@eddyskywalker/dsh-chatgpt-subscription 0.65.0 → 0.66.1

package/CHANGELOG.md

@@ -2,6 +2,53 @@
 
 ## Unreleased
 
+## 0.66.1 - 2026-11-09
+
+- **[kimi-code / 号池] 瞬时失败保持账号亲和，只有明确的账号级限制才冷却换号**
+  - 普通 429 背压与上游引擎过载不再被当成本账号额度耗尽：重试沿用原账号，不冷却、不轮换；只有响应正文能够证明账号额度或账号限制时，才进入冷却并选择其他账号。流内错误与响应头错误共用同一套分类，避免换号破坏前缀缓存亲和。
+  - 持久化有界、非机密的最近请求诊断，卡片显示实际账号、是否切号、请求结果、输入/缓存读写/输出 token 及 reasoning 占比；高推理占比时明确提示成本主要来自推理，而不是缓存失效。
+
+- **[kimi-code / 工具调用] 拦截失控参数流，并允许同一步在原账号引导恢复一次**
+  - 两种线上协议都在工具执行前拦截超过线路字节预算的参数，以及失控的重复键参数族；`ask_user_question`、大型程序与普通工具分别沿用适合其用途的限制，原始失控参数不会写入恢复提示。
+  - 第一次命中专用守卫错误后，在失败 attempt 之后追加最小合法 JSON 提示，由同一个 turn/step 立即重试，并钉住原账号以保留缓存；第二次仍失控则交还默认错误处理，不形成无限恢复循环。
+
+- **[cache-warmup / 诊断] 扩大安全结果预览范围，区分前缀延续与新增输入**
+  - 有明确通过汇总且无失败证据的测试日志、严格识别根形状的 PTC 工具目录，以及超过当前预算且能可靠结构化的 grep/glob 结果可生成 preview；完整名称、错误集合与汇总按策略保留，完整原文仍通过 `ref://` 补读。混合证据、无法可靠解析或可能丢失关键事实的内容继续逐字保留。
+  - Kimi 缓存卡片新增「最近前缀延续」，分别展示最近请求缓存读、新增输入与按 256-token 对齐基线估算的保留比例；该指标仅用于诊断，不声明上游内部缓存契约。
+
+- **[glm-coding-plan] 模型目录构建不再写账号池**
+  - 目录只需要区域信息，不再调用真实请求选号入口；按号池 revision 与指定账号缓存只读区域快照，避免一次模型选择器刷新被放大成重复 DPAPI 解密、凭据读取和 `lastUsedAt` 回写。未登录或读取失败时不缓存失败结果，登录后仍可重试。
+
+- **[client] 修复设置枢纽主按钮悬停与登录行视觉层级**
+  - 共享 hover 洗色不再匹配 `.dsha-btn-primary`，避免「保存设置」「添加账号」等主按钮悬停时底色消失、白字落在近白背景上；Antigravity 与 Kimi 两份相同样式同步修复。
+  - 删除登录行对所有按钮的无差别 info 填充，保留「添加账号」主按钮与「设备码登录」「导入本地登录」次要按钮的视觉层级，布局规则不变。
+
+- **验证**：完整类型检查通过；清洁构建后全量 Vitest 为 384 个文件通过 / 2 个跳过（6992 项通过 / 9 项跳过 / 0 失败）；`npm pack --dry-run` 为 977 个文件、约 4.3 MB。按钮修复以 v0.66.0 与当前版本在相同 DOM、760×300 截图区域及浅色主题下完成 Before/After 复核。
+
+## 0.66.0 - 2026-11-09
+
+- **[shared / cache] Kimi Code 的缓存与运行时优化沉淀为跨线路共享能力，并按上游证据启用**
+  - **共享机制**：新增缓存策略注册表、静态前缀稳定化、冷启动请求协调、账号缓存亲和、持久化活动/图片状态、推理方言学习、失败冷却边界与 hosted-media 单飞内核。Codex、Antigravity、Step、Claude、MiniMax 获得静态前缀与 in-flight 冷启动协调；Cline、OpenDesign、Qoder 可学习并回放实际 reasoning 字段；Claude/MiniMax 的粘性图片状态可跨重启恢复。
+  - **全线低风险收益**：重复工具调用检测与分级提醒覆盖受控线路；大工具结果 `ref://` 预览及 `cache_warmup_result` 精确补读不再绑死 Kimi。只有 Kimi 保留第 12 次强停，MiniMax 保留官方错误族、ABAB 与轮询规则，其他线路不继承强停。
+  - **证据门控**：新增 `toolRunawayGuard`、`cacheWarmupProviders` 与非 volatile 的 `promptCacheOverrides`。Kimi 的 25 分钟 TTL、等待续热、账号级亲和与 `ms://` 协议保持原语义；其他线路只有在明确 TTL、账号隔离或上传协议证据时才开启对应行为。Step 另获官方 5 分钟 TTL 的缓存过期提示与可选亲和选号。
+  - **回归与界面**：新增策略、前缀、协调、活动、推理、媒体和 UI 测试；设置区新增工具失控治理开关，中英俄三语同步，并在原 GUI 同视口/主题完成 Before/After 复核。强制类型检查、构建、打包检查与全量 Vitest 均通过（6952 例通过 / 9 跳过 / 0 失败）。
+
+- **[kimi-code / 号池] 多账号轮换按「前缀缓存还热着」选账号；冷启动组键补上账号维度**
+  - **问题**：服务端的前缀缓存是**按账号**各存一份，而内核 `selectAccount` 只看冷却、令牌与最近使用，永远看不到「这条前缀在哪个账号上还热着」——于是轮换可能把请求发给一个从没见过这段前缀的账号。更糟的是冷启动组键 `region:model:systemPromptHash:toolsHash[:staticPrefixHash]` **不含账号**，插件会在另一个账号烤热之后误判「这组是热的」，跟随者白等 2.5 秒预算，而它那条请求本来就不会命中。
+  - **改法**（`src/host/common/account-pool.ts`、`src/host/kimi-code/{account-pool,adapter,cache-affinity}.ts`、`src/index.ts`）：内核新增 `AccountSelectionHint` 与 `getEffectiveAccount(..., selection)`——顺序是**钉住账号 > 请求偏好 > 轮换策略**，偏好里不可用/已尝试的 id 直接跳过，全跳过则回落策略（辅助工具的 `getCredentialAccount` 不消费偏好，读凭据不动轮换）；`coldStartGroupKey` 追加 `|accountScope`，并新增 `prefixGroupKey` / `isWarmPrefixGroup` 让两处共用同一个「热」定义；新模块 `cache-affinity.ts` 把取舍写成纯函数（剩余额度取**最紧窗口**、`=20%` 仍可用、配额未知按「没证明不足」处理、同份额保持池序）；适配器在取凭据之前算一次偏好，只对真的热着的账号读配额。新增默认开启的 volatile 开关 `cacheAffinity`，设置页「进展更新」一格加一行，三语齐备，`DSH_KIMI_CODE_CACHE_AFFINITY` 可覆盖。
+  - **刻意不做**：不把亲和做成第四种轮换策略（用户不必为了收益改策略）；不因亲和失败而报错（热账号不可用就照旧回落）；组键变更会让现存会话各自冷启动一次，这是一次性成本。
+  - **回归**：新增 `test/kimi-code-cache-affinity.test.ts` 15 例（最紧窗口、20% 边界、未知配额、冷账号永不入列、稳定排序、开关解析与环境变量）；`test/account-pool-core.test.ts` 3 例（偏好压过策略、钉住压过偏好、偏好不可用时回落）；`test/kimi-code-adapter.test.ts` 3 例（组键按账号区分、跨账号不再共享冷启动闸门、热账号被优先选中且开关关闭后回到池的选择）——其中「跨账号不再共享闸门」做了反向验证：把组键改回不带账号，该用例立刻失败。
+
+- **[kimi-code / 诊断] 卡片显示「开局继承」：这个会话的第一条请求有多少 token 是继承来的**
+  - **背景**：跨会话前缀继承是真实存在的（同账号的另一个会话写下的头段，25 分钟内可被新会话读回），但它此前只体现在滚动命中率里，看不出「开局那一两千到上万 token 是谁付的」。同一台机器上的实测：新会话在另一会话结束 3 分钟后开工，首轮读到 14,080 / 21,527（65%）；隔 5 小时续跑则只剩 2,304（14%）。
+  - **改法**（`src/host/kimi-code/{session-cache-state,adapter,routes}.ts`、`src/shared/kimi-code-contracts.ts`、`src/client/kimi-code/`）：持久记录新增 `firstRequestCachedTokens` / `firstRequestPromptTokens`，只在会话首轮写入且之后原样保留（滚动合计会淹没这个数字）；DTO 增 `opening`，卡片在「最近前缀状态」下方显示 `开局继承 14,080 / 21,527 tok（65.4%）`，无数据时不画这一行（旧记录与单账号首轮冷启动都属于「没有可说的」）。
+  - **回归**：`test/kimi-code-adapter.test.ts` 1 例（首轮写入、次轮不覆盖）；`test/kimi-code-k31-card-ui.test.tsx` 2 例（有数据时的三个数字、无数据时不画该行）。
+
+- **[kimi-code / 结论归档] 「把全局与工作区 AGENTS.md 拆成两条以提高跨工作区继承」经证伪，不实施**
+  - **依据**：宿主 `@deepseek-ai/dsh-agent-instructions` 的 `render.ts` 把每个指令文件渲染成 `Instructions from: <path>\n\n<content>`，全部 section 拼进**同一条** `<system-reminder>`，且 `USER_GLOBAL_DIRECTORIES` 决定全局段本来就在前；本仓库 mapper 的 `orderStaticPrefixMessages` 只按 tier 重排，不改变块内字节顺序。因此拆成两条消息后，跨工作区能匹配到的前缀长度**一个字节都不会变**。
+  - **还有一个反效果**：现有重排按 tier 只提升**首次出现**（`seenTiers` / `hoistedByTier`），第二条 tier-2 会被当成 rank 5 挤到用户消息之后——照原方案做会实实在在变差。
+  - **仍成立的部分**：工具说明与 `code_run` 恒定说明段本来就在 system prompt 里、已包含在共享的约 14k 头段内；要扩大共享头段只能让更多「全工作区相同」的内容排在第一个工作区相关字节之前，而目前除了语言覆盖已没有这类内容。
+
 ## 0.65.0 - 2026-11-08
 
 - **[antigravity] 出站请求身份、版本清单跟随与控制面契约对齐官方客户端**

package/lib/client.js

@@ -5072,6 +5072,22 @@ window.__ModuleLoader__.load({
 			unknownMetric: "暂无",
 			cacheExpired: "本会话已空闲超过缓存有效期，下一轮将重新处理约 {tokens} 个 token。",
 			cacheDrift: "最近前缀状态",
+			cacheOpening: "开局继承",
+			cacheOpeningValue: "{cached} / {tokens} tok（{ratio}）",
+			cacheContinuation: "最近前缀延续",
+			cacheContinuationValue: "{cached} / {expected} tok（{ratio}）· 新增 {fresh} tok",
+			cacheLastRequest: "最近请求",
+			cacheRequestSuccess: "成功",
+			cacheRequestError: "失败",
+			cacheRequestAborted: "已中止",
+			cacheAccountChanged: "（已切号）",
+			cacheLastInput: "本次输入",
+			cacheLastInputValue: "新增 {fresh} · 缓存读 {cached} · 缓存写 {written}",
+			cacheLastOutput: "本次输出",
+			cacheLastReasoning: "其中 reasoning",
+			cacheReasoningValue: "{tokens} tok（占输出 {ratio}）",
+			cacheReasoningCost: "本步成本主要来自推理，并非缓存失效。",
+			metricNotReported: "未报告",
 			cacheDriftFirst: "首个请求（缓存为空）",
 			cacheDriftStable: "稳定（前缀未变）",
 			cacheDriftSystemPrompt: "系统提示已变化（缓存失效）",
@@ -5243,6 +5259,22 @@ window.__ModuleLoader__.load({
 			cacheWarmupSettled: "Steady state ({hits}/{required}) · cap {chars} chars",
 			unknownMetric: "n/a",
 			cacheDrift: "Last prefix state",
+			cacheOpening: "Opening inheritance",
+			cacheOpeningValue: "{cached} / {tokens} tok ({ratio})",
+			cacheContinuation: "Latest prefix continuity",
+			cacheContinuationValue: "{cached} / {expected} tok ({ratio}) · {fresh} fresh tok",
+			cacheLastRequest: "Latest request",
+			cacheRequestSuccess: "success",
+			cacheRequestError: "error",
+			cacheRequestAborted: "aborted",
+			cacheAccountChanged: " (account changed)",
+			cacheLastInput: "Input this request",
+			cacheLastInputValue: "{fresh} fresh · {cached} cache read · {written} cache write",
+			cacheLastOutput: "Output this request",
+			cacheLastReasoning: "Reasoning included",
+			cacheReasoningValue: "{tokens} tok ({ratio} of output)",
+			cacheReasoningCost: "Reasoning, not a cache miss, drove this request cost.",
+			metricNotReported: "not reported",
 			cacheDriftFirst: "First request (cache cold)",
 			cacheDriftStable: "Stable (prefix unchanged)",
 			cacheDriftSystemPrompt: "System prompt changed (cache invalidated)",
@@ -5837,6 +5869,7 @@ window.__ModuleLoader__.load({
 		* number this build cannot read at all falls back to "暂无".
 		*/
 		function tokenCount(value, t) {
+			if (value === null) return t.metricNotReported;
 			return typeof value === "number" && Number.isFinite(value) ? value.toLocaleString() : t.unknownMetric;
 		}
 		/**
@@ -5866,6 +5899,64 @@ window.__ModuleLoader__.load({
 				default: return String(cause);
 			}
 		}
+		function KimiLatestRequestRows(input) {
+			const { request, status, t } = input;
+			const account = status.accounts?.find((entry) => entry.id === request.accountId)?.alias ?? request.accountId ?? t.unknownMetric;
+			const outcome = request.outcome === "success" ? t.cacheRequestSuccess : request.outcome === "aborted" ? t.cacheRequestAborted : t.cacheRequestError;
+			const usage = request.usage;
+			const reasoning = usage?.reasoningTokens;
+			const reasoningRatio = usage !== void 0 && reasoning !== null && reasoning !== void 0 && usage.outputTokens > 0 ? reasoning / usage.outputTokens : null;
+			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)(react_jsx_runtime.Fragment, { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+				className: "dsha-row",
+				children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+					className: "dsha-label",
+					children: t.cacheLastRequest
+				}), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("span", {
+					className: "dsha-value",
+					children: [
+						account,
+						request.accountChanged ? t.cacheAccountChanged : "",
+						" · ",
+						outcome
+					]
+				})]
+			}), usage !== void 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsxs)(react_jsx_runtime.Fragment, { children: [
+				/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+					className: "dsha-row",
+					children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+						className: "dsha-label",
+						children: t.cacheLastInput
+					}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+						className: "dsha-value",
+						children: t.cacheLastInputValue.replace("{fresh}", tokenCount(usage.freshTokens, t)).replace("{cached}", tokenCount(usage.cacheReadTokens, t)).replace("{written}", tokenCount(usage.cacheWriteTokens, t))
+					})]
+				}),
+				/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+					className: "dsha-row",
+					children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+						className: "dsha-label",
+						children: t.cacheLastOutput
+					}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+						className: "dsha-value",
+						children: tokenCount(usage.outputTokens, t)
+					})]
+				}),
+				/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+					className: "dsha-row",
+					children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+						className: "dsha-label",
+						children: t.cacheLastReasoning
+					}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+						className: "dsha-value",
+						children: reasoning == null ? t.metricNotReported : reasoningRatio === null ? tokenCount(reasoning, t) : t.cacheReasoningValue.replace("{tokens}", tokenCount(reasoning, t)).replace("{ratio}", `${(reasoningRatio * 100).toFixed(1)}%`)
+					})]
+				}),
+				reasoning != null && reasoning >= 1e3 && reasoningRatio !== null && reasoningRatio >= .5 && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+					className: "dsha-notice",
+					children: t.cacheReasoningCost
+				})
+			] })] });
+		}
 		function KimiCodeSection({ onModelChange, loadModelDirectory }) {
 			const [status, setStatus] = (0, react.useState)(null);
 			const [loading, setLoading] = (0, react.useState)(true);
@@ -6436,6 +6527,31 @@ window.__ModuleLoader__.load({
 											children: driftLabel(status.cache.lastDrift, t)
 										})]
 									}),
+									status.cache.opening !== void 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+										className: "dsha-row",
+										children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+											className: "dsha-label",
+											children: t.cacheOpening
+										}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+											className: "dsha-value",
+											children: t.cacheOpeningValue.replace("{cached}", status.cache.opening.cachedTokens.toLocaleString()).replace("{tokens}", status.cache.opening.promptTokens.toLocaleString()).replace("{ratio}", `${(status.cache.opening.hitRatio * 100).toFixed(1)}%`)
+										})]
+									}),
+									status.cache.continuation !== void 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+										className: "dsha-row",
+										children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+											className: "dsha-label",
+											children: t.cacheContinuation
+										}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+											className: "dsha-value",
+											children: t.cacheContinuationValue.replace("{cached}", status.cache.continuation.cachedTokens.toLocaleString()).replace("{expected}", status.cache.continuation.expectedCachedTokens.toLocaleString()).replace("{fresh}", status.cache.continuation.freshTokens.toLocaleString()).replace("{ratio}", `${(status.cache.continuation.retentionRatio * 100).toFixed(1)}%`)
+										})]
+									}),
+									status.cache.lastRequest !== void 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsx)(KimiLatestRequestRows, {
+										status,
+										request: status.cache.lastRequest,
+										t
+									}),
 									/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
 										className: "dsha-row",
 										children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
@@ -12286,6 +12402,7 @@ window.__ModuleLoader__.load({
 			cacheRequestParamsOnHint: "已打开能力位：请求会带上缓存参数。该字段未在 Step 文档中说明，官方对 Step 端点默认也不发。",
 			cachePricingNote: "单价未文档化（官方模型表里四个 cost 键全是 0），因此本页只展示 token 维度，不展示任何金额。",
 			cacheEmpty: "本进程还没有任何请求回报过用量，跑一次对话后再看。",
+			cacheExpiredHint: "缓存可能已过期：距上次请求约 {minutes} 分钟，下次会重新处理前缀。",
 			cacheHitRatioUnavailable: "—"
 		};
 		const en$8 = {
@@ -12453,6 +12570,7 @@ window.__ModuleLoader__.load({
 			cacheRequestParamsOnHint: "The capability is on: requests carry cache parameters. Step does not document them, and the official client does not send them to Step either.",
 			cachePricingNote: "Unit prices are undocumented (all four cost keys are 0 in the official model table), so this page shows TOKENS ONLY and no money figure at all.",
 			cacheEmpty: "No request on this machine has reported usage yet — run a conversation and look again.",
+			cacheExpiredHint: "The cache has likely expired: about {minutes} minutes since the last request, so the next turn will reprocess the prefix.",
 			cacheHitRatioUnavailable: "—"
 		};
 		const dictionaries$8 = {
@@ -13597,6 +13715,11 @@ window.__ModuleLoader__.load({
 									})
 								]
 							}),
+							status?.cacheHint !== void 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+								className: "dsha-muted",
+								"data-dsh-part": "step-cache-hint",
+								children: t.cacheExpiredHint.replace("{minutes}", String(Math.floor(status.cacheHint.idleMs / 6e4)))
+							}),
 							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
 								className: "dsha-pref-row",
 								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("strong", { children: t.cacheRequestParams }), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
@@ -21656,15 +21779,19 @@ window.__ModuleLoader__.load({
 			settingsEnabledName: "进展更新",
 			settingsEnabledNote: "让 agent 在对话中发布进展更新；本插件的所有线路都可用",
 			settingsCacheWarmupName: "缓存预热",
-			settingsCacheWarmupNote: "Kimi Code 专用：投影大型日志结果，原文可通过 ref 按需补读；非 Kimi 线路开关无影响",
+			settingsCacheWarmupNote: "按线路配置：投影大型日志结果，原文可通过 ref 按需补读；只影响之后的新结果",
 			settingsRetrievalName: "检索结果投影",
-			settingsRetrievalNote: "Kimi Code 专用：导航任务中投影文件、网页与搜索，编辑审查保留原文；非 Kimi 线路开关无影响",
+			settingsRetrievalNote: "按线路配置：导航任务中投影文件、网页与搜索，编辑审查保留原文；只影响之后的新结果",
 			settingsCacheWaitWarmupName: "等待缓存续热",
-			settingsCacheWaitWarmupNote: "Kimi Code 专用：等待子代理期间每满 20 分钟自动触发一次轻量续热，只要子代理未完成就持续保活；非 Kimi 线路开关无影响",
+			settingsCacheWaitWarmupNote: "等待子代理期间按缓存 TTL 轻量续热；默认用于 Kimi，其他线路需明确 TTL 并显式启用",
+			settingsCacheAffinityName: "账号缓存亲和",
+			settingsCacheAffinityNote: "多账号轮换时优先选前缀仍热的账号，低额度时让位；默认用于 Kimi，其他线路需证明账号级缓存并显式启用",
 			settingsPlanModePipelineName: "计划模式管线",
 			settingsPlanModePipelineNote: "澄清访谈收敛后自动进入计划模式出计划，再交审批（仅调度模式生效）；关闭后该模式会请用户手动输入 /plan",
 			settingsDispatchExecutionGuardName: "调度执行守卫",
 			settingsDispatchExecutionGuardNote: "执行期拦截子代理越权调用主控工具及等待子代理时的纯 sleep 空等；不改工具声明面，零损耗保护 Kimi 缓存",
+			settingsToolRunawayGuardName: "工具失控治理",
+			settingsToolRunawayGuardNote: "同一工具与参数连续重复时分级提醒；Kimi 保留第 12 次强停，其他线路只提醒",
 			settingsTeamPolicyCacheGuardName: "智能体团队策略段",
 			settingsTeamPolicyCacheGuardNote: "把内置 Agent Teams 的策略段从系统提示词摘掉，改为开启团队协作时以尾部消息投递一次；会话中途切开关不再刷掉整段前缀缓存（默认开启）",
 			settingsThresholdName: "催促阈值",
@@ -21705,15 +21832,19 @@ window.__ModuleLoader__.load({
 			settingsEnabledName: "Progress updates",
 			settingsEnabledNote: "Let the agent post progress updates in the conversation; every line of this plugin can use it",
 			settingsCacheWarmupName: "Cache warmup",
-			settingsCacheWarmupNote: "Kimi Code only: project large log results and read the original on demand through a ref; no effect on other lines",
+			settingsCacheWarmupNote: "Per-line control: project large log results and read the original through a ref on demand; only new results are affected",
 			settingsRetrievalName: "Retrieval projection",
-			settingsRetrievalNote: "Kimi Code only: project files, pages, and searches for navigation; keep source verbatim for editing and review; no effect on other lines",
+			settingsRetrievalNote: "Per-line control: project files, pages, and searches for navigation; keep source verbatim for editing and review; only new results are affected",
 			settingsCacheWaitWarmupName: "Wait-time cache keepalive",
-			settingsCacheWaitWarmupNote: "Kimi Code only: trigger a lightweight warmup every 20 minutes while waiting for subagents so the prefix cache stays warm until they finish; no effect on other lines",
+			settingsCacheWaitWarmupNote: "Keep the prefix warm while subagents run, using the line TTL; on by default for Kimi, other lines need a known TTL and explicit opt-in",
+			settingsCacheAffinityName: "Account cache affinity",
+			settingsCacheAffinityNote: "Prefer the account whose prefix is still warm, but yield when quota runs low; on by default for Kimi, other lines need proven account-scoped caching and explicit opt-in",
 			settingsPlanModePipelineName: "Plan-mode pipeline",
 			settingsPlanModePipelineNote: "After the clarification interview converges, enter plan mode, write the plan and send it for approval (dispatch mode only); turning it off makes that mode ask the user to type /plan",
 			settingsDispatchExecutionGuardName: "Dispatch execution guard",
 			settingsDispatchExecutionGuardNote: "Block subagents from calling dispatcher-only tools at execution time and reject pure sleep waits while child agents run; keeps tool declarations untouched to protect the Kimi prefix cache",
+			settingsToolRunawayGuardName: "Tool runaway guard",
+			settingsToolRunawayGuardNote: "Remind the agent when the same tool call repeats; Kimi keeps its 12th-call handoff, while other lines receive reminders only",
 			settingsTeamPolicyCacheGuardName: "Agent Teams policy section",
 			settingsTeamPolicyCacheGuardNote: "Lift the built-in Agent Teams policy section out of the system prompt and deliver it as one tail message when team collaboration is switched on, so toggling mid-session no longer rebuilds the whole prefix cache (on by default)",
 			settingsThresholdName: "Nudge threshold",
@@ -22645,6 +22776,22 @@ window.__ModuleLoader__.load({
 			unknownMetric: "нет данных",
 			cacheExpired: "Сессия простаивала дольше времени жизни кеша, поэтому следующий ход заново обработает около {tokens} токенов.",
 			cacheDrift: "Состояние последнего префикса",
+			cacheOpening: "Наследование на старте",
+			cacheOpeningValue: "{cached} / {tokens} ток. ({ratio})",
+			cacheContinuation: "Непрерывность последнего префикса",
+			cacheContinuationValue: "{cached} / {expected} ток. ({ratio}) · новых: {fresh} ток.",
+			cacheLastRequest: "Последний запрос",
+			cacheRequestSuccess: "успешно",
+			cacheRequestError: "ошибка",
+			cacheRequestAborted: "прервано",
+			cacheAccountChanged: " (аккаунт сменился)",
+			cacheLastInput: "Ввод запроса",
+			cacheLastInputValue: "новых: {fresh} · чтение кеша: {cached} · запись кеша: {written}",
+			cacheLastOutput: "Вывод запроса",
+			cacheLastReasoning: "В том числе reasoning",
+			cacheReasoningValue: "{tokens} ток. ({ratio} вывода)",
+			cacheReasoningCost: "Стоимость этого шага обусловлена рассуждениями, а не промахом кеша.",
+			metricNotReported: "не сообщено",
 			cacheDriftFirst: "Первый запрос (кеш пуст)",
 			cacheDriftStable: "Стабильно (префикс не менялся)",
 			cacheDriftSystemPrompt: "Системный промпт изменился (кеш сброшен)",
@@ -23127,15 +23274,19 @@ window.__ModuleLoader__.load({
 			settingsEnabledName: "Ход работы",
 			settingsEnabledNote: "Разрешить агенту публиковать ход работы в диалоге; доступно на всех линиях этого плагина",
 			settingsCacheWarmupName: "Прогрев кеша",
-			settingsCacheWarmupNote: "Только Kimi Code: проецировать большие журналы и читать оригинал по ref при необходимости; на других линиях переключатель ничего не меняет",
+			settingsCacheWarmupNote: "По линиям: проецировать большие журналы и читать оригинал по ref при необходимости; затрагиваются только новые результаты",
 			settingsRetrievalName: "Проекция результатов поиска",
-			settingsRetrievalNote: "Только Kimi Code: проецировать файлы, страницы и поиск при навигации; сохранять код при правке и ревью; на других линиях переключатель ничего не меняет",
+			settingsRetrievalNote: "По линиям: проецировать файлы, страницы и поиск при навигации; сохранять код при правке и ревью; затрагиваются только новые результаты",
 			settingsCacheWaitWarmupName: "Поддержание кеша при ожидании",
-			settingsCacheWaitWarmupNote: "Только Kimi Code: каждые 20 минут ожидания субагентов отправлять лёгкий прогрев, пока субагенты работают, чтобы префиксный кеш не остывал; на других линиях переключатель ничего не меняет",
+			settingsCacheWaitWarmupNote: "Поддерживать префиксный кеш во время ожидания субагентов по TTL линии; по умолчанию для Kimi, другие линии требуют известный TTL и явного включения",
+			settingsCacheAffinityName: "Сродство кеша учётных записей",
+			settingsCacheAffinityNote: "Предпочитать учётную запись с ещё тёплым префиксом, уступая при низкой квоте; по умолчанию для Kimi, другие линии требуют доказанного разделения кеша по учётным записям и явного включения",
 			settingsPlanModePipelineName: "Конвейер режима плана",
 			settingsPlanModePipelineNote: "После схождения уточняющего интервью автоматически войти в режим плана, подготовить план и отправить его на утверждение (только в режиме диспетчера); если выключено, этот режим попросит пользователя ввести /plan вручную",
 			settingsDispatchExecutionGuardName: "Защита выполнения диспетчера",
 			settingsDispatchExecutionGuardNote: "Блокировать вызовы инструментов диспетчера субагентами во время выполнения и запрещать пустое ожидание sleep при активных дочерних агентах; не меняет объявления инструментов и сохраняет префиксный кеш Kimi",
+			settingsToolRunawayGuardName: "Защита от циклов инструментов",
+			settingsToolRunawayGuardNote: "Напоминать агенту о повторяющихся вызовах; Kimi сохраняет передачу управления на 12-м вызове, другие линии получают только напоминания",
 			settingsTeamPolicyCacheGuardName: "Раздел политики Agent Teams",
 			settingsTeamPolicyCacheGuardNote: "Убрать встроенный раздел политики Agent Teams из системного промпта и передавать его одним сообщением в хвосте при включении командной работы, чтобы переключение посреди сессии больше не пересобирало весь префиксный кеш (включено по умолчанию)",
 			settingsThresholdName: "Порог напоминания",
@@ -23439,6 +23590,7 @@ window.__ModuleLoader__.load({
 			cacheRequestParamsOnHint: "Возможность включена: запросы несут параметры кеша. Step их не документирует, и официальный клиент для Step их тоже не отправляет.",
 			cachePricingNote: "Цены за единицу не документированы (все четыре ключа cost в официальной таблице моделей равны 0), поэтому страница показывает только токены и никаких сумм.",
 			cacheEmpty: "Ни один запрос на этой машине ещё не сообщил расход — проведите диалог и посмотрите снова.",
+			cacheExpiredHint: "Кеш, вероятно, истёк: с последнего запроса прошло около {minutes} мин., поэтому следующий шаг снова обработает префикс.",
 			cacheHitRatioUnavailable: "—"
 		};
 		//#endregion
@@ -23961,7 +24113,13 @@ window.__ModuleLoader__.load({
 .dsha-value{font-size:13px;min-width:0;overflow-wrap:anywhere;text-align:right;color:var(--dsw-alias-label-primary)}
 .dsha-actions{display:flex;flex-wrap:wrap;gap:8px;justify-content:flex-end;padding-top:12px}
 .dsha-btn{background:var(--dsw-alias-bg-layer-2);border:0.5px solid var(--dsw-alias-border-l2);border-radius:999px;corner-shape:round;color:var(--dsw-alias-label-primary);cursor:pointer;font:inherit;font-size:13px;line-height:1;padding:8px 13px;white-space:nowrap}
-.dsha-btn:hover:not(:disabled){background:var(--dsw-alias-interactive-bg-hover)}
+/* The hover tint belongs to the unfilled button only: it is a translucent wash
+   over the neutral fill, and letting it match .dsha-btn-primary would paint the
+   filled button with the wash while the label stayed white — the primary fill
+   is (0,1,0) and this rule is (0,3,0), so nothing else could win it back. The
+   exclusion is written into the selector, like the shared chip's
+   :not([aria-pressed='true']), so a later filled variant cannot reintroduce it. */
+.dsha-btn:hover:not(:disabled):not(.dsha-btn-primary){background:var(--dsw-alias-interactive-bg-hover)}
 .dsha-btn:disabled{cursor:default;opacity:.5}
 .dsha-btn-primary{background:var(--dsw-alias-button-info-fill,#397ee8);border-color:transparent;color:var(--dsw-alias-button-info-label,#fff)}
 .dsha-btn-primary:hover:not(:disabled){opacity:.9}
@@ -24439,7 +24597,13 @@ window.__ModuleLoader__.load({
 .dsha-value{font-size:13px;min-width:0;overflow-wrap:anywhere;text-align:right;color:var(--dsw-alias-label-primary)}
 .dsha-actions{display:flex;flex-wrap:wrap;gap:8px;justify-content:flex-end;padding-top:12px}
 .dsha-btn{background:var(--dsw-alias-bg-layer-2);border:0.5px solid var(--dsw-alias-border-l2);border-radius:999px;corner-shape:round;color:var(--dsw-alias-label-primary);cursor:pointer;font:inherit;font-size:13px;line-height:1;padding:8px 13px;white-space:nowrap}
-.dsha-btn:hover:not(:disabled){background:var(--dsw-alias-interactive-bg-hover)}
+/* The hover tint belongs to the unfilled button only: it is a translucent wash
+   over the neutral fill, and letting it match .dsha-btn-primary would paint the
+   filled button with the wash while the label stayed white — the primary fill
+   is (0,1,0) and this rule is (0,3,0), so nothing else could win it back. The
+   exclusion is written into the selector, like the shared chip's
+   :not([aria-pressed='true']), so a later filled variant cannot reintroduce it. */
+.dsha-btn:hover:not(:disabled):not(.dsha-btn-primary){background:var(--dsw-alias-interactive-bg-hover)}
 .dsha-btn:disabled{cursor:default;opacity:.5}
 .dsha-btn-primary{background:var(--dsw-alias-button-info-fill,#397ee8);border-color:transparent;color:var(--dsw-alias-button-info-label,#fff)}
 .dsha-btn-primary:hover:not(:disabled){opacity:.9}
@@ -25353,12 +25517,14 @@ window.__ModuleLoader__.load({
 .dshglm-model-meta{color:var(--dsw-alias-label-tertiary);font-size:11px;margin-left:5px}
 .dsha-badge.glm-adopted{background:color-mix(in srgb,var(--dsw-alias-button-info-fill,#397ee8) 16%,var(--dsw-alias-bg-layer-1));border-color:color-mix(in srgb,var(--dsw-alias-button-info-fill,#397ee8) 45%,var(--dsw-alias-border-l2));color:var(--dsw-alias-label-primary)}
 
-/* 跨线路共享的登录入口按钮行：两区登录入口把多个「添加账号」按钮排进同一行，
-   按钮沿用基础 .dsha-btn 的形状，只在这里统一换成信息色填充。原先是另一张卡片的
-   样式表带出来的，那张卡片下线后这份声明移到这里，选择器与取值一字未改。 */
+/* 跨线路共享的登录入口按钮行：两区登录入口把多个按钮排进同一行，行本身只负责
+   换行与间距。按钮的主次由各自的 class 决定 —— 主按钮带 .dsha-btn-primary，取信息
+   色填充；次要入口（设备码登录、导入本地登录）保持中性外观，这样一行里才读得出
+   「主入口 + 另外两种进入方式」。这里原先统一涂成信息色，是为了绕开
+   .dsha-btn:hover 盖掉主按钮填充的问题（见 antigravity / kimi-code 两张表里的
+   :not(.dsha-btn-primary)）；通用 hover 修好后它只剩副作用：把 .dsha-btn 的次要
+   按钮也画成主按钮，故删除。 */
 .dsha-account-add-actions{display:flex;flex-wrap:wrap;gap:8px}
-.dsha-account-add-actions .dsha-btn{background:var(--dsw-alias-button-info-fill,#397ee8);border-color:transparent;color:var(--dsw-alias-button-info-label,#fff)}
-.dsha-account-add-actions .dsha-btn:hover:not(:disabled){opacity:.9;background:var(--dsw-alias-button-info-fill,#397ee8)}
 `, STYLE_ID$3);
 		}
 		//#endregion
@@ -55503,6 +55669,37 @@ window.__ModuleLoader__.load({
 				]
 			});
 		}
+		/** 账号缓存亲和那一行的标记：号池两层里靠上的那层带勾（优先用缓存还热的账号）。 */
+		function CacheAffinityMark() {
+			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("svg", {
+				viewBox: "0 0 20 20",
+				width: "20",
+				height: "20",
+				fill: "none",
+				stroke: "currentColor",
+				strokeWidth: "1.6",
+				strokeLinecap: "round",
+				strokeLinejoin: "round",
+				"aria-hidden": "true",
+				children: [
+					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("rect", {
+						x: "3.5",
+						y: "3.8",
+						width: "13",
+						height: "5",
+						rx: "1.6"
+					}),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("rect", {
+						x: "3.5",
+						y: "11.2",
+						width: "13",
+						height: "5",
+						rx: "1.6"
+					}),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("path", { d: "m6.2 6.3 1.2 1.2 2.1-2.3" })
+				]
+			});
+		}
 		/** 计划模式管线那一行的标记：一份带勾的计划（计划先出、再交审批）。 */
 		function PlanModeMark() {
 			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("svg", {
@@ -55643,8 +55840,10 @@ window.__ModuleLoader__.load({
 			const cacheWarmup = value.cacheWarmup !== false;
 			const retrieval = value.cacheWarmupRetrieval !== false;
 			const cacheWaitWarmup = value.cacheWaitWarmup !== false;
+			const cacheAffinity = value.cacheAffinity !== false;
 			const planModePipeline = value.planModePipeline !== false;
 			const dispatchExecutionGuard = value.dispatchExecutionGuard !== false;
+			const toolRunawayGuard = value.toolRunawayGuard !== false;
 			const teamPolicyCacheGuard = value.teamPolicyCacheGuard !== false;
 			const threshold = typeof value.progressUpdatesNudgeThreshold === "number" ? value.progressUpdatesNudgeThreshold : 8;
 			const language = value.progressUpdatesLanguage === "zh" || value.progressUpdatesLanguage === "en" || value.progressUpdatesLanguage === "ru" ? value.progressUpdatesLanguage : "auto";
@@ -55804,6 +56003,27 @@ window.__ModuleLoader__.load({
 								}
 							})
 						}),
+						/* @__PURE__ */ (0, react_jsx_runtime.jsx)(CardRow, {
+							part: "settings-row-cache-affinity",
+							brand: "#1f7a52",
+							mark: /* @__PURE__ */ (0, react_jsx_runtime.jsx)(CacheAffinityMark, {}),
+							name: translate("settingsCacheAffinityName"),
+							note: failure?.row === "cacheAffinity" ? failure.text : translate("settingsCacheAffinityNote"),
+							danger: failure?.row === "cacheAffinity",
+							enabled: cacheAffinity,
+							children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+								type: "button",
+								role: "switch",
+								"aria-checked": cacheAffinity,
+								"aria-label": translate("settingsCacheAffinityName"),
+								className: "dsh-progress-switch",
+								"data-dsh-part": "settings-cache-affinity-switch",
+								disabled: !writable,
+								onClick: () => {
+									write("cacheAffinity", "cacheAffinity", !cacheAffinity);
+								}
+							})
+						}),
 						/* @__PURE__ */ (0, react_jsx_runtime.jsx)(CardRow, {
 							part: "settings-row-plan-mode-pipeline",
 							brand: "#566179",
@@ -55846,6 +56066,27 @@ window.__ModuleLoader__.load({
 								}
 							})
 						}),
+						/* @__PURE__ */ (0, react_jsx_runtime.jsx)(CardRow, {
+							part: "settings-row-tool-runaway-guard",
+							brand: "#a15c34",
+							mark: /* @__PURE__ */ (0, react_jsx_runtime.jsx)(DispatchGuardMark, {}),
+							name: translate("settingsToolRunawayGuardName"),
+							note: failure?.row === "toolRunawayGuard" ? failure.text : translate("settingsToolRunawayGuardNote"),
+							danger: failure?.row === "toolRunawayGuard",
+							enabled: toolRunawayGuard,
+							children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+								type: "button",
+								role: "switch",
+								"aria-checked": toolRunawayGuard,
+								"aria-label": translate("settingsToolRunawayGuardName"),
+								className: "dsh-progress-switch",
+								"data-dsh-part": "settings-tool-runaway-guard-switch",
+								disabled: !writable,
+								onClick: () => {
+									write("toolRunawayGuard", "toolRunawayGuard", !toolRunawayGuard);
+								}
+							})
+						}),
 						/* @__PURE__ */ (0, react_jsx_runtime.jsx)(CardRow, {
 							part: "settings-row-team-policy-cache-guard",
 							brand: "#8a5a9e",
@@ -56836,8 +57077,6 @@ window.__ModuleLoader__.load({
 			installPluginStyle("workbuddy", `
 .dsh-wb-mono{font-family:ui-monospace,SFMono-Regular,Consolas,monospace;font-size:11px}
 .dsha-account-add-actions{display:flex;flex-wrap:wrap;gap:8px}
-.dsha-account-add-actions .dsha-btn{background:var(--dsw-alias-button-info-fill,#397ee8);border-color:transparent;color:var(--dsw-alias-button-info-label,#fff)}
-.dsha-account-add-actions .dsha-btn:hover:not(:disabled){opacity:.9;background:var(--dsw-alias-button-info-fill,#397ee8)}
 /* The whole-pool state banner. It renders BEFORE the account list on purpose:
    a paused pool means nothing is routable, and a list of green account cards
    above that sentence reads as "available" whatever the banner says. */