@serviceme/devtools-shared 0.4.8 → 0.4.10

package/dist/index.d.mts

@@ -341,6 +341,12 @@ interface ProviderMutationPayload {
     enabled?: boolean;
     /** See {@link ProviderConfig.minimaxBillingType}. */
     minimaxBillingType?: "token_plan" | "pay_as_you_go";
+    /**
+     * See {@link ProviderConfig.visionMode} (2026-08-20 — surfaced in
+     * the ProvidersTab form). Omit to leave unchanged on update /
+     * default to `"native"` on add.
+     */
+    visionMode?: ProviderVisionMode;
 }
 /** Which vendor API to call for usage data. */
 type ProviderUsageKind = "minimax" | "deepseek" | "kimi" | "zhipu" | "stepfun" | "siliconflow" | "openrouter" | "novita";
@@ -568,15 +574,16 @@ declare function lookupModelMetadata(modelId: string): CuratedModelMetadata | un
  * platform), `api.minimax.io` maps to USD (the global platform),
  * everything else falls back to USD (the global default). The
  * match is exact-host so a typo in the hostname never silently
- * flips currency. The earlier 2026-08-18 draft also mapped
- * `api.deepseek.com` to CNY, but that was wrong: DeepSeek's
- * international `.com` endpoint publishes USD prices for global
- * accounts (per https://api-docs.deepseek.com/quick_start/pricing/,
- * which lists both $ and ¥ on the same page, with the $ block
- * corresponding to the .com endpoint). The CNY-priced alternative
- * is the China-domestic region, not the international `.com`
- * host, so the currency now follows the project default (USD)
- * via the catch-all below.
+ * flips currency.
+ *
+ * `api.deepseek.com` maps to CNY (2026-08-20 product decision —
+ * user request). History: the 8/18 draft put it in CNY by mistake
+ * (per the $ block on the English pricing page), 8/19 moved it to
+ * the USD catch-all, and 8/20 moved it BACK to CNY deliberately —
+ * our user base bills in ¥ on this endpoint (the official zh-cn
+ * page publishes the ¥ table for it), and the picker should show
+ * the price those users actually pay. The DeepSeek entries pin the
+ * PEAK tier (conservative worst-case; off-peak is exactly half).
  */
 declare function currencyForBaseUrl(baseUrl: string): "USD" | "CNY";
 
@@ -600,7 +607,7 @@ declare function currencyForBaseUrl(baseUrl: string): "USD" | "CNY";
  *   - agnes: OpenAI-compatible at `https://apihub.agnes-ai.com/v1`
  *     → routed to OpenAIAdapter
  *   - medalsoft: OpenAI-compatible internal gateway at
- *     `https://llm.proxy.alio.wang/v1` → routed to OpenAIAdapter
+ *     `https://nexus.servicemecloud.com/v1` → routed to OpenAIAdapter
  *     (empty starter model list — populate via "Fetch from API")
  *
  * Earlier iterations of the minimax default 401'd on the team; do

package/dist/index.d.ts

@@ -341,6 +341,12 @@ interface ProviderMutationPayload {
     enabled?: boolean;
     /** See {@link ProviderConfig.minimaxBillingType}. */
     minimaxBillingType?: "token_plan" | "pay_as_you_go";
+    /**
+     * See {@link ProviderConfig.visionMode} (2026-08-20 — surfaced in
+     * the ProvidersTab form). Omit to leave unchanged on update /
+     * default to `"native"` on add.
+     */
+    visionMode?: ProviderVisionMode;
 }
 /** Which vendor API to call for usage data. */
 type ProviderUsageKind = "minimax" | "deepseek" | "kimi" | "zhipu" | "stepfun" | "siliconflow" | "openrouter" | "novita";
@@ -568,15 +574,16 @@ declare function lookupModelMetadata(modelId: string): CuratedModelMetadata | un
  * platform), `api.minimax.io` maps to USD (the global platform),
  * everything else falls back to USD (the global default). The
  * match is exact-host so a typo in the hostname never silently
- * flips currency. The earlier 2026-08-18 draft also mapped
- * `api.deepseek.com` to CNY, but that was wrong: DeepSeek's
- * international `.com` endpoint publishes USD prices for global
- * accounts (per https://api-docs.deepseek.com/quick_start/pricing/,
- * which lists both $ and ¥ on the same page, with the $ block
- * corresponding to the .com endpoint). The CNY-priced alternative
- * is the China-domestic region, not the international `.com`
- * host, so the currency now follows the project default (USD)
- * via the catch-all below.
+ * flips currency.
+ *
+ * `api.deepseek.com` maps to CNY (2026-08-20 product decision —
+ * user request). History: the 8/18 draft put it in CNY by mistake
+ * (per the $ block on the English pricing page), 8/19 moved it to
+ * the USD catch-all, and 8/20 moved it BACK to CNY deliberately —
+ * our user base bills in ¥ on this endpoint (the official zh-cn
+ * page publishes the ¥ table for it), and the picker should show
+ * the price those users actually pay. The DeepSeek entries pin the
+ * PEAK tier (conservative worst-case; off-peak is exactly half).
  */
 declare function currencyForBaseUrl(baseUrl: string): "USD" | "CNY";
 
@@ -600,7 +607,7 @@ declare function currencyForBaseUrl(baseUrl: string): "USD" | "CNY";
  *   - agnes: OpenAI-compatible at `https://apihub.agnes-ai.com/v1`
  *     → routed to OpenAIAdapter
  *   - medalsoft: OpenAI-compatible internal gateway at
- *     `https://llm.proxy.alio.wang/v1` → routed to OpenAIAdapter
+ *     `https://nexus.servicemecloud.com/v1` → routed to OpenAIAdapter
  *     (empty starter model list — populate via "Fetch from API")
  *
  * Earlier iterations of the minimax default 401'd on the team; do

package/dist/index.js

@@ -170,7 +170,7 @@ var PROVIDER_BASE_URL_PRESETS = {
     { 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" }],
+  medalsoft: [{ label: "\u5185\u90E8\u7F51\u5173", baseUrl: "https://nexus.servicemecloud.com/v1" }],
   "vscode-builtin": []
 };
 function getProviderBaseUrlPresets(type) {
@@ -190,7 +190,18 @@ var PROVIDER_CACHE_CONTROL_METADATA = {
     // OpenAI-compat prompt_cache_key
   },
   kimi: {},
-  zhipu: {},
+  // 2026-08-20 — Zhipu joins the cache-aware set (learnings doc §9
+  // task #6). Zhipu's OpenAI-compatible endpoint performs automatic
+  // prefix caching on stable request prefixes; the `prompt_cache_key`
+  // header the OpenAI adapter writes for cache-aware types gives the
+  // upstream a stable per-conversation bucket (mirrors the reference's
+  // Coding-Plan prompt-cache behaviour). The Anthropic-protocol
+  // `/api/anthropic` routes are unaffected — that path uses the
+  // AnthropicAdapter's own 4-breakpoint `cache_control` logic and
+  // ignores this flag.
+  zhipu: {
+    supportsCacheControl: true
+  },
   stepfun: {},
   siliconflow: {},
   openrouter: {},
@@ -264,20 +275,20 @@ var PRIMARY_METADATA = {
     detail: "Fast, general-purpose model",
     imageInput: true,
     toolCalling: true,
-    // 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 },
+    // Per https://api-docs.deepseek.com/zh-cn/quick_start/pricing
+    // (re-fetched 2026-08-20): peak/off-peak tiered pricing since
+    // the 2026-08-17 调价. Pinned the PEAK rate (2026-08-20 product
+    // decision — user request): DeepSeek is surfaced as a CNY
+    // provider for China accounts, whose working hours
+    // (09:00–12:00 + 14:00–18:00 Beijing) fall entirely inside
+    // the peak window — peak is the conservative worst-case
+    // estimate (cost never exceeds it; off-peak turns are billed
+    // at exactly half).
+    //   Peak:     $0.014 cache hit / $0.44 input / $1.32 output
+    //             ¥0.1  cache hit / ¥3   input / ¥9   output
+    //   Off-peak: exactly 1/2 of peak (¥1.5 / ¥4.5 / ¥0.05).
+    pricingUSD: { input: 0.44, output: 1.32, cacheRead: 0.014 },
+    pricingCNY: { input: 3, output: 9, cacheRead: 0.1 },
     priceCategory: "low",
     // Official docs (api-docs.deepseek.com/quick_start/pricing, fetched
     // 2026-07-27): "THINKING MODE: Supports both non-thinking and
@@ -293,20 +304,13 @@ var PRIMARY_METADATA = {
     detail: "Most capable reasoning model",
     imageInput: true,
     toolCalling: true,
-    // 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 },
+    // Same source + same PEAK-tier decision as deepseek-v4-flash
+    // (2026-08-20): conservative worst-case estimate; off-peak is
+    // exactly half (¥4.5 / ¥13.5 / ¥0.15).
+    //   Peak: $0.044 cache hit / $1.32 input / $3.96 output
+    //         ¥0.3  cache hit / ¥9   input / ¥27  output
+    pricingUSD: { input: 1.32, output: 3.96, cacheRead: 0.044 },
+    pricingCNY: { input: 9, output: 27, cacheRead: 0.3 },
     priceCategory: "low",
     thinkingSchema: "thinkingEnabled",
     maxInputTokens: 655360,
@@ -511,12 +515,15 @@ var PRIMARY_METADATA = {
     // `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
-    // input, so a guess would mislead budget estimates).
-    pricingUSD: { input: 1.4, output: 4.4, cacheRead: null },
-    pricingCNY: { input: 10, output: 31, cacheRead: null },
+    // Vendor pricing (corrected 2026-08-20): bigmodel.cn/pricing
+    // now lists GLM-5.3 at the SAME rate as GLM-5.2 —
+    //   ¥8 input / ¥28 output / ¥2 cache hit per 1M tokens
+    // The previous CNY 10/31 was a mis-read of the Decrypt launch
+    // article (2026-08-14). USD comes from Z.ai's independently-
+    // set international list ($1.4 / $4.4 / $0.26 — not the ×0.14
+    // CNY mirror), matching the reference consts.ts USD block.
+    pricingUSD: { input: 1.4, output: 4.4, cacheRead: 0.26 },
+    pricingCNY: { input: 8, output: 28, cacheRead: 2 },
     priceCategory: "high",
     maxInputTokens: 1e6,
     maxOutputTokens: 128e3
@@ -536,8 +543,10 @@ var PRIMARY_METADATA = {
     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 },
+    // architecture as the base model. USD = the standard ×0.14 mirror
+    // of 6/24/1.3 (2026-08-20 — fixed a copy-paste slip that had
+    // cloned glm-4.6's 0.6/2.2/0.11 USD block here, a ×0.10 rate).
+    pricingUSD: { input: 0.84, output: 3.36, cacheRead: 0.182 },
     pricingCNY: { input: 6, output: 24, cacheRead: 1.3 },
     priceCategory: "medium",
     maxInputTokens: 2e5,
@@ -800,6 +809,62 @@ var PRIMARY_METADATA = {
     maxInputTokens: 128e3,
     maxOutputTokens: 8192
   },
+  // GLM-4.6V-Flash — the FULLY-FREE vision model on
+  // bigmodel.cn/pricing's 视觉理解 tab (every column 免费,
+  // 2026-08-18 snapshot). This is the exact model the GLM-for-
+  // copilot reference uses as its transparent vision proxy
+  // (`DEFAULT_GLM_VISION_MODEL_ID` in its consts.ts — "the
+  // extension accepts images for text models through the
+  // GLM-4.6V-Flash vision proxy"). Zero cost makes it the
+  // natural candidate for our future ProxyVisionAdapter (v2).
+  // Token caps mirror the reference's entry (96K input + 32K
+  // output = the 128K shared window). Not in the zhipu preset —
+  // multimodal models aren't exposed via the public
+  // `/v1/chat/completions` models list (same trim rationale as
+  // glm-5v-turbo / glm-4.5v); users add it by hand.
+  "glm-4.6v-flash": {
+    detail: "GLM-4.6V Flash \u2014 \u5B8C\u5168\u514D\u8D39\u89C6\u89C9\u7406\u89E3\u6A21\u578B",
+    imageInput: true,
+    // Reference consts.ts declares GLM_TOOLS_LIMIT (128) tool
+    // calling for this model — the flash vision tier keeps the
+    // full function-call surface.
+    toolCalling: true,
+    // Thinking switch (auto-think family per the reference's
+    // `requiresThinkingParam: true`; no reasoning_effort —
+    // 4.6 < 5.2 threshold).
+    thinkingSchema: "reasoningEffort",
+    // Free tier — input / output / cache hit all 0 per
+    // bigmodel.cn 2026-08-18. USD mirrors CNY at 0 rather than
+    // inventing a rate.
+    pricingUSD: { input: 0, output: 0, cacheRead: 0 },
+    pricingCNY: { input: 0, output: 0, cacheRead: 0 },
+    priceCategory: "low",
+    maxInputTokens: 98304,
+    maxOutputTokens: 32768
+  },
+  // GLM-4.6V-FlashX — the 快速版 (quick) paid sibling of the free
+  // glm-4.6v-flash. Per bigmodel.cn/pricing 视觉理解 tab
+  // (2026-08-18):
+  //   [0, 32K)    — ¥0.15 input / ¥1.5 output / ¥0.03 cache hit
+  //   [32K, 128K) — ¥0.3 input  / ¥3 output   / ¥0.03 cache hit
+  // The LOWER tier is pinned per the project's multi-tier
+  // convention. Not in the zhipu preset (same multimodal trim
+  // rationale as glm-4.6v-flash). Token caps are a best-guess
+  // from sibling glm-4.6v (128K context / 8K output) — Zhipu
+  // publishes no dedicated spec for the flashx vision tier.
+  "glm-4.6v-flashx": {
+    detail: "GLM-4.6V FlashX \u2014 \u5FEB\u901F\u7248\u89C6\u89C9\u7406\u89E3\uFF082-tier pricing\uFF09",
+    imageInput: true,
+    // 快速版 convention: the flashx siblings (glm-4.7-flashx,
+    // glm-4.5-airx) drop tool calling.
+    toolCalling: false,
+    thinkingSchema: "reasoningEffort",
+    pricingUSD: { input: 0.021, output: 0.21, cacheRead: 42e-4 },
+    pricingCNY: { input: 0.15, output: 1.5, cacheRead: 0.03 },
+    priceCategory: "low",
+    maxInputTokens: 128e3,
+    maxOutputTokens: 8192
+  },
   "glm-5v-turbo": {
     detail: "GLM-5V Turbo \u2014 \u591A\u6A21\u6001 Coding \u6A21\u578B\uFF082-tier pricing\uFF09",
     imageInput: true,
@@ -952,9 +1017,13 @@ var PRIMARY_METADATA = {
     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 },
+    // SiliconFlow 官方价格页（https://siliconflow.cn/pricing，
+    // re-fetched 2026-08-20）：¥1.8 input / ¥10.8 output per 1M
+    // tokens，无缓存价格列。此前 ¥1.6/¥12.8 来自发布新闻稿，
+    // 已过时（输入上调、输出下调）。USD = CNY ÷7 镜像取两位
+    // 小数（0.26 / 1.54）。
+    pricingUSD: { input: 0.26, output: 1.54, cacheRead: null },
+    pricingCNY: { input: 1.8, output: 10.8, 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）。
@@ -1024,7 +1093,19 @@ function currencyForBaseUrl(baseUrl) {
     // 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") {
+    hostname === "dev.bigmodel.cn" || // DeepSeek — 2026-08-20 product decision (user request):
+    // surface the domestic ¥ prices. The `.com` endpoint
+    // serves both regions, but our user base bills in CNY on
+    // it (the official zh-cn pricing page publishes the ¥
+    // table for this endpoint), so the picker now shows the
+    // CNY block. This REVERSES the 2026-08-19 mapping that
+    // left it in the USD catch-all ("$ block corresponds to
+    // the .com endpoint") — with a CNY-billed user base, the
+    // ¥ block is the relevant one. The pricingCNY block in
+    // MODEL_METADATA pins the PEAK tier for these entries
+    // (conservative worst-case; see the deepseek-v4-flash
+    // comment).
+    hostname === "api.deepseek.com") {
       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" || // Z.ai / Zhipu international. Billed in USD per the
@@ -1437,7 +1518,7 @@ var BUILTIN_PROVIDER_PRESETS = {
     // 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",
+    baseUrl: "https://nexus.servicemecloud.com/v1",
     models: []
   }
 };

package/dist/index.mjs

@@ -95,7 +95,7 @@ var PROVIDER_BASE_URL_PRESETS = {
     { 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" }],
+  medalsoft: [{ label: "\u5185\u90E8\u7F51\u5173", baseUrl: "https://nexus.servicemecloud.com/v1" }],
   "vscode-builtin": []
 };
 function getProviderBaseUrlPresets(type) {
@@ -115,7 +115,18 @@ var PROVIDER_CACHE_CONTROL_METADATA = {
     // OpenAI-compat prompt_cache_key
   },
   kimi: {},
-  zhipu: {},
+  // 2026-08-20 — Zhipu joins the cache-aware set (learnings doc §9
+  // task #6). Zhipu's OpenAI-compatible endpoint performs automatic
+  // prefix caching on stable request prefixes; the `prompt_cache_key`
+  // header the OpenAI adapter writes for cache-aware types gives the
+  // upstream a stable per-conversation bucket (mirrors the reference's
+  // Coding-Plan prompt-cache behaviour). The Anthropic-protocol
+  // `/api/anthropic` routes are unaffected — that path uses the
+  // AnthropicAdapter's own 4-breakpoint `cache_control` logic and
+  // ignores this flag.
+  zhipu: {
+    supportsCacheControl: true
+  },
   stepfun: {},
   siliconflow: {},
   openrouter: {},
@@ -189,20 +200,20 @@ var PRIMARY_METADATA = {
     detail: "Fast, general-purpose model",
     imageInput: true,
     toolCalling: true,
-    // 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 },
+    // Per https://api-docs.deepseek.com/zh-cn/quick_start/pricing
+    // (re-fetched 2026-08-20): peak/off-peak tiered pricing since
+    // the 2026-08-17 调价. Pinned the PEAK rate (2026-08-20 product
+    // decision — user request): DeepSeek is surfaced as a CNY
+    // provider for China accounts, whose working hours
+    // (09:00–12:00 + 14:00–18:00 Beijing) fall entirely inside
+    // the peak window — peak is the conservative worst-case
+    // estimate (cost never exceeds it; off-peak turns are billed
+    // at exactly half).
+    //   Peak:     $0.014 cache hit / $0.44 input / $1.32 output
+    //             ¥0.1  cache hit / ¥3   input / ¥9   output
+    //   Off-peak: exactly 1/2 of peak (¥1.5 / ¥4.5 / ¥0.05).
+    pricingUSD: { input: 0.44, output: 1.32, cacheRead: 0.014 },
+    pricingCNY: { input: 3, output: 9, cacheRead: 0.1 },
     priceCategory: "low",
     // Official docs (api-docs.deepseek.com/quick_start/pricing, fetched
     // 2026-07-27): "THINKING MODE: Supports both non-thinking and
@@ -218,20 +229,13 @@ var PRIMARY_METADATA = {
     detail: "Most capable reasoning model",
     imageInput: true,
     toolCalling: true,
-    // 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 },
+    // Same source + same PEAK-tier decision as deepseek-v4-flash
+    // (2026-08-20): conservative worst-case estimate; off-peak is
+    // exactly half (¥4.5 / ¥13.5 / ¥0.15).
+    //   Peak: $0.044 cache hit / $1.32 input / $3.96 output
+    //         ¥0.3  cache hit / ¥9   input / ¥27  output
+    pricingUSD: { input: 1.32, output: 3.96, cacheRead: 0.044 },
+    pricingCNY: { input: 9, output: 27, cacheRead: 0.3 },
     priceCategory: "low",
     thinkingSchema: "thinkingEnabled",
     maxInputTokens: 655360,
@@ -436,12 +440,15 @@ var PRIMARY_METADATA = {
     // `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
-    // input, so a guess would mislead budget estimates).
-    pricingUSD: { input: 1.4, output: 4.4, cacheRead: null },
-    pricingCNY: { input: 10, output: 31, cacheRead: null },
+    // Vendor pricing (corrected 2026-08-20): bigmodel.cn/pricing
+    // now lists GLM-5.3 at the SAME rate as GLM-5.2 —
+    //   ¥8 input / ¥28 output / ¥2 cache hit per 1M tokens
+    // The previous CNY 10/31 was a mis-read of the Decrypt launch
+    // article (2026-08-14). USD comes from Z.ai's independently-
+    // set international list ($1.4 / $4.4 / $0.26 — not the ×0.14
+    // CNY mirror), matching the reference consts.ts USD block.
+    pricingUSD: { input: 1.4, output: 4.4, cacheRead: 0.26 },
+    pricingCNY: { input: 8, output: 28, cacheRead: 2 },
     priceCategory: "high",
     maxInputTokens: 1e6,
     maxOutputTokens: 128e3
@@ -461,8 +468,10 @@ var PRIMARY_METADATA = {
     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 },
+    // architecture as the base model. USD = the standard ×0.14 mirror
+    // of 6/24/1.3 (2026-08-20 — fixed a copy-paste slip that had
+    // cloned glm-4.6's 0.6/2.2/0.11 USD block here, a ×0.10 rate).
+    pricingUSD: { input: 0.84, output: 3.36, cacheRead: 0.182 },
     pricingCNY: { input: 6, output: 24, cacheRead: 1.3 },
     priceCategory: "medium",
     maxInputTokens: 2e5,
@@ -725,6 +734,62 @@ var PRIMARY_METADATA = {
     maxInputTokens: 128e3,
     maxOutputTokens: 8192
   },
+  // GLM-4.6V-Flash — the FULLY-FREE vision model on
+  // bigmodel.cn/pricing's 视觉理解 tab (every column 免费,
+  // 2026-08-18 snapshot). This is the exact model the GLM-for-
+  // copilot reference uses as its transparent vision proxy
+  // (`DEFAULT_GLM_VISION_MODEL_ID` in its consts.ts — "the
+  // extension accepts images for text models through the
+  // GLM-4.6V-Flash vision proxy"). Zero cost makes it the
+  // natural candidate for our future ProxyVisionAdapter (v2).
+  // Token caps mirror the reference's entry (96K input + 32K
+  // output = the 128K shared window). Not in the zhipu preset —
+  // multimodal models aren't exposed via the public
+  // `/v1/chat/completions` models list (same trim rationale as
+  // glm-5v-turbo / glm-4.5v); users add it by hand.
+  "glm-4.6v-flash": {
+    detail: "GLM-4.6V Flash \u2014 \u5B8C\u5168\u514D\u8D39\u89C6\u89C9\u7406\u89E3\u6A21\u578B",
+    imageInput: true,
+    // Reference consts.ts declares GLM_TOOLS_LIMIT (128) tool
+    // calling for this model — the flash vision tier keeps the
+    // full function-call surface.
+    toolCalling: true,
+    // Thinking switch (auto-think family per the reference's
+    // `requiresThinkingParam: true`; no reasoning_effort —
+    // 4.6 < 5.2 threshold).
+    thinkingSchema: "reasoningEffort",
+    // Free tier — input / output / cache hit all 0 per
+    // bigmodel.cn 2026-08-18. USD mirrors CNY at 0 rather than
+    // inventing a rate.
+    pricingUSD: { input: 0, output: 0, cacheRead: 0 },
+    pricingCNY: { input: 0, output: 0, cacheRead: 0 },
+    priceCategory: "low",
+    maxInputTokens: 98304,
+    maxOutputTokens: 32768
+  },
+  // GLM-4.6V-FlashX — the 快速版 (quick) paid sibling of the free
+  // glm-4.6v-flash. Per bigmodel.cn/pricing 视觉理解 tab
+  // (2026-08-18):
+  //   [0, 32K)    — ¥0.15 input / ¥1.5 output / ¥0.03 cache hit
+  //   [32K, 128K) — ¥0.3 input  / ¥3 output   / ¥0.03 cache hit
+  // The LOWER tier is pinned per the project's multi-tier
+  // convention. Not in the zhipu preset (same multimodal trim
+  // rationale as glm-4.6v-flash). Token caps are a best-guess
+  // from sibling glm-4.6v (128K context / 8K output) — Zhipu
+  // publishes no dedicated spec for the flashx vision tier.
+  "glm-4.6v-flashx": {
+    detail: "GLM-4.6V FlashX \u2014 \u5FEB\u901F\u7248\u89C6\u89C9\u7406\u89E3\uFF082-tier pricing\uFF09",
+    imageInput: true,
+    // 快速版 convention: the flashx siblings (glm-4.7-flashx,
+    // glm-4.5-airx) drop tool calling.
+    toolCalling: false,
+    thinkingSchema: "reasoningEffort",
+    pricingUSD: { input: 0.021, output: 0.21, cacheRead: 42e-4 },
+    pricingCNY: { input: 0.15, output: 1.5, cacheRead: 0.03 },
+    priceCategory: "low",
+    maxInputTokens: 128e3,
+    maxOutputTokens: 8192
+  },
   "glm-5v-turbo": {
     detail: "GLM-5V Turbo \u2014 \u591A\u6A21\u6001 Coding \u6A21\u578B\uFF082-tier pricing\uFF09",
     imageInput: true,
@@ -877,9 +942,13 @@ var PRIMARY_METADATA = {
     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 },
+    // SiliconFlow 官方价格页（https://siliconflow.cn/pricing，
+    // re-fetched 2026-08-20）：¥1.8 input / ¥10.8 output per 1M
+    // tokens，无缓存价格列。此前 ¥1.6/¥12.8 来自发布新闻稿，
+    // 已过时（输入上调、输出下调）。USD = CNY ÷7 镜像取两位
+    // 小数（0.26 / 1.54）。
+    pricingUSD: { input: 0.26, output: 1.54, cacheRead: null },
+    pricingCNY: { input: 1.8, output: 10.8, 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）。
@@ -949,7 +1018,19 @@ function currencyForBaseUrl(baseUrl) {
     // 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") {
+    hostname === "dev.bigmodel.cn" || // DeepSeek — 2026-08-20 product decision (user request):
+    // surface the domestic ¥ prices. The `.com` endpoint
+    // serves both regions, but our user base bills in CNY on
+    // it (the official zh-cn pricing page publishes the ¥
+    // table for this endpoint), so the picker now shows the
+    // CNY block. This REVERSES the 2026-08-19 mapping that
+    // left it in the USD catch-all ("$ block corresponds to
+    // the .com endpoint") — with a CNY-billed user base, the
+    // ¥ block is the relevant one. The pricingCNY block in
+    // MODEL_METADATA pins the PEAK tier for these entries
+    // (conservative worst-case; see the deepseek-v4-flash
+    // comment).
+    hostname === "api.deepseek.com") {
       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" || // Z.ai / Zhipu international. Billed in USD per the
@@ -1362,7 +1443,7 @@ var BUILTIN_PROVIDER_PRESETS = {
     // 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",
+    baseUrl: "https://nexus.servicemecloud.com/v1",
     models: []
   }
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@serviceme/devtools-shared",
-  "version": "0.4.8",
+  "version": "0.4.10",
   "description": "Shared webview↔extension message contracts and cross-package data models used by SERVICEME.",
   "license": "SEE LICENSE IN LICENSE.md",
   "repository": {