dsh-xdpool 2.4.6 → 2.5.0

package/CHANGELOG.md

@@ -1,6 +1,141 @@
 # Changelog
 本项目的所有重要变更都会记录在此文件。
 版本号遵循 [语义化版本](https://semver.org/lang/zh-CN/)。
+
+## 2.5.0 — 十张 issue 一起收：失效账号、第三条指纹、CodeArts 上限、保存回滚、侧栏开关
+
+> #68 到 #83 攒了十来张单子，其中三张（#72/#73/#74）是同一类毛病：代码把一个「没数据」和「数据是空」当成了同一件事。这批按这个线索统一收拾，另外把社区两个 PR 收了进来。
+
+### 最要紧的三条
+
+| 症状 | 真实原因 |
+| --- | --- |
+| 一个失效账号吃掉整次请求的 8 次重试，最后报「API 密钥无效」 | 以「刷新拿到了新 token」判定账号仍可用。有一类账号 refresh 长期有效、chat 恒回 403，判死条件因此恒假 |
+| buddy 某一轮被中断后，之后每一轮都失败（`a message has empty content`） | 中断留下的 assistant 消息只有 reasoning、没有正文，序列化时空字符串被原样发到 wire，国内网关 400/11151 且不在重试集合里 |
+| zcode 全部模型勾选后，每次重启又变回全禁用 | 用户全开时黑名单是空对象，与「从未选过」无法区分，启动补静音据此又把它们全关一遍 |
+
+### #68 失效账号耗尽重试预算
+
+池车道以「刷新成功」当「账号可用」。对 refresh 长期有效而业务侧恒 403 的账号，`recovered` 与 `accessToken !== before` 恒真 ⇒ 永不 `penalizeCredentialDead` ⇒ `available()` 不排除 ⇒ `priority` 下 `acquire()` 一直返回同一个 ⇒ 8 次预算全花在它身上。
+
+改法：单次请求内只允许该账号换一次票（`refreshedThisRequest`），第二次仍报 `session_dead` 就直接判死并轮到下一个账号。压缩重试路径同步。取「第二次」而不是把惩罚提到刷新之前，是为了仍给真正过期的 token 一次机会。
+
+### #70 侧栏入口开关
+
+新增 `sidebarRow` 配置（默认开）。**只**控制左侧栏那一行：设置页的 XD Pool 卡片始终注册，模型池不受影响 —— 关掉入口不能把开关本身也锁在门外。开关写在卡片里，`apply()` 与卡片读同一份设置镜像，因此两边不会给出不同答案；宿主发布了变更通知的平台上无需重启。
+
+### #72 订阅包余额虚高 30 倍
+
+`fetchCredits` 只把 `CapacityType === 4` 当月包，而实际有三种：`1`=一次性包、`2`=**周期订阅包**（12 期 × 4000，`CapacityRemain` 是整期剩余、`CycleCapacityRemain` 才是当期剩余）、`4`=月包。`type=2` 落进「一次性」分支，整期额度被当余额。三账号实测：官方 3,399.8 / 插件 102,396 → 修复后 3,366；另两个只有月包的账号（3,049 / 2,771）不变，所以长期没被发现。
+
+`expiresAtMs` 对 `type=2` 取 `DeductionEndTime`，**不能用 `CycleEndTime`** —— 它每月都到，会让「会过期」天天报警。空包只在一次性包时丢弃，月包当期耗尽要留着显示「0/500 · 下月回满」。
+
+Refs #72。
+
+### #73 中断回合留下的空 assistant 消息
+
+`serializeMessages2` 出口加 `isBlankWireMessage` / `dropBlankWireMessages`，`serializeMessages4` 同步。三点取舍都用真实请求验证过：`user` 空消息不动（两网关都容忍）；不伪造占位内容（会污染模型看到的对话）；**有 `tool_calls` 的 assistant 消息绝不删**（它是 tool 结果的配对锚点，删了造孤儿）。
+
+实测矩阵（国内）：空 assistant 400/11151；空 system 同样 400/11151；空 user 200；`content:null` + `tool_calls` 200；删掉空 assistant 后正常 SSE。国际网关对全部形态都 200，所以故障面目前只在 buddy 车道。
+
+Refs #73。
+
+### #74 全开渠道每次重启被重新禁用
+
+`setModelDisabled` 打开模型时是**删键**而不是写 `false`，于是「用户全开」与「从未选择」在磁盘上是同一个空对象。新增持久标记 `selectionTouched`：四个用户入口（`setModelDisabled` / `setModelsDisabled` / `clearModelsDisabled` / `clearDisabledModels`）都置位，启动补静音只在标记缺失时执行，且自身写黑名单时传 `{ mark: false }` 不置位。
+
+影响面是「所有模型都打开」的渠道，目录越小越必然（zcode 2 个、gemini/minimax/raccoon 同理）；部分启用的渠道免疫，这解释了为什么只有 zcode 反复被报。
+
+Refs #74。
+
+### #75 保存模型选择后界面回滚 30 秒
+
+`saveModels` 收尾是 `setDraftFor(region, undefined)`，而它的语义是「退回池子最后报告的状态」，那份状态靠 30 秒轮询刷新 —— 于是保存后卡片画的是**保存前**的清单，用户在这 30 秒里再点一次保存，写入的就是回滚后的旧清单。另外 `settingsScope.set` 被声明成 `Promise<void>`，把宿主 revision 冲突时返回的 `false` 丢掉了，被拒绝也提示「已保存」。
+
+改法：先 `refresh(region)` 再清草稿（重读失败则保留草稿）；`set` 返回类型放宽到 `boolean | void`，`=== false` 时保留草稿、报「设置已在别处被修改」、不显示已保存。
+
+Refs #75。
+
+### #82 RPC 桥惰性解析
+
+2.4.6 的 CHANGELOG 声称这条已修，但发布产物里 `installPackRpc` 仍只捕获一次。现已改为每次 `callPack` 前 `if (packRpcBridge === undefined) packRpcBridge = resolvePackRpcBridge()`：`connection` 就绪晚于 `apply()` 时，桥不会再永久为空、退化成裸 `fetch('/api/channel-pack')` 吃 401 而让账号池整片空白。
+
+Refs #82。
+
+### 同批合并的社区 PR
+
+- **#81（[@jmglsi](https://github.com/jmglsi)）** 第三条内容指纹：`Relative paths resolve from the workspace root` → `are resolved against the workspace root`。用陈述句的过去分词形式，语义、长度、换行位置都不变，被网关字符串表收录的子串消失。
+- **#84（[@Refeain](https://github.com/Refeain)）** CodeArts 输出上限见下节。
+
+### #69 / #71 尚未定性
+
+`#69 CodeArts 提示密钥无效` 只有一句话。它至少有两条来源：本条 `InferHub.4001.200`（SSE 帧里的 200，不是 4xx）与真正的 401/403。前者已随 #83 修掉；后者需要失败时的原始响应才能确认。
+`#71 kimi k3.1 pro 没有响应` 同样只有一句话，且目录里没有 `kimi-k3.1-pro` 这个 id（有 `kimi-k3-1`）。请在 issue 里补一下：是空回复、卡住不动、还是有报错文案。
+
+### 详情：CodeArts 的输出上限是上限，不是默认值
+
+> 华为这条链接 65536 能用、131072 就空流，代码里也一直写着这个数字。但它是 `options.maxTokens ?? 65536` —— 只兜底，不设界。
+
+### 触发条件：宿主把别的 provider 的额度原样递过来
+
+宿主从模型目录里取 `maxTokens` 填进 `options.maxTokens`，而那份目录里 384000 是 deepseek-v4-pro、fireworks、huggingface 这些**真的能到 384000** 的条目的正常值。CodeArts 自己没声明上限时，这个 384000 就原样发出去了。
+
+后果不是 4xx：后端回 **HTTP 200**，把错误装在 SSE 帧里 ——
+
+```
+InferHub.4001.200 The request param is invalid, Please check it
+```
+
+适配器把它映射成 `INVALID_REQUEST`。用户看到的是「这个账号好好的，突然发不出话」，而新建会话发「你好」就失败，第一反应必然是密钥坏了 —— 这也是 #69「CodeArts 提示密钥无效」的一条候选根因（#69 正文只有一句话，无法确认是这条还是 403 那条，故 #83 里保留判断）。
+
+### 实测（2026-10-25，失败会话的原文 system/user + 全部 133 个 tools 直连，只改 max_tokens）
+
+| max_tokens | 结果 |
+| --- | --- |
+| 384000 | HTTP 200 + SSE 错误帧 `InferHub.4001.200` → INVALID_REQUEST |
+| 131072 | HTTP 200，空流 |
+| 65536 | 正常流式回复 |
+
+### 改法：三处对齐，线上请求做 clamp
+
+- `listModels()` 每个模型带 `maxTokens`，宿主没有理由再自己猜一个更大的
+- `resolveModel()` 带 `defaultMaxTokens`，`catalogFactsFor()` 会把它当 `fact.maxTokens` 交给宿主 —— 这是宿主唯一能知道上限的通道
+- 线上 `max_tokens: Math.min(options.maxTokens ?? CODEARTS_MAX_OUTPUT_TOKENS, CODEARTS_MAX_OUTPUT_TOKENS)`，调用方给更大就收回，给更小仍保留
+
+TRAE 与 Cline 早就这么做了（`clampTraeMaxTokens` / `clampClineMaxTokens`，也都写 `resolved.defaultMaxTokens`），CodeArts 是唯一的例外。
+
+- 回归：`tests/codearts-max-tokens.test.ts` 新增 5 例（源码断言，与 `tests/codearts-resign-headers.test.ts` 同形状 —— `pack.js` 是另一个仓库的构建产物，无法 import 驱动）
+- 反向对照：撤掉 `pack.js` 的补丁后 5 例全红，补上全绿
+
+Refs #83, #69。
+
+### 详情：第三条内容指纹：文件引用指引里的 “resolve from the workspace root”
+
+> 2.4.1/2.4.2 修好的两个指纹管了一周，10 月 28 日又整片 400/11128。这次被点名的不是产品名，而是一句普通说明。
+
+### 触发点在 read 工具的文件引用指引
+
+DSH 在 agent 有 `read` 工具时会贡献一段文件引用指引，开头是「Tokens prefixed with @ are paths the user explicitly referenced.」，第二句是 **「Relative paths resolve from the workspace root」**。把当次真实系统提示（约 41k 字符）整段发上去稳定 400/11128，逐段二分到 204 字符命中这段，再逐句验证：两句**单独发都是 200**，两句同现才 400；改其中任意一处小改动也 200 —— 所以这是内容指纹，不是长度规则、也不是头部规则。
+
+这也解释了为什么它看起来「不像」前两条：前两条是识别别的客户端的文案，这一条只是被网关的字符串表收进去了，句子本身没有任何产品名。
+
+### 改法：把陈述句改成同一句的过去分词形式
+
+`Relative paths resolve from the workspace root` → `Relative paths are resolved against the workspace root`。语义、长度量级、换行位置都不变，被列表收录的子串消失。同一账号上整段 41k 系统提示由 400 转 200；渠道车道与池车道同步补上，`pack.js` 的三处 `[xdpool-local]` 标记保持原样。
+
+- 池车道：`BANNED_SYSTEM_PHRASES` 增加第三行
+- 渠道车道：`vendor/channel-pack/pack.js` 的同一个 `replace` 链增加第三段
+- 回归：`tests/system-fingerprint.test.ts` 新增 5 例（含「三处指纹同现」整段用例）；`tests/chat-identity-headers.test.ts` 对已发布 bundle 的字节断言补上第三段，pack 刷新若把它冲掉会变红
+
+### 其他
+- 全套 79 文件 1028 用例（1026 通过 + 2 跳过）。
+
+### 验证
+
+- 全套 **80 文件 / 1033 用例**（含 #81 的 5 例、#84 的 5 例、本批新增）
+- `pack.js` 每次改动后 `node --check` 通过
+- server `tsc` 0 错；client `tsc` 的 6 条是既存问题（`accounts.ts` / `context-budget.ts` / `free-upstream.ts` 的 Node 类型不匹配），改前改后数量一致
+
 ## 2.4.6 — 扫码账号每次打开都空白：一份数据卡在两个细节里
 
 > 2.4.4/2.4.5 之后侧栏能进、能滚了，但扫码过的渠道在账号池里经常整片空白，重开也一样。磁盘上账号是齐的（`state.json` 里 15 条），所以问题全在读取链。

package/lib/bin.js

@@ -339,10 +339,39 @@ const CHAT_IDENTITY_HEADERS = Object.freeze({
 * rewrite is applied to assistant history too — one past reply carrying a
 * fingerprint would otherwise fail every later request in the conversation.
 *
+* A THIRD FINGERPRINT (measured against the live domestic gateway 2026-10-28)
+*
+* The file-reference guidance DSH contributes whenever the agent has a `read`
+* tool reintroduced the same 400/11128 after the two rewrites above shipped:
+*
+*   Tokens prefixed with @ are paths the user explicitly referenced.
+*   Relative paths resolve from the workspace root; absolute paths identify
+*   files or directories on the host.
+*
+* Ordinary phrasing throughout — the blocked region names no product — so the
+* fingerprint is not "this prompt is not WorkBuddy's" but a literal substring
+* list that happens to include this sentence. Neither sentence is the trigger
+* on its own: each 200s in isolation, and the pair 400s; the smallest edit
+* anywhere in the pair clears it, which is what marks it as content and not
+* prompt length, and the matching half is the second sentence ("Tokens
+* prefixed with @" alone stayed clean while the "Relative paths…" line was the
+* one whose text the gateway recognised).
+*
+* The rewrite turns the indicative into the same statement's participle —
+* "Relative paths resolve from the workspace root" becomes "Relative paths are
+* resolved against the workspace root". Meaning, length class and line breaks
+* are preserved; the substring the gateway lists is gone. Verified at 200 with
+* the full runtime system prompt (≈41k chars) that reproduced the 400 on the
+* same account, and with the pair alone.
+*
 * The list is an ARRAY because these fingerprints are appended to over time; a
 * future rejection should be one more row here, not a new code path.
 */
-const BANNED_SYSTEM_PHRASES = Object.freeze([[/powered by DeepSeek Harness/gi, "running on DeepSeek Harness"], [/DeepSeek Harness Web GUI/gi, "DeepSeek Harness web interface"]]);
+const BANNED_SYSTEM_PHRASES = Object.freeze([
+	[/powered by DeepSeek Harness/gi, "running on DeepSeek Harness"],
+	[/DeepSeek Harness Web GUI/gi, "DeepSeek Harness web interface"],
+	[/Relative paths resolve from the workspace root/gi, "Relative paths are resolved against the workspace root"]
+]);
 /**
 * Rewrite fingerprinted system/assistant content, leaving everything else
 * byte-identical.
@@ -1325,14 +1354,17 @@ var WorkBuddyUpstreamClient = class {
 			if (typeof raw !== "object" || raw === null) continue;
 			const account = raw;
 			const num = (key) => typeof account[key] === "number" ? account[key] : 0;
-			const monthly = num("CapacityType") === 4;
-			const size = monthly ? num("CycleCapacitySize") : num("CapacitySize");
-			const remain = monthly ? num("CycleCapacityRemain") : num("CapacityRemain");
+			const capacityType = num("CapacityType");
+			const monthly = capacityType === 4;
+			const periodic = monthly || capacityType === 2;
+			const size = periodic ? num("CycleCapacitySize") : num("CapacitySize");
+			const remain = periodic ? num("CycleCapacityRemain") : num("CapacityRemain");
 			const capped = remain < 0 ? 0 : remain;
 			const cycleEndMs = parseDate(account["CycleEndTime"]);
-			const expiresAtMs = monthly ? void 0 : parseDate(account["ExpiredTime"]) ?? cycleEndMs;
+			const expiresAtMs = monthly ? void 0 : capacityType === 2 ? parseDate(account["DeductionEndTime"]) : parseDate(account["ExpiredTime"]) ?? cycleEndMs;
 			const refreshAtMs = monthly ? cycleEndMs === void 0 ? void 0 : cycleEndMs + 1e3 : void 0;
-			if (!monthly && (capped <= 0 || expiresAtMs !== void 0 && expiresAtMs <= Date.now())) continue;
+			if (expiresAtMs !== void 0 && expiresAtMs <= Date.now()) continue;
+			if (!periodic && capped <= 0) continue;
 			total += capped;
 			if (expiresAtMs !== void 0) {
 				if (nearestExpiryMs === void 0 || expiresAtMs < nearestExpiryMs) nearestExpiryMs = expiresAtMs;
@@ -11274,7 +11306,8 @@ z.object({
 	modelSelectionCn: asVolatile(modelSelectionSchema.description("Model selection for the domestic gateway")),
 	modelSelectionGlobal: asVolatile(modelSelectionSchema.description("Model selection for the international gateway")),
 	automation: asVolatile(automationSchema.description("Daily points automation (activity report, task claiming, check-in)")),
-	automationEarnings: asVolatile(z.any().description("Automation earnings ledger (written by the scheduler)"))
+	automationEarnings: asVolatile(z.any().description("Automation earnings ledger (written by the scheduler)")),
+	sidebarRow: asVolatile(z.boolean().default(true).description("Show the XD Pool entry in the left sidebar (the Settings page keeps working either way)"))
 });
 /** Assemble the runtime objects without registering anything. */
 /**

package/lib/client.js

@@ -7,6 +7,44 @@ window.__ModuleLoader__.load({
 		//#endregion
 		let react = require("react");
 		let react_jsx_runtime = require("react/jsx-runtime");
+		//#region src/client/sidebar-row.ts
+		/**
+		* The `sidebarRow` switch, read and written from one place.
+		*
+		* Two independent owners need this flag: the page card draws the checkbox, and
+		* `apply()` decides whether to register the sidebar row before that card has ever
+		* mounted. Both therefore have to read the same source of truth — the settings
+		* mirror the host exposes — instead of one asking the server and the other
+		* guessing.
+		*
+		* The scope is typed structurally here rather than imported from `PoolCard.tsx`:
+		* this module is pulled into both tsconfigs (the server one has no JSX setting),
+		* and the only thing it needs from that component is `getSnapshot().value`.
+		*
+		* The default is ON, and the read is fail-open on purpose: "the mirror has no
+		* opinion yet" and "the user turned it off" must not be the same answer, or a
+		* host that has not populated the mirror would silently hide a feature nobody
+		* switched off. That is also why only the literal `false` disables the row.
+		*/
+		/** Settings field the switch writes (declared in the plugin Config schema). */
+		const SIDEBAR_ROW_FIELD = "sidebarRow";
+		/**
+		* Whether the sidebar entry should exist.
+		*
+		* Never throws: a locked-down host makes `getSnapshot()` unusable, and the
+		* documented degradation there is the settings page, not a missing entry.
+		*/
+		function readSidebarRow(scope) {
+			if (scope === void 0) return true;
+			try {
+				const value = scope.getSnapshot().value;
+				if (value === null || typeof value !== "object") return true;
+				return value[SIDEBAR_ROW_FIELD] !== false;
+			} catch {
+				return true;
+			}
+		}
+		//#endregion
 		//#region src/status-paths.ts
 		/**
 		* Node-free constants and types shared by the Host and browser halves of the
@@ -518,6 +556,7 @@ window.__ModuleLoader__.load({
 .dsm-xdpool-usage-model{font-size:12px;font-weight:600;word-break:break-all}
 .dsm-xdpool-usage-numbers{color:var(--dsw-alias-label-tertiary,#999);font-size:11px;line-height:17px;word-break:break-word}
 .dsm-xdpool-usage-note{margin:0;color:var(--dsw-alias-label-tertiary,#999);font-size:11px;line-height:17px}
+.dsm-xdpool-sidebar-row-head{display:flex;align-items:center;justify-content:space-between;gap:12px;margin-bottom:6px}
 .dsm-xdpool-models-actions{display:flex;align-items:center;gap:8px;flex:none}
 /* "Built-in list (offline)" chip: a warning tint, since the user is looking at
    a SHORTER roster than the gateway offers and may wonder where models went. */
@@ -9008,6 +9047,7 @@ window.__ModuleLoader__.load({
 		* message.
 		*/
 		async function callPack(method, payload) {
+			if (packRpcBridge === void 0) packRpcBridge = resolvePackRpcBridge();
 			if (packRpcBridge !== void 0) try {
 				const result = await packRpcBridge.call(method, payload);
 				if (result !== null && typeof result === "object") {
@@ -12368,7 +12408,7 @@ window.__ModuleLoader__.load({
 		*
 		* @module dsh-xdpool/version
 		*/
-		const POOL_VERSION = "2.4.6";
+		const POOL_VERSION = "2.5.0";
 		//#endregion
 		//#region src/client/PoolCard.tsx
 		/**
@@ -12664,6 +12704,23 @@ window.__ModuleLoader__.load({
 		/** Render pool health, per-account credits/cooldown, and the model directory. */
 		function PoolCard({ t, settingsScope }) {
 			const settingsWritable = settingsScope?.getSnapshot().writable === true;
+			/**
+			* Whether the plugin owns a row in the left sidebar (issue #70).
+			*
+			* Seeded from, and kept in step with, the settings mirror rather than from the
+			* pool status: the row is registered by `apply()`, before any card exists, and
+			* both sides must answer the same question the same way.
+			*/
+			const [sidebarRow, setSidebarRowState] = (0, react.useState)(() => readSidebarRow(settingsScope));
+			const [sidebarRowBusy, setSidebarRowBusy] = (0, react.useState)(false);
+			(0, react.useEffect)(() => {
+				const unsubscribe = settingsScope?.subscribe?.(() => {
+					setSidebarRowState(readSidebarRow(settingsScope));
+				});
+				return () => {
+					unsubscribe?.();
+				};
+			}, [settingsScope]);
 			/** Which region tab is showing. A CN-only install never leaves this. */
 			const [activeRegion, setActiveRegion] = (0, react.useState)("cn");
 			/**
@@ -13324,11 +13381,18 @@ window.__ModuleLoader__.load({
 					const imageModelIds = Object.entries(local).filter(([, e]) => e.images).map(([id]) => id);
 					const contextBudgets = {};
 					for (const [id, entry] of Object.entries(local)) if (entry.budget !== void 0) contextBudgets[id] = entry.budget;
-					await write.call(settingsScope, MODEL_SELECTION_KEY[region], {
+					if (await write.call(settingsScope, MODEL_SELECTION_KEY[region], {
 						enabledModelIds,
 						imageModelIds,
 						contextBudgets
-					});
+					}) === false) {
+						if (mounted.current) setError(t?.("row.modelsConflict") ?? "The settings changed elsewhere; review and save again");
+						return;
+					}
+					if (await refresh(region) === void 0) {
+						if (mounted.current) setFlash(t?.("row.modelsSaved") ?? "Saved");
+						return;
+					}
 					setDraftFor(region, void 0);
 					if (mounted.current) setFlash(t?.("row.modelsSaved") ?? "Saved");
 				} catch (cause) {
@@ -13365,6 +13429,35 @@ window.__ModuleLoader__.load({
 				}
 			};
 			/**
+			* Turn the sidebar row on or off (issue #70).
+			*
+			* Written through the same settings field `apply()` reads, so the two agree
+			* without a new endpoint. The row itself is owned by the bundle registration,
+			* so the mirror's change notification is what makes this take effect without a
+			* restart on hosts that publish one; everywhere else it lands on next launch,
+			* which the hint text says rather than implies.
+			*/
+			const setSidebarRow = async (next) => {
+				const write = settingsScope?.set;
+				if (write === void 0) {
+					setError(t?.("row.modelsSaveError", { message: "settings scope is read-only" }) ?? "settings scope is read-only");
+					return;
+				}
+				setSidebarRowBusy(true);
+				setFlash(void 0);
+				try {
+					if (await write.call(settingsScope, "sidebarRow", next) === false) {
+						if (mounted.current) setError(t?.("row.modelsConflict") ?? "The settings changed elsewhere; review and save again");
+						return;
+					}
+					if (mounted.current) setSidebarRowState(next);
+				} catch (cause) {
+					if (mounted.current) setError(String(cause));
+				} finally {
+					if (mounted.current) setSidebarRowBusy(false);
+				}
+			};
+			/**
 			* Switch the daily-points automation on or off.
 			*
 			* The whole `automation` object is written as one key, because that is how the
@@ -13639,6 +13732,30 @@ window.__ModuleLoader__.load({
 							})
 						]
 					}),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("section", {
+						className: "dsm-xdpool-card dsm-xdpool-sidebar-row",
+						"aria-label": t?.("row.sidebarRowTitle") ?? "Sidebar entry",
+						children: [/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+							className: "dsm-xdpool-sidebar-row-head",
+							children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("h3", {
+								className: "dsm-xdpool-usage-title",
+								children: t?.("row.sidebarRowTitle") ?? "Sidebar entry"
+							}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+								type: "button",
+								role: "switch",
+								"aria-checked": sidebarRow,
+								disabled: !settingsWritable || sidebarRowBusy,
+								className: `dsm-xdpool-auto-switch${sidebarRow ? " dsm-xdpool-auto-switch-on" : ""}`,
+								onClick: () => {
+									setSidebarRow(!sidebarRow);
+								},
+								children: sidebarRowBusy ? t?.("row.autoBusy") ?? "Saving…" : sidebarRow ? t?.("row.autoOn") ?? "On" : t?.("row.autoOff") ?? "Off"
+							})]
+						}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+							className: "dsm-xdpool-usage-note",
+							children: sidebarRow ? t?.("row.sidebarRowOn") ?? "XD Pool has its own row in the left sidebar. Turn it off to open the pool from Settings instead." : t?.("row.sidebarRowOff") ?? "The sidebar row is hidden; the pool still opens from Settings → XD Pool, and models are unaffected."
+						})]
+					}),
 					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("section", {
 						className: "dsm-xdpool-regions",
 						children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
@@ -14959,6 +15076,10 @@ window.__ModuleLoader__.load({
 			"row.modelsSaving": "Saving…",
 			"row.modelsDiscard": "Discard",
 			"row.modelsSaved": "Model selection saved",
+			"row.sidebarRowTitle": "Sidebar entry",
+			"row.sidebarRowOn": "XD Pool has its own row in the left sidebar. Turn it off to open the pool from Settings instead.",
+			"row.sidebarRowOff": "The sidebar row is hidden; the pool still opens from Settings → XD Pool, and models are unaffected.",
+			"row.modelsConflict": "The settings changed elsewhere; review and save again",
 			"row.modelsSaveError": "Could not save: {message}",
 			"row.modelsEmpty": "No model enabled — enable at least one before saving.",
 			"row.catalogOffline": "built-in list (offline)",
@@ -15470,6 +15591,10 @@ window.__ModuleLoader__.load({
 			"row.modelsSaving": "保存中…",
 			"row.modelsDiscard": "放弃修改",
 			"row.modelsSaved": "模型选择已保存",
+			"row.sidebarRowTitle": "侧栏入口",
+			"row.sidebarRowOn": "左侧栏会给 XD Pool 留一行入口。关掉后改为从「设置 → XD Pool」进入，模型池不受影响。",
+			"row.sidebarRowOff": "侧栏入口已隐藏；仍可从「设置 → XD Pool」进入，模型池不受影响。",
+			"row.modelsConflict": "设置已在别处被修改，请确认后重新保存",
 			"row.modelsSaveError": "保存失败：{message}",
 			"row.modelsEmpty": "至少要启用一个模型才能保存。",
 			"row.catalogOffline": "内置列表（离线）",
@@ -15921,12 +16046,29 @@ window.__ModuleLoader__.load({
 				}, PoolCard));
 				installNavIcon(ctx, () => t("row.navLabel"));
 				try {
-					ctx.slots.inject("sidebar.panellist", () => ctx.slots.register({
-						name: "sidebar.panellist",
-						id: "xdpool",
-						order: 20,
-						label: () => t("row.navLabel")
-					}, PanelIcon));
+					let disposeRow;
+					const syncSidebarRow = () => {
+						const wanted = readSidebarRow(settingsScope);
+						if (wanted && disposeRow === void 0) disposeRow = ctx.slots.register({
+							name: "sidebar.panellist",
+							id: "xdpool",
+							order: 20,
+							label: () => t("row.navLabel")
+						}, PanelIcon);
+						else if (!wanted && disposeRow !== void 0) {
+							disposeRow();
+							disposeRow = void 0;
+						}
+					};
+					ctx.slots.inject("sidebar.panellist", () => {
+						syncSidebarRow();
+						const unsubscribe = settingsScope?.subscribe?.(syncSidebarRow);
+						return () => {
+							unsubscribe?.();
+							disposeRow?.();
+							disposeRow = void 0;
+						};
+					});
 					ctx.slots.inject("main", () => ctx.slots.register({
 						name: "main",
 						key: "xdpool",

package/lib/index.js

@@ -390,10 +390,39 @@ const CHAT_IDENTITY_HEADERS = Object.freeze({
 * rewrite is applied to assistant history too — one past reply carrying a
 * fingerprint would otherwise fail every later request in the conversation.
 *
+* A THIRD FINGERPRINT (measured against the live domestic gateway 2026-10-28)
+*
+* The file-reference guidance DSH contributes whenever the agent has a `read`
+* tool reintroduced the same 400/11128 after the two rewrites above shipped:
+*
+*   Tokens prefixed with @ are paths the user explicitly referenced.
+*   Relative paths resolve from the workspace root; absolute paths identify
+*   files or directories on the host.
+*
+* Ordinary phrasing throughout — the blocked region names no product — so the
+* fingerprint is not "this prompt is not WorkBuddy's" but a literal substring
+* list that happens to include this sentence. Neither sentence is the trigger
+* on its own: each 200s in isolation, and the pair 400s; the smallest edit
+* anywhere in the pair clears it, which is what marks it as content and not
+* prompt length, and the matching half is the second sentence ("Tokens
+* prefixed with @" alone stayed clean while the "Relative paths…" line was the
+* one whose text the gateway recognised).
+*
+* The rewrite turns the indicative into the same statement's participle —
+* "Relative paths resolve from the workspace root" becomes "Relative paths are
+* resolved against the workspace root". Meaning, length class and line breaks
+* are preserved; the substring the gateway lists is gone. Verified at 200 with
+* the full runtime system prompt (≈41k chars) that reproduced the 400 on the
+* same account, and with the pair alone.
+*
 * The list is an ARRAY because these fingerprints are appended to over time; a
 * future rejection should be one more row here, not a new code path.
 */
-const BANNED_SYSTEM_PHRASES = Object.freeze([[/powered by DeepSeek Harness/gi, "running on DeepSeek Harness"], [/DeepSeek Harness Web GUI/gi, "DeepSeek Harness web interface"]]);
+const BANNED_SYSTEM_PHRASES = Object.freeze([
+	[/powered by DeepSeek Harness/gi, "running on DeepSeek Harness"],
+	[/DeepSeek Harness Web GUI/gi, "DeepSeek Harness web interface"],
+	[/Relative paths resolve from the workspace root/gi, "Relative paths are resolved against the workspace root"]
+]);
 /**
 * Rewrite fingerprinted system/assistant content, leaving everything else
 * byte-identical.
@@ -1485,14 +1514,17 @@ var WorkBuddyUpstreamClient = class {
 			if (typeof raw !== "object" || raw === null) continue;
 			const account = raw;
 			const num = (key) => typeof account[key] === "number" ? account[key] : 0;
-			const monthly = num("CapacityType") === 4;
-			const size = monthly ? num("CycleCapacitySize") : num("CapacitySize");
-			const remain = monthly ? num("CycleCapacityRemain") : num("CapacityRemain");
+			const capacityType = num("CapacityType");
+			const monthly = capacityType === 4;
+			const periodic = monthly || capacityType === 2;
+			const size = periodic ? num("CycleCapacitySize") : num("CapacitySize");
+			const remain = periodic ? num("CycleCapacityRemain") : num("CapacityRemain");
 			const capped = remain < 0 ? 0 : remain;
 			const cycleEndMs = parseDate(account["CycleEndTime"]);
-			const expiresAtMs = monthly ? void 0 : parseDate(account["ExpiredTime"]) ?? cycleEndMs;
+			const expiresAtMs = monthly ? void 0 : capacityType === 2 ? parseDate(account["DeductionEndTime"]) : parseDate(account["ExpiredTime"]) ?? cycleEndMs;
 			const refreshAtMs = monthly ? cycleEndMs === void 0 ? void 0 : cycleEndMs + 1e3 : void 0;
-			if (!monthly && (capped <= 0 || expiresAtMs !== void 0 && expiresAtMs <= Date.now())) continue;
+			if (expiresAtMs !== void 0 && expiresAtMs <= Date.now()) continue;
+			if (!periodic && capped <= 0) continue;
 			total += capped;
 			if (expiresAtMs !== void 0) {
 				if (nearestExpiryMs === void 0 || expiresAtMs < nearestExpiryMs) nearestExpiryMs = expiresAtMs;
@@ -19296,6 +19328,7 @@ function createWorkBuddyShim(options) {
 		const conversationKey = conversationKeyOf(poolRaw);
 		const tried = [];
 		let last;
+		const refreshedThisRequest = /* @__PURE__ */ new Set();
 		const turnId = randomBytes(8).toString("hex");
 		const usageContext = usageRecorder === void 0 ? void 0 : {
 			recorder: usageRecorder,
@@ -19350,6 +19383,12 @@ function createWorkBuddyShim(options) {
 				message: result.message
 			};
 			if (result.kind === "session_dead") {
+				if (refreshedThisRequest.has(account.id)) {
+					pool.penalizeCredentialDead(account.id);
+					logger?.warn(`dsh-xdpool: ${account.label} still refused after a token refresh (attempt ${attempt + 1}/${maxAttempts}); rotating`);
+					continue;
+				}
+				refreshedThisRequest.add(account.id);
 				logger?.warn(`dsh-xdpool: ${account.label} session dead; refreshing token and retrying`);
 				const before = account.credential.accessToken;
 				if (!await pool.refreshAccount(account.id, { force: true }) || account.credential.accessToken === before) {
@@ -19607,6 +19646,7 @@ async function recoverFromContextOverrun(options) {
 	};
 	const retryBody = client.buildChatBody(parsed, compacted, region ?? "cn");
 	const tried = [];
+	const refreshedThisRequest = /* @__PURE__ */ new Set();
 	for (let attempt = 0; attempt < maxAttempts; attempt += 1) {
 		if (controller.signal.aborted) return {
 			ok: false,
@@ -19640,6 +19680,11 @@ async function recoverFromContextOverrun(options) {
 			detail: "prompt still exceeded the window after compaction"
 		};
 		if (result.kind === "session_dead") {
+			if (refreshedThisRequest.has(account.id)) {
+				pool.penalizeCredentialDead(account.id);
+				continue;
+			}
+			refreshedThisRequest.add(account.id);
 			const before = account.credential.accessToken;
 			if (!await pool.refreshAccount(account.id, { force: true }) || account.credential.accessToken === before) pool.penalizeCredentialDead(account.id);
 			continue;
@@ -20699,7 +20744,8 @@ const Config = z.object({
 	modelSelectionCn: asVolatile(modelSelectionSchema.description("Model selection for the domestic gateway")),
 	modelSelectionGlobal: asVolatile(modelSelectionSchema.description("Model selection for the international gateway")),
 	automation: asVolatile(automationSchema.description("Daily points automation (activity report, task claiming, check-in)")),
-	automationEarnings: asVolatile(z.any().description("Automation earnings ledger (written by the scheduler)"))
+	automationEarnings: asVolatile(z.any().description("Automation earnings ledger (written by the scheduler)")),
+	sidebarRow: asVolatile(z.boolean().default(true).description("Show the XD Pool entry in the left sidebar (the Settings page keeps working either way)"))
 });
 /** Live API, published for the CLI. */
 let api;
@@ -21339,7 +21385,7 @@ function apply(ctx, config = {}) {
 			}) };
 		}).finally(() => {
 			if (stopped) return;
-			adoptChannelPackAccounts(ctx, core.pool, ctx.logger);
+			adoptChannelPackAccounts(scoped, core.pool, ctx.logger);
 		});
 	});
 	mountBillionContext(ctx, ctx.logger).then((outcome) => {

package/package.json

@@ -2,7 +2,7 @@
   "name": "dsh-xdpool",
   "displayName": "DSH XD Pool",
   "description": "One DSH model pool for every free quota you have: merge all locally signed-in WorkBuddy accounts with auto-failover, sign in to 13 vendors by QR/web, aggregate 76 vendors' free tiers, plus a usage dashboard, model routing, a local OpenAI-compatible gateway and a points-automation suite.",
-  "version": "2.4.6",
+  "version": "2.5.0",
   "license": "MIT",
   "author": "XDTrees",
   "homepage": "https://github.com/XDTrees/dsh-xdpool#readme",

package/vendor/channel-pack/pack.js

@@ -29453,6 +29453,15 @@ var channelPackSchema = Schema.object({
   // **必须显式声明**：SettingsStore 的 `scope.replace()` 是整体覆盖，
   // schema 里没声明的键会被 schemastery 剥掉，写回时就此静默丢失。
   contextBudgets: Schema.dict(Schema.any()).default({}),
+  // ── 本地扩展（dsh-xdpool，2026-10-26）────────────────────────────────
+  // 记录「用户亲手动过这个 provider 的模型选择」，用来把「用户把模型全部
+  // 打开」和「用户从未选择过」区分开。`disabledModels` 本身做不到这件事：
+  // 打开模型时我们**删除**键而不是写 false（见 setModelDisabled 注释），
+  // 所以「全开」与「从未碰过」都表现为空表。启动补静音原先拿
+  // `disabledModelsFor(provider).size === 0` 当判据，于是把「全开」当成
+  // 「从未选择」，每次重启又把整目录静音回去 —— 目录越小的渠道越必然中招
+  // （zcode 只有两个模型，用户勾满之后每次重启都变回全禁用）。
+  selectionTouched: Schema.dict(Schema.any()).default({}),
   loomyPermanentLocked: Schema.boolean().default(false),
   gatewayEnabled: Schema.boolean().default(true),
   providerOrder: Schema.array(Schema.string()).default([])
@@ -29511,6 +29520,14 @@ function sanitizeContextBudgets(raw) {
   }
   return result;
 }
+function sanitizeSelectionTouched(raw) {
+  if (typeof raw !== "object" || raw === null || Array.isArray(raw)) return {};
+  const result = {};
+  for (const [provider, value] of Object.entries(raw)) {
+    if (value === true && provider.length > 0) result[provider] = true;
+  }
+  return result;
+}
 function sanitizeAccounts(raw) {
   if (!Array.isArray(raw)) return [];
   const accounts = [];
@@ -29542,6 +29559,7 @@ var SettingsStore = class {
       accounts: sanitizeAccounts(value.accounts),
       disabledModels: sanitizeDisabledModels(value.disabledModels),
       contextBudgets: sanitizeContextBudgets(value.contextBudgets),
+    selectionTouched: sanitizeSelectionTouched(value.selectionTouched),
       // 只是**镜像**（权威表在 permanent-locks.json）；老文档没这个键 → false。
       loomyPermanentLocked: value.loomyPermanentLocked === true,
       // 老文档没这个键 → 启用，与升级前行为一致。
@@ -29557,6 +29575,9 @@ var SettingsStore = class {
       // 预算表丢失的后果只是「下次选择器回到原生窗口」，用户重选一次即可
       // （纯偏好、可逆），故接受老版本重写时丢掉本键。
       contextBudgets: state.contextBudgets ?? {},
+      // 丢了这个标记的后果是「下次启动把该渠道重新静音一次」，用户再勾一遍
+      // 就会重新置位，属可逆，故与 contextBudgets 同样接受老版本重写时丢掉。
+      selectionTouched: state.selectionTouched ?? {},
       // 镜像字段由 AccountPool 与独立文档**同源写出**：同机上只认这个字段的
       // 旧版本代码（其它 profile）读它、也会原样写回它，故两边不会脱节。
       loomyPermanentLocked: state.loomyPermanentLocked === true,
@@ -29661,6 +29682,7 @@ var FileStore = class {
         accounts: sanitizeAccounts(value.accounts),
         disabledModels: sanitizeDisabledModels(value.disabledModels),
         contextBudgets: sanitizeContextBudgets(value.contextBudgets),
+    selectionTouched: sanitizeSelectionTouched(value.selectionTouched),
         // 只是镜像（权威表在 permanent-locks.json）；老文档没这个键 → false。
         loomyPermanentLocked: value.loomyPermanentLocked === true,
         // 老文档没这个键 → 启用，与升级前行为一致。
@@ -29705,6 +29727,9 @@ var FileStore = class {
         accounts,
         disabledModels: {},
         contextBudgets: {},
+        // 从凭据文件重建的文档不含任何选择历史 → 空表，让补静音照常生效
+        // （这些账号确实是「从未选择过」，正是补静音要照顾的情形）。
+        selectionTouched: {},
         // 凭据文件里没有任何锁定信息 → 镜像写 false（权威表另有其文档）。
         loomyPermanentLocked: false,
         // 恢复出的文档本来就不含任何开关信息 → 启用（与全新安装一致）。
@@ -29856,6 +29881,17 @@ var AccountPool = class {
    * 把窗口收窄到 200K 这类值」，不改变上游真实能力。
    */
   contextBudgetCache = {};
+  /**
+   * 「用户动过该 provider 的模型选择」标记的进程内权威副本。
+   *
+   * 存在的唯一理由是把「用户把模型全部打开」与「用户从未选择过」区分开：
+   * 打开模型时黑名单里**删除**键（见 {@link setModelDisabled}），所以「全开」
+   * 与「从未碰过」的 `disabledModels` 都是空表，单看黑名单无法分辨，启动补
+   * 静音就会把「全开」当成「从未选择」再静音一次。
+   *
+   * ⚠️ 只记录已置位的 provider（缺键 = 从未选择）。
+   */
+  selectionTouchedCache = {};
   /**
    * 「锁定永久积分」开关表的**权威进程内副本**（与 {@link cache} 同理）。
    *
@@ -29907,6 +29943,7 @@ var AccountPool = class {
       this.cache = state.accounts;
       this.modelCache = state.disabledModels;
       this.contextBudgetCache = state.contextBudgets ?? {};
+      this.selectionTouchedCache = state.selectionTouched ?? {};
       this.gatewayEnabledCache = state.gatewayEnabled !== false;
       this.providerOrderCache = state.providerOrder ?? [];
     }
@@ -29984,6 +30021,7 @@ var AccountPool = class {
         accounts: this.cache,
         disabledModels: this.modelCache,
         contextBudgets: this.contextBudgetCache,
+        selectionTouched: this.selectionTouchedCache,
         ...this.lockFields(),
         // ⚠️ 必须与账号、黑名单取自**同一时刻**的快照（见本方法的注释）：
         // 分两次快照会让同一份文档里的字段互相矛盾。
@@ -30105,6 +30143,9 @@ var AccountPool = class {
     else delete perProvider[modelId];
     if (Object.keys(perProvider).length === 0) delete next[provider];
     else next[provider] = perProvider;
+    // 先置位再落盘：`writeModels` 排队的快照里已带上标记，一次写入同时完成
+    // 「改黑名单」和「记录用户动过选择」，不会出现只落了一半的状态。
+    this.selectionTouchedCache = { ...this.selectionTouchedCache, [provider]: true };
     await this.writeModels(next);
   }
   /**
@@ -30131,6 +30172,29 @@ var AccountPool = class {
     this.ensureLoaded();
     return { ...this.contextBudgetCache[provider] ?? {} };
   }
+  /**
+   * 该 provider 的模型选择是否被用户动过。
+   *
+   * 启动补静音据此判断「要不要补」：动过就尊重用户的当前状态（哪怕是全开，
+   * 表现为空黑名单），没动过才补静音。
+   */
+  selectionTouched(provider) {
+    this.ensureLoaded();
+    return this.selectionTouchedCache[provider] === true;
+  }
+  /**
+   * 置位「用户动过该 provider 的模型选择」并落盘。
+   *
+   * ⚠️ 幂等：已置位直接返回，不产生写入 —— 否则每次勾选都要重写整份文档。
+   */
+  async touchSelection(provider) {
+    this.ensureLoaded();
+    if (this.selectionTouchedCache[provider] === true) return;
+    const next = { ...this.selectionTouchedCache, [provider]: true };
+    this.selectionTouchedCache = next;
+    if (this.store.kind === "memory") return;
+    await this.queueStoreSave();
+  }
   /**
    * 设置/清除单个模型的上下文预算。
    *
@@ -30160,13 +30224,18 @@ var AccountPool = class {
    * 注意这与 {@link clearDisabledModels} **不对称** —— 后者的语义是「清空」，
    * 即使传入空列表也仍有事可做（详见该方法注释）。
    */
-  async setModelsDisabled(provider, modelIds) {
+  async setModelsDisabled(provider, modelIds, options) {
     if (modelIds.length === 0) return;
     this.ensureLoaded();
     const next = { ...this.modelCache };
     const perProvider = { ...next[provider] ?? {} };
     for (const id of modelIds) perProvider[id] = true;
     next[provider] = perProvider;
+    // @param options.mark - 默认 true（用户动作）；自动静音传 false，见
+    // touchSelection 的说明 —— 它不是「用户动过选择」。
+    if (options?.mark !== false) {
+      this.selectionTouchedCache = { ...this.selectionTouchedCache, [provider]: true };
+    }
     await this.writeModels(next);
   }
   /**
@@ -30199,6 +30268,9 @@ var AccountPool = class {
     const next = { ...this.modelCache };
     if (Object.keys(nextPerProvider).length === 0) delete next[provider];
     else next[provider] = nextPerProvider;
+    // 注意这条路径也会让黑名单变成空表（用户把该组全开），正是「全开」最
+    // 容易发生的场景，必须同样置位，否则启动补静音会把它们再关回去。
+    this.selectionTouchedCache = { ...this.selectionTouchedCache, [provider]: true };
     await this.writeModels(next);
   }
   /**
@@ -30265,6 +30337,8 @@ var AccountPool = class {
     if (this.modelCache[provider] === void 0) return;
     const next = { ...this.modelCache };
     delete next[provider];
+    // 「全部打开」的最高频入口，必须置位（见 selectionTouched 注释）。
+    this.selectionTouchedCache = { ...this.selectionTouchedCache, [provider]: true };
     await this.writeModels(next);
   }
   /** 持久化模型黑名单（同时更新进程内权威副本）。 */
@@ -32970,6 +33044,13 @@ var DsmlContentExtractor = class {
     return { text: remaining, reasoning: "" };
   }
 };
+// CodeArts rejects an oversized max_tokens with an HTTP-200 SSE error frame
+// (`InferHub.4001.200 The request param is invalid`) rather than a 4xx, so it
+// surfaces to the user as an opaque INVALID_REQUEST. 65536 is the measured
+// ceiling; 131072 already produces an empty stream. Keep the advertised default
+// and the final wire request on that same ceiling - TRAE and Cline do the same
+// with clampTraeMaxTokens / clampClineMaxTokens.
+var CODEARTS_MAX_OUTPUT_TOKENS = 65536;
 var CodeArtsAdapter = class extends LlmAdapter {
   constructor(options) {
     super();
@@ -33043,6 +33124,7 @@ var CodeArtsAdapter = class extends LlmAdapter {
         id: m.id,
         name: m.name,
         inputModalities: modalitiesFor(m.id),
+        maxTokens: CODEARTS_MAX_OUTPUT_TOKENS,
         ...contextWindow === void 0 ? {} : { contextWindow }
       };
     });
@@ -33052,7 +33134,15 @@ var CodeArtsAdapter = class extends LlmAdapter {
     const remoteModel = this.remoteModels?.find((m) => m.id === model);
     const name2 = remoteModel?.name ?? model;
     const contextWindow = contextWindowFor(model);
-    const resolved = { provider, id: model, name: name2, inputModalities: modalitiesFor(model) };
+    const resolved = {
+      provider,
+      id: model,
+      name: name2,
+      inputModalities: modalitiesFor(model),
+      // catalogFactsFor() forwards this as fact.maxTokens; without it the host
+      // has no way to learn the ceiling and passes its own catalog value through.
+      defaultMaxTokens: CODEARTS_MAX_OUTPUT_TOKENS
+    };
     if (contextWindow !== void 0) resolved.context = { contextWindow };
     resolved.reasoning = {
       efforts: [
@@ -33134,9 +33224,11 @@ var CodeArtsAdapter = class extends LlmAdapter {
       // 大文件 write 工具参数（如 1000-2000+ 行文档）需要数万 token 的生成空间，
       // 若沿用后端默认输出上限，参数 JSON 会在中途被截断成非法 JSON，harness
       // 工具校验报 `invalid arguments: "arguments" must be an object`。
-      // 参考实现 e2e 实测：65536 可用，131072 反而触发空流被后端拒绝；
-      // 显式传入的 options.maxTokens 优先，未设置时默认 65536。
-      max_tokens: options.maxTokens ?? 65536,
+      // 参考实现 e2e 实测：65536 可用，131072 反而触发空流被后端拒绝。
+      // 因此这是上限而不只是默认值：调用方传来的更大值（宿主模型目录里别的
+      // provider 带 384000）会被原样发出并触发 INVALID_REQUEST，必须收回；
+      // 调用方主动设的较小值保留。
+      max_tokens: Math.min(options.maxTokens ?? CODEARTS_MAX_OUTPUT_TOKENS, CODEARTS_MAX_OUTPUT_TOKENS),
       ...wireTools !== void 0 && wireTools.length > 0 ? { tools: wireTools } : {}
     });
     const url = `${CHAT_API_BASE}/chat/completions`;
@@ -34694,6 +34786,37 @@ function contentToText2(content) {
   return content.filter((block) => typeof block === "object" && block !== null && block.type === "text").map((block) => String(block.text)).join("");
 }
 var TOOL_RESULT_IMAGE_TEXT = "Attached image(s) from tool result:";
+/**
+ * LOCAL ADAPTATION: a wire message that carries no payload at all.
+ *
+ * The domestic gateway rejects `{ role: "assistant", content: "" }` with
+ * 400 / code 11151 "a message has empty content". That shape is reachable
+ * without the model doing anything wrong: a turn interrupted while the model
+ * was still thinking leaves an assistant message holding only reasoning, and
+ * the serializers below emit it with an empty body because the domestic
+ * gateway does not read `reasoning_content`. One such message then poisons
+ * every later turn of that conversation, and retrying cannot help because the
+ * offending message is history.
+ *
+ * `user` messages are deliberately NOT filtered: both gateways accept an empty
+ * user message (verified against live endpoints), and dropping one would erase
+ * a turn boundary the model is reading.
+ */
+function isBlankWireMessage(entry) {
+  if (entry === null || typeof entry !== "object") return false;
+  if (entry.role !== "assistant" && entry.role !== "system") return false;
+  // A tool call IS content: it anchors the matching tool result, so removing
+  // the message would orphan that result and fail differently.
+  if (Array.isArray(entry.tool_calls) && entry.tool_calls.length > 0) return false;
+  const content = entry.content;
+  if (typeof content === "string") return content.trim().length === 0;
+  if (Array.isArray(content)) return content.length === 0;
+  return false;
+}
+function dropBlankWireMessages(wire) {
+  const kept = wire.filter((entry) => !isBlankWireMessage(entry));
+  return kept.length === wire.length ? wire : kept;
+}
 function serializeMessages2(messages, imageUrls) {
   const wire = [];
   let pendingToolImages = [];
@@ -34772,7 +34895,7 @@ function serializeMessages2(messages, imageUrls) {
     }
   }
   flushToolImages();
-  return wire;
+  return dropBlankWireMessages(wire);
 }
 function errorMessage2(error) {
   if (error instanceof Error) return error.message;
@@ -35326,9 +35449,13 @@ var BuddyAdapter = class _BuddyAdapter extends LlmAdapter2 {
     // pair (measured on a live account 2026-10-21; the pool's own copy of this
     // list is BANNED_SYSTEM_PHRASES in src/upstream.ts). Runs AFTER the prepend
     // above, because that is where the host's system text lands.
+    // The third entry is the file-reference guidance DSH adds whenever the
+    // agent has a `read` tool: the "Relative paths…" half is the substring the
+    // gateway lists, and the sentence names no product, so it is rewritten for
+    // its text rather than for claiming another client (measured 2026-10-28).
     for (const message of messages) {
       if ((message.role === "system" || message.role === "assistant") && typeof message.content === "string") {
-        message.content = message.content.replace(/powered by DeepSeek Harness/gi, "running on DeepSeek Harness").replace(/DeepSeek Harness Web GUI/gi, "DeepSeek Harness web interface");
+        message.content = message.content.replace(/powered by DeepSeek Harness/gi, "running on DeepSeek Harness").replace(/DeepSeek Harness Web GUI/gi, "DeepSeek Harness web interface").replace(/Relative paths resolve from the workspace root/gi, "Relative paths are resolved against the workspace root");
       }
     }
     const tools = options.tools?.map((tool) => ({
@@ -38225,7 +38352,7 @@ function serializeMessages4(messages, imageUrls) {
     }
   }
   flushToolImages();
-  return wire;
+  return dropBlankWireMessages(wire);
 }
 function errorMessage4(error) {
   if (error instanceof Error) return error.message;
@@ -68155,12 +68282,19 @@ function apply(ctx, config = {}) {
         all.map((account) => account.provider).filter((provider) => typeof provider === "string" && provider !== "" && !POOL_OWNED_PROVIDERS.has(provider))
       );
       for (const provider of providers) {
+        // 「黑名单为空」有两种可能：用户从未选择过，或者用户把模型**全部打开**了
+        // （打开时我们删除键，所以两者都是空表）。单看黑名单无法分辨，一律补静音
+        // 就会把用户特意全开的渠道每次重启又关回去 —— 目录越小的渠道越必然中招。
+        // 只有「用户没动过选择」才补。
+        if (pool.selectionTouched(provider)) continue;
         if (pool.disabledModelsFor(provider).size !== 0) continue;
         const catalog = channelPackModelAdaptersRef?.[provider]?.listAllModels();
         if (!Array.isArray(catalog) || catalog.length === 0) continue;
-        await pool.setModelsDisabled(provider, catalog.map((model) => model.id));
+        // mark: false —— 自动静音不是「用户动过选择」，不能让它把标记置位，
+        // 否则此后用户全开也会因为标记已在而被永久跳过后续逻辑。
+        await pool.setModelsDisabled(provider, catalog.map((model) => model.id), { mark: false });
         ctx.logger.info(
-          `[channel-pack] ${provider} 启动补静音（已有账号但黑名单为空）：已默认静音 ${catalog.length} 个模型（可在账号池页勾选启用）`
+          `[channel-pack] ${provider} 启动补静音（已有账号但从未选择过模型）：已默认静音 ${catalog.length} 个模型（可在账号池页勾选启用）`
         );
       }
     } catch (error) {
@@ -68174,15 +68308,19 @@ function apply(ctx, config = {}) {
     const isFirstForProvider = pool.listAccountsByProvider(entry.provider).length === 0;
     const hadEmptyBlocklist = pool.disabledModelsFor(entry.provider).size === 0;
     const result = await rawAddAccount(entry);
-    if ((isFirstForProvider || hadEmptyBlocklist) && !POOL_OWNED_PROVIDERS.has(entry.provider)) {
+    // 同上：`hadEmptyBlocklist` 分不清「从未选择」与「用户全开」，再加一道
+    // `selectionTouched` 判据，否则给已有账号的渠道新加一个号就会把用户刚打开的
+    // 整目录模型全部关掉。
+    const untouched = !pool.selectionTouched(entry.provider);
+    if ((isFirstForProvider || hadEmptyBlocklist) && untouched && !POOL_OWNED_PROVIDERS.has(entry.provider)) {
       void Promise.resolve().then(async () => {
         try {
           const adapters = channelPackModelAdaptersRef;
           const catalog = adapters?.[entry.provider]?.listAllModels();
           if (!Array.isArray(catalog) || catalog.length === 0) return;
-          await pool.setModelsDisabled(entry.provider, catalog.map((model) => model.id));
+          await pool.setModelsDisabled(entry.provider, catalog.map((model) => model.id), { mark: false });
           ctx.logger.info(
-            `[channel-pack] ${entry.provider} ${isFirstForProvider ? "首次登录" : "补静音（黑名单为空）"}：已默认静音 ${catalog.length} 个模型（可在账号池页勾选启用）`
+            `[channel-pack] ${entry.provider} ${isFirstForProvider ? "首次登录" : "补静音（从未选择过）"}：已默认静音 ${catalog.length} 个模型（可在账号池页勾选启用）`
           );
         } catch (error) {
           ctx.logger.warn(