dsh-connect-trae 4.4.0 → 4.4.1

package/CHANGELOG.md

@@ -1,5 +1,60 @@
 # Changelog
 
+## 4.4.1 (2026-10-22)
+
+> **4.4.0 的修复自己藏掉了 `kimi-k2-0905`。这个版本把它捞回来，并补上 4.4.0 缺的实测。**
+>
+> 一条自我纠正。4.4.0 发布后作者对提升的模型做了**真实调用**实测，实测把 4.4.0 的
+> 一个缺陷暴露出来：为了防止重名孪生行，4.4.0 加了「显示名已被占用就跳过」的规则 ——
+> 这条规则把 `kimi-k2-0905` 也丢掉了，而它**是另一个真模型**（168K 上下文，实测可调用）。
+> 于是报告人 `yuantao313` 抱怨的那一类「带日期的快照 ID」在 `kimi-k2-0905` 上**依然拿不到** ——
+> 成因从「合并只遍历骨架」换成了「去重按名字判断」，**症状没变**。
+> 详见 [`docs/ISSUE46_DIAGNOSIS.md`](https://github.com/dingminhua/dsh-connect-trae/blob/main/docs/ISSUE46_DIAGNOSIS.md) 第 9 节。
+
+### 修复
+
+- **重名判断改为「能力测试」而非「名字测试」**（`fix(catalog)`，issue #46）。
+  `claimedNames` 记录每个显示名下**最大的**上下文窗口；一个重名行只有在**严格更大于**
+  该最大值时才提升，并被追加 id 以便区分（两行都渲染成 `Kimi-K2-0905` 的 picker 是在撒谎）。
+  - **实测依据**（2026-10-22，CN 凭据）：`kimi-k2` 是 28000 而 `kimi-k2-0905` 是 168000，
+    名字相同、能力差 6 倍，且 `kimi-k2-0905` 在 `solo_coder` 与 `chat_v3` 下**都能调用**。
+  - 反面对照：`glm-4.7` 与 `glm-4.7-auto` 元数据**完全相同**（100000/16000/`chat_v3`），
+    是真重复，仍然只留一行；`glm-5.2_advisor_default`（168000）比 skeleton 的 `glm-5.2`
+    （200000）**更小**，是内部改道，仍被排除。
+  - **实测结果**：目录 **42 → 46 行**，提升行 7 → 11。新增
+    `kimi-k2-0905`、`minimax-m2-auto`、`minimax-m2.1-auto`、`glm-4.6-auto`
+    （全部为 168K / `solo_coder`，显示名形如 `MiniMax-M2 (minimax-m2-auto)`）。
+  - 比较基准是**最强的占用者**，因此与 wire 名单的顺序无关（有回归测试钉住）。
+
+### 测试补强
+
+- **982 例测试全绿**（`pnpm run test`），新增 **4 例**：重名但更强者被提升且名字带 id、
+  重名且相等/更弱者仍被排除、`-auto` 孪生按能力取舍、比较基准取最强占用者。
+
+### 实测记录（4.4.0 遗留边界现已关闭）
+
+`4.4.0` 的 CHANGELOG 曾记「新提升的 7 个模型未逐个实测可调用性」。现已实测，**全部通过**
+（真调用，走插件自己的 bridge，每个一次最小 `ping`）：
+
+| 模型 | function | 结果 |
+| --- | --- | --- |
+| `seed-code-pro-0430` / `glm-5-turbo` | `solo_work_remote` | AVAILABLE |
+| `glm-4.7` / `glm-4.6` / `minimax-m2` / `minimax-m2.1` / `qwen3-coder` | `chat_v3` | AVAILABLE |
+| `kimi-k2` | `inline_chat` · `chat_v3` · `solo_work_remote` | 三个 function 全部 AVAILABLE |
+| 4.4.1 新增的 4 行 | `solo_coder` | 全部 AVAILABLE |
+
+合计 21 次真实调用，消耗 **约 11 额度**（约 0.43/次）。
+因此 `4.4.0` 中「`kimi-k2` 归属 `inline_chat`、可调用性尚未取证」这条边界**一并关闭**。
+
+### 已知边界
+
+- fix A 多问 function 带出的 `doubao_1_8`（`Doubao-Seed-1.8`）**实测 4001 拒绝**：它没有
+  `wireFunction`、也没有上下文与输出元数据，是条残缺的骨架行。按 issue #19「全部呈现 +
+  如实报错」的策略这是合规行为（选中会得到明确的上游原文错误），但它确实是一条新增的、
+  调不通的行 —— 记在这里，不假装它可用。
+- 取证仍在维护者本机的 CN 账号上完成。报告人账号上是否真有
+  `deepseek-v4.1-flash-1017` 这一行，仍需他升级后确认（本机 roster 上无此行）。
+
 ## 4.4.0 (2026-10-22)
 
 > **模型表不再系统性排除「带日期的快照配置」；账号池开关不再被模型目录的保存冲掉。**

package/lib/index.d.ts

@@ -385,7 +385,9 @@ interface TraeWireModel {
  *    name has nothing to show in the picker.
  *
  * The display-name-collision test lives in the merge itself (it needs the
- * Remote rows to compare against); see {@link mergeTraeModelSources}.
+ * Remote rows to compare against); see {@link mergeTraeModelSources}. It is a
+ * CAPABILITY test there, not a name test: a collision only suppresses a row
+ * when the equally-named row is at least as capable.
  *
  * The survivors are exactly the models a user would recognise and the IDE
  * offers but the plugin hid: `kimi-k2` / `kimi-k2-0905`, `glm-4.6`,

package/lib/index.js

@@ -210,7 +210,9 @@ function applyImageSelection(models, selected) {
 *    name has nothing to show in the picker.
 *
 * The display-name-collision test lives in the merge itself (it needs the
-* Remote rows to compare against); see {@link mergeTraeModelSources}.
+* Remote rows to compare against); see {@link mergeTraeModelSources}. It is a
+* CAPABILITY test there, not a name test: a collision only suppresses a row
+* when the equally-named row is at least as capable.
 *
 * The survivors are exactly the models a user would recognise and the IDE
 * offers but the plugin hid: `kimi-k2` / `kimi-k2-0905`, `glm-4.6`,
@@ -316,15 +318,61 @@ function mergeTraeModelSources(remote, wire) {
 		});
 	}
 	const remoteIds = new Set(remote.map((model) => displayKey(model.id)));
-	const remoteNames = new Set(remote.map((model) => displayKey(model.name)).filter((name) => name !== ""));
+	/**
+	* Display name -> the LARGEST context window claimed under it.
+	*
+	* A name collision does NOT by itself mean "alias". Two rows sharing a
+	* display name are sometimes genuinely different models, and the only
+	* upstream signal that separates them is capability. Measured 2026-10-22 on
+	* a live CN credential:
+	*
+	*  - `kimi-k2` (28000) and `kimi-k2-0905` (168000) BOTH display as
+	*    `Kimi-K2-0905` — and both ANSWER (verified: `kimi-k2-0905` answers
+	*    under `solo_coder` and `chat_v3`). Dropping the second one hid a real
+	*    168k model behind an equally-named 28k one, which is exactly the
+	*    "dated snapshot the IDE offers, the plugin does not" complaint of
+	*    issue #46.
+	*  - `minimax-m2` (100000, `chat_v3`) and `minimax-m2-auto` (168000,
+	*    `solo_coder`) both display as `MiniMax-M2` — again a bigger config
+	*    behind an equally-named smaller one.
+	*  - `glm-4.7` and `glm-4.7-auto` both display as `GLM-4.7` with IDENTICAL
+	*    metadata (100000/16000/`chat_v3`) — a true duplicate, and promoting it
+	*    would list one model twice for no gain.
+	*  - `glm-5.2_advisor_default` (168000) is SMALLER than the skeleton's
+	*    `glm-5.2` (200000), so it is an internal re-routing of a model the user
+	*    already has in full — not worth a second row.
+	*
+	* So the rule is: a collision skips the row UNLESS the new row carries a
+	* STRICTLY LARGER context window, in which case the user is being offered
+	* capability the equally-named row does not have and hiding it would cost
+	* them something real. Names are tracked with their best context so the
+	* comparison is always against the strongest claimant, whatever the order.
+	*/
+	const claimedNames = /* @__PURE__ */ new Map();
+	const claim = (name, contextWindow) => {
+		const key = displayKey(name);
+		if (key === "") return;
+		claimedNames.set(key, Math.max(claimedNames.get(key) ?? 0, contextWindow ?? 0));
+	};
+	for (const model of remote) claim(model.name, model.contextWindow);
 	for (const wireModel of wire) {
 		if (remoteIds.has(displayKey(wireModel.id))) continue;
 		if (!isPromotableWireModel(wireModel)) continue;
-		if (remoteNames.has(displayKey(wireModel.name))) continue;
-		remoteNames.add(displayKey(wireModel.name));
+		const collides = claimedNames.has(displayKey(wireModel.name));
+		if (collides && (wireModel.contextWindow ?? 0) <= (claimedNames.get(displayKey(wireModel.name)) ?? 0)) continue;
+		/**
+		* A surviving collision gets the id appended, because the whole point of
+		* promoting it is that the user can TELL IT APART from the equally-named
+		* row: two rows rendering as `MiniMax-M2` is a picker that lies about
+		* having one model. The id is the `config_name` the bridge sends, so it is
+		* also the most actionable disambiguator available.
+		*/
+		const name = collides ? `${wireModel.name} (${wireModel.id})` : wireModel.name;
+		claim(wireModel.name, wireModel.contextWindow);
+		claim(name, wireModel.contextWindow);
 		result.push({
 			id: wireModel.id,
-			name: wireModel.name,
+			name,
 			...wireModel.contextWindow === void 0 ? {} : { contextWindow: wireModel.contextWindow },
 			...wireModel.maxTokens === void 0 ? {} : { maxTokens: wireModel.maxTokens },
 			...wireModel.creditMultiplier === void 0 ? {} : { creditMultiplier: wireModel.creditMultiplier },

package/package.json

@@ -2,7 +2,7 @@
   "name": "dsh-connect-trae",
   "displayName": "DSH Connect Trae",
   "description": "把本机登录的 Trae 模型接入 DeepSeek Harness：国内版与国际版双供应商并行，提供用量/积分概览与每日签到领取。",
-  "version": "4.4.0",
+  "version": "4.4.1",
   "author": "dingminhua",
   "copyright": "Copyright (c) 2026 LaoDing",
   "license": "MIT",