dsh-workbuddy-xdpool 1.4.0 → 1.4.1

package/CHANGELOG.md

@@ -4,6 +4,23 @@
 
 版本号遵循 [语义化版本](https://semver.org/lang/zh-CN/)。
 
+## 1.4.1 (2026-09-24)
+
+### 修了：hy3 超长对话还是会弹 400
+
+1.4.0 让溢出消息能被 Harness 认出、并给 shim 加了本地压缩，但 hy3 仍然会爆。根因是
+**窗口值写错了**：catalog 里 hy3 的 `contextWindow` 填的是 `200_000`，而它真实上限远小于
+此。shim 压缩时按「估算 token 数的一半」设预算（约 100K），DSH 压力触发也按 200K 算，
+两侧都以为 200K 够用，所以压缩到 100K 仍远超真实窗口，重试再次溢出。
+
+修法：
+- catalog 里 hy3 的窗口改为真实值 `32_000`（同步把 `maxOutputTokens` 从 128K 调到 8K）；
+- `recoverFromContextOverrun` 不再用「估算一半」，而是用模型**真实**窗口算预算
+  （`realWindow * 0.8 - 2048`），并在调用处把 catalog 的窗口传进去。
+
+这样无论 DSH 外层还是 shim 内层，都按真实窗口压缩，不会再把压过的 prompt 发到真实窗口之外。
+- 新增 `context-budget-real-window.test.ts` 钉住这个回归（含「hy3 不再声明 200K」）。
+
 ## 1.4.0 (2026-09-24)
 
 ### 长对话不再撞死在上下文上限上

package/lib/bin.js

@@ -2261,8 +2261,8 @@ const FALLBACK_WORKBUDDY_MODELS = [
 	{
 		id: "hy3",
 		name: "Hy3",
-		contextWindow: 2e5,
-		maxOutputTokens: 128e3,
+		contextWindow: 32e3,
+		maxOutputTokens: 8e3,
 		supportsImages: true
 	},
 	{

package/lib/index.js

@@ -2368,8 +2368,8 @@ const FALLBACK_WORKBUDDY_MODELS = [
 	{
 		id: "hy3",
 		name: "Hy3",
-		contextWindow: 2e5,
-		maxOutputTokens: 128e3,
+		contextWindow: 32e3,
+		maxOutputTokens: 8e3,
 		supportsImages: true
 	},
 	{
@@ -4495,7 +4495,8 @@ function createWorkBuddyShim(options) {
 				logger,
 				client,
 				pool,
-				maxAttempts
+				maxAttempts,
+				contextWindow: modelId === void 0 ? void 0 : catalog.current().find((m) => m.id === modelId)?.contextWindow
 			});
 			if (recovered.ok) {
 				await serveSuccessfulStream(res, recovered.account, recovered.result, logger, refreshBalance, pool);
@@ -4599,7 +4600,8 @@ async function recoverFromContextOverrun(options) {
 		detail: "request carried no usable messages"
 	};
 	const overrunTokens = estimateMessagesTokens(messages);
-	const budget = Math.max(512, Math.floor(overrunTokens / 2));
+	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`);
 	let summary;
 	let compacted = messages;

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": "1.4.0",
+  "version": "1.4.1",
   "license": "MIT",
   "author": "XDTrees",
   "homepage": "https://github.com/XDTrees/dsh-workbuddy-xdpool#readme",