow.cn/v1"
+      ),
+      buildPresetModel(
+        "deepseek-ai/DeepSeek-V4-Flash",
+        "DeepSeek V4 Flash (via SiliconFlow)",
+        "https://api.siliconflow.cn/v1"
+      ),
       // 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)"),
+      buildPresetModel(
+        "zai-org/GLM-5.2",
+        "GLM-5.2 (via SiliconFlow)",
+        "https://api.siliconflow.cn/v1"
+      ),
       // 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)"),
+      buildPresetModel(
+        "Qwen/Qwen3.6-35B-A3B",
+        "Qwen3.6-35B-A3B (via SiliconFlow)",
+        "https://api.siliconflow.cn/v1"
+      ),
       // Kimi K2.7-Code (2026-06-12) — Moonshot coding 旗舰
-      buildPresetModel("moonshotai/Kimi-K2.7-Code", "Kimi K2.7 Code (via SiliconFlow)")
+      buildPresetModel(
+        "moonshotai/Kimi-K2.7-Code",
+        "Kimi K2.7 Code (via SiliconFlow)",
+        "https://api.siliconflow.cn/v1"
+      )
     ]
   },
   openrouter: {
     displayName: "OpenRouter",
     baseUrl: "https://openrouter.ai/api/v1",
-    models: [buildPresetModel("openrouter/auto", "OpenRouter Auto")]
+    models: [
+      buildPresetModel("openrouter/auto", "OpenRouter Auto", "https://openrouter.ai/api/v1")
+    ]
   },
   novita: {
     displayName: "Novita",
@@ -890,16 +1327,76 @@ var BUILTIN_PROVIDER_PRESETS = {
     // 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)"),
+      buildPresetModel(
+        "deepseek/deepseek-v4-pro",
+        "DeepSeek V4 Pro (via Novita)",
+        "https://api.novita.ai/openai/v1"
+      ),
+      buildPresetModel(
+        "deepseek/deepseek-v4-flash",
+        "DeepSeek V4 Flash (via Novita)",
+        "https://api.novita.ai/openai/v1"
+      ),
       // 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)"),
+      buildPresetModel("zai/glm-5.2", "GLM-5.2 (via Novita)", "https://api.novita.ai/openai/v1"),
+      buildPresetModel("zai/glm-5.1", "GLM-5.1 (via Novita)", "https://api.novita.ai/openai/v1"),
       // Kimi K3 (2026-07-16 API, 2026-07-27 开源) — 1M context, 2.8T MoE
-      buildPresetModel("moonshotai/kimi-k3", "Kimi K3 (via Novita)")
+      buildPresetModel(
+        "moonshotai/kimi-k3",
+        "Kimi K3 (via Novita)",
+        "https://api.novita.ai/openai/v1"
+      )
     ]
+  },
+  medalsoft: {
+    displayName: "Medalsoft",
+    // Medalsoft internal LLM gateway (公司内部代理) — OpenAI-compatible
+    // `/v1/chat/completions`. The gateway forwards to upstream vendors
+    // (GLM / DeepSeek / Kimi / ...), so the model catalogue is dynamic
+    // and NOT curated here: the preset ships an EMPTY starter list and
+    // the user populates it via "Fetch from API" (`GET /v1/models`) in
+    // the ProvidersTab. `buildPresetModel`'s fail-loudly contract is
+    // why we don't guess ids — a curated `MODEL_METADATA` entry only
+    // exists for vendor-native ids, not the gateway's routing table.
+    //
+    // NOTE: models fetched from the gateway carry the upstream model
+    // ids, so the curated metadata (pricing / thinking dropdown /
+    // token caps) still resolves via `MODEL_METADATA` after the fetch.
+    baseUrl: "https://llm.proxy.alio.wang/v1",
+    models: []
   }
 };
+var PRESET_MODEL_DISPLAY_NAMES = (() => {
+  const map = {};
+  for (const preset of Object.values(BUILTIN_PROVIDER_PRESETS)) {
+    for (const model of preset.models) {
+      if (model.displayName !== void 0) {
+        map[model.id] = model.displayName;
+      }
+    }
+  }
+  return Object.freeze(map);
+})();
+function getPresetModelDisplayName(id) {
+  const explicit = PRESET_MODEL_DISPLAY_NAMES[id];
+  if (explicit !== void 0) {
+    return explicit;
+  }
+  const meta = MODEL_METADATA[id];
+  if (meta === void 0) {
+    return void 0;
+  }
+  const detail = meta.detail;
+  if (typeof detail !== "string" || detail.trim() === "") {
+    return void 0;
+  }
+  const dashIndex = detail.indexOf(" \u2014 ");
+  if (dashIndex === -1) {
+    return detail;
+  }
+  const head = detail.slice(0, dashIndex).trim();
+  return head === "" ? void 0 : head;
+}
 function getBuiltinProviderPreset(type) {
   switch (type) {
     case "minimax":
@@ -911,6 +1408,7 @@ function getBuiltinProviderPreset(type) {
     case "siliconflow":
     case "openrouter":
     case "novita":
+    case "medalsoft":
       return BUILTIN_PROVIDER_PRESETS[type];
     default:
       return null;
@@ -1334,6 +1832,8 @@ var WebviewMessageType = /* @__PURE__ */ ((WebviewMessageType2) => {
   WebviewMessageType2["SetProviderOrder"] = "setProviderOrder";
   WebviewMessageType2["TestProvider"] = "testProvider";
   WebviewMessageType2["ProviderTestResultMessage"] = "providerTestResult";
+  WebviewMessageType2["TestProviderModel"] = "testProviderModel";
+  WebviewMessageType2["ProviderTestModelResultMessage"] = "providerTestModelResult";
   WebviewMessageType2["DefaultProviderChanged"] = "defaultProviderChanged";
   WebviewMessageType2["FetchProviderModels"] = "fetchProviderModels";
   WebviewMessageType2["FetchProviderModelsResult"] = "fetchProviderModelsResult";
@@ -1424,8 +1924,12 @@ export {
   GetCachedServerUrl,
   GetServerProxyState,
   GetUtilityModels,
+  LISTABLE_PRESET_MODELS,
   LogLevel,
   MODEL_METADATA,
+  NAMESPACE_ALIASES,
+  NAMESPACE_ALIAS_FAMILY,
+  PRESET_MODEL_FAMILIES,
   PROVIDER_BASE_URL_PRESETS,
   PROVIDER_CACHE_CONTROL_METADATA,
   ServerProxyStateResponse,
@@ -1438,21 +1942,27 @@ export {
   __internal,
   asAbortSignal,
   buildGitHubLocalEmail,
+  buildPresetModel,
   checkGitHubOrgMembership,
   createConsoleLogger,
   currencyForBaseUrl,
+  effectiveAdapterType,
   fetchGitHubUser,
   getBuiltinProviderPreset,
   getGitHubOrgMembership,
+  getPresetModelDisplayName,
   getProviderBaseUrlPresets,
   isGitHubLocalEmail,
   isProviderCacheControlAware,
   isValidCanonicalSlug,
+  listPresetModelGroups,
   lookupModelMetadata,
   normalizeCanonicalSlug,
   normalizeErrorForLog,
   normalizeGitUrl,
   parsePayload,
+  protocolForBaseUrl,
   resolvePrimaryEmail,
-  safeJson
+  safeJson,
+  unionProviderModelWithPreset
 };