@mars-sea/dsh-commandcode-provider 0.14.2 → 0.14.4

package/CHANGELOG.md

@@ -4,6 +4,44 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.14.4] - 2026-10-20
+
+### Changed
+
+- **同步 2026-10-20 的目录侧漂移（CLI 产物不变，仍是 1.84.0）。** 官方 CLI 的 npm `latest` 与变更日志本轮都没有新内容（`1.84.0` 于 2026-10-16T18:29Z 发布，日志顶部仍停在它），但实时模型目录在 10-19 晚变了：默认视图（`GET /provider/v1/models`）93 → 94 个 id，`?view=client` 101 → 102。快照随默认视图投影：`KNOWN_PLANS` 93 → 94、`KNOWN_EFFORTS` 65 → 66、`KNOWN_IMAGE_MODELS` 70 → 72、`KNOWN_DEALS` 7 → 7、`KNOWN_NON_ZDR_MODELS` 25 → 25；`COMMAND_CODE_CLI_VERSION` 与运行时代码不变（纯数据同步）。审计状态 `incomplete` 是预期形态：3 个 error 全是「目录比 1.84.0 产物新」的目录侧单源行（MiniMax M3.1 Flash 两行 + 既有的 `reflection/beam:free`），不是抓取失败。
+
+### Added
+
+- **MiniMax M3.1 Flash：付费 `minimax/minimax-m3.1-flash` 与限时免费 `minimax/minimax-m3.1-flash:free`。** 100 万上下文、文本 + 图像（Vision）、五档推理强度（`low`／`medium`／`high`／`xhigh`／`max`）、ZDR 覆盖、Go 起（全部套餐可见）。付费行 $0.15 输入／$0.60 输出／$0.015 缓存读（每百万 token，Go $6／GOAT $20／Pro $30 月额度）；免费行 100% off，促销期为 `2026-10-19T11:00Z` → `2026-10-21T10:45Z`（北京时间 10-19 19:00 → 10-21 18:45），官方说明「限时、无每日请求上限、账户需有 $1 余额」。**免费行是 `KNOWN_DEALS` 里第一个带 `expiresAt` 的免费促销**：到期后 FREE 徽标由 `dealLabel()` 自动消失，`:free` 行仍按 id 后缀 0 价，不会把过期折扣当成现价。
+
+- **定价页渲染表的「同一显示名多行」解析。** 上游自该模型上线起把付费行与免费行用同一个显示名 `MiniMax M3.1 Flash` 发布，`scripts/sync-model-prices.mjs` 原先「一个显示名一行」的唯一键假设因此让 `--check` 直接失败（`价格表模型名称重复，无法确定对应行：minimax m3.1 flash`）。现改为：渲染行按显示名分组，再按「免费 / 数值」形态唯一配对——同名同形态仍然拒绝生成，配到数值行后照旧逐项交叉校验三档费率与划线价。价格表随之新增一行 `minimax-m3.1-flash`（92 → 93 行）。同时给 `scripts/sync-output-limits.mjs` 的供应商映射补上小写 `minimax` 命名空间：新模型的目录 id 用的是小写前缀，缺这一项会退回聚合共识，把付费 id 的输出上限丢给运行时学习；补齐后两个 id 都落到供应商行（512000），表 78 → 80 行。
+
+### Removed
+
+- **Stealth Cat Jam 预览（`stealth/cat-jam:free`）从模型表下线。** 2026-10-20 检查时两个目录视图、定价页、模型文档页与 Go 套餐页都不再服务该行；`KNOWN_PLANS`／`KNOWN_EFFORTS`／`KNOWN_IMAGE_MODELS`／`KNOWN_DEALS` 移除，`tests/adapter.test.ts` 的免费模型排序夹具改用 Glyph Cluster。**保留在 `KNOWN_NON_ZDR_MODELS`**：1.84.0 产物的 ZDR 排除集仍点名它，而该表是「artifact 排除集 ∪ 目录 `zdr:false`」的并集，删条目会让老会话的 ZDR 判定偏乐观（判据见[《决策记录》第七章第 13 条](docs/决策记录.md#七上游快照同步边界)）。集合仍为 25 条。
+
+## [0.14.3] - 2026-10-19
+
+### Changed
+
+- **发布包不再携带 sourcemap，体积减少约 16%（tarball 2.9 MB → 2.5 MB，解包 5.1 MB → 3.5 MB；文件数 18 → 16）。** `tsdown.config.ts` 的 `sourcemap: true` 会产出 `lib/index.js.map` 与 `lib/client.js.map`，而 `files` 直接收整个 `lib`，两份 map（1.63 MiB）因此进了 tarball。它们没有任何读取方：`lib/client.js` 是 WebView 直接执行的 CJS 包，`//# sourceMappingURL` 只是尾部注释；`scripts/verify-engine-load.mjs` 的审计用正则匹配 `import`／`require` **语句**，不匹配注释。本地 `npm run build` 照旧产出 map，调试能力不受影响。**用户侧无需做任何事**，运行时行为与公开导出面零变化（`lib/index.d.ts` 的 `export { ... }` 一行未改）。
+
+- **长会话的历史分析报告更快（约 −28%～−33%）。** `buildAnalytics` 原先有两处对「事实 × 模型」的重复扫描：每模型小计对整表过滤 N 次，场景矩阵对每个（事实 × 目录模型）都查一次候选价格行（96 个模型共 480,000 次查找，查的是同一批行）。现改为先按模型分桶一次，并把定价算式收进 `cost-facts.requestCost()`、让场景循环把查价提到内层循环之外；`requestRates` 增加可选的高峰时段 memo（由调用方持有，不会活得比它的价格表更久）。两套独立测量给出同一量级：PR 作者脚本 5,000 条事实 × 96 个目录模型 42.17 ms → 30.36 ms，维护者本机用另一套合成数据复测 14.87 ms → 9.90 ms。**输出逐位不变**：分桶顺序与原 `Set` 顺序一致，排序本来就是按唯一 id 的全序（维护者另用 3,000 条确定性事实 × 96 个模型对比两版 JSON 输出，147,944 字节完全一致）。
+
+- **图片请求里同一张图只编码一次 base64。** `readImageRequest()` 对每个（附件，编码目标）返回同一个缓存对象，但一条请求里同一张图会出现在多条历史消息中；三个 wire 编解码器原先各自按「出现次数」执行 `Buffer.from(version.data).toString('base64')`（1 MiB 载荷 0.738 ms／次）。现改为以该缓存对象为键的 `WeakMap`，保存原始 base64 与 data URL，副本生命周期跟随宿主自己的版本缓存。命中约 21 ns。
+
+- **「查看模型用量、成本与性能分析」入口移到「账户用量」分组标题行**，紧邻「刷新」。它原先裸放在页面说明段与第一个分组之间，既无外边距也不属于任何分组，视觉上像是游离在版式之外；移到「账户用量」标题行后与服务着同一批用量／成本数据的「刷新」并列，也复用了该行既有的 `cc-spacer` 布局，不新增样式类。分析入口仍只在 Host 暴露面板接缝时出现，行为不变（切到面板的「模型分析」页签并打开面板）。
+
+- **多账户轮换说明精简 2 处冗余用词**（中文 47 → 43 字，英文同步）。`账户操作立即生效。当前账户达到用量限额或密钥失效时，请求会按列表顺序自动切换到下一个可用账户。` → 去掉「用量」与「自动」后为：`账户操作立即生效。当前账户达到限额或密钥失效时，请求会按列表顺序切换到下一个可用账户。` 「按列表顺序切换」本身已表达自动性，`达到限额` 在账户上下文里仍是额度含意；英文侧 `usage limit` → `limit`，`switch` 原本就没有 `automatically`。**已知取舍**：同页单账户文案 `accountsHint` 仍用「用量限额」，两条说明的术语因此不完全一致。
+
+### Added
+
+- **新增 `.gitattributes`：全部文本文件统一按 LF 检出与提交。** 仓库此前没有该文件，Windows 贡献者按 CRLF 检出后，sourcemap 的 `sourcesContent` 会内嵌带 CRLF 的源码，重建出的 `.map` 与提交里的不再逐字节相同，CI 的 `git diff --exit-code -- lib/` 守卫因此误报「产物过期」（2026-10-17 那次 CI 失败的根因）。实测对照：无该文件时 `core.autocrlf=true` 检出后 `src/adapter.ts` 有 3,482 行 CRLF、`lib/index.js` 有 11,817 行；加上后两者均为 0。图片与压缩包按 `binary` 排除在行尾转换之外。**本次加入时已跟踪文件 100% 是 LF，`git add --renormalize .` 未改写任何文件。**
+
+### Removed
+
+- 两份 sourcemap（`lib/index.js.map`、`lib/client.js.map`）不再纳入版本控制。它们是 tsdown 产物里唯一在 CRLF 工作树上无法逐字节重建的文件，跟踪它们只会让「产物过期」守卫依赖工作树的行尾；移出 tarball 与版本控制后，该守卫在任意平台都只反映源码与产物的真实一致性。
+
 ## [0.14.2] - 2026-10-18
 
 ### Changed

package/lib/client.js

@@ -3237,7 +3237,7 @@ window.__ModuleLoader__.load({
 			cardRouteActive: "已启用",
 			cardLoadingHint: "正在读取 Command Code 配置…",
 			cardRegistrationHint: "此卡片由 Command Code 插件提供，正在等待设置页装配，稍后即可显示完整内容。",
-			accountsRotationHint: "账户操作立即生效。当前账户达到用量限额或密钥失效时，请求会按列表顺序自动切换到下一个可用账户。",
+			accountsRotationHint: "账户操作立即生效。当前账户达到限额或密钥失效时，请求会按列表顺序切换到下一个可用账户。",
 			accountModeAuto: "使用方式：自动切换（按列表顺序）",
 			accountModePinned: "使用方式：固定使用「{name}」，不可用时仍会自动切换",
 			accountStatusPinned: "已固定",
@@ -3463,7 +3463,7 @@ window.__ModuleLoader__.load({
 			cardRouteActive: "Active",
 			cardLoadingHint: "Loading the Command Code configuration…",
 			cardRegistrationHint: "This card is contributed by the Command Code plugin and is waiting for the settings page to finish assembling.",
-			accountsRotationHint: "Account actions apply immediately. When the current account hits its usage limit or its key stops working, requests switch to the next usable account in list order.",
+			accountsRotationHint: "Account actions apply immediately. When the current account hits its limit or its key stops working, requests switch to the next usable account in list order.",
 			accountModeAuto: "Mode: switch automatically (list order)",
 			accountModePinned: "Mode: pinned to \"{name}\", still switching when it is unavailable",
 			accountStatusPinned: "Pinned",
@@ -3704,7 +3704,7 @@ window.__ModuleLoader__.load({
 		}
 		//#endregion
 		//#region package.json
-		var version = "0.14.2";
+		var version = "0.14.4";
 		var repository = {
 			"type": "git",
 			"url": "git+https://github.com/Mars-Sea/dsh-commandcode-provider.git"
@@ -5554,7 +5554,7 @@ window.__ModuleLoader__.load({
 		* each, with its status and quota inline — so one account's facts no longer
 		* have to be cross-read across separate surfaces.
 		*/
-		function AccountsCard({ t, state, usage, login, disabled, actions, onRefresh, createEnrollment }) {
+		function AccountsCard({ t, state, usage, login, disabled, actions, onRefresh, openAnalytics, createEnrollment }) {
 			const [adding, setAdding] = (0, react.useState)(false);
 			const [enrollment] = (0, react.useState)(createEnrollment);
 			const snapshot = (0, react.useSyncExternalStore)(enrollment.store.subscribe, enrollment.store.getSnapshot);
@@ -5582,6 +5582,12 @@ window.__ModuleLoader__.load({
 								children: t("accountsTitle")
 							}),
 							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", { className: "cc-spacer" }),
+							openAnalytics ? /* @__PURE__ */ (0, react_jsx_runtime.jsx)(_deepseek_ai_dsh_client_ui_primitives.Button, {
+								variant: "ghost",
+								size: "sm",
+								onClick: openAnalytics,
+								children: t("openAnalytics")
+							}) : null,
 							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
 								type: "button",
 								className: "cc-linkButton",
@@ -5929,12 +5935,6 @@ window.__ModuleLoader__.load({
 						className: "cc-intro",
 						children: t("intro")
 					}),
-					props.openAnalytics ? /* @__PURE__ */ (0, react_jsx_runtime.jsx)(_deepseek_ai_dsh_client_ui_primitives.Button, {
-						variant: "ghost",
-						size: "sm",
-						onClick: props.openAnalytics,
-						children: t("openAnalytics")
-					}) : null,
 					!state.writable ? /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
 						className: "cc-readOnly",
 						role: "status",
@@ -5948,6 +5948,7 @@ window.__ModuleLoader__.load({
 						disabled,
 						actions,
 						onRefresh: props.refreshUsage,
+						openAnalytics: props.openAnalytics,
 						createEnrollment: props.createEnrollment
 					}),
 					/* @__PURE__ */ (0, react_jsx_runtime.jsx)(ModelsCard, {
@@ -8376,10 +8377,16 @@ window.__ModuleLoader__.load({
 		* today, so the combination becoming real is visible before it reaches a user.
 		* This ordering is a latent choice, not a verified upstream rule.
 		*/
-		function requestRates(price, at, contextTokens, table) {
+		function requestRates(price, at, contextTokens, table, peakHourCache) {
 			const tier = price.contextTiers?.find((tier) => tier.maxContext === void 0 || contextTokens <= tier.maxContext);
 			if (tier !== void 0) return tier;
-			return price.peak !== void 0 && at !== null && peakHour(at, table.peakHours) ? price.peak : price;
+			if (price.peak === void 0 || at === null) return price;
+			let peak = peakHourCache?.get(at);
+			if (peak === void 0) {
+				peak = peakHour(at, table.peakHours);
+				peakHourCache?.set(at, peak);
+			}
+			return peak ? price.peak : price;
 		}
 		/** Cache version and wire guard: changing bands/rates must refold historical groups. */
 		function pricingKey(table) {

package/lib/index.d.ts

@@ -207,7 +207,13 @@ interface CommandCodeConnectionOptions {
  * text-only request never depends on the attachment seam.
  */
 type ResolveAttachments = () => AttachmentStore | undefined;
-/** What the rotation hook knows about the request it is rotating within. */
+/**
+ * What the rotation hook knows about the request it is rotating within.
+ *
+ * Not exported: the only consumer is the `rotateApiKey` shape below, so the
+ * name never reaches the published `.d.ts` and the package's `exports` map
+ * leaves no deep-import path to it.
+ */
 interface AccountRotationContext {
   /** Every API key this request has already used, just-rejected key included. */
   tried: readonly string[];
@@ -265,13 +271,13 @@ interface CommandCodeAdapterDeps<C extends CommandCodeConnectionOptions = Comman
    */
   imageOffload?: SurfaceImagePolicy;
 }
-/** Account identity from `/alpha/whoami`. */
+/** Account identity from `/alpha/whoami`; reached only through `CommandCodeUsageReport`. */
 interface CommandCodeAccount {
   id: string;
   name: string;
   userName: string;
 }
-/** Usage summary from `/alpha/usage/summary`. */
+/** Usage summary from `/alpha/usage/summary`; reached only through `CommandCodeUsageReport`. */
 interface CommandCodeUsage {
   totalCount: number;
   totalCost: number;
@@ -326,7 +332,7 @@ interface CommandCodeCredits {
     resetAt: number;
   };
 }
-/** Subscription plan state from `/alpha/billing/subscriptions`. */
+/** Subscription plan state from `/alpha/billing/subscriptions`; reached only through `CommandCodeUsageReport`. */
 interface CommandCodePlan {
   /** Raw subscription plan id (e.g. `individual-pro`); empty when unreported. */
   planId: string;
@@ -1041,8 +1047,10 @@ declare const KNOWN_IMAGE_MODELS: ReadonlySet<string>;
  * `minimax/minimax-m3-free` is the one entry the public catalog's default view
  * (`/provider/v1/models`) does not serve through the picker (it is CLI/pricing-visible
  * and hidden from the picker); it stays listed because the CLI carries it and a Go-plan
- * request can still name it. Both this row and `stealth/cat-jam:free` ARE served
- * by the `?view=client` catalog, which is where the 1.82.4 verdict comes from.
+ * request can still name it. At the 1.84.0 sync both this row and
+ * `stealth/cat-jam:free` were also served by the `?view=client` catalog, which
+ * is where the 1.82.4 verdict came from; the catalog stopped serving cat-jam on
+ * 2026-10-20, so the artifact's exclusion set is now its only evidence.
  *
  * command-code@1.82.4 moved the authority for this verdict: `knownModelSupportsZdr`
  * no longer negates a bundled Set, it returns `resolveKnownCatalogModel({model,
@@ -1080,6 +1088,12 @@ declare const KNOWN_IMAGE_MODELS: ReadonlySet<string>;
  * UNION of the artifact's exclusion set and the catalog's `zdr:false` list —
  * neither source alone is complete, which is why this set is not simply "what
  * the CLI says".
+ *
+ * The 2026-10-20 catalog-only sync changes no membership: `stealth/cat-jam:free`
+ * left both catalog views (its `zdr:false` verdict is no longer served), but the
+ * 1.84.0 artifact's exclusion set still names it and this set is the union, so
+ * membership stays at 25. The MiniMax M3.1 Flash launch pair is ZDR-covered
+ * (`zdr:true` on both ids) and joins no set.
  */
 declare const KNOWN_NON_ZDR_MODELS: ReadonlySet<string>;
 /**

package/lib/index.js

@@ -38,10 +38,30 @@ function peakHour(at, windows) {
 * today, so the combination becoming real is visible before it reaches a user.
 * This ordering is a latent choice, not a verified upstream rule.
 */
-function requestRates(price, at, contextTokens, table) {
+function requestRates(price, at, contextTokens, table, peakHourCache) {
 	const tier = price.contextTiers?.find((tier) => tier.maxContext === void 0 || contextTokens <= tier.maxContext);
 	if (tier !== void 0) return tier;
-	return price.peak !== void 0 && at !== null && peakHour(at, table.peakHours) ? price.peak : price;
+	if (price.peak === void 0 || at === null) return price;
+	let peak = peakHourCache?.get(at);
+	if (peak === void 0) {
+		peak = peakHour(at, table.peakHours);
+		peakHourCache?.set(at, peak);
+	}
+	return peak ? price.peak : price;
+}
+/**
+* Price one request's token buckets against an already-resolved price row.
+*
+* Split out of the report's per-model lookup so a caller that prices the same
+* candidate model against every fact — the analytics scenario matrix — resolves
+* the price row once per model instead of once per (fact, model) pair. The
+* arithmetic lives here alone so the two callers cannot drift.
+*/
+function requestCost(tokens, price, at, table, peakHourCache) {
+	if (price.free === true) return 0;
+	const r = requestRates(price, at, tokens.uncachedInputTokens + tokens.cacheReadTokens + tokens.cacheWriteTokens, table, peakHourCache);
+	if (tokens.cacheWriteTokens > 0 && r.cacheWriteCost === void 0) return void 0;
+	return (tokens.uncachedInputTokens * r.inputCost + tokens.outputTokens * r.outputCost + tokens.cacheReadTokens * r.cacheReadCost + tokens.cacheWriteTokens * (r.cacheWriteCost ?? 0)) / 1e6;
 }
 /** Cache version and wire guard: changing bands/rates must refold historical groups. */
 function pricingKey(table) {
@@ -103,9 +123,13 @@ const PLAN_ORDER = Object.fromEntries(PLAN_TIER_ORDER.map((tier, index) => [tier
 * subscription-plan labels, deals, and hourly (peak/off-peak) pricing.
 *
 * Everything here is synced from official sources — the command-code CLI
-* bundle's model table (`dist/cli.mjs`, re-verified at command-code@1.84.0) and
-* the official plan/pricing/model docs; see the dsh-commandcode-upstream skill
-* for the extraction procedures. Keeping the snapshot in its own module
+* bundle's model table (`dist/cli.mjs`, re-verified at command-code@1.84.0), the
+* live model catalog (`GET /provider/v1/models?view=client`, the authority for
+* capability fields since 1.82.4), and the official plan/pricing/model docs; see
+* the dsh-commandcode-upstream skill for the extraction procedures. The
+* 2026-10-20 sync is catalog-only: the CLI artifact stayed at 1.84.0 while the
+* catalog added the MiniMax M3.1 Flash launch pair and dropped the Stealth Cat
+* Jam preview. Keeping the snapshot in its own module
 * confines those frequent sync diffs here: src/adapter.ts holds only the stable
 * wire/runtime logic. The read helpers live here too, for the same reason.
 *
@@ -446,6 +470,20 @@ const KNOWN_EFFORTS = {
 		"medium",
 		"high"
 	],
+	"minimax/minimax-m3.1-flash": [
+		"low",
+		"medium",
+		"high",
+		"xhigh",
+		"max"
+	],
+	"minimax/minimax-m3.1-flash:free": [
+		"low",
+		"medium",
+		"high",
+		"xhigh",
+		"max"
+	],
 	"gpt-6.1-sol": [
 		"low",
 		"medium",
@@ -475,13 +513,6 @@ const KNOWN_EFFORTS = {
 		"medium",
 		"high",
 		"xhigh"
-	],
-	"stealth/cat-jam:free": [
-		"low",
-		"medium",
-		"high",
-		"xhigh",
-		"max"
 	]
 };
 /**
@@ -551,6 +582,8 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 	"meta/muse-spark-open",
 	"nvidia/nemotron-3.5-super-vl",
 	"mistral/mistral-large-4",
+	"minimax/minimax-m3.1-flash",
+	"minimax/minimax-m3.1-flash:free",
 	"moonshotai/Kimi-K2.5",
 	"moonshotai/Kimi-K2.6",
 	"moonshotai/Kimi-K2.7-Code",
@@ -597,8 +630,10 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 * `minimax/minimax-m3-free` is the one entry the public catalog's default view
 * (`/provider/v1/models`) does not serve through the picker (it is CLI/pricing-visible
 * and hidden from the picker); it stays listed because the CLI carries it and a Go-plan
-* request can still name it. Both this row and `stealth/cat-jam:free` ARE served
-* by the `?view=client` catalog, which is where the 1.82.4 verdict comes from.
+* request can still name it. At the 1.84.0 sync both this row and
+* `stealth/cat-jam:free` were also served by the `?view=client` catalog, which
+* is where the 1.82.4 verdict came from; the catalog stopped serving cat-jam on
+* 2026-10-20, so the artifact's exclusion set is now its only evidence.
 *
 * command-code@1.82.4 moved the authority for this verdict: `knownModelSupportsZdr`
 * no longer negates a bundled Set, it returns `resolveKnownCatalogModel({model,
@@ -636,6 +671,12 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 * UNION of the artifact's exclusion set and the catalog's `zdr:false` list —
 * neither source alone is complete, which is why this set is not simply "what
 * the CLI says".
+*
+* The 2026-10-20 catalog-only sync changes no membership: `stealth/cat-jam:free`
+* left both catalog views (its `zdr:false` verdict is no longer served), but the
+* 1.84.0 artifact's exclusion set still names it and this set is the union, so
+* membership stays at 25. The MiniMax M3.1 Flash launch pair is ZDR-covered
+* (`zdr:true` on both ids) and joins no set.
 */
 const KNOWN_NON_ZDR_MODELS = /* @__PURE__ */ new Set([
 	"MiniMaxAI/MiniMax-M3",
@@ -832,6 +873,8 @@ const MODEL_OUTPUT_TOKEN_LIMITS = /* @__PURE__ */ new Map([
 	["meta/muse-spark-1.3", 131072],
 	["meta/muse-spark-1.3-contributor", 131072],
 	["meta/muse-wonder-1.0", 131072],
+	["minimax/minimax-m3.1-flash", 512e3],
+	["minimax/minimax-m3.1-flash:free", 512e3],
 	["MiniMaxAI/MiniMax-M2.5", 131072],
 	["MiniMaxAI/MiniMax-M2.7", 131072],
 	["MiniMaxAI/MiniMax-M3", 512e3],
@@ -930,6 +973,8 @@ const KNOWN_PLANS = {
 	"meta/muse-spark-1.3-contributor": "go",
 	"meta/muse-spark-open": "go",
 	"mistral/mistral-large-4": "go",
+	"minimax/minimax-m3.1-flash": "go",
+	"minimax/minimax-m3.1-flash:free": "go",
 	"moonshotai/Kimi-K2.5": "go",
 	"moonshotai/Kimi-K2.6": "go",
 	"moonshotai/Kimi-K2.7-Code": "go",
@@ -942,7 +987,6 @@ const KNOWN_PLANS = {
 	"stepfun/Step-3.5-Flash": "go",
 	"stepfun/Step-3.7-Flash": "go",
 	"stepfun/Step-5-Preview": "go",
-	"stealth/cat-jam:free": "go",
 	"stealth/glyph-cluster:free": "go",
 	"tencent/hy3-paid": "go",
 	"tencent/hy4-preview": "go",
@@ -1129,6 +1173,11 @@ function modelVisibleForAnyAccount(modelId, accounts) {
 }
 const KNOWN_DEALS = {
 	"MiniMaxAI/MiniMax-M3": { label: "50% off" },
+	"minimax/minimax-m3.1-flash:free": {
+		label: "FREE",
+		free: true,
+		expiresAt: "2026-10-21T10:45:00Z"
+	},
 	"xiaomi/mimo-v2.5-pro": { label: "99% off" },
 	"xiaomi/mimo-v2.5": { label: "98% off" },
 	"poolside/laguna-s-2.1-free": {
@@ -1139,10 +1188,6 @@ const KNOWN_DEALS = {
 		label: "FREE",
 		free: true
 	},
-	"stealth/cat-jam:free": {
-		label: "FREE",
-		free: true
-	},
 	"stealth/glyph-cluster:free": {
 		label: "FREE",
 		free: true
@@ -2352,6 +2397,19 @@ const MODEL_PRICE_ROWS = [
 			pro: 57
 		}
 	},
+	{
+		id: "minimax-m3.1-flash",
+		rates: [
+			.15,
+			.6,
+			.015
+		],
+		allowance: {
+			go: 6,
+			goat: 20,
+			pro: 30
+		}
+	},
 	{
 		id: "mistral-large-4",
 		rates: [
@@ -3257,6 +3315,7 @@ function foldRequests(session, log) {
 	}
 	return facts;
 }
+/** The zeroed accumulator `aggregate` adds into. Internal: nothing outside this module reads it. */
 const emptyMetrics = () => ({
 	sessions: 0,
 	requests: 0,
@@ -3277,17 +3336,13 @@ const emptyMetrics = () => ({
 	decodeMs: 0,
 	decodeTokens: 0
 });
-function priceRequest(f, id, table, prices) {
+function priceRequest(f, id, table, prices, peakHourCache) {
 	if (!f.tokens) return void 0;
 	const price = prices.get(id);
 	if (!price) return void 0;
-	if (price.free === true) return 0;
-	const context = f.tokens.uncachedInputTokens + f.tokens.cacheReadTokens + f.tokens.cacheWriteTokens;
-	const r = requestRates(price, f.at, context, table);
-	if (f.tokens.cacheWriteTokens > 0 && r.cacheWriteCost === void 0) return void 0;
-	return (f.tokens.uncachedInputTokens * r.inputCost + f.tokens.outputTokens * r.outputCost + f.tokens.cacheReadTokens * r.cacheReadCost + f.tokens.cacheWriteTokens * (r.cacheWriteCost ?? 0)) / 1e6;
+	return requestCost(f.tokens, price, f.at, table, peakHourCache);
 }
-function aggregate(facts, table, prices) {
+function aggregate(facts, table, prices, peakHourCache) {
 	const m = emptyMetrics();
 	m.sessions = new Set(facts.map((f) => f.session)).size;
 	for (const f of facts) {
@@ -3298,7 +3353,7 @@ function aggregate(facts, table, prices) {
 			m.usageSamples++;
 			for (const k of COST_TOKEN_KEYS) m.tokens[k] += f.tokens[k];
 		}
-		const cost = priceRequest(f, f.model, table, prices);
+		const cost = priceRequest(f, f.model, table, prices, peakHourCache);
 		if (cost !== void 0) {
 			m.pricedRequests++;
 			m.cost += cost;
@@ -3323,13 +3378,21 @@ function buildAnalytics(facts, query, table, catalog, now = Date.now()) {
 	const selected = query.mode === "range" ? facts.filter((f) => f.at >= query.from && f.at < query.to) : facts;
 	const prices = /* @__PURE__ */ new Map();
 	for (const p of table.models) for (const id of [p.id, p.slug]) if (!prices.has(id)) prices.set(id, p);
-	const totals = aggregate(selected, table, prices);
-	const models = [...new Set(selected.map((f) => f.model))].map((id) => ({
+	const peakHourCache = /* @__PURE__ */ new Map();
+	const totals = aggregate(selected, table, prices, peakHourCache);
+	const byModel = /* @__PURE__ */ new Map();
+	for (const f of selected) {
+		const bucket = byModel.get(f.model);
+		if (bucket === void 0) byModel.set(f.model, [f]);
+		else bucket.push(f);
+	}
+	const models = [...byModel].map(([id, bucket]) => ({
 		id,
-		...aggregate(selected.filter((f) => f.model === id), table, prices)
+		...aggregate(bucket, table, prices, peakHourCache)
 	})).sort((a, b) => b.cost - a.cost || a.id.localeCompare(b.id));
 	const scenarios = catalog.map((model) => {
 		let cost = 0, pricedRequests = 0, excludedRequests = 0, unknownRequests = 0;
+		const price = prices.get(model.id);
 		for (const f of selected) {
 			if (!f.tokens) {
 				unknownRequests++;
@@ -3340,7 +3403,7 @@ function buildAnalytics(facts, query, table, catalog, now = Date.now()) {
 				excludedRequests++;
 				continue;
 			}
-			const priced = priceRequest(f, model.id, table, prices);
+			const priced = price === void 0 ? void 0 : requestCost(f.tokens, price, f.at, table, peakHourCache);
 			if (priced === void 0) unknownRequests++;
 			else {
 				cost += priced;
@@ -6307,9 +6370,16 @@ const DEFAULT_API_BASE = "https://api.commandcode.ai";
 * under the historical name so existing importers are unaffected.
 */
 const DEFAULT_GENERATE_MAX_TOKENS = MAX_OUTPUT_TOKEN_LIMIT;
+/** Catalog fetch budget. Internal: both call sites are in this module. */
 const MODELS_TIMEOUT_MS = 1e4;
 /** How long the picker's plan-filter billing facts stay cached before refetching. */
 const BILLING_ACCESS_TTL_MS = 3e5;
+/**
+* How long a learned CLI-only protocol preference stays cached per account.
+* After this TTL a request may probe Provider API again, so an upgraded Go
+* account can recover without a restart.
+*/
+const PROTOCOL_CACHE_TTL_MS = 9e5;
 /** The entry-plan tier weight (`individual-go` in KNOWN_SUBSCRIPTION_PLANS). */
 const GO_TIER_WEIGHT = 0;
 /** Hard cap on account rotations within one request (one attempt per distinct key). */
@@ -6960,6 +7030,31 @@ function priceRequestImage(block, model, imageCapable) {
 	};
 }
 /**
+* Base64 payloads keyed by the cached request version itself.
+*
+* `readImageRequest()` returns one codec object per (attachment, encoder
+* target) and `requestImages()` reuses that same object for every occurrence in
+* one generation, but a history that carries an image also replays it in every
+* later message. Encoding per occurrence therefore re-base64s the same
+* multi-megabyte buffer once per replay — the single most expensive
+* per-occurrence step of an image request. Keying on the object makes an
+* occurrence a map hit, and the WeakMap ties the copy's lifetime to the
+* harness's own version cache rather than to this plugin.
+*/
+const encodedRequestImages = /* @__PURE__ */ new WeakMap();
+function encodedRequestImage(version) {
+	let encoded = encodedRequestImages.get(version);
+	if (encoded === void 0) {
+		const base64 = Buffer.from(version.data).toString("base64");
+		encoded = {
+			base64,
+			dataUrl: `data:${version.mediaType};base64,${base64}`
+		};
+		encodedRequestImages.set(version, encoded);
+	}
+	return encoded;
+}
+/**
 * Convert one image reference to the final `/alpha/generate` wire format.
 * The official CLI accepts a base64 source internally, then toWireMessages()
 * sends a data URL plus mimeType. Use the request version's media type because
@@ -6969,7 +7064,7 @@ async function imageToCommandCode(ref, readImage) {
 	const version = await readImage(ref);
 	return {
 		type: "image",
-		image: `data:${version.mediaType};base64,${Buffer.from(version.data).toString("base64")}`,
+		image: encodedRequestImage(version).dataUrl,
 		mimeType: version.mediaType
 	};
 }
@@ -7102,10 +7197,9 @@ async function messagesToCC(messages, readImage) {
 * `{ type: 'image_url', image_url: { url: 'data:...;base64,...' } }`.
 */
 async function imageToOpenAI(ref, readImage) {
-	const version = await readImage(ref);
 	return {
 		type: "image_url",
-		image_url: { url: `data:${version.mediaType};base64,${Buffer.from(version.data).toString("base64")}` }
+		image_url: { url: encodedRequestImage(await readImage(ref)).dataUrl }
 	};
 }
 /**
@@ -7172,7 +7266,7 @@ async function imageToMessages(ref, readImage) {
 		source: {
 			type: "base64",
 			media_type: version.mediaType,
-			data: Buffer.from(version.data).toString("base64")
+			data: encodedRequestImage(version).base64
 		}
 	};
 }
@@ -8174,7 +8268,7 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 	/** Cached protocol decision for a key, or undefined when expired/unknown. */
 	cachedProtocol(apiBase, apiKey) {
 		const hit = this.protocolCache.get(this.accountCacheKey(apiBase, apiKey));
-		if (hit === void 0 || Date.now() - hit.at >= 9e5) return void 0;
+		if (hit === void 0 || Date.now() - hit.at >= PROTOCOL_CACHE_TTL_MS) return void 0;
 		return hit.protocol;
 	}
 	rememberProtocol(apiBase, apiKey, protocol) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mars-sea/dsh-commandcode-provider",
-  "version": "0.14.2",
+  "version": "0.14.4",
   "description": "Unofficial DeepSeek Harness LLM provider plugin for Command Code, ported from pi-commandcode-provider (MIT). Registers the 'commandcode' provider route with a Models-page card and live model catalog.",
   "icon": "./icon.svg",
   "type": "module",
@@ -39,6 +39,7 @@
   },
   "files": [
     "lib",
+    "!lib/**/*.map",
     "assets",
     "icon.svg",
     "locale",