dsh-connect-trae 4.6.3 → 4.6.5

package/CHANGELOG.md

@@ -1,5 +1,90 @@
 # Changelog
 
+## 4.6.5 (2026-10-30)
+
+> **一条没带上下文窗口的目录行，会让整个 `trae` 区域不可用。** 失效的不是
+> 那条模型，而是**整条路由**——DSH 逐模型校验，且校验失败是 provider 级的。
+> 插件早就写好了 `FALLBACK_CONTEXT_WINDOW` 作为「最后防线」，但它被赋给了一个
+> 本插件**从不会走**的字段，于是从未生效过。这一版把回退落到模型行本身。
+
+### 修复
+
+- **一条没带上下文窗口的目录行不再拖垮整个区域**（#70，贡献者
+  [@beibing173](https://github.com/beibing173)）。
+  - **现象**：模型选择器里 `trae` 整个区域加载失败，报
+    `adapter returned invalid context metadata for provider "trae" model "doubao_1.8"`。
+    失败粒度是 **provider 级**，不是单模型级。
+  - **根因**：`FALLBACK_CONTEXT_WINDOW` 只被赋给 profile 的
+    `defaultContextWindow`，而那个字段由**内核自己的目录解析步骤**读取——本插件
+    从不走那条路径。它用 `createProvider({ models })` 手工装配 pi-ai provider，
+    `Models.getModel()` 直接返回 `toPiModel` 产出的原始行，**不做任何补默认值**；
+    而 `toPiModel` 在 `contextWindow === undefined` 时**省略**该字段。于是
+    `PiAiAdapter.modelInfo()` 解析出 `context: { contextWindow: undefined }`，
+    DSH 拒绝，且拒绝的是整条路由。常量写了，没有生效点。
+  - **修法**：把回退**落到行本身**。新增 `servedContextWindow()`，`toPiModel`
+    现在**总是**输出 `contextWindow`：行自己的正整数窗口原样保留，其余一切
+    （缺失、`NaN`、`0`、负数、小数）回退到 `FALLBACK_CONTEXT_WINDOW`——因为
+    DSH 只接受正整数。profile 上的 `defaultContextWindow` 保留（内核路径仍会读），
+    但注释改成与事实一致：真正的保证在 `toPiModel`，那只是保险带。
+  - **为什么小数也要回退**：目录解析器的 `finitePositive` 只校验 `isFinite` 与
+    `> 0`，所以上游若发布 `context_window_tokens.dev = 200000.5`，它会原样进到
+    适配器；DSH 对非整数的拒绝与缺失完全相同。这条此前没有任何用例锁住。
+  - **测试**：`tests/context-window-fallback.spec.ts` 走宿主的真实路径
+    （`adapter.current()` → `profiles` → `models.getModel()`，以及
+    `adapter.resolveModel()`），覆盖未定尺寸行、混合目录逐行可解析、行自带窗口
+    不被覆盖，以及小数/零/负数/NaN 四种畸形窗口一律回退。
+    **变异验证**：把 `toPiModel` 的改动回退，8 例中 7 例转红。
+  - **不变的部分**：不改变任何已有窗口的行；不改变名称、reasoning、图像能力、
+    计费倍率等元数据；兜底目录本身仍然每行手写真实窗口（`tests/catalog.spec.ts`
+    继续锁这个不变量）。本修复只是让「万一漏了一条」退化成一条保守尺寸的模型，
+    而不是死掉一个区域。
+
+## 4.6.4 (2026-10-28)
+
+> **缓存命中恒为 0% 不是插件丢了数字，是上游没有分配缓存池。** Trae 的
+> `metadata` 事件里 `prompt_cache_pool_record` 非空才代表这次请求挂了前缀
+> 缓存池；实测 CN/AI 双区、4 个账号、7 个模型、4 种 function，该字段
+> **100% 为 `null`**。没有池，`cache_read_input_tokens` 只能是 0 —— 0% 是
+> 正确读数。本版把这个事实变成可观测的，用户不必再怀疑插件吞了数据。
+
+### 修复
+
+- **上游的缓存池事实与管线名不再被丢弃**（#67）。
+  - **现象**：无论简单还是复杂提问，缓存命中恒为 0%，十个会话一致；
+    同时 tps 在 60–80 之间浮动。
+  - **根因**：不是插件。issue #10 的转发修复仍在（`lib/index.js` 含
+    `cached_tokens`，同版本下 `workbuddy`/`dsh-opencode-go` 命中显示正常），
+    是上游 `llm_utils_chat` 改由 `llm_raw_chat_v2` 管线、火山方舟 `ep-…`
+    端点承载后，不再为这些请求分配前缀缓存池。当初产出 5248/9216 的
+    `scripts/probe-cache-convention.mjs` 原样重跑，现在三轮全是 0。
+  - **修法**：`sse.ts` 新增 `diagnostic` 事件类型，解码 `metadata`
+    （`prompt_cache_pool_record`）与 `timing_cost`（`name` 管线名、
+    `provider_model_name`）；`solo-bridge.ts` 通过可选 `log` 上报
+    `promptCachePoolRecord` / `pipeline` / `providerModelName`；
+    `index.ts` 四个 `TraeSoloBridge` 构造点接入 `ctx.logger.warn`。
+    诊断事件不进答案文本、不改 usage 映射，**对外行为不变**。
+  - **为什么区分 `null` 与「字段不存在」**：`null` 是上游明确表示「没挂池」，
+    字段缺失只是没这条信息；两者混为一谈就又回到「看不出是谁的问题」。
+
+### 修复（顺带守住一个回归）
+
+- **`metadata` 不能被 peek 当成「上游已应答」**。它现在是被认识的事件类型，
+  若据此停止 peek，紧跟其后的拒绝（如 `1005` 未订阅）会退化成截断的 200 流
+  —— 正是 issue #19 的故障形态。`openTraeSource` 显式把 `diagnostic` 排除在
+  「已应答」之外。
+
+### 测试
+
+- `tests/sse.spec.ts`：`metadata` / `timing_cost` 解码为 `diagnostic`，
+  保留 `prompt_cache_pool_record: null`，且区分字段缺失。
+- `tests/solo-bridge.spec.ts`：诊断事实上报（含 `null` 原样保留、事件不泄漏进
+  答案）；`metadata` 之后紧跟的 `1005` 仍转成 402 `hard_credit`。
+
+### 说明
+
+- 诊断细节与完整证据链见 `docs/ISSUE67_DIAGNOSIS.md`（不随 npm 包发布）。
+- 上游若开始分配缓存池，`cached_tokens` 会自动恢复显示，插件侧无需再改。
+
 ## 4.6.3 (2026-10-27)
 
 > **国际版美国区账号不会再被当成新加坡区。** 它的凭据明明写着

package/lib/index.d.ts

@@ -3426,6 +3426,29 @@ type TraeStreamEvent = {
    */
   cacheReadTokens?: number;
   cacheWriteTokens?: number;
+} | {
+  /**
+   * The upstream's own account of how the request was served: `metadata`
+   * (acceptance, plus `prompt_cache_pool_record`) and `timing_cost` (the
+   * pipeline and the backing provider model).
+   *
+   * These events were previously folded into `unknown` and dropped. That
+   * mattered for issue #67: `prompt_cache_pool_record` is the ONLY upstream
+   * signal that says whether a prompt-cache pool was even attached to the
+   * request, so with it dropped a "0% cache" report could not be told apart
+   * from a plugin that lost the numbers. See docs/ISSUE67_DIAGNOSIS.md.
+   */
+  type: 'diagnostic';
+  /**
+   * `metadata.prompt_cache_pool_record`. `null` is a real, observed value
+   * (100% of measured requests on 2026-10-28) and means the upstream attached
+   * no cache pool; it is kept distinct from "the field was absent".
+   */
+  promptCachePoolRecord?: unknown;
+  /** `timing_cost.name`, e.g. `llm_raw_chat_v2`. */
+  pipeline?: string;
+  /** `timing_cost.provider_model_name` — the model that actually answered. */
+  providerModelName?: string;
 } | {
   type: 'done';
   finishReason: string;
@@ -3569,7 +3592,21 @@ declare class TraeSoloBridge implements TraeUpstreamClient {
   private readonly upstream;
   private readonly catalog?;
   private readonly wireResolver?;
-  constructor(upstream: TraeUpstreamClient, catalog?: Pick<TraeCatalog, "current"> | undefined, wireResolver?: TraeWireResolver | undefined);
+  /**
+   * Optional sink for the upstream's own diagnostics (`metadata` /
+   * `timing_cost`). Issue #67: without it, `prompt_cache_pool_record: null`
+   * was invisible, so "the upstream attached no cache pool" and "the plugin
+   * lost the numbers" looked identical to a user reading a 0% cache rate.
+   */
+  private readonly log?;
+  constructor(upstream: TraeUpstreamClient, catalog?: Pick<TraeCatalog, "current"> | undefined, wireResolver?: TraeWireResolver | undefined,
+  /**
+   * Optional sink for the upstream's own diagnostics (`metadata` /
+   * `timing_cost`). Issue #67: without it, `prompt_cache_pool_record: null`
+   * was invisible, so "the upstream attached no cache pool" and "the plugin
+   * lost the numbers" looked identical to a user reading a 0% cache rate.
+   */
+  log?: ((message: string, detail?: unknown) => void) | undefined);
   chatStream(bodyJson: string, signal?: AbortSignal): Promise<TraeChatResult>;
 }
 //#endregion

package/lib/index.js

@@ -866,8 +866,30 @@ const REQUEST_IMAGE_BUDGETS = {
 * window of its own. Every fallback row and every live Trae row is expected to
 * state its real window; this is the backstop that keeps one unsized row from
 * failing the entire provider route (`INVALID_MODEL_CONTEXT`).
+*
+* APPLIED IN {@link toPiModel}, NOT ONLY ON THE PROFILE. The profile's
+* `defaultContextWindow` is read by the kernel's catalog-resolution step, which
+* this plugin never runs: it builds its pi-ai provider BY HAND through
+* `createProvider({ models })`, and `Models.getModel()` then returns the raw
+* `toPiModel` row with no defaulting applied anywhere. Leaving the fallback on
+* the profile alone is what produced
+* `adapter returned invalid context metadata for provider "trae" model
+* "doubao_1.8"` — a row the live directory published without a window reached
+* DSH unsized and took the whole `trae` route down.
 */
 const FALLBACK_CONTEXT_WINDOW = 2e5;
+/**
+* The window a served row is actually advertised with.
+*
+* Positive integers pass through untouched — a row's own window is never
+* second-guessed. Anything else (missing, `NaN`, `0`, a negative or a
+* fractional count from a malformed directory entry) falls back, because DSH
+* accepts ONLY a positive integer and rejects the entire provider otherwise.
+*/
+function servedContextWindow(info) {
+	const window = info.contextWindow;
+	return Number.isInteger(window) && window > 0 ? window : FALLBACK_CONTEXT_WINDOW;
+}
 function toPiModel(info, baseUrl, providerId, withRate = true) {
 	return {
 		id: info.id,
@@ -906,7 +928,7 @@ thinkingLevelMap: {
 			xhigh: info.reasoningEfforts.xhigh ?? null,
 			max: info.reasoningEfforts.max ?? null
 		} },
-		...info.contextWindow === void 0 ? {} : { contextWindow: info.contextWindow },
+		contextWindow: servedContextWindow(info),
 		...info.maxTokens === void 0 ? {} : { maxTokens: info.maxTokens },
 		compat: { supportsReasoningEffort: info.reasoningEfforts !== void 0 }
 	};
@@ -5581,6 +5603,12 @@ function decodeTraeEvent(event) {
 		...typeof record["cache_read_input_tokens"] === "number" ? { cacheReadTokens: record["cache_read_input_tokens"] } : {},
 		...typeof record["cache_creation_input_tokens"] === "number" ? { cacheWriteTokens: record["cache_creation_input_tokens"] } : {}
 	};
+	if (event.event === "metadata" || event.event === "timing_cost") return {
+		type: "diagnostic",
+		...record["prompt_cache_pool_record"] === void 0 ? {} : { promptCachePoolRecord: record["prompt_cache_pool_record"] },
+		...typeof record["name"] === "string" ? { pipeline: record["name"] } : {},
+		...typeof record["provider_model_name"] === "string" ? { providerModelName: record["provider_model_name"] } : {}
+	};
 	if (event.event === "done" || typeof record["finish_reason"] === "string" && record["response"] === void 0) return {
 		type: "done",
 		finishReason: typeof record["finish_reason"] === "string" ? record["finish_reason"] : "stop"
@@ -5747,7 +5775,7 @@ function bridgeTraeSoloStream(response, model) {
 	return bridgeTraeSource(source, model);
 }
 /** {@link bridgeTraeSoloStream} over an already-opened byte stream. */
-function bridgeTraeSource(source, model) {
+function bridgeTraeSource(source, model, log) {
 	const id = `chatcmpl-${randomUUID().replaceAll("-", "").slice(0, 24)}`;
 	const created = Math.floor(Date.now() / 1e3);
 	const decoder = new TextDecoder();
@@ -5780,6 +5808,15 @@ function bridgeTraeSource(source, model) {
 					if (failure !== void 0) upstreamError = new Error(describeTraeFailure(failure));
 					return;
 				}
+				if (decoded.type === "diagnostic") {
+					if (decoded.promptCachePoolRecord !== void 0 || decoded.pipeline !== void 0) log?.("dsh-connect-trae: upstream request diagnostics", {
+						model,
+						promptCachePoolRecord: decoded.promptCachePoolRecord,
+						pipeline: decoded.pipeline,
+						providerModelName: decoded.providerModelName
+					});
+					return;
+				}
 				if (decoded.type === "delta") {
 					const delta = {};
 					if (decoded.text !== "") delta["content"] = decoded.text;
@@ -5932,7 +5969,7 @@ async function openTraeSource(response) {
 						message: describeTraeFailure(failure)
 					};
 				}
-				if (decoded.type !== "unknown") answered = true;
+				if (decoded.type !== "unknown" && decoded.type !== "diagnostic") answered = true;
 			}
 			if (answered) break;
 		}
@@ -5956,12 +5993,12 @@ async function openTraeSource(response) {
 * The sibling {@link bridgeTraeSoloStream} stays synchronous and always answers
 * 200; it is the primitive, and this is the policy the bridge class uses.
 */
-async function bridgeTraeSoloStreamOrFail(response, model) {
+async function bridgeTraeSoloStreamOrFail(response, model, log) {
 	const opened = await openTraeSource(response);
 	if (!opened.ok) return opened;
 	return {
 		ok: true,
-		response: bridgeTraeSource(opened.source, model)
+		response: bridgeTraeSource(opened.source, model, log)
 	};
 }
 /**
@@ -6002,10 +6039,12 @@ var TraeSoloBridge = class {
 	upstream;
 	catalog;
 	wireResolver;
-	constructor(upstream, catalog, wireResolver) {
+	log;
+	constructor(upstream, catalog, wireResolver, log) {
 		this.upstream = upstream;
 		this.catalog = catalog;
 		this.wireResolver = wireResolver;
+		this.log = log;
 	}
 	async chatStream(bodyJson, signal) {
 		let model = "glm-5.2";
@@ -6093,7 +6132,7 @@ var TraeSoloBridge = class {
 		}
 		const result = await this.upstream.chatStream(prepared, signal);
 		if (!result.ok) return result;
-		return await bridgeTraeSoloStreamOrFail(result.response, model);
+		return await bridgeTraeSoloStreamOrFail(result.response, model, this.log);
 	}
 };
 //#endregion
@@ -11189,7 +11228,7 @@ function apply(ctx, config) {
 				route,
 				log: (message, detail) => ctx.logger.warn(message, detail)
 			});
-			const delegating = new TraeDelegatingUpstreamClient(new TraeSoloBridge(client, catalog, wireResolver));
+			const delegating = new TraeDelegatingUpstreamClient(new TraeSoloBridge(client, catalog, wireResolver, (message, detail) => ctx.logger.warn(message, detail)));
 			routeClients[provider] = {
 				provider,
 				route,
@@ -11230,7 +11269,7 @@ function apply(ctx, config) {
 							runtime: resolved.runtime
 						}
 					}),
-					solo: new TraeSoloBridge(primary.client, catalog, wireResolver),
+					solo: new TraeSoloBridge(primary.client, catalog, wireResolver, (message, detail) => ctx.logger.warn(message, detail)),
 					endpoint: `${REGION_GATEWAYS.cn.chat}/api/ide/v2/llm_raw_chat`,
 					edition: resolved.identity.edition,
 					identity: {
@@ -11627,7 +11666,7 @@ function apply(ctx, config) {
 				identity: () => stack.identity(),
 				route: stack.routeOf(providerOfPanel(panel)),
 				log: (message, detail) => ctx.logger.warn(message, detail)
-			}), stack.catalog, stack.wireResolver)
+			}), stack.catalog, stack.wireResolver, (message, detail) => ctx.logger.warn(message, detail))
 		};
 	}
 	/**
@@ -11709,7 +11748,7 @@ function apply(ctx, config) {
 				identity: () => stack.identity(),
 				route: stack.routeOf(providerOfPanel(panel)),
 				log: (message, detail) => ctx.logger.warn(message, detail)
-			}), stack.catalog, stack.wireResolver);
+			}), stack.catalog, stack.wireResolver, (message, detail) => ctx.logger.warn(message, detail));
 			const chat = (body, signal) => bridge.chatStream(body, signal);
 			bridges.set(accountId, chat);
 			return chat;

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": "4.6.3",
+  "version": "4.6.5",
   "author": "dingminhua",
   "copyright": "Copyright (c) 2026 LaoDing",
   "license": "MIT",