dsh-workbuddy-xdpool 1.7.6 → 1.7.7

package/CHANGELOG.md

@@ -4,6 +4,83 @@
 
 版本号遵循 [语义化版本](https://semver.org/lang/zh-CN/)。
 
+## 1.7.7 — 别再让你为了一个「等一会儿」去重新登录
+
+> 报错说「API 密钥无效」，于是你跑去重新登录 —— 但密钥根本没问题，只是那个模型暂时限流了。
+
+### 问题一：限流被说成了「密钥无效」
+
+**你会看到**：国际版用 hy4 时，偶尔弹出「API 密钥无效」。你重新登录、换账号，都不一定管用。
+
+**为什么**：看日志就清楚了（时间连着三行）：
+
+```
+14:50:22  qingche526  rate-limited on hy4-preview-f  → cooling + rotating
+14:50:23  aosi526     rate-limited on hy4-preview-f  → cooling + rotating
+14:50:25  401: "no WorkBuddy credential found; sign in on the desktop app"
+```
+
+两个账号的 hy4-preview-f 都被限流**冷却**了（一小时），于是池子里暂时没有可用账号。可插件判断「有没有被限流」用的是**本次请求内**的临时记录 —— 上次请求造成的冷却，它看不见。于是它以为「一个账号都没有 → 你没登录」→ 报 **401**。
+
+DSH 把 401 显示成「API 密钥无效」。**你被派去修一个根本没坏的东西。**
+
+**怎么修的**：不再靠本次请求的临时记录猜，而是**直接问账号池「为什么现在没有可用账号」**，然后按真实原因报对应的错：
+
+| 真实情况 | 以前 | 现在 |
+|---|---|---|
+| 模型限流中 | 401 密钥无效 | **429 该模型限流中**（等一会儿即可）|
+| 账号全被关/移出 | 401 密钥无效 | **403 账号全部已关闭**（去卡片打开）|
+| 积分低于保留线 | 401 密钥无效 | **402 积分不足** |
+| 确实没登录 | 401 | 401（这个才是对的）|
+
+四种情况**需要完全相反的处理**，以前却都推给你「重新登录」。现在只有真的该重新登录时才这么说。
+
+### 问题二：上下文「压缩」把目标设得比原文还大
+
+日志里抓到的：
+
+```
+context overrun on deepseek-v4.1-flash (~791793 tokens); compacting to ~797952
+                                  ↑ 原文 79.1 万                    ↑ 目标 79.7 万
+```
+
+**压缩目标比要压缩的东西还大** —— 等于什么都没压，重试当然又超限。而日志看起来还挺像"我处理过了"。
+
+**为什么**：目标值是按**目录里写的窗口**算的（`deepseek-v4.1-flash` 标称 1M），取 80% → `797952`。但**标称 1M 是宣传值，不是真实生效上限** —— 否则它就不会被拒了。
+
+**怎么修的**：**正在超限的那个请求本身就是证据**。既然它太大了，目标就必须明显小于它。现在取「按窗口算」和「按原文一半」的较小值：
+
+```
+原文 791793 → 目标 395896      （真正压掉一半）
+```
+
+### 顺便：这类报错终于有线索了
+
+上游的 `model_param_invalid` 里 **`param` 字段是空的** —— 它只说「参数不符合要求」，不说是哪个参数。你之前贴的那条 11133 就是这样，等于没有线索。
+
+现在遇到 400 时，日志会记下**请求的形状**（顶层参数名、各角色的消息条数、工具数量、内容字符数）：
+
+```
+rejected the request as invalid (model deepseek-v4.1-flash) —
+keys=[max_tokens messages model reasoning_effort stream temperature tool_choice tools top_k top_p]
+messages=122 (system:1 user:61 assistant:30 tool:30) tools=77 contentChars=481233
+```
+
+**只有形状，没有内容**（对话是你的，不该进日志）。下次再遇到，日志里就能看出是哪个参数/哪个维度不对。
+
+### 我排查时的事
+
+我一开始想直接复现 11133，试了：采样参数（`top_k`/`repetition_penalty`/`temperature`/`top_p`）、推理档位（`minimal` 到 `max`，还试了无效档位）、各种消息形状（回放 `reasoning_content`、工具调用链）、消息数量（最多 362 条）、77 个工具 + 长会话、超长 prompt ——**国内版和国际版全部返回 200**。
+
+**我没能复现它**，所以没有针对 11133 写"修复"（那会是猜测）。改的都是**从日志里坐实的**问题：那两个是确凿的。这个 11133 加上刚加的形状日志，下次出现时就能定位了。
+
+### 验证
+
+- 类型检查全绿，测试 **325 项全绿**（+12）。
+- 新增 `tests/pool-availability.test.ts`（7 项）：池能正确区分「无人登录 / 模型限流 / 账号全关 / 积分保留 / 可用」、**限流时返回 429 且文案不含「sign in」**、真没登录时仍是 401、**不把另一个 region 的账号算进来**、过期冷却自动视为可用。
+- 新增 `tests/context-budget-budget.test.ts`（5 项）：**目录窗口远大于原文时目标必须更小**（用日志里那组真实数字）、窗口未知时减半、窗口小于原文时也不放大、目标不为零。
+- **反向验证**：把 429 分支禁用、退回原来的 401 → 测试变红。
+
 ## 1.7.6 — 一个读不出来的凭据文件，不再带走整个账号池
 
 > 你切了 4 个账号登录，界面显示 4 个；重启之后只剩 2 个 —— 不是账号没了，是它们被一个文件拖累了。

package/lib/bin.js

@@ -2846,6 +2846,70 @@ var WorkBuddyAccountPool = class {
 			return true;
 		});
 	}
+	/**
+	* Why no account is available right now, for an accurate error.
+	*
+	* The pool can be empty for reasons that need OPPOSITE remedies: nobody is
+	* signed in (the user must sign in), every account is rate-limited (the user
+	* must wait, and retrying later works), or every account was switched off /
+	* ignored (the user must re-enable one). Reporting all of them as "no
+	* credential, sign in" sent users to re-authenticate over a temporary 429 —
+	* observed as an "API key invalid" panel for a model that was merely cooling.
+	*
+	* Counts are over the region's accounts, since a provider only ever sees its
+	* own gateway.
+	*/
+	unavailableReason(modelId, region) {
+		const now = Date.now();
+		const inRegion = this.accounts.filter((account) => region === void 0 || regionOf(account.credential.domain) === region);
+		if (inRegion.length === 0) return {
+			total: 0,
+			cooling: 0,
+			disabled: 0,
+			reason: "empty"
+		};
+		let cooling = 0;
+		let disabled = 0;
+		for (const account of inRegion) {
+			if (this.disabledIds.has(account.id)) {
+				disabled += 1;
+				continue;
+			}
+			const modelCooling = modelId !== void 0 && (account.modelCooldowns[modelId] ?? 0) > now;
+			if (account.cooldownUntilMs > now || modelCooling) {
+				cooling += 1;
+				continue;
+			}
+			const reserve = this.creditReserves.get(account.id);
+			if (reserve !== void 0 && reserve > 0) {
+				const balance = this.creditBalances.get(account.id);
+				if (balance !== void 0 && balance <= reserve) return {
+					total: inRegion.length,
+					cooling,
+					disabled,
+					reason: "reserve"
+				};
+			}
+		}
+		if (cooling > 0) return {
+			total: inRegion.length,
+			cooling,
+			disabled,
+			reason: "cooling"
+		};
+		if (disabled > 0) return {
+			total: inRegion.length,
+			cooling,
+			disabled,
+			reason: "disabled"
+		};
+		return {
+			total: inRegion.length,
+			cooling,
+			disabled,
+			reason: "none"
+		};
+	}
 	/** Round-robin: the legacy cursor walk, kept for the distribution that asks for it. */
 	pickRoundRobin(pool) {
 		const index = this.cursor % pool.length;

package/lib/index.d.ts

@@ -989,6 +989,25 @@ export declare class WorkBuddyAccountPool {
    * model, e.g. CLI diagnostics).
    */
   private available;
+  /**
+   * Why no account is available right now, for an accurate error.
+   *
+   * The pool can be empty for reasons that need OPPOSITE remedies: nobody is
+   * signed in (the user must sign in), every account is rate-limited (the user
+   * must wait, and retrying later works), or every account was switched off /
+   * ignored (the user must re-enable one). Reporting all of them as "no
+   * credential, sign in" sent users to re-authenticate over a temporary 429 —
+   * observed as an "API key invalid" panel for a model that was merely cooling.
+   *
+   * Counts are over the region's accounts, since a provider only ever sees its
+   * own gateway.
+   */
+  unavailableReason(modelId?: string, region?: WorkBuddyRegion): {
+    total: number;
+    cooling: number;
+    disabled: number;
+    reason: 'empty' | 'cooling' | 'disabled' | 'reserve' | 'none';
+  };
   /** Round-robin: the legacy cursor walk, kept for the distribution that asks for it. */
   private pickRoundRobin;
   /**

package/lib/index.js

@@ -2954,6 +2954,70 @@ var WorkBuddyAccountPool = class {
 			return true;
 		});
 	}
+	/**
+	* Why no account is available right now, for an accurate error.
+	*
+	* The pool can be empty for reasons that need OPPOSITE remedies: nobody is
+	* signed in (the user must sign in), every account is rate-limited (the user
+	* must wait, and retrying later works), or every account was switched off /
+	* ignored (the user must re-enable one). Reporting all of them as "no
+	* credential, sign in" sent users to re-authenticate over a temporary 429 —
+	* observed as an "API key invalid" panel for a model that was merely cooling.
+	*
+	* Counts are over the region's accounts, since a provider only ever sees its
+	* own gateway.
+	*/
+	unavailableReason(modelId, region) {
+		const now = Date.now();
+		const inRegion = this.accounts.filter((account) => region === void 0 || regionOf(account.credential.domain) === region);
+		if (inRegion.length === 0) return {
+			total: 0,
+			cooling: 0,
+			disabled: 0,
+			reason: "empty"
+		};
+		let cooling = 0;
+		let disabled = 0;
+		for (const account of inRegion) {
+			if (this.disabledIds.has(account.id)) {
+				disabled += 1;
+				continue;
+			}
+			const modelCooling = modelId !== void 0 && (account.modelCooldowns[modelId] ?? 0) > now;
+			if (account.cooldownUntilMs > now || modelCooling) {
+				cooling += 1;
+				continue;
+			}
+			const reserve = this.creditReserves.get(account.id);
+			if (reserve !== void 0 && reserve > 0) {
+				const balance = this.creditBalances.get(account.id);
+				if (balance !== void 0 && balance <= reserve) return {
+					total: inRegion.length,
+					cooling,
+					disabled,
+					reason: "reserve"
+				};
+			}
+		}
+		if (cooling > 0) return {
+			total: inRegion.length,
+			cooling,
+			disabled,
+			reason: "cooling"
+		};
+		if (disabled > 0) return {
+			total: inRegion.length,
+			cooling,
+			disabled,
+			reason: "disabled"
+		};
+		return {
+			total: inRegion.length,
+			cooling,
+			disabled,
+			reason: "none"
+		};
+	}
 	/** Round-robin: the legacy cursor walk, kept for the distribution that asks for it. */
 	pickRoundRobin(pool) {
 		const index = this.cursor % pool.length;
@@ -5548,6 +5612,40 @@ function isContextTooLong(body) {
 	if (/exceeds?\s+(the\s+)?(model\s+)?context\s+(window|limit)/iu.test(body)) return true;
 	return false;
 }
+/**
+* A content-free description of a request, for the log when the upstream
+* rejects it.
+*
+* `model_param_invalid` arrives with an EMPTY `param` field, so the error alone
+* says nothing about which parameter was rejected — the user is left with "the
+* request parameters do not meet the current model requirements" and no way to
+* act. Recording the shape (top-level keys, message count by role, tool count,
+* approximate prompt size) makes such a rejection diagnosable from the log.
+*
+* Deliberately SHAPE ONLY: no message content, no tool schemas, no tokens. It
+* ends up in a log file, and the conversation it describes is the user's.
+*/
+function requestShape(raw) {
+	try {
+		const parsed = JSON.parse(raw);
+		const keys = Object.keys(parsed).sort();
+		const messages = Array.isArray(parsed["messages"]) ? parsed["messages"] : [];
+		const byRole = /* @__PURE__ */ new Map();
+		let chars = 0;
+		for (const entry of messages) {
+			if (typeof entry !== "object" || entry === null) continue;
+			const message = entry;
+			const role = typeof message["role"] === "string" ? message["role"] : "?";
+			byRole.set(role, (byRole.get(role) ?? 0) + 1);
+			chars += JSON.stringify(message["content"] ?? "").length;
+		}
+		const roles = [...byRole.entries()].map(([role, n]) => `${role}:${n}`).join(" ");
+		const tools = Array.isArray(parsed["tools"]) ? parsed["tools"].length : 0;
+		return `keys=[${keys.join(" ")}] messages=${messages.length}${roles === "" ? "" : ` (${roles})`} tools=${tools} contentChars=${chars}`;
+	} catch {
+		return "body was not parseable JSON";
+	}
+}
 function readBody(req) {
 	return new Promise((resolve, reject) => {
 		const chunks = [];
@@ -5686,17 +5784,26 @@ function createWorkBuddyShim(options) {
 		}
 		const tried = [];
 		let last;
-		let exhaustedByRateLimit = false;
 		for (let attempt = 0; attempt < maxAttempts; attempt += 1) {
 			if (controller.signal.aborted) return;
 			const account = await pool.acquire(modelId, region);
 			if (account === void 0) {
-				if (exhaustedByRateLimit && last !== void 0) {
+				const why = pool.unavailableReason(modelId, region);
+				if (why.reason === "cooling") {
 					const subject = modelId === void 0 ? "every WorkBuddy account is rate-limited" : `every account is rate-limited for model ${modelId}`;
-					writeOpenAIError(res, KIND_STATUS[last.kind], last.kind, `${subject} (tried ${tried.length}: ${tried.join(", ")}); resets at the upstream window — ${last.message.slice(0, 200)}`);
+					const upstream = last === void 0 ? "" : ` — ${last.message.slice(0, 200)}`;
+					writeOpenAIError(res, 429, "soft_rate", `${subject} (${why.cooling}/${why.total} cooling, tried ${tried.length})${tried.length === 0 ? "" : `: ${tried.join(" → ")}`}${upstream}`);
+					return;
+				}
+				if (why.reason === "disabled") {
+					writeOpenAIError(res, 403, "no_enabled_account", `all ${why.total} WorkBuddy account(s) for this gateway are switched off or ignored; re-enable one on the pool card`);
+					return;
+				}
+				if (why.reason === "reserve") {
+					writeOpenAIError(res, 402, "reserve_floor", `every WorkBuddy account is at or below its credit reserve (${why.total} account(s)); lower the reserve on the pool card or top up credits`);
 					return;
 				}
-				writeOpenAIError(res, 401, "not_signed_in", "no WorkBuddy credential found; sign in on the desktop app (or set WORKBUDDY_AUTH_FILE)");
+				writeOpenAIError(res, 401, "not_signed_in", why.total === 0 ? "no WorkBuddy credential found for this gateway; sign in on the desktop app (or set WORKBUDDY_AUTH_FILE)" : `no usable WorkBuddy credential among ${why.total} account(s) for this gateway; sign in again on the desktop app`);
 				return;
 			}
 			tried.push(account.label);
@@ -5720,8 +5827,8 @@ function createWorkBuddyShim(options) {
 				logger?.warn(`dsh-workbuddy-xdpool: ${account.label} has no credits left (attempt ${attempt + 1}/${maxAttempts}); rotating`);
 				continue;
 			}
+			if (result.kind === "client") logger?.warn(`dsh-workbuddy-xdpool: ${account.label} rejected the request as invalid (model ${modelId ?? "(none)"}) — ${requestShape(raw)}; upstream: ${result.message.slice(0, 300)}`);
 			if (result.kind !== "soft_rate") break;
-			exhaustedByRateLimit = true;
 			pool.penalize(account.id, parseRateLimitReset(result.message), modelId);
 			logger?.warn(`dsh-workbuddy-xdpool: ${account.label} rate-limited on ${modelId ?? "(no model)"} (attempt ${attempt + 1}/${maxAttempts}); rotating`);
 		}
@@ -5844,8 +5951,10 @@ async function recoverFromContextOverrun(options) {
 	};
 	const overrunTokens = estimateMessagesTokens(messages);
 	const realWindow = options.contextWindow;
-	const budget = realWindow !== void 0 && realWindow > 0 ? Math.max(512, Math.floor(realWindow * .8) - 2048) : Math.max(512, Math.floor(overrunTokens / 2));
-	logger?.warn(`dsh-workbuddy-xdpool: context overrun on ${modelId ?? "(no model)"} (~${overrunTokens} tokens); compacting to ~${budget} and retrying once`);
+	const fromWindow = realWindow !== void 0 && realWindow > 0 ? Math.max(512, Math.floor(realWindow * .8) - 2048) : Number.POSITIVE_INFINITY;
+	const fromOverrun = Math.max(512, Math.floor(overrunTokens / 2));
+	const budget = Math.min(fromWindow, fromOverrun);
+	logger?.warn(`dsh-workbuddy-xdpool: context overrun on ${modelId ?? "(no model)"} (~${overrunTokens} tokens, catalog window ${realWindow ?? "unknown"}); compacting to ~${budget} and retrying once`);
 	let summary;
 	let compacted = messages;
 	let compactionDetail = "";

package/package.json

@@ -2,7 +2,7 @@
   "name": "dsh-workbuddy-xdpool",
   "displayName": "DSH WorkBuddy XD Pool",
   "description": "Merge every locally signed-in WorkBuddy account into DeepSeek Harness as one auto-failing-over model pool (multi-account rotation, live credits, daily check-in and model catalog).",
-  "version": "1.7.6",
+  "version": "1.7.7",
   "license": "MIT",
   "author": "XDTrees",
   "homepage": "https://github.com/XDTrees/dsh-workbuddy-xdpool#readme",