@mokuyoaxis/dsh-iris 0.3.0-alpha.1 → 0.3.0-beta.1

package/docs/VIDEO_FLOWS.md

@@ -0,0 +1,89 @@
+# 文生视频 · 持久流程与资源交付
+
+这是 **0.3.0-beta.1 新增功能**。使用准确beta版本或npm `beta`通道；旧alpha.1没有下面的 `--video`、视频 Web 表单与 MCP 分组。原有 CLI / DSH 视频入口继续可用。
+
+协议按实际模型选择。`wan2.2-t2v-plus`等旧系列应对照官方[万相2.1–2.6文生视频接口](https://help.aliyun.com/zh/model-studio/legacy-wan-text-to-video-api-reference)；官方当前主视频页已转向wan2.7新协议，不能仅凭同品牌模型套用其字段或声称已有实现覆盖。
+
+## 新建专用工作区
+
+以下路径必须是新的空 Flow 根：
+
+```bash
+node bin/dsh-iris.js flow init --data-root /absolute/new-video-core
+node bin/dsh-iris.js execution init --video --data-root /absolute/new-video-core
+node bin/dsh-iris.js web serve --data-root /absolute/new-video-core \
+  --provider-config /absolute/private/providers.json --allow-execution \
+  --auth-store /absolute/private/video-auth \
+  --connection-store /absolute/private/video-connections \
+  --allow-connection-management --allow-mcp
+```
+
+`--video` 显式建立 `creation-requests/v4`，同时包含原图片、语音与转写流程。不能与 `--audio` 或 `--image-generation` 混用；已有 v0–v3 根不升级、迁移或降级。旧版本在请求区门禁拒绝 v4；不要让旧版本 CLI 写入此根。
+
+初始化、读取目录、创建连接与保存视频意图均不生成。只有明确的首次继续调用模型，可能产生费用。模型池仍来自真实配置和既有用途规则；此固定流程要求用户选择一个实际 `modelRef`，不会跨请求轮询或悄悄切换模型。
+
+beta.1修复统一检查明确的t2v/i2v/s2v/r2v模型家族：文生与图生候选不互选，固定文生视频Flow拒绝已知i2v/s2v/r2v模型。CLI/DSH和显式视频重试使用相同规则。没有这些家族标记的兼容模型ID延续原有t2v/i2v路径，实际支持仍由协议和上游决定；本片不新增模型元数据格式或迁移目录。
+
+独立 CLI `flow create/continue` 已补齐视频候选装配（beta.1修复），无需启动 Web：
+
+```bash
+node bin/dsh-iris.js flow create --data-root /absolute/new-video-core \
+  --provider-config /absolute/private/providers.json \
+  --input '{"recipe":"video-generate/v0","model_ref":"provider::t2v-model","prompt":"鸢尾花随风轻摆","duration":5}'
+node bin/dsh-iris.js flow continue run_0123456789abcdef01234567 \
+  --data-root /absolute/new-video-core --provider-config /absolute/private/providers.json
+```
+
+Run ID 使用创建命令的实际返回值。首次继续仅提交一次；异步未结束时再次明确继续只观察原 Task，不重提。该固定流程仍仅文生视频；独立 `task retry` 的模式和媒体输入规则见[CLI 重试](HEADLESS_CLI.md#检查-task-与-artifact)，不会因此扩展本流程范围。
+
+## 输入、任务和交付边界
+
+- 固定配方为 `video-generate/v0`：明确 `modelRef`、`prompt`，可选 `size` 和数值型 `duration`（1–60 秒）。Prompt 最多20,000字符、尺寸最多64字符；具体参数支持由实际模型决定，目录不等于实测。
+- 首片仅文生视频；不接受首帧、音频、S2V、任意供应商 JSON、本机路径或外部结果 URL。原 CLI / DSH 的图生视频和 S2V 入口不受影响。
+- 创建预留唯一 Run/Task，并持久保存创建 UUID 和输入摘要；不立即提交。继续保存独立 Operation UUID，一次有限调用只提交或观察原 Task，受理未知不重新提交。
+- 服务重启、刷新、历史读取不自动生成或查询供应商。异步任务需明确点击“查看原任务”再观察一次；停止等待只终止本地操作，不代表远端取消、退款或未计费。
+- 原 Task 已成功但落盘失败时接回原交付；已有完整缓存可在无 Provider 的重启后修复。保留原 Run/Task/Operation/Artifact 身份。
+- 视频为单一结果，无图片选择步骤。交付和读取限 MP4/WebM、最多 **32 MiB**；API 验证保存记录、大小及 SHA-256。没有完整编码探测、Range 或边下边播，客户端收到完整 Blob 后播放；真实模型的编码兼容性仍需验收。
+- 类型或大小不符时保留原 Task 与已有交付证据，明确报错，不重新生成。工作台不输出供应商签名 URL、密钥或服务器存储路径。
+
+beta.1修复会先检查 Manifest、媒体类型及实际普通文件大小，再读取固定流程的结果。超出32 MiB时保留已成功的原 Task/Artifact；重启后再次继续仍明确拒绝，不整读该文件或重新生成。这个上限属于固定 Flow/Core Web/MCP 视频读取入口；通用 Provider Task 的视频下载上限仍为1 GiB。DSH同源作品媒体GET按流发送、HEAD不分配完整文件缓冲，两者在发送前保留分块SHA-256校验；没有新增Range或编码探测。
+
+Web 的 **创作 → 生成视频** 使用实际候选，参数默认折叠；结果提供原视频播放器和下载。**作品 → 视频** 点击“载入视频”才读取，不自动播放；离页、换分类和退出释放媒体。撤销执行权限的只读服务仍可按读取授权播放已有结果。浏览器恢复表只存身份，不存 Prompt 或凭据；丢失创建回包后接回原意图，不自动首次提交。
+
+## HTTP 与 MCP
+
+HTTP 路径均以 `/api/v0` 开头，沿用 Bearer、Host/Origin、连接功能和具体工具授权：
+
+| 入口 | 含义 |
+|---|---|
+| GET/HEAD `/video-models` | 真实池内文生视频候选；不探测，无密钥/端点 |
+| POST `/videos` | JSON 意图，稳定 UUIDv4 `Idempotency-Key`；同 Key 同输入重放，不同输入409 |
+| GET/HEAD `/artifacts/:id/video` | 授权后的原 MP4/WebM |
+| GET/HEAD `/runs/:id/download` | 单一结果原字节及匹配扩展名 |
+| Run / Operation 的既有入口 | 创建查询、有限继续、原操作查询和本地停止 |
+
+JSON 正文最多64 KiB。v0–v3 根不开放此媒体组，旧连接和默认 image-edit 不自动扩权。MCP 由用户明确选择 `video-generate`，再分别选择 read/create/execute/stop 和需要的具体工具：
+
+| 工具 | 行为 |
+|---|---|
+| `iris_core_video_models` | 读取实际候选 |
+| `iris_core_create_video` | 保存意图，尚不生成 |
+| `iris_core_video` | 校验并引用已保存的原视频，不调用模型 |
+
+继续、查询和停止复用原工具，并按原 Run 媒体组校验。MCP 无原生 video 内容块，视频工具返回标准 `resource_link`，URI 为 `iris://video/ARTIFACT_ID`。客户端通过标准 `resources/read` 获取 Base64 BlobResourceContents；每次读取重新检查 Core 当前授权。仅选择视频读取工具才注册资源模板，缩权或撤销后旧进程也无法继续读取。连接专属 Skill 由实际工具生成视频指导。
+
+客户端可能只显示引用，不提供播放器；拿到引用不等于看过视频。DSH 原生共享桥暂不投影视频资源，选择视频组时明确拒绝启动；使用通用 MCP、Core Web，或原 DSH 视频工具。
+
+## 验证范围
+
+以下测试和`scripts/`开发验收命令从源码仓库根运行，不随npm包分发。定向用例：`tests/video-flow.mjs`、`tests/video-api.mjs`、`tests/video-recovery.mjs`。真实 Chromium 场景：
+
+```bash
+node scripts/verify-local-web-browser.mjs \
+  --browser-executable /absolute/path/to/chromium \
+  --output /absolute/new-video-evidence --video
+```
+
+覆盖实际小视频解码/播放、原字节下载、重启接回、丢失创建回包、停止、只读、窄屏和暗色；标准 SDK 的 HTTP/stdio 2025/2026 路径验证资源与权限。这些回归使用受控 Provider。
+
+另于2026-11-01完成百炼北京 `wan2.2-t2v-plus` 的真实单例：832×480、5秒H.264/MP4，1次生成提交、2次原任务查询；隔离服务重启接回同一Task、正式MCP资源读取、真实Chromium播放/下载与SHA一致均通过。未修改日常连接或查询实际账单。该结果仅覆盖这个模型、参数和当时环境，**不代表所有上游、账号、编码或客户端播放器均已实测**。

package/docs/VISION_MODEL.md

@@ -48,4 +48,6 @@ DSH rc.2 会压缩宿主附件。看图、OCR 和定位允许这种宿主规范
 
 [长图 OCR](OCR_MODEL.md)、[定位、拼图摘要与自述](COMPOSITE_VISION.md) 均已消费共享端口与整次预算；各自另验分块、bbox 解析、一张拼图一次生成及可选 Core 转写。legacy `vision.js` / Host `visionModel.analyze()` 兼容实现仍保留，现有业务入口不再调用。Model Port 的迁移切片本身不新增公开 SDK export 或生产依赖；0.2.1 提供 [视觉 CLI](VISION_CLI.md) 和可选 `visionInput`。复杂图像质量与 Android 浏览器仍需独立验证。
 
+beta.1另修复旧 `adapters.visionStream` 兼容入口：累计SSE正文最多2 MiB，声明长度与实际字节均检查；外部取消及超时在发送前、收到响应后、逐块及结束时保留失败，已收到文字也不返回部分成功。主业务仍使用上文具有严格终态校验的Model Port，不把旧入口的宽松解析等同于该契约。
+
 契约与迁移说明见 [Model Port](MODEL_PORT_CONTRACT.md) 和 [DSH → Core](DSH_CORE_MIGRATION.md)。

package/docs/WEB_PROTOTYPE.md

@@ -2,7 +2,7 @@
 
 这是面向 **0.3.0** 的交互预览，包含创作、作品和连接草稿三个页面。首轮布局已围绕大图预览、紧凑参数面板和默认折叠的详情整理，仍用于验证操作方向，尚未定稿。
 
-Demo 可在源码树中独立预览，无需 DSH；使用内存演示与内置样图，不连接模型、不计费，也不写入 Core 数据。0.3 alpha源码的真实工作台已接入固定图片编辑及[独立文生图](IMAGE_GENERATION.md)；npm 0.2.7 不含该页面，也不携带 prototypes/ 和预览脚本。
+Demo 可在源码树中独立预览，无需 DSH；使用内存演示与内置样图，不连接模型、不计费，也不写入 Core 数据。0.3 beta的真实工作台已接入固定图片编辑及[独立文生图](IMAGE_GENERATION.md)；npm 0.2.7 不含该页面，也不携带 prototypes/ 和预览脚本。
 
 ## 启动预览
 
@@ -46,9 +46,9 @@ node scripts/preview-workbench.mjs
 
 原型位于 `prototypes/workbench/`，使用原生 HTML/CSS/JavaScript，复用 Iris 图标和示例图片，没有新框架或生产依赖。运行详情和工具预览默认折叠，关闭显示“展开”、打开显示“收起”；提供浅/深色、键盘焦点与 320/390px 布局。
 
-0.2.7 已有 [默认只读 API](LOCAL_API.md)及显式开启的 [本地上传/Run 创建](LOCAL_CREATION.md)，带持久请求意图和身份接回；0.3 alpha源码另有 [Operation 与有限执行服务](CORE_SERVICE.md)、[版本约束选择](LOCAL_SELECTION.md)及真实创作页面。此 Demo 仍独立使用内存样例，其布局、提示词和样例效果用于说明方向，不代表模型能力；真实页面的恢复、故障与发布门禁继续独立验收。
+0.2.7 已有 [默认只读 API](LOCAL_API.md)及显式开启的 [本地上传/Run 创建](LOCAL_CREATION.md)，带持久请求意图和身份接回；0.3 beta另有 [Operation 与有限执行服务](CORE_SERVICE.md)、[版本约束选择](LOCAL_SELECTION.md)及真实创作页面。此 Demo 仍独立使用内存样例，其布局、提示词和样例效果用于说明方向，不代表模型能力；真实页面的恢复、故障与发布门禁继续独立验收。
 
-[本地 API 契约](LOCAL_API_CONTRACT.md)区分已发布的只读/本地创建，以及未发布的有限执行、本地停止和 owner 持久选择。相同 Run 的 Task 复用、HTTP 创建幂等与选择的版本门禁分别验证；替换演示传输时仍须单独验证浏览器写入、原身份恢复与无 DSH 的完整流程。
+[本地 API 契约](LOCAL_API_CONTRACT.md)区分已发布的只读/本地创建，以及0.3提供的有限执行、本地停止和 owner 持久选择。相同 Run 的 Task 复用、HTTP 创建幂等与选择的版本门禁分别验证；替换演示传输时仍须单独验证浏览器写入、原身份恢复与无 DSH 的完整流程。
 
 源码树可运行下面的检查；浏览器检查需显式提供已经安装的 Chromium，不下载或安装浏览器。无沙箱仅在受控测试环境确有需要时显式追加 `--no-sandbox`，不改变生产浏览器策略。
 

package/docs/WEB_RECOVERY.md

@@ -1,4 +1,4 @@
-# 工作台请求核对与刷新恢复（0.3 alpha）
+# 工作台请求核对与刷新恢复（0.3 beta）
 
 本页描述源码中的 W1b 恢复及 W1c 页面接入。npm 0.2.7 不含这些新增接口与浏览器客户端；[真实工作台](LOCAL_WEB.md)已接入明确创作、选择和本地停止，“操作核对”默认折叠，自动恢复始终只读。
 
@@ -15,7 +15,7 @@ GET /api/v0/creation-requests?kind=flow-create&key=<original-lowercase-UUIDv4>
 {contractVersion:0, kind, state:"prepared"|"succeeded", result}
 ```
 
-创建类型另包括 `generation-create`；新 v3 音频工作区增加 `audio-import`、`speech-create`、`transcription-create`，返回同样的有限身份，不返回音频或朗读正文。音频恢复与大小边界见[音频流程](AUDIO_FLOWS.md)。
+创建类型另包括 `generation-create`；新 v3 音频工作区增加 `audio-import`、`speech-create`、`transcription-create`，返回同样的有限身份，不返回音频或朗读正文。音频恢复与大小边界见[音频流程](AUDIO_FLOWS.md)。beta.1的 v4 根另有 `video-create`，同样只保存原 UUID/Run 身份，不存 Prompt；查询或找回已创建意图不自动首次生成，见[视频流程](VIDEO_FLOWS.md)。
 
 `result` 为 `{artifactId}`、`{runId}` 或 `null`。查询会核对原意图与已有结果；prepared 尚无目标时返回 null，已有目标核验通过后才返回引用。它不补写完成回执、不重建目标、不观察 Task 或调用 Provider。已推进或选择的 Run 不退回初始状态。没有原请求为 404；原结果不可用或证据损坏为 409，并保留原记录。
 

package/docs/WORKBENCH_ARTIFACTS.md

@@ -36,6 +36,12 @@ Core静态PNG/JPEG/WebP图片卡提供“改图”。填写修改指令，可选
 
 目录查询复用已提交的 Manifest 与派生 Index，只读取元数据和文件大小，不反复读取视频等媒体的完整字节、不重建索引或取得 Core writer 租约。打开媒体、看图/OCR、CLI inspect/list/export 保留原有完整内容哈希核验；目录里显示的哈希不表示本次浏览已经重新核验了全部文件内容。
 
+beta.1把可兼容读取的早期Core提交记录单列为 `legacyCoreSkipped`，不计入损坏/缺失的 `droppedCore`。这些作品仍可按原ID检查或导出；需要编入当前目录时，由用户显式运行 `artifact rebuild --data-root /absolute/core`，该写操作可能升级早期记录。浏览和刷新不自动迁移，也不为分类重新哈希媒体；结构、大小或索引信息不符仍按不可读取处理。
+
+DSH Core快照也复用这条元数据路径：兼容早期条目单列`legacySkipped`，坏条目单列`dropped`，正常作品继续显示。新作品提交对早期条目复用已校验索引中的摘要与大小，不读取历史媒体；缺少或不匹配的摘要要求显式重建，不补造内容证据。重建和库存刷新在写入前标记索引待刷新，失败后普通列表明确报错，成功发布新索引才清除标记；原作品记录和按ID读取保持。
+
+beta.1中，DSH作品媒体GET先分块核验内容，再流式发送；HEAD同样验证哈希但不读取完整文件到内存。客户端断开或实例关闭会排空并关闭媒体流，原作品不变；不增加Range支持。图库索引未完成刷新时会明确报错并由Doctor提示，不以缺项的旧列表冒充完整目录；保留现场后可显式执行 `artifact rebuild`。
+
 OCR 动作 JSON 附 `status/totalChunks/successfulChunks/failedChunks/skippedChunks`，便于界面表达完整度。识别正文不自动保存为 Core Artifact。
 
 ## 选择、详情与批量下载

package/docs/WORKBENCH_LOGIN.md

@@ -1,6 +1,6 @@
-# 工作台访问口令（0.3 alpha）
+# 工作台访问口令（0.3 beta）
 
-普通 `web serve` 使用一个访问口令，不要求账号、邮箱或读取 Token 文件。**稳定 npm 0.2.7 没有 `web serve`；下文针对0.3.0-alpha.1。** 源码树可把 `dsh-iris` 替换为 `node bin/dsh-iris.js`。
+普通 `web serve` 使用一个访问口令，不要求账号、邮箱或读取 Token 文件。**稳定 npm 0.2.7 没有 `web serve`；下文针对0.3.0-beta.1。** 源码树可把 `dsh-iris` 替换为 `node bin/dsh-iris.js`。
 
 ## 首次设置与日常使用
 

package/docs/releases/0.3.0-beta.1.md

@@ -0,0 +1,34 @@
+# Iris 0.3.0-beta.1: routing, scheduling and durable video
+
+The first 0.3 beta builds on alpha.1's standalone Core workbench, generic MCP, ComfyUI and explicit plugins. It freezes the [minimal plugin contracts and main user flows](https://github.com/mokuyoaxis/dsh-iris/blob/v0.3.0-beta.1/docs/BETA_SCOPE.md) within this beta scope. The broader SDK remains experimental.
+
+```bash
+npm install @mokuyoaxis/dsh-iris@0.3.0-beta.1
+```
+
+The npm `beta` channel selects this release. Stable `latest` remains **0.2.7**, and `alpha` retains **0.3.0-alpha.1**. Follow the [Core quickstart](https://github.com/mokuyoaxis/dsh-iris/blob/v0.3.0-beta.1/docs/CORE_QUICKSTART.md) for a new workspace.
+
+## Added since alpha.1
+
+- **Explicit network resources and routing.** Manage HTTP/HTTPS proxy nodes and fixed or ordered pools, bind a Provider, and inspect the route recorded for a task. An owner-only Web/CLI/API Doctor separates local configuration from an explicit bounded GET check. Unbound Providers inherit the actual process network environment; Iris does not configure Clash, control system TUN or infer geographic suitability.
+- **Opt-in bounded media scheduling.** One physical root still has one writer. Network waits can overlap while short writes remain serialized, with separate HTTP concurrency, dispatch-rate and unfinished-remote-task limits. The default remains serial. Unknown acceptance never triggers automatic resubmission.
+- **Durable text-to-video in Core Web and MCP.** A new v4 root adds actual model candidates, size/duration inputs, original Run/Task/Operation recovery, local stop and MP4/WebM playback/download up to 32 MiB. MCP exposes standard resource links and checks authorization on each read. Image-to-video, S2V, music and vision retain their existing CLI/DSH entries.
+- **Environment credential references.** Explicit `apiKeyEnv` configuration stores a variable name and resolves its value inside Core. Inline keys remain supported. Conflicting inline/env/DSH sources are rejected; workflow reports and model-directory imports protect captured credentials. No automatic `.env` loading or host-configuration scanning is added.
+- **Independent plugin starters.** The package includes a business/page/PNG-preview starter and a Wan asynchronous Provider starter. Provider factory context v1 lends Core's network port and resolved connection; Core owns scheduling, lifecycle, persistent tasks and storage. Plugin identities, exact dependencies and client grants remain explicit.
+- **Clearer MCP onboarding and plugin status.** The workbench connects selection/authorization, configuration export and connection checks. Generic stdio/Streamable HTTP remain available with explicit client presets. Configuration inspection, protocol handshake and actual Agent use are separate facts.
+
+## Reliability fixes
+
+Paid submissions do not automatically follow redirects or replay ambiguous requests. Queue rejection stays local to the affected Operation; accepted-task recovery preserves the original Attempt. Source-artifact permissions, multipart/Request interoperability, bounded MCP HTTP load, network-client retirement, media size checks and audio container metadata have been corrected.
+
+Artifact index failures remain visible and recoverable. Normal commits refresh metadata without hashing the entire media library; explicit rebuild and media reads retain content verification. DSH screenshot delivery confirms that the requested Iris HTML route actually served the page before accepting a PNG. Video model families, transcription retries and the browser-only DSH client type declaration have also been aligned.
+
+## Compatibility and validation
+
+Existing roots and grants are **not migrated or expanded automatically**. Video requires an explicitly initialized new v4 root; network bindings use a separate versioned configuration. Keep compatible readers and a complete backup when trying a newer format. DSH's existing exact host-version range remains unchanged. The Node.js minimum is **22.1.0**: isolated testing reproduced native fetch body-cancellation failures on 22.0.0, including without Iris, while 22.1.0 and current Node 22 pass. Package engines, Doctor and the minimum CI job use the same baseline.
+
+Release checks cover the Ubuntu/Windows and minimum/current Node 22 matrix, actual tarball consumers, strict NodeNext public types and Chromium workbench scenarios. Controlled Providers and local proxies are distinguished from real accounts. Selected Alibaba Cloud image/video/audio paths have separate real-account evidence; this does not certify every model, account, GPU, browser or harness. Live ComfyUI GPU inference, automatic quotas/GC, weighted model/proxy rotation and a plugin marketplace remain outside this beta.
+
+## 中文摘要
+
+beta.1 主要新增网络出口池与Doctor、有界并发调度、Core Web/MCP文生视频、环境凭据引用，以及随包分发的两类独立插件模板；同时完善MCP接入流程与插件状态，修复这轮审计确认的凭据、提交/恢复、媒体限额、索引和DSH集成问题。使用npm `beta`或准确版本安装，稳定版仍为0.2.7。最低Node为22.1.0；旧根不自动迁移，视频需显式新建v4根；最小契约收口不代表完整SDK或全部上游已验收。

package/docs/screenshots.md

@@ -2,9 +2,9 @@
 
 ## Core 工作台开发预览 / Core workbench development preview
 
-2026-10-29 重拍，使用当前未发布源码、独立演示数据根及 Chrome for Testing 153.0.8010.12（Linux ARM64）。全新浏览器 profile 只使用内置明亮/暗色主题；流程与作品是此前从零使用演练已完成记录的独立副本，没有读取用户背景、真实供应商凭据或日常浏览器会话。
+2026-10-29 重拍，使用alpha.1发布前的源码、独立演示数据根及 Chrome for Testing 153.0.8010.12（Linux ARM64）。全新浏览器 profile 只使用内置明亮/暗色主题；流程与作品是此前从零使用演练已完成记录的独立副本，没有读取用户背景、真实供应商凭据或日常浏览器会话。
 
-Recaptured on 2026-10-29 from the current unreleased source in Chrome for Testing 153.0.8010.12 on Linux ARM64. A fresh browser profile uses only the built-in light/dark themes. The separate demo data root contains copies of completed records from the earlier first-use walkthrough. No personal background, real provider credentials or daily browser session is used.
+Recaptured on 2026-10-29 from the pre-alpha.1 source in Chrome for Testing 153.0.8010.12 on Linux ARM64. A fresh browser profile uses only the built-in light/dark themes. The separate demo data root contains copies of completed records from the earlier first-use walkthrough. No personal background, real provider credentials or daily browser session is used.
 
 | 创作与下载 / Studio and download | 完整六类总览 / All six capabilities |
 |---|---|

package/lib/actions.js

@@ -30,7 +30,7 @@ import { formatOcrResult } from './ocr.js';
 import { runOcrRequest } from './ocr-model-routing.js';
 import { ffmpegAvailable, extractFrames, probeVideo } from './media-probe.js';
 import { ensurePrivateDir, writePrivateFile } from './private-storage.js';
-import { selectMediaProtocol, selectImageProtocol, providerForImageModel, supportsImageEdit, supportsModelCapability, dashscopeTtsProtocolIssue, dashscopeTtsInputIssue } from './provider-protocol.js';
+import { selectMediaProtocol, selectImageProtocol, providerForImageModel, providerApiKey, supportsImageEdit, supportsModelCapability, dashscopeTtsProtocolIssue, dashscopeTtsInputIssue } from './provider-protocol.js';
 import { musicCandidatesFromCatalog, providerTaskBinding } from './provider-catalog.js';
 import { invokeProviderOperation, providerPollTaskView, prepareProviderInput } from './provider-adapter.js';
 import { createConfiguredProviderAdapter as createProviderAdapter } from './provider-adapters.js';
@@ -43,6 +43,7 @@ import { ModelPortError } from './model-port-contract.js';
 import { compileComfyWorkflow, projectComfyWorkflow } from './comfyui-workflow.js';
 import { createComfyUiClient } from './comfyui-client.js';
 import { inspectComfyWorkflow } from './comfyui-inspection.js';
+import { createProviderSecretScrubber } from './provider-credentials.js';
 
 const PROVIDER_ACTIONS = new Set(['image', 'image_edit', 'video', 'music', 'tts', 'transcribe', 'look', 'relook',
   'locate', 'ocr', 'media_summarize', 'task_reobserve', 'task_redeliver', 'task_manual_retry',
@@ -1104,7 +1105,7 @@ register('providers_list', async (host, args) => {
       ...selectMediaProtocol(p),
       visionInput: p.visionInput || null,
       directory: modelDirectorySummary(p),
-      apiKeyHint: p.dshProvider ? '由 DSH 管理' : String(p.apiKey || '').length <= 8 ? '****' : String(p.apiKey).slice(0, 3) + '****' + String(p.apiKey).slice(-4),
+      apiKeyHint: p.dshProvider ? '由 DSH 管理' : p.apiKeyEnv ? '由环境变量提供' : String(p.apiKey || '').length <= 8 ? '****' : String(p.apiKey).slice(0, 3) + '****' + String(p.apiKey).slice(-4),
       models: models.providerModelsWithDirectory(p).map((m) => ({
         id: m.id,
         capabilities: m.capabilities,
@@ -1154,11 +1155,11 @@ register('providers_check_comfy_workflow', async (_host, args, { signal }) => {
   const file = path.join(store.irisHome(), 'providers.json'), before = fs.readFileSync(file);
   const workflow = compileComfyWorkflow(args.workflow !== undefined ? args.workflow : route.comfyWorkflow,
     { workflow_parameters: args.workflow_parameters });
+  const scrub = createProviderSecretScrubber([route]);
   const report = await inspectComfyWorkflow(createComfyUiClient(route), workflow, { signal });
   signal?.throwIfAborted();
   if (!fs.readFileSync(file).equals(before)) throw Object.assign(new Error('检查期间配置已变化，请重新检查'), { code: 'IRIS_CONFIG_CHANGED' });
-  const key = route.apiKey;
-  return { ok: true, report: key ? JSON.parse(JSON.stringify(report).split(key).join('[redacted]')) : report };
+  return { ok: true, report: scrub(report) };
 });
 
 /** 新增/更新供应商 */

package/lib/adapters.js

@@ -8,6 +8,7 @@
  * 所有函数只收显式 key；错误统一 throw Error(人话)。
  */
 import fs from 'node:fs';
+import { inheritedProviderFetch } from './provider-network.js';
 import path from 'node:path';
 import { Readable } from 'node:stream';
 import { pipeline } from 'node:stream/promises';
@@ -146,10 +147,10 @@ function transportFetchError(err, response) {
 }
 
 /** 只包装网络读取；HTTP 业务错误和图片协议校验由调用方处理。 */
-async function fetchImageJson(url, options) {
+async function fetchImageJson(url, options, fetchPort) {
   let response;
   try {
-    response = await fetch(url, options);
+    response = await fetchPort(url, options);
     return { res: response, body: await readJson(response, response.ok ? PROVIDER_RESPONSE_LIMITS.mediaJson : PROVIDER_RESPONSE_LIMITS.json) };
   } catch (err) {
     if (err instanceof ProviderContractError) throw err;
@@ -191,11 +192,12 @@ export function applyImageSize({ protocol, mode, size }) {
 }
 
 /** 新版多模态同步协议：响应中直接返回图片 URL。 */
-export async function generateImageMultimodal({ key, baseUrl, model, prompt, image, size, n = 1, timeoutMs = T_SYNC_GEN, signal }) {
+export async function generateImageMultimodal({ key, baseUrl, model, prompt, image, size, n = 1, timeoutMs = T_SYNC_GEN, signal, fetchPort = inheritedProviderFetch }) {
   const apiBase = dashscopeApiBase(baseUrl);
   const requestSize = applyImageSize({ protocol: 'dashscope', mode: 'multimodal-sync', size });
   const { res, body } = await fetchImageJson(`${apiBase}/services/aigc/multimodal-generation/generation`, {
     method: 'POST',
+    redirect: 'error',
     headers: authHeaders(key),
     signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({
@@ -206,7 +208,7 @@ export async function generateImageMultimodal({ key, baseUrl, model, prompt, ima
       ] }] },
       parameters: { ...(requestSize ? { size: requestSize } : {}), n: Number(n) || 1 }
     })
-  });
+  }, fetchPort);
   if (!res.ok) throw providerError(res.status, body, '生成失败', res.headers);
   const urls = [];
   for (const choice of (body && body.output && body.output.choices) || []) {
@@ -226,15 +228,16 @@ export async function startImageGeneration(opts) {
   }
   return { remoteTaskId: await submitImage(opts), urls: [] };
 }
-export async function submitImage({ key, baseUrl, model, prompt, size, n = 1, timeoutMs = T_SUBMIT, signal }) {
+export async function submitImage({ key, baseUrl, model, prompt, size, n = 1, timeoutMs = T_SUBMIT, signal, fetchPort = inheritedProviderFetch }) {
   const apiBase = dashscopeApiBase(baseUrl);
   const requestSize = applyImageSize({ protocol: 'dashscope', mode: 'legacy-async', size });
   const { res, body } = await fetchImageJson(`${apiBase}/services/aigc/text2image/image-synthesis`, {
     method: 'POST',
+    redirect: 'error',
     headers: { ...authHeaders(key), 'X-DashScope-Async': 'enable' },
     signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({ model, input: { prompt }, parameters: { ...(requestSize ? { size: requestSize } : {}), n: Number(n) || 1 } })
-  });
+  }, fetchPort);
   if (!res.ok) throw providerError(res.status, body, '提交失败', res.headers);
   const taskId = body && body.output && body.output.task_id;
   if (!taskId) throw new Error('提交成功但未返回 task_id');
@@ -250,7 +253,7 @@ export async function submitImage({ key, baseUrl, model, prompt, size, n = 1, ti
  * - 其他（t2v/i2v）：POST /services/aigc/video-generation/video-synthesis，
  *   imgDataUrl 支持公网 URL 或 data:image base64。
  */
-export async function submitVideo({ key, baseUrl, model, prompt, imgDataUrl, size = '1280*720', duration, audioUrl, resolution, timeoutMs = T_SUBMIT, signal }) {
+export async function submitVideo({ key, baseUrl, model, prompt, imgDataUrl, size = '1280*720', duration, audioUrl, resolution, timeoutMs = T_SUBMIT, signal, fetchPort = inheritedProviderFetch }) {
   const apiBase = dashscopeApiBase(baseUrl);
   const isS2V = /s2v/i.test(model || '');
   let url, input, parameters;
@@ -267,11 +270,12 @@ export async function submitVideo({ key, baseUrl, model, prompt, imgDataUrl, siz
     url = `${apiBase}/services/aigc/video-generation/video-synthesis`;
     input = { prompt };
     if (imgDataUrl) input.img_url = imgDataUrl;
-    parameters = { size };
+    parameters = { size: typeof size === 'string' ? size.replace(/^(\d+)[xX](\d+)$/, '$1*$2') : size };
     if (duration) parameters.duration = Number(duration);
   }
-  const res = await fetch(url, {
+  const res = await fetchPort(url, {
     method: 'POST',
+    redirect: 'error',
     headers,
     signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({ model, input, parameters })
@@ -289,7 +293,7 @@ export async function submitVideo({ key, baseUrl, model, prompt, imgDataUrl, siz
  * 返回 oss:// 形式 URL；调用模型时请求头须带 X-DashScope-OssResourceResolve: enable
  * （submitVideo 的 s2v 分支已自动处理）。
  */
-export async function uploadTempFile({ key, baseUrl, model, filePath, bytes, filename, timeoutMs = T_UPLOAD, signal }) {
+export async function uploadTempFile({ key, baseUrl, model, filePath, bytes, filename, timeoutMs = T_UPLOAD, signal, fetchPort = inheritedProviderFetch }) {
   if ((filePath === undefined) === (bytes === undefined)) throw new TypeError('上传须且只能提供本地路径或明确字节');
   const name = filePath !== undefined ? path.basename(filePath) : filename;
   if (typeof name !== 'string' || !name || name.length > 128 || /[\/\\\0]/.test(name)
@@ -298,7 +302,7 @@ export async function uploadTempFile({ key, baseUrl, model, filePath, bytes, fil
   const u = new URL(`${apiBase}/uploads`);
   u.searchParams.set('action', 'getPolicy');
   u.searchParams.set('model', model);
-  const pres = await fetch(u, { headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal) });
+  const pres = await fetchPort(u, { headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal) });
   const pbody = await readJson(pres);
   if (!pres.ok || !pbody || !pbody.data) throw providerError(pres.status, pbody, '获取上传凭证失败', pres.headers);
   const pol = pbody.data;
@@ -312,7 +316,8 @@ export async function uploadTempFile({ key, baseUrl, model, filePath, bytes, fil
   fd.append('key', keyPath);
   fd.append('success_action_status', '200');
   fd.append('file', new Blob([bytes ?? fs.readFileSync(filePath)]), name);
-  const up = await fetch(pol.upload_host, { method: 'POST', body: fd, signal: timeoutSignal(timeoutMs, signal) });
+  const up = await fetchPort(pol.upload_host, { method: 'POST', redirect: 'error', body: fd, signal: timeoutSignal(timeoutMs, signal) });
+  await up.body?.cancel();
   if (up.status !== 200) throw new Error(`临时上传失败：HTTP ${up.status}`);
   return 'oss://' + keyPath;
 }
@@ -320,9 +325,9 @@ export async function uploadTempFile({ key, baseUrl, model, filePath, bytes, fil
 const TERMINAL = new Set(['SUCCEEDED', 'FAILED', 'CANCELED', 'UNKNOWN']);
 
 /** @returns {{done:boolean, ok:boolean, urls:string[], message?:string}} */
-export async function pollTask({ key, baseUrl, remoteTaskId, timeoutMs = T_POLL, signal }) {
+export async function pollTask({ key, baseUrl, remoteTaskId, timeoutMs = T_POLL, signal, fetchPort = inheritedProviderFetch }) {
   const apiBase = dashscopeApiBase(baseUrl);
-  const res = await fetch(`${apiBase}/tasks/${encodeURIComponent(remoteTaskId)}`, {
+  const res = await fetchPort(`${apiBase}/tasks/${encodeURIComponent(remoteTaskId)}`, {
     headers: authHeaders(key),
     signal: timeoutSignal(timeoutMs, signal)
   });
@@ -363,10 +368,11 @@ export async function pollTask({ key, baseUrl, remoteTaskId, timeoutMs = T_POLL,
 }
 
 /* ---------------- DashScope：qwen-tts（同步） ---------------- */
-export async function synthesizeTts({ key, baseUrl, model = 'qwen-tts-latest', text, voice = 'Cherry', timeoutMs = T_SYNC_GEN, signal }) {
+export async function synthesizeTts({ key, baseUrl, model = 'qwen-tts-latest', text, voice = 'Cherry', timeoutMs = T_SYNC_GEN, signal, fetchPort = inheritedProviderFetch }) {
   const apiBase = dashscopeApiBase(baseUrl);
-  const res = await fetch(`${apiBase}/services/aigc/multimodal-generation/generation`, {
+  const res = await fetchPort(`${apiBase}/services/aigc/multimodal-generation/generation`, {
     method: 'POST',
+    redirect: 'error',
     headers: authHeaders(key),
     signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({ model, input: { text, voice } })
@@ -389,15 +395,16 @@ export async function synthesizeTts({ key, baseUrl, model = 'qwen-tts-latest', t
 }
 
 /* ---------------- OpenAI Images 兼容（同步 base64/url） ---------------- */
-export async function openAiGenerateImage({ key, baseUrl, model, prompt, size = '1024x1024', n = 1, timeoutMs = T_SYNC_GEN, signal }) {
+export async function openAiGenerateImage({ key, baseUrl, model, prompt, size = '1024x1024', n = 1, timeoutMs = T_SYNC_GEN, signal, fetchPort = inheritedProviderFetch }) {
   const base = String(baseUrl || '').replace(/\/+$/, '');
   const requestSize = applyImageSize({ protocol: 'openai-images', size });
   const { res, body } = await fetchImageJson(`${base}/images/generations`, {
     method: 'POST',
+    redirect: 'error',
     headers: authHeaders(key),
     signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({ model, prompt, size: requestSize, n: Number(n) || 1 })
-  });
+  }, fetchPort);
   if (!res.ok) throw providerError(res.status, body, '生成失败', res.headers);
   const items = (body && Array.isArray(body.data) && body.data) || [];
   const out = [];
@@ -457,7 +464,7 @@ function chatImageItems(body, label = '聊天生图') {
   return items;
 }
 
-export async function openAiChatGenerateImage({ key, baseUrl, model, prompt, image, size, n = 1, timeoutMs = T_SYNC_GEN, signal }) {
+export async function openAiChatGenerateImage({ key, baseUrl, model, prompt, image, size, n = 1, timeoutMs = T_SYNC_GEN, signal, fetchPort = inheritedProviderFetch }) {
   if (typeof size === 'string' && size.trim()) {
     throw new ProviderContractError('聊天生图协议不支持 size 参数；请在提示词中描述所需尺寸', {
       stage: 'validate', category: 'invalid_request', acceptance: 'not_accepted'
@@ -469,16 +476,16 @@ export async function openAiChatGenerateImage({ key, baseUrl, model, prompt, ima
     { type: 'image_url', image_url: { url: 'data:' + image.mediaType + ';base64,' + Buffer.from(image.bytes).toString('base64') } }
   ] : prompt;
   const { res, body } = await fetchImageJson(`${base}/chat/completions`, {
-    method: 'POST', headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal),
+    method: 'POST', redirect: 'error', headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({ model, stream: false, messages: [{ role: 'user', content }],
       ...(Number(n) > 1 ? { n: Number(n) } : {}) })
-  });
+  }, fetchPort);
   if (!res.ok) throw providerError(res.status, body, '聊天生图失败', res.headers);
   return chatImageItems(body);
 }
 
 /** 同步 Responses 生图：主模型由 model 指定，图片工具模型沿用服务端默认。 */
-export async function openAiResponsesGenerateImage({ key, baseUrl, model, prompt, size, n = 1, timeoutMs = T_SYNC_GEN, signal }) {
+export async function openAiResponsesGenerateImage({ key, baseUrl, model, prompt, size, n = 1, timeoutMs = T_SYNC_GEN, signal, fetchPort = inheritedProviderFetch }) {
   if (Number(n) !== 1) {
     throw new ProviderContractError('Responses 生图当前只支持 n=1', {
       stage: 'validate', category: 'invalid_request', acceptance: 'not_accepted'
@@ -487,11 +494,11 @@ export async function openAiResponsesGenerateImage({ key, baseUrl, model, prompt
   const base = String(baseUrl || '').replace(/\/+$/, '');
   const requestSize = applyImageSize({ protocol: 'openai-responses-images', size });
   const { res, body } = await fetchImageJson(`${base}/responses`, {
-    method: 'POST', headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal),
+    method: 'POST', redirect: 'error', headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({ model, input: prompt, stream: false, store: false, background: false,
       tools: [{ type: 'image_generation', action: 'generate', ...(requestSize ? { size: requestSize } : {}) }],
       tool_choice: { type: 'image_generation' } })
-  });
+  }, fetchPort);
   if (!res.ok) throw providerError(res.status, body, 'Responses 生图失败', res.headers);
   if (body?.status !== 'completed') throw imageResponseError('Responses 生图未返回同步完成结果；禁止自动重提');
   const output = Array.isArray(body.output) ? body.output : [];
@@ -515,7 +522,7 @@ export async function openAiResponsesGenerateImage({ key, baseUrl, model, prompt
  * 普通 OpenAI 兼容端点维持 GET /models 与字符串 id 返回形状。
  * @returns {Promise<Array<string|{id:string,capabilities?:string[]}>>}
  */
-export async function listModels({ key, baseUrl, timeoutMs = T_SUBMIT, signal }) {
+export async function listModels({ key, baseUrl, timeoutMs = T_SUBMIT, signal, fetchPort = inheritedProviderFetch }) {
   const base = String(baseUrl || '').trim().replace(/\/+$/, '');
   if (!base) throw new Error('iris: 供应商缺 baseUrl，无法发现模型');
   const dashscope = isDashScopeBaseUrl(base);
@@ -530,7 +537,7 @@ export async function listModels({ key, baseUrl, timeoutMs = T_SUBMIT, signal })
       url.searchParams.set('page_no', String(page));
       url.searchParams.set('page_size', '100');
     }
-    const res = await fetch(url, { headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal) });
+    const res = await fetchPort(url, { headers: authHeaders(key), signal: timeoutSignal(timeoutMs, signal) });
     const body = await readJson(res);
     if (!res.ok) throw providerError(res.status, body, '模型发现失败', res.headers);
     if (!body || body.success === false) throw new ProviderContractError('模型目录响应无效，已保留上次完整目录', {
@@ -593,14 +600,17 @@ export async function listModels({ key, baseUrl, timeoutMs = T_SUBMIT, signal })
  * @param {AbortSignal} [signal] 取消传播
  * @returns {Promise<string>} 完整回答文本
  */
-export async function visionStream({ key, baseUrl, model = 'qwen-vl-plus', prompt, imageDataUrl, signal, timeoutMs = T_VISION }) {
+export async function visionStream({ key, baseUrl, model = 'qwen-vl-plus', prompt, imageDataUrl, signal, timeoutMs = T_VISION, fetchPort = inheritedProviderFetch }) {
   const base = String(baseUrl || '').trim().replace(/\/+$/, '');
+  const requestSignal = timeoutSignal(timeoutMs, signal);
+  requestSignal.throwIfAborted();
   let res;
   try {
-    res = await fetch(`${base}/chat/completions`, {
+    res = await fetchPort(`${base}/chat/completions`, {
       method: 'POST',
+      redirect: 'error',
       headers: authHeaders(key),
-      signal: timeoutSignal(timeoutMs, signal),
+      signal: requestSignal,
       body: JSON.stringify({
         model,
         stream: true,
@@ -616,8 +626,10 @@ export async function visionStream({ key, baseUrl, model = 'qwen-vl-plus', promp
       })
     });
   } catch (err) {
+    requestSignal.throwIfAborted();
     throw networkError(err); // 网络层失败归类为 network
   }
+  if (requestSignal.aborted) { await res.body?.cancel(); requestSignal.throwIfAborted(); }
   if (!res.ok) {
     const text = await readProviderText(res, { maxBytes: PROVIDER_RESPONSE_LIMITS.json, stage: 'response', acceptance: 'unknown' });
     let body = null;
@@ -627,8 +639,8 @@ export async function visionStream({ key, baseUrl, model = 'qwen-vl-plus', promp
   let full = '';
   const decoder = new TextDecoder();
   let buffer = '';
-  for await (const chunk of res.body) {
-    if (signal && signal.aborted) break;
+  for await (const chunk of boundedProviderBody(res, { maxBytes: PROVIDER_RESPONSE_LIMITS.json, stage: 'response', acceptance: 'unknown' })) {
+    requestSignal.throwIfAborted();
     buffer += decoder.decode(chunk, { stream: true });
     let idx;
     while ((idx = buffer.indexOf('\n')) >= 0) {
@@ -646,6 +658,7 @@ export async function visionStream({ key, baseUrl, model = 'qwen-vl-plus', promp
       }
     }
   }
+  requestSignal.throwIfAborted();
   return full.trim();
 }
 
@@ -660,15 +673,15 @@ export function extFromUrl(url, fallback = 'bin') {
   return fallback;
 }
 
-export async function downloadTo(url, filePath, { timeoutMs = T_DOWNLOAD, signal, maxBytes = PROVIDER_RESPONSE_LIMITS.video } = {}) {
-  const res = await fetch(url, { signal: timeoutSignal(timeoutMs, signal) });
+export async function downloadTo(url, filePath, { timeoutMs = T_DOWNLOAD, signal, maxBytes = PROVIDER_RESPONSE_LIMITS.video, fetchPort = inheritedProviderFetch } = {}) {
+  ensurePrivateDir(path.dirname(filePath));
+  const tmp = privateSibling(filePath, 'download');
+  const res = await fetchPort(url, { signal: timeoutSignal(timeoutMs, signal) });
   if (!res.ok) {
     await res.body?.cancel();
     throw Object.assign(new Error(`下载结果失败：HTTP ${res.status}`), { stage: 'download', category: 'provider', httpStatus: res.status, acceptance: 'accepted' });
   }
   if (!res.body) throw new Error('下载结果失败：响应没有内容');
-  ensurePrivateDir(path.dirname(filePath));
-  const tmp = privateSibling(filePath, 'download');
   try {
     const output = fs.createWriteStream(tmp, { flags: 'wx', mode: 0o600 });
     const closed = new Promise(resolve => output.once('close', resolve));
@@ -688,7 +701,7 @@ export async function downloadTo(url, filePath, { timeoutMs = T_DOWNLOAD, signal
  * @param {{key:string, model:string, audioUrl:string}} opts audioUrl 为 oss:// 或公网 URL
  * @returns {Promise<string>} remoteTaskId
  */
-export async function submitTranscription({ key, baseUrl, model = 'qwen-audio-3.0-asr-flash-filetrans', audioUrl, timeoutMs = T_SUBMIT, signal }) {
+export async function submitTranscription({ key, baseUrl, model = 'qwen-audio-3.0-asr-flash-filetrans', audioUrl, timeoutMs = T_SUBMIT, signal, fetchPort = inheritedProviderFetch }) {
   const apiBase = dashscopeApiBase(baseUrl);
   const headers = { ...authHeaders(key), 'X-DashScope-Async': 'enable' };
   if (/^oss:\/\//.test(audioUrl || '')) headers['X-DashScope-OssResourceResolve'] = 'enable';
@@ -696,8 +709,9 @@ export async function submitTranscription({ key, baseUrl, model = 'qwen-audio-3.
   const input = /^qwen3-asr-.*filetrans/i.test(model || '')
     ? { file_url: audioUrl }
     : { file_urls: [audioUrl] };
-  const res = await fetch(`${apiBase}/services/audio/asr/transcription`, {
+  const res = await fetchPort(`${apiBase}/services/audio/asr/transcription`, {
     method: 'POST',
+    redirect: 'error',
     headers,
     signal: timeoutSignal(timeoutMs, signal),
     body: JSON.stringify({ model, input })
@@ -728,9 +742,9 @@ export function transcriptionText(payload) {
 }
 
 /** 轮询 ASR 任务；正式结果位于 results[].transcription_url 指向的 JSON。 */
-export async function pollTranscriptionTask({ key, baseUrl, remoteTaskId, timeoutMs = T_POLL, signal }) {
+export async function pollTranscriptionTask({ key, baseUrl, remoteTaskId, timeoutMs = T_POLL, signal, fetchPort = inheritedProviderFetch }) {
   const apiBase = dashscopeApiBase(baseUrl);
-  const res = await fetch(`${apiBase}/tasks/${encodeURIComponent(remoteTaskId)}`, {
+  const res = await fetchPort(`${apiBase}/tasks/${encodeURIComponent(remoteTaskId)}`, {
     headers: authHeaders(key),
     signal: timeoutSignal(timeoutMs, signal)
   });
@@ -751,8 +765,11 @@ export async function pollTranscriptionTask({ key, baseUrl, remoteTaskId, timeou
       if (typeof result.transcription_text === 'string') text += result.transcription_text;
       else if (typeof result.text === 'string') text += result.text;
       if (status === 'SUCCEEDED' && result.subtask_status !== 'FAILED' && typeof result.transcription_url === 'string') {
-        const transcriptRes = await fetch(result.transcription_url, { signal: timeoutSignal(timeoutMs, signal) });
-        if (!transcriptRes.ok) throw new Error(`下载转写结果失败：HTTP ${transcriptRes.status}`);
+        const transcriptRes = await fetchPort(result.transcription_url, { signal: timeoutSignal(timeoutMs, signal) });
+        if (!transcriptRes.ok) {
+          await transcriptRes.body?.cancel();
+          throw new Error(`下载转写结果失败：HTTP ${transcriptRes.status}`);
+        }
         const transcriptBody = await readJson(transcriptRes);
         if (!transcriptBody) throw new Error('下载转写结果失败：响应不是 JSON');
         text += transcriptionText(transcriptBody);