@mokuyoaxis/dsh-iris 0.1.4 → 0.2.1

package/lib/actions.js

@@ -1,4 +1,5 @@
 'use strict';
+import { visionInputLimits } from './vision-image-input.js';
 /**
  * Iris 动作执行层（阶段 5 GUI 操作面板）。
  *
@@ -11,6 +12,7 @@
  */
 import fs from 'node:fs';
 import path from 'node:path';
+import crypto from 'node:crypto';
 import * as adapters from './adapters.js';
 import * as store from './config.js';
 import * as cap from './capability.js';
@@ -18,25 +20,35 @@ import * as models from './models.js';
 import * as tasks from './tasks.js';
 import * as artifacts from './artifacts.js';
 import { registerMedia, mediaLinksOf, webBase } from './media.js';
-import { buildVisionBackendsFromHost, askWithBackends, testVisionCapability, RED_TEST_IMAGE } from './vision.js';
+import { RED_TEST_IMAGE } from './vision.js';
+import { runVisionRequest, readVisionImageFile, probeVisionModel, visionImageFromDataUrl } from './vision-model-routing.js';
 import { HOST_CONTRACT_VERSION, HOST_PORTS, defineHostAdapter, hasHostPort, requireHostPort } from './host-contract.js';
-import { cropImage, pixelDiff, imageDimensions } from './pixels.js';
-import { locateObject } from './locate.js';
-import { longOcr } from './ocr.js';
+import { coreSnapshotForDsh, cropForDsh, diffForDsh, framesForDsh, summaryFramesForDsh, visionImageForDsh, htmlForDsh, inspectProviderTaskForDsh, submitProviderTaskForDsh, watchProviderTaskForDsh } from './dsh-core-adapter.js';
+import { runLocateRequest, runSummaryRequest } from './composite-vision-routing.js';
+import { normalizeSummarySource, summaryFrameRecords, summaryTimelineLabel } from './summary-input.js';
+import { normalizeVisionImageSource } from './vision-input.js';
+import { formatOcrResult } from './ocr.js';
+import { runOcrRequest } from './ocr-model-routing.js';
 import { ffmpegAvailable, extractFrames, probeVideo } from './media-probe.js';
-import { buildContactSheet, summarizeMedia } from './summarize.js';
 import { ensurePrivateDir, writePrivateFile } from './private-storage.js';
-import { inferMediaProtocol } from './provider-protocol.js';
-import { ProviderContractError, submitWithAcceptanceBoundary } from './provider-contract.js';
-import { invokeProviderOperation, providerPollTaskView } from './provider-adapter.js';
-import { createConfiguredProviderAdapter } from './provider-adapters.js';
+import { selectMediaProtocol, selectImageProtocol, providerForImageModel } from './provider-protocol.js';
+import { providerTaskBinding } from './provider-catalog.js';
+import { invokeProviderOperation, providerPollTaskView, prepareProviderInput } from './provider-adapter.js';
+import { createConfiguredProviderAdapter as createProviderAdapter } from './provider-adapters.js';
 import { deriveUserState } from './task-semantics.js';
+import { normalizeGenerationInput } from './generation-input.js';
+import { videoTaskCandidates } from './video-input.js';
 
 const MEDIA_TYPES = {
   '.png': 'image/png', '.jpg': 'image/jpeg', '.jpeg': 'image/jpeg',
   '.webp': 'image/webp', '.gif': 'image/gif'
 };
 
+function createConfiguredProviderAdapter(provider, options = {}) {
+  return createProviderAdapter(provider, { ...options,
+    onRateLimit: options.allowRateLimited ? undefined : store.recordRateLimit });
+}
+
 const EMPTY_ACTION_HOST = defineHostAdapter({
   id: 'iris-hostless',
   ports: {},
@@ -53,21 +65,8 @@ function actionHost(value) {
   throw new TypeError('Iris Action 只接受有效的 Host Adapter，不接受原始 ctx');
 }
 
-/* ---------------- 视觉效果自述（辅助 iris_draw_image） ---------------- */
-async function describeFirstImage(host, ref, absPath, originalPrompt) {
-  try {
-    const mt = MEDIA_TYPES[path.extname(absPath).toLowerCase()] || 'image/png';
-    const buf = fs.readFileSync(absPath);
-    const dataUrl = `data:${mt};base64,${buf.toString('base64')}`;
-    const backends = buildVisionBackendsFromHost(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION) });
-    const { answer } = await askWithHealth(backends, {
-      question: `用不超过两句话描述这张图片的主题与构图。生成该图的提示词是：「${String(originalPrompt).slice(0, 200)}」`,
-      imageDataUrl: dataUrl,
-      signal: undefined
-    });
-    return answer.slice(0, 300);
-  } catch (_) { return ''; }
-}
+// 工作台清空可选文本框会发送空字符串；在 Host 边界解释为未提供，保持默认值。
+function optionalActionText(value) { return value === '' ? undefined : value; }
 
 /* ---------------- IO 目录 ---------------- */
 function outputDir() {
@@ -89,24 +88,49 @@ function throwIfAborted(signal) {
  */
 function selectProviders(capability, label, requestedModelRef) {
   let providers = store.pickAllFor(capability);
-  const parsedRef = models.parseModelRef(String(requestedModelRef || '').trim());
-  if (parsedRef) {
+  const requested = String(requestedModelRef || '').trim();
+  if (requested) {
+    const pool = models.modelPool(store.providers());
+    const selected = models.resolvePoolModel(pool, requested, capability);
+    if (selected?.rateLimited) throw Object.assign(new Error(store.modelHealth(selected.providerId, selected.id, capability).retryAt
+      ? 'iris: 指定模型正在限流冷却，请稍后使用或手动实测'
+      : 'iris: 指定模型额度或预算耗尽，请手动实测成功后再使用'), {
+      code: 'IRIS_PROVIDER_MODEL_RATE_LIMITED'
+    });
     const field = models.CAP_FIELD[capability];
-    providers = providers.filter((p) => p.id === parsedRef.providerId && (!field || p[field] === parsedRef.modelId));
+    providers = selected ? providers.filter((p) => p.id === selected.providerId && p[field] === selected.id) : [];
     if (!providers.length) throw new Error('iris: 指定的模型复合引用不可用或不具备' + label + '能力');
   }
   if (!providers.length) throw new Error('iris: 没有可用' + label + '供应商——先在 Iris 设置里添加、标注并分配模型');
   return providers;
 }
 
+function configuredModel(provider, capability, requested) {
+  const ref = models.resolveModelRef(provider, capability, requested);
+  if (!ref) throw new Error('iris: 未配置可用模型，请发现或手工添加并标注对应能力');
+  return models.parseModelRef(ref).modelId;
+}
+
+function routeSelectionReason(provider, capability, requested) {
+  if (String(requested || '').trim()) return 'explicit';
+  const field = models.CAP_FIELD[capability];
+  const selectedRef = field && provider?.[field]
+    ? models.modelRef(provider.id, provider[field]) : '';
+  return models.selectionReasonForModel(
+    models.modelPool(store.providers()),
+    capability,
+    store.assignments()[capability],
+    selectedRef
+  ) || 'pool';
+}
+
 function imageRoute(provider, args) {
-  const protocol = provider.mediaProtocol || inferMediaProtocol(provider.baseUrl);
-  const parsed = models.parseModelRef(String(args.model || ''));
-  const model = args.model
-    ? (parsed?.modelId || String(args.model))
-    : (provider.imageModel || (protocol === 'openai-images' ? 'gpt-image-1' : 'wan2.2-t2i-flash'));
+  const model = configuredModel(provider, cap.CAPABILITIES.IMAGE, args.model);
+  provider = providerForImageModel(provider, model);
+  const protocol = selectMediaProtocol(provider).mediaProtocol;
   return {
     provider, protocol, model, modelRef: models.modelRef(provider.id, model),
+    selectionReason: routeSelectionReason(provider, cap.CAPABILITIES.IMAGE, args.model),
     adapter: createConfiguredProviderAdapter(provider, { protocol })
   };
 }
@@ -168,164 +192,110 @@ function recordVisionHealth(errors) {
   }
 }
 
-async function askWithHealth(backends, request) {
+async function askSingleVisionWithHealth(host, request) {
   try {
-    const result = await askWithBackends(backends, request);
+    const result = await runVisionRequest(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION), onRateLimit: store.recordRateLimit, ...request });
     recordVisionHealth(result.errors);
-    if (result.via === 'selfstack') {
-      recordProviderHealthSafe(result.backendId, result.model, cap.CAPABILITIES.VISION, {
-        ok: true, category: 'success', note: '视觉请求成功'
-      });
-    }
+    if (result.via === 'selfstack') recordProviderHealthSafe(result.backendId, result.model, cap.CAPABILITIES.VISION, {
+      ok: true, category: 'success', note: '视觉请求成功'
+    });
     return result;
-  } catch (error) {
-    recordVisionHealth(error && error.errors);
-    throw error;
-  }
+  } catch (error) { recordVisionHealth(error?.errors); throw error; }
 }
 
-async function deliverCompletedImage(taskId, value, signal, providerAdapter) {
-  const names = [];
-  try {
-    tasks.beginDelivery(taskId);
-    if (value && value.kind === 'urls') {
-      for (const url of value.items || []) {
-        const name = taskId + '-' + names.length + '.' + adapters.extFromUrl(url, 'png');
-        await invokeProviderOperation(providerAdapter, 'download', {
-          artifact: { kind: 'remote-url', url }, targetPath: path.join(outputDir(), name), signal
-        });
-        names.push(name);
-      }
-    } else if (value && value.kind === 'openai') {
-      for (const item of value.items || []) {
-        const name = taskId + '-' + names.length + '.png';
-        const target = path.join(outputDir(), name);
-        if (item && item.b64) writePrivateFile(target, Buffer.from(item.b64, 'base64'));
-        else if (item && item.url) await invokeProviderOperation(providerAdapter, 'download', {
-          artifact: { kind: 'remote-url', url: item.url }, targetPath: target, signal
-        });
-        else throw new Error('同步图片结果缺少 b64 或 URL');
-        names.push(name);
-      }
-    } else {
-      throw new Error('未知同步图片结果结构');
-    }
-    if (!names.length) throw new Error('生成已成功，但没有可交付的图片');
-    for (const name of names) registerMedia(taskId, path.join(outputDir(), name));
-    tasks.completeDelivery(taskId, names);
-    return tasks.get(taskId);
-  } catch (error) {
-    let recorded;
-    try {
-      recorded = tasks.failDelivery(taskId, error, names);
-    } catch (persistError) {
-      throw taskActionError(taskId, '图像已生成，但交付状态无法可靠落盘；禁止重新生成', {
-        message: String((persistError && persistError.message) || persistError)
+function recordCoreTaskAttemptHealth(task, capability) {
+  const label = capability === cap.CAPABILITIES.VIDEO ? '视频' : '图片';
+  for (const attempt of task?.attempts || []) {
+    const parsed = models.parseModelRef(attempt.model);
+    if (!parsed) continue;
+    if (attempt.acceptance === 'accepted') {
+      recordProviderHealthSafe(attempt.providerId, parsed.modelId, capability, {
+        ok: true, category: 'success',
+        note: attempt.resultKind === 'completed' ? `同步${label}生成成功` : `${label}任务已受理`
+      });
+    } else if (attempt.error) {
+      recordProviderHealthSafe(attempt.providerId, parsed.modelId, capability, {
+        ok: false, category: attempt.error.category, httpStatus: attempt.error.httpStatus,
+        note: attempt.error.safeMessage
       });
     }
-    throw taskActionError(taskId, '图像已生成，但产物交付失败：' + recorded.error, recorded.lastError);
   }
 }
 
-/** v0.1.3 图片能力统一链路：异步、同步及混合候选共用一个 Task。 */
-async function submitImageV2(routes, prompt, args, signal) {
-  const task = tasks.createV2({
-    cap: 'image',
-    prompt: prompt.slice(0, 4000),
-    size: args.size || '',
-    count: Number(args.n || 1),
-    ...(args.model ? { requestedModelRef: String(args.model) } : {}),
-    ...(args._retryOf ? { retryOf: String(args._retryOf) } : {})
-  });
-  const candidates = routes.map((route) => ({
-    id: route.provider.id,
-    model: route.modelRef,
-    route,
-    submit: async () => {
-      if (signal && signal.aborted) {
-        throw new ProviderContractError('提交前收到取消信号', {
-          stage: 'submit', category: 'aborted', acceptance: 'unknown'
-        });
-      }
-      return observeProviderCapability(route, cap.CAPABILITIES.IMAGE, () =>
-        invokeProviderOperation(route.adapter, 'submit', {
-          capability: 'image',
-          model: route.model,
-          input: { prompt, size: args.size, n: args.n },
-          signal
-        }));
-    }
-  }));
-
-  const submitted = await submitWithAcceptanceBoundary(candidates, { capability: 'image' }, {
-    beforeAttempt: ({ providerId, model }) => {
-      const route = routes.find((item) => item.provider.id === providerId && item.modelRef === model);
-      if (!route) throw new Error('候选路由与 Attempt 身份不一致');
-      return tasks.beginAttempt(task.id, {
-        providerId,
-        providerName: route.provider.name,
-        model,
-        protocol: route.protocol
-      });
-    },
-    afterResult: (snapshot) => tasks.recordAttemptResult(task.id, snapshot)
+/** DSH 图片入口：同步和异步候选共用一份 Core Task/Attempt/Artifact 事实。 */
+async function submitCoreImage(routes, providerInput, signal, sourceArtifactId) {
+  const submitted = await submitProviderTaskForDsh({
+    capability: 'image',
+    candidates: routes.map((route) => ({
+      adapter: route.adapter,
+      model: route.modelRef,
+      selectionReason: route.selectionReason,
+      providerBinding: providerTaskBinding(route.provider)
+    })),
+    providerInput,
+    sourceArtifactId,
+    signal
   });
-
-  if (submitted.localError) {
-    throw taskActionError(task.id, '提交结果未能可靠落盘，已停止后续候选', submitted.localError);
+  const task = submitted.task;
+  recordCoreTaskAttemptHealth(task, cap.CAPABILITIES.IMAGE);
+  if (submitted.localError) throw taskActionError(task.id, 'Core 提交结果未能可靠落盘，已停止后续候选', submitted.localError);
+  if (task.acceptance === 'unknown') throw taskActionError(task.id, '供应商受理状态未知，禁止自动重提', task.lastError);
+  if (task.acceptance === 'not_accepted') throw taskActionError(task.id, '所有候选均明确未受理', task.attempts);
+  if (task.outcome === 'succeeded' && task.deliveryState === 'failed') {
+    throw taskActionError(task.id, '图像已生成，但 Core Artifact 交付失败；禁止重新生成', task.lastError);
   }
-  if (!submitted.candidate) {
-    const lastError = submitted.result && submitted.result.error;
-    tasks.finalizeNoAcceptance(task.id, lastError || '所有候选均明确未受理');
-    throw taskActionError(task.id, '所有候选均明确未受理', submitted.attempts);
-  }
-  if (submitted.result.kind === 'acceptance_unknown') {
-    const safe = submitted.result.error && submitted.result.error.safeMessage;
-    throw taskActionError(task.id, '供应商受理状态未知，禁止自动重提' + (safe ? '：' + safe : ''), submitted.result.error);
+  const selected = routes.find((route) => route.provider.id === task.providerId && route.modelRef === task.modelRef);
+  if (!selected) throw taskActionError(task.id, 'Core Task 的供应商身份无法映射回当前路由', task);
+  if (task.acceptance === 'accepted' && task.remoteTaskId
+      && ['none', 'unknown'].includes(task.outcome) && task.phase !== 'terminal') {
+    await watchProviderTaskForDsh({ taskId: task.id, signal });
+  } else if (task.outcome !== 'succeeded' || task.deliveryState !== 'ready' || !task.artifactIds.length) {
+    throw taskActionError(task.id, '图片路由没有得到可观察任务或可交付产物；事实已保留在 Core，未回退重提', task);
   }
+  return { task, provider: selected.provider, model: selected.model };
+}
 
-  const route = submitted.candidate.route;
-  if (submitted.result.kind === 'completed') {
-    const delivered = await deliverCompletedImage(task.id, submitted.result.value, signal, route.adapter);
-    return { task: delivered, provider: route.provider, model: route.model, remoteTaskId: null };
-  }
-  if (submitted.result.kind !== 'accepted') {
-    throw taskActionError(task.id, '图片提交结果无法识别，已停止后续候选', submitted.result);
+/**
+ * DSH 视频入口（E 阶段）：t2v/i2v 共用 Core Task/Attempt/Artifact 事实轴，
+ * 不共用图片状态机——视频走自己的交付 Profile（video/mp4 / generated-video）。
+ * s2v 在写前 Attempt 之后按账号上传首帧和音频，交付复用视频 Profile。
+ */
+async function submitCoreVideo(routes, frame, prompt, args, signal) {
+  const normalized = normalizeGenerationInput('video', {
+    ...(prompt ? { prompt } : {}),
+    ...(frame ? { first_frame_path: frame.absPath } : {}),
+    ...(args.audio_path ? { audio_path: args.audio_path, resolution: args.resolution } : {}),
+    size: optionalActionText(args.size),
+    duration: args.duration
+  }, { allowVideoPaths: true });
+  const submitted = await submitProviderTaskForDsh({
+    capability: 'video',
+    candidates: videoTaskCandidates(routes.map((route) => ({
+      adapter: route.adapter,
+      model: route.modelRef,
+      selectionReason: route.selectionReason,
+      providerBinding: providerTaskBinding(route.provider)
+    })), normalized),
+    providerInput: normalized.providerInput,
+    signal
+  });
+  const task = submitted.task;
+  recordCoreTaskAttemptHealth(task, cap.CAPABILITIES.VIDEO);
+  if (submitted.localError) throw taskActionError(task.id, 'Core 提交结果未能可靠落盘，已停止后续候选', submitted.localError);
+  if (task.acceptance === 'unknown') throw taskActionError(task.id, '供应商受理状态未知，禁止自动重提', task.lastError);
+  if (task.acceptance === 'not_accepted') throw taskActionError(task.id, '所有候选均明确未受理', task.attempts);
+  if (task.outcome === 'succeeded' && task.deliveryState === 'failed') {
+    throw taskActionError(task.id, '视频已生成，但 Core Artifact 交付失败；禁止重新生成', task.lastError);
   }
-
-  try {
-    tasks.activateWatch(task.id);
-  } catch (error) {
-    throw taskActionError(task.id, '远端已受理，但本地盯守状态落盘失败；禁止自动重提', { message: String(error.message || error) });
+  const selected = routes.find((route) => route.provider.id === task.providerId && route.modelRef === task.modelRef);
+  if (!selected) throw taskActionError(task.id, 'Core Task 的供应商身份无法映射回当前路由', task);
+  if (task.acceptance === 'accepted' && task.remoteTaskId
+      && ['none', 'unknown'].includes(task.outcome) && task.phase !== 'terminal') {
+    await watchProviderTaskForDsh({ taskId: task.id, signal });
+  } else if (task.outcome !== 'succeeded' || task.deliveryState !== 'ready' || !task.artifactIds.length) {
+    throw taskActionError(task.id, '视频路由没有得到可观察任务或可交付产物；事实已保留在 Core，未回退重提', task);
   }
-  tasks.watch(tasks.get(task.id), {
-    poll: ({ remoteTaskId, signal: pollSignal }) => observeProviderCapability(
-      route, cap.CAPABILITIES.IMAGE,
-      () => invokeProviderOperation(route.adapter, 'poll', {
-        capability: 'image', remoteTaskId, signal: pollSignal
-      }),
-      { recordSuccess: false }
-    ),
-    intervalMs: 2500,
-    onSuccess: async (record, result) => {
-      const names = [];
-      for (const url of result.urls || []) {
-        const name = record.id + '-' + names.length + '.' + adapters.extFromUrl(url, 'png');
-        await invokeProviderOperation(route.adapter, 'download', {
-          artifact: { kind: 'remote-url', url }, targetPath: path.join(outputDir(), name)
-        });
-        names.push(name);
-      }
-      for (const name of names) registerMedia(record.id, path.join(outputDir(), name));
-      return names;
-    }
-  });
-  return {
-    task: tasks.get(task.id),
-    provider: route.provider,
-    model: route.model,
-    remoteTaskId: submitted.result.remoteTaskId
-  };
+  return { task, provider: selected.provider, model: selected.model, mode: selected.mode };
 }
 
 function findGeneratedAttachment(attachmentId) {
@@ -339,12 +309,26 @@ function findGeneratedAttachment(attachmentId) {
   return null;
 }
 
-function resolveFirstFrameFile(args) {
+async function resolveFirstFrameFile(host, args, signal) {
   if (args.first_frame_attachment_id) {
     const hit = findGeneratedAttachment(String(args.first_frame_attachment_id).trim());
-    if (!hit) throw new Error('iris: 找不到该 first_frame_attachment_id');
-    if (!fs.existsSync(hit.absPath)) throw new Error('iris: 首帧图片的本地缓存已被清理');
-    return hit;
+    if (hit) {
+      if (!fs.existsSync(hit.absPath)) throw new Error('iris: 首帧图片的本地缓存已被清理');
+      return hit;
+    }
+    const sessions = requireHostPort(host, 'sessions', '读取首帧附件');
+    const attachments = requireHostPort(host, 'attachments', '读取首帧附件');
+    const ref = await sessions.findImageAttachment(args.session_id, String(args.first_frame_attachment_id).trim());
+    if (!ref) throw new Error('iris: 找不到本会话中的 first_frame_attachment_id');
+    const stored = await attachments.readImage(ref, signal);
+    signal?.throwIfAborted();
+    const mediaType = stored.mediaType || ref.mediaType;
+    const extension = Object.keys(MEDIA_TYPES).find(key => MEDIA_TYPES[key] === mediaType);
+    if (!extension) throw new Error('iris: 不支持的首帧图片格式');
+    const bytes = Buffer.from(stored.data);
+    const file = path.join(store.irisHome(), 'uploads', 'frame-' + crypto.createHash('sha256').update(bytes).digest('hex') + extension);
+    if (!fs.existsSync(file)) writePrivateFile(file, bytes, { flag: 'wx' });
+    return { absPath: file, mediaType };
   }
   if (args.first_frame_path) {
     const p = String(args.first_frame_path).trim();
@@ -357,292 +341,46 @@ function resolveFirstFrameFile(args) {
   return null;
 }
 
-function fileDataUrl(file) {
-  return 'data:' + file.mediaType + ';base64,' + fs.readFileSync(file.absPath).toString('base64');
-}
-
 function videoRoute(provider, args, frame) {
-  const parsed = models.parseModelRef(String(args.model || ''));
-  const model = args.model ? (parsed?.modelId || String(args.model)) : (provider.videoModel || 'wan2.2-t2v-flash');
-  const isS2V = /s2v/i.test(model);
+  const model = configuredModel(provider, cap.CAPABILITIES.VIDEO, args.model);
+  const protocol = selectMediaProtocol(provider).mediaProtocol;
+  // B-7：必须区分两件事——
+  //   s2vModel  : 该模型【本身】是否是数字人模型（决定它能否接受 audio_path）
+  //   isS2V     : 本次【请求】是否应走 s2v 链路（显式 audio_path 意图 或 模型本身是 s2v）
+  // 旧实现只有 isS2V 且只看模型名，导致 audio_path + 非 s2v 模型被静默降级为 t2v。
+  const s2vModel = /s2v/i.test(model);
+  const isS2V = Boolean(String(args.audio_path || '').trim()) || s2vModel;
   const mode = isS2V ? 's2v' : (frame ? 'i2v' : 't2v');
   return {
-    provider, model, modelRef: models.modelRef(provider.id, model), isS2V, mode, protocol: 'dashscope',
-    adapter: createConfiguredProviderAdapter(provider, { protocol: 'dashscope' })
+    provider, model, modelRef: models.modelRef(provider.id, model), isS2V, s2vModel, mode, protocol,
+    adapter: createConfiguredProviderAdapter(provider, { protocol }),
+    selectionReason: routeSelectionReason(provider, cap.CAPABILITIES.VIDEO, args.model)
   };
 }
 
-function beforeGenerationError(error, stage) {
-  const category = error && (error.category === 'network' || error instanceof TypeError)
-    ? 'network'
-    : (stage === 'validate' ? 'invalid_request' : 'local_io');
-  return new ProviderContractError(String((error && error.message) || error || '提交前准备失败'), {
-    stage,
-    category,
-    acceptance: 'not_accepted',
-    retryable: category === 'network'
-  });
-}
-
-async function submitVideoV2(routes, frame, prompt, args, signal) {
-  const task = tasks.createV2({
-    cap: 'video', prompt: prompt.slice(0, 4000),
-    ...(args.model ? { requestedModelRef: String(args.model) } : {}),
-    ...(args._retryOf ? { retryOf: String(args._retryOf) } : {})
-  });
-  const candidates = routes.map((route) => ({
-    id: route.provider.id,
-    model: route.modelRef,
-    route,
-    submit: async (_input, context) => {
-      const attemptId = context.attempt.id;
-      let submitArgs;
-      if (route.isS2V) {
-        if (!frame) throw beforeGenerationError(new Error('iris: s2v 数字人需要首帧'), 'validate');
-        const audioPath = String(args.audio_path || '').trim();
-        if (!audioPath || !path.isAbsolute(audioPath)) {
-          throw beforeGenerationError(new Error('iris: s2v 的 audio_path 必须是绝对路径'), 'validate');
-        }
-        if (!fs.existsSync(audioPath)) throw beforeGenerationError(new Error('iris: 音频文件不存在'), 'validate');
-        tasks.markAttemptStage(task.id, attemptId, 'upload');
-        try {
-          const imageUrl = await adapters.uploadTempFile({
-            key: route.provider.apiKey, baseUrl: route.provider.baseUrl,
-            model: route.model, filePath: frame.absPath, signal
-          });
-          const audioUrl = await adapters.uploadTempFile({
-            key: route.provider.apiKey, baseUrl: route.provider.baseUrl,
-            model: route.model, filePath: audioPath, signal
-          });
-          submitArgs = { imgDataUrl: imageUrl, audioUrl, resolution: args.resolution || '480P' };
-        } catch (error) {
-          throw beforeGenerationError(error, 'upload');
-        }
-      } else {
-        if (!prompt) throw beforeGenerationError(new Error('iris: 文生/图生视频需要 prompt'), 'validate');
-        tasks.markAttemptStage(task.id, attemptId, 'prepare');
-        try {
-          submitArgs = {
-            prompt,
-            imgDataUrl: frame ? fileDataUrl(frame) : undefined,
-            size: args.size || '1280*720',
-            duration: args.duration ? Number(args.duration) : undefined
-          };
-        } catch (error) {
-          throw beforeGenerationError(error, 'prepare');
-        }
-      }
-      tasks.markAttemptStage(task.id, attemptId, 'submit');
-      return observeProviderCapability(route, cap.CAPABILITIES.VIDEO, () =>
-        invokeProviderOperation(route.adapter, 'submit', {
-          capability: 'video', model: route.model, input: submitArgs, signal
-        }));
-    }
-  }));
-
-  const submitted = await submitWithAcceptanceBoundary(candidates, { capability: 'video' }, {
-    beforeAttempt: ({ providerId, model }) => {
-      const route = routes.find((item) => item.provider.id === providerId && item.modelRef === model);
-      if (!route) throw new Error('候选路由与 Attempt 身份不一致');
-      const attempt = tasks.beginAttempt(task.id, {
-        providerId, providerName: route.provider.name, model, protocol: route.protocol
-      });
-      tasks.update(task.id, {
-        mode: route.mode,
-        ...(route.isS2V ? { resolution: args.resolution || '480P' } : {
-          size: args.size || '1280*720', ...(args.duration ? { duration: Number(args.duration) } : {})
-        })
-      });
-      return attempt;
-    },
-    afterResult: (snapshot) => tasks.recordAttemptResult(task.id, snapshot)
-  });
-
-  if (submitted.localError) {
-    throw taskActionError(task.id, '视频提交结果未能可靠落盘，已停止后续候选', submitted.localError);
-  }
-  if (!submitted.candidate) {
-    const lastError = submitted.result && submitted.result.error;
-    tasks.finalizeNoAcceptance(task.id, lastError || '所有视频候选均明确未受理');
-    throw taskActionError(task.id, '所有视频候选均明确未受理', submitted.attempts);
-  }
-  if (submitted.result.kind !== 'accepted') {
-    const safe = submitted.result.error && submitted.result.error.safeMessage;
-    throw taskActionError(task.id, '视频受理状态未知，禁止自动重提' + (safe ? '：' + safe : ''), submitted.result.error);
-  }
-
-  const route = submitted.candidate.route;
-  tasks.activateWatch(task.id);
-  tasks.watch(tasks.get(task.id), {
-    poll: ({ remoteTaskId: id, signal: pollSignal }) => observeProviderCapability(
-      route, cap.CAPABILITIES.VIDEO,
-      () => invokeProviderOperation(route.adapter, 'poll', {
-        capability: 'video', remoteTaskId: id, signal: pollSignal
-      }),
-      { recordSuccess: false }
-    ),
-    intervalMs: 6000,
-    onSuccess: async (record, result) => {
-      const names = [];
-      for (const url of result.urls || []) {
-        const name = record.id + '-' + names.length + '.' + adapters.extFromUrl(url, 'mp4');
-        await invokeProviderOperation(route.adapter, 'download', {
-          artifact: { kind: 'remote-url', url }, targetPath: path.join(outputDir(), name)
-        });
-        names.push(name);
-      }
-      for (const name of names) registerMedia(record.id, path.join(outputDir(), name));
-      return names;
-    }
-  });
-  return { task: tasks.get(task.id), provider: route.provider, model: route.model, remoteTaskId: submitted.result.remoteTaskId };
-}
-
-function simpleModelRoute(provider, args, field, fallback, protocol = 'dashscope') {
-  const parsed = models.parseModelRef(String(args.model || ''));
-  const model = args.model ? (parsed?.modelId || String(args.model)) : (provider[field] || fallback);
+function simpleModelRoute(provider, args, capability) {
+  const protocol = selectMediaProtocol(provider).mediaProtocol;
+  const model = configuredModel(provider, capability, args.model);
   return {
     provider, model, modelRef: models.modelRef(provider.id, model), protocol,
-    adapter: createConfiguredProviderAdapter(provider, { protocol })
+    adapter: createConfiguredProviderAdapter(provider, { protocol }),
+    selectionReason: routeSelectionReason(provider, capability, args.model)
   };
 }
 
-async function submitTranscriptionV2(routes, audioPath, args, signal) {
-  const task = tasks.createV2({
-    cap: 'transcribe', prompt: 'audio transcribe: ' + path.basename(audioPath),
-    ...(args.model ? { requestedModelRef: String(args.model) } : {})
-  });
-  const candidates = routes.map((route) => ({
-    id: route.provider.id,
-    model: route.modelRef,
-    route,
-    submit: async (_input, context) => {
-      tasks.markAttemptStage(task.id, context.attempt.id, 'upload');
-      let audioUrl;
-      try {
-        audioUrl = await adapters.uploadTempFile({
-          key: route.provider.apiKey, baseUrl: route.provider.baseUrl,
-          model: route.model, filePath: audioPath, signal
-        });
-      } catch (error) {
-        throw beforeGenerationError(error, 'upload');
-      }
-      tasks.markAttemptStage(task.id, context.attempt.id, 'submit');
-      return observeProviderCapability(route, cap.CAPABILITIES.TRANSCRIBE, () =>
-        invokeProviderOperation(route.adapter, 'submit', {
-          capability: 'transcribe', model: route.model, input: { audioUrl }, signal
-        }));
-    }
-  }));
-  const submitted = await submitWithAcceptanceBoundary(candidates, { capability: 'transcribe' }, {
-    beforeAttempt: ({ providerId, model }) => {
-      const route = routes.find((item) => item.provider.id === providerId && item.modelRef === model);
-      if (!route) throw new Error('候选路由与 Attempt 身份不一致');
-      return tasks.beginAttempt(task.id, {
-        providerId, providerName: route.provider.name, model, protocol: route.protocol
-      });
-    },
-    afterResult: (snapshot) => tasks.recordAttemptResult(task.id, snapshot)
-  });
-  if (submitted.localError) throw taskActionError(task.id, '转写提交结果未能可靠落盘，已停止后续候选', submitted.localError);
-  if (!submitted.candidate) {
-    tasks.finalizeNoAcceptance(task.id, submitted.result?.error || '所有转写候选均明确未受理');
-    throw taskActionError(task.id, '所有转写候选均明确未受理', submitted.attempts);
-  }
-  if (submitted.result.kind !== 'accepted') {
-    const safe = submitted.result.error && submitted.result.error.safeMessage;
-    throw taskActionError(task.id, '转写受理状态未知，禁止自动重提' + (safe ? '：' + safe : ''), submitted.result.error);
-  }
-  const route = submitted.candidate.route;
-  tasks.activateWatch(task.id);
-  tasks.watch(tasks.get(task.id), {
-    poll: ({ remoteTaskId: id, signal: pollSignal }) => observeProviderCapability(
-      route, cap.CAPABILITIES.TRANSCRIBE,
-      () => invokeProviderOperation(route.adapter, 'poll', {
-        capability: 'transcribe', remoteTaskId: id, signal: pollSignal
-      }),
-      { recordSuccess: false }
-    ),
-    intervalMs: 2500,
-    allowEmptyResult: true,
-    onSuccess: async (record, result) => {
-      tasks.update(record.id, { transcribeText: (result.urls && result.urls.join('')) || '' });
-      return [];
-    }
-  });
-  return { task: tasks.get(task.id), provider: route.provider, model: route.model, remoteTaskId: submitted.result.remoteTaskId };
-}
 
-async function submitTtsV2(routes, text, args, signal) {
-  const task = tasks.createV2({
-    cap: 'tts', prompt: text.slice(0, 500),
-    ...(args.model ? { requestedModelRef: String(args.model) } : {})
-  });
-  const candidates = routes.map((route) => ({
-    id: route.provider.id,
-    model: route.modelRef,
-    route,
-    submit: async (_input, context) => {
-      tasks.markAttemptStage(task.id, context.attempt.id, 'submit');
-      return observeProviderCapability(route, cap.CAPABILITIES.TTS, () =>
-        invokeProviderOperation(route.adapter, 'submit', {
-          capability: 'tts', model: route.model, input: { text, voice: route.voice }, signal
-        }));
-    }
-  }));
-  const submitted = await submitWithAcceptanceBoundary(candidates, { capability: 'tts' }, {
-    beforeAttempt: ({ providerId, model }) => {
-      const route = routes.find((item) => item.provider.id === providerId && item.modelRef === model);
-      if (!route) throw new Error('候选路由与 Attempt 身份不一致');
-      const attempt = tasks.beginAttempt(task.id, {
-        providerId, providerName: route.provider.name, model, protocol: route.protocol
-      });
-      tasks.update(task.id, { voice: route.voice });
-      return attempt;
-    },
-    afterResult: (snapshot) => tasks.recordAttemptResult(task.id, snapshot)
-  });
-  if (submitted.localError) throw taskActionError(task.id, '语音合成结果未能可靠落盘，已停止后续候选', submitted.localError);
-  if (!submitted.candidate) {
-    tasks.finalizeNoAcceptance(task.id, submitted.result?.error || '所有语音候选均明确未受理');
-    throw taskActionError(task.id, '所有语音候选均明确未受理', submitted.attempts);
-  }
-  if (submitted.result.kind !== 'completed') {
-    const safe = submitted.result.error && submitted.result.error.safeMessage;
-    throw taskActionError(task.id, '语音合成受理状态未知，禁止自动重提' + (safe ? '：' + safe : ''), submitted.result.error);
-  }
-  const route = submitted.candidate.route;
-  const audio = submitted.result.value || {};
-  const names = [];
-  try {
-    tasks.beginDelivery(task.id);
-    const name = task.id + '.' + (audio.audioUrl ? adapters.extFromUrl(audio.audioUrl, 'wav') : 'wav');
-    const target = path.join(outputDir(), name);
-    if (audio.audioUrl) await invokeProviderOperation(route.adapter, 'download', {
-      artifact: { kind: 'remote-url', url: audio.audioUrl }, targetPath: target, signal
-    });
-    else if (audio.audioB64) writePrivateFile(target, Buffer.from(audio.audioB64, 'base64'));
-    else throw new Error('语音合成成功响应缺少音频数据');
-    names.push(name);
-    registerMedia(task.id, target);
-    tasks.completeDelivery(task.id, names);
-  } catch (error) {
-    const recorded = tasks.failDelivery(task.id, error, names);
-    throw taskActionError(task.id, '语音已合成，但产物交付失败：' + recorded.error, recorded.lastError);
-  }
-  return { task: tasks.get(task.id), provider: route.provider, model: route.model };
-}
 
 /** 重新观察/交付与启动恢复共享的远端查询和本地落盘依赖。 */
 export function taskPollDeps(provider, task) {
   if (!provider || !task) return null;
   if (!['image', 'video', 'transcribe'].includes(task.cap)) return null;
   const protocol = task.protocol || (task.cap === 'image'
-    ? (provider.mediaProtocol || inferMediaProtocol(provider.baseUrl))
+    ? selectMediaProtocol(provider).mediaProtocol
     : 'dashscope');
   let providerAdapter;
   try {
-    providerAdapter = createConfiguredProviderAdapter(provider, { protocol });
+    providerAdapter = createConfiguredProviderAdapter(provider, { protocol,
+      modelId: models.parseModelRef(String(task.modelRef || task.model || ''))?.modelId || String(task.model || '') });
   } catch (_) {
     return null;
   }
@@ -724,65 +462,185 @@ export async function runAction(host, name, args, { signal } = {}) {
 /* ============ 动作实现 ============ */
 
 /* ---------- 🖼️ 画图 ---------- */
-register('image', async (host, args, { signal }) => {
-  const prompt = String(args.prompt || '').trim();
-  if (!prompt) throw new Error('iris: prompt 不能为空');
+async function imageAction(host, args, { signal }, editing = false) {
+  if (editing && !args.source_artifact_id) throw new Error('iris: 改图必须提供 source_artifact_id');
+  const { providerInput, sourceArtifactId } = normalizeGenerationInput('image', {
+    prompt: args.prompt, size: optionalActionText(args.size), n: args.n,
+    ...(editing ? { source_artifact_id: args.source_artifact_id } : {})
+  });
 
   const imageProviders = selectProviders(cap.CAPABILITIES.IMAGE, '图像', args.model);
   const routes = imageProviders.map((provider) => imageRoute(provider, args));
-  const submitted = await submitImageV2(routes, prompt, args, signal);
+  const submitted = await submitCoreImage(routes, providerInput, signal, sourceArtifactId);
+  const firstArtifactId = submitted.task.artifactIds[0];
+  const ready = submitted.task.outcome === 'succeeded'
+    && submitted.task.deliveryState === 'ready' && Boolean(firstArtifactId);
   return {
-    ok: true, taskId: submitted.task.id, model: submitted.model, providerId: submitted.provider.id,
-    remoteTaskId: submitted.remoteTaskId,
-    text: submitted.remoteTaskId
-      ? '[iris] 图像任务已提交（' + submitted.model + '，task: ' + submitted.task.id + '）'
-      : '[iris] 图像任务已完成（' + submitted.model + '，task: ' + submitted.task.id + '）'
+    ok: true, storage: 'core', taskId: submitted.task.id, model: submitted.model,
+    providerId: submitted.provider.id, remoteTaskId: submitted.task.remoteTaskId || null,
+    artifactIds: submitted.task.artifactIds,
+    ...(sourceArtifactId ? { sourceArtifactId } : {}),
+    ...(ready ? { imageUrl: '/iris/api/core/artifact/' + encodeURIComponent(firstArtifactId) + '/media' } : {}),
+    text: ready
+      ? '[iris] 图像任务已完成（' + submitted.model + '，Core task: ' + submitted.task.id + '）'
+      : '[iris] 图像任务已提交并由 Core 后台观察（' + submitted.model + '，Core task: ' + submitted.task.id + '）'
   };
-});
+}
+register('image', imageAction);
+register('image_edit', (host, args, context) => imageAction(host, args, context, true));
 
 /* ---------- 🎬 视频生成 ---------- */
 register('video', async (host, args, { signal }) => {
   const prompt = String(args.prompt || '').trim();
-  if (!prompt && !args.audio_path) throw new Error('iris: prompt 不能为空');
-  const frame = resolveFirstFrameFile(args);
+  const frame = await resolveFirstFrameFile(host, args, signal);
   const providers = selectProviders(cap.CAPABILITIES.VIDEO, '视频', args.model);
   const routes = providers.map((provider) => videoRoute(provider, args, frame));
-  const submitted = await submitVideoV2(routes, frame, prompt, args, signal);
+  // 请求子类型决定候选过滤；池中存在 s2v 不影响普通视频。
+  const hasAudioPath = Boolean(String(args.audio_path || '').trim());
+  const s2v = hasAudioPath || (args.model && /s2v/i.test(String(args.model)));
+  // B-7：audio_path 与选中模型必须语义一致。若用户提供了音频却选中非 s2v 模型，
+  // 旧实现会落入 else 分支静默丢弃 audio_path 并生成普通视频（可能产生非预期计费）。
+  // 这里在受理边界与任何网络调用之前明确拒绝，不猜测用户意图。
+  if (hasAudioPath && args.model && routes.some((route) => !route.s2vModel)) {
+    const mismatched = routes.filter((route) => !route.s2vModel).map((route) => route.model).join(', ');
+    throw new Error('iris: audio_path 只能配合支持 s2v 数字人的模型；当前候选 ' + mismatched
+      + ' 不是 s2v 模型。请改用 s2v 模型（如 wan2.2-s2v），或移除 audio_path 走文生/图生视频。');
+  }
+  const coreRoutes = routes.filter((route) => route.isS2V === Boolean(s2v));
+  if (!coreRoutes.length) {
+    throw new Error(s2v ? 'iris: 未配置可用 s2v 数字人模型' : 'iris: 当前视频候选只有 s2v 数字人模型；请使用 audio_path 的数字人入口');
+  }
+  const submitted = await submitCoreVideo(coreRoutes, frame, prompt, args, signal);
+  const firstArtifactId = submitted.task.artifactIds[0];
+  const ready = submitted.task.outcome === 'succeeded'
+    && submitted.task.deliveryState === 'ready' && Boolean(firstArtifactId);
   return {
-    ok: true, taskId: submitted.task.id, model: submitted.model, mode: submitted.task.mode,
-    providerId: submitted.provider.id, remoteTaskId: submitted.remoteTaskId,
-    text: '[iris] 视频任务已提交（' + submitted.model + '，task: ' + submitted.task.id + '）'
+    ok: true, storage: 'core', taskId: submitted.task.id, model: submitted.model, mode: submitted.mode,
+    providerId: submitted.provider.id, remoteTaskId: submitted.task.remoteTaskId || null,
+    artifactIds: submitted.task.artifactIds,
+    ...(ready ? { videoUrl: '/iris/api/core/artifact/' + encodeURIComponent(firstArtifactId) + '/media' } : {}),
+    text: ready
+      ? '[iris] 视频任务已完成（' + submitted.model + '，Core task: ' + submitted.task.id + '）'
+      : '[iris] 视频任务已提交并由 Core 后台观察（' + submitted.model + '，Core task: ' + submitted.task.id + '）'
   };
 });
 
+/**
+ * DSH 语音合成入口（E2）：TTS 是同步完成型——submit 内 completed，同一次调用
+ * 直接交付 generated-audio Artifact（无 remote 长轮询）。整条候选链共用一个
+ * voice（Core providerInput 静态；legacy 曾按候选各自取 provider.ttsVoice）。
+ */
+async function submitCoreTts(routes, normalizedInput, signal) {
+  const voice = normalizedInput.voice || String(routes[0].provider.ttsVoice || '').trim();
+  const { providerInput } = normalizeGenerationInput('tts', {
+    text: normalizedInput.text, ...(voice ? { voice } : {})
+  });
+  const submitted = await submitProviderTaskForDsh({
+    capability: 'tts',
+    candidates: routes.map((route) => ({
+      adapter: route.adapter,
+      model: route.modelRef,
+      selectionReason: route.selectionReason,
+      providerBinding: providerTaskBinding(route.provider)
+    })),
+    providerInput,
+    signal
+  });
+  const task = submitted.task;
+  recordCoreTaskAttemptHealth(task, cap.CAPABILITIES.TTS);
+  if (submitted.localError) throw taskActionError(task.id, 'Core 提交结果未能可靠落盘，已停止后续候选', submitted.localError);
+  if (task.acceptance === 'unknown') throw taskActionError(task.id, '语音合成受理状态未知，禁止自动重提', task.lastError);
+  if (task.acceptance === 'not_accepted') throw taskActionError(task.id, '所有候选均明确未受理', task.attempts);
+  if (task.outcome === 'succeeded' && task.deliveryState === 'failed') {
+    throw taskActionError(task.id, '语音已合成，但 Core Artifact 交付失败；禁止重新生成', task.lastError);
+  }
+  if (task.outcome !== 'succeeded' || task.deliveryState !== 'ready' || !task.artifactIds.length) {
+    throw taskActionError(task.id, '语音合成没有得到可交付产物；事实已保留在 Core，未回退重提', task);
+  }
+  const selected = routes.find((route) => route.provider.id === task.providerId && route.modelRef === task.modelRef);
+  if (!selected) throw taskActionError(task.id, 'Core Task 的供应商身份无法映射回当前路由', task);
+  return { task, provider: selected.provider, model: selected.model };
+}
+
 /* ---------- 🔊 语音合成 ---------- */
 register('tts', async (host, args, { signal }) => {
-  const text = String(args.text || '').trim();
-  if (!text) throw new Error('iris: text 不能为空');
+  const { providerInput } = normalizeGenerationInput('tts', {
+    text: args.text, voice: optionalActionText(args.voice)
+  });
   const providers = selectProviders(cap.CAPABILITIES.TTS, '语音', args.model);
-  const routes = providers.map((provider) => ({
-    ...simpleModelRoute(provider, args, 'ttsModel', 'qwen-tts-latest'),
-    voice: args.voice || provider.ttsVoice || 'Cherry'
-  }));
-  const submitted = await submitTtsV2(routes, text, args, signal);
+  const routes = providers.map((provider) => simpleModelRoute(provider, args, cap.CAPABILITIES.TTS));
+  const submitted = await submitCoreTts(routes, providerInput, signal);
+  const firstArtifactId = submitted.task.artifactIds[0];
   return {
-    ok: true, taskId: submitted.task.id, model: submitted.model, providerId: submitted.provider.id,
-    text: '[iris] 语音已合成（' + submitted.model + '，task: ' + submitted.task.id + '）'
+    ok: true, storage: 'core', taskId: submitted.task.id, model: submitted.model,
+    providerId: submitted.provider.id,
+    artifactIds: submitted.task.artifactIds,
+    audioUrl: '/iris/api/core/artifact/' + encodeURIComponent(firstArtifactId) + '/media',
+    text: '[iris] 语音已合成（' + submitted.model + '，Core task: ' + submitted.task.id + '）'
   };
 });
 
+/**
+ * DSH 转写入口（E3）：上传型异步——音频先在 Host 边界上传到首选候选的临时存储
+ * （签名 URL 只进 providerInput，绝不落 Core 记录），再经受理边界提交；观察/接管
+ * 与人工控制面复用同一 Task/Attempt 轴；正文物化为 text/plain transcript Artifact。
+ * 已知边界：临时音频 URL 属于首选候选的账号，failover 到不同账号时如实记录失败。
+ */
+async function submitCoreTranscribe(routes, normalized, signal) {
+  let providerInput = normalized.providerInput;
+  if (normalized.audioPath) {
+    const prepared = await prepareProviderInput(routes[0].adapter, {
+      model: routes[0].model, filePath: normalized.audioPath, signal
+    });
+    providerInput = normalizeGenerationInput('transcribe', { audio_url: prepared.url }).providerInput;
+  }
+  const submitted = await submitProviderTaskForDsh({
+    capability: 'transcribe',
+    candidates: routes.map((route) => ({
+      adapter: route.adapter,
+      model: route.modelRef,
+      selectionReason: route.selectionReason,
+      providerBinding: providerTaskBinding(route.provider)
+    })),
+    providerInput,
+    signal
+  });
+  const task = submitted.task;
+  recordCoreTaskAttemptHealth(task, cap.CAPABILITIES.TRANSCRIBE);
+  if (submitted.localError) throw taskActionError(task.id, 'Core 提交结果未能可靠落盘，已停止后续候选', submitted.localError);
+  if (task.acceptance === 'unknown') throw taskActionError(task.id, '转写受理状态未知，禁止自动重提', task.lastError);
+  if (task.acceptance === 'not_accepted') throw taskActionError(task.id, '所有候选均明确未受理', task.attempts);
+  if (task.outcome === 'succeeded' && task.deliveryState === 'failed') {
+    throw taskActionError(task.id, '转写已完成，但 Core Artifact 交付失败；禁止重新生成', task.lastError);
+  }
+  const selected = routes.find((route) => route.provider.id === task.providerId && route.modelRef === task.modelRef);
+  if (!selected) throw taskActionError(task.id, 'Core Task 的供应商身份无法映射回当前路由', task);
+  if (task.acceptance === 'accepted' && task.remoteTaskId
+      && ['none', 'unknown'].includes(task.outcome) && task.phase !== 'terminal') {
+    await watchProviderTaskForDsh({ taskId: task.id, signal });
+  } else if (task.outcome !== 'succeeded' || task.deliveryState !== 'ready' || !task.artifactIds.length) {
+    throw taskActionError(task.id, '转写路由没有得到可观察任务或可交付产物；事实已保留在 Core，未回退重提', task);
+  }
+  return { task, provider: selected.provider, model: selected.model };
+}
+
 /* ---------- 👁 看图 ---------- */
 register('look', async (host, args, { signal }) => {
+  const source = normalizeVisionImageSource(args);
   const p = String(args.image_path || '').trim();
-  if (!p || !path.isAbsolute(p)) throw new Error('iris: image_path 必须是绝对路径');
   const mt = MEDIA_TYPES[path.extname(p).toLowerCase()];
-  if (!mt || mt === 'image/gif') throw new Error('iris: 不支持的图片格式');
-  if (!fs.existsSync(p)) throw new Error('iris: 图片不存在');
-  const dataUrl = `data:${mt};base64,${fs.readFileSync(p).toString('base64')}`;
-  const backends = buildVisionBackendsFromHost(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION), model: args.model });
-  const { answer, via, model } = await askWithHealth(backends, { question: (args.question || '详细描述这张图片。'), imageDataUrl: dataUrl, signal });
+  if (source === 'image_path') {
+    if (!p || !path.isAbsolute(p)) throw new Error('iris: image_path 必须是绝对路径');
+    if (!mt || mt === 'image/gif') throw new Error('iris: 不支持的图片格式');
+    if (!fs.existsSync(p)) throw new Error('iris: 图片不存在');
+  }
+  const { answer, via, model } = await askSingleVisionWithHealth(host, { model: args.model,
+    question: String(args.question || '').trim() || '详细描述这张图片。',
+    prepareImage: operationSignal => source === 'artifact_id'
+      ? visionImageForDsh({ artifactId: args.artifact_id, signal: operationSignal }) : readVisionImageFile(p, mt, operationSignal), signal });
   const viaLabel = via === 'selfstack' ? 'iris 自持栈' : 'DSH 全局视觉模型';
-  return { ok: true, text: `[iris] 看图回答（${model} · ${viaLabel}）：\n${answer}` };
+  return { ok: true, ...(args.artifact_id ? { artifactId: args.artifact_id } : {}),
+    text: `[iris] 看图回答（${model} · ${viaLabel}）：\n${answer}` };
 });
 
 /* ---------- ✂️ 裁剪 ---------- */
@@ -790,79 +648,121 @@ register('crop', async (host, args, { signal }) => {
   const p = String(args.image_path || '').trim();
   if (!p || !path.isAbsolute(p)) throw new Error('iris: image_path 必须是绝对路径');
   if (!fs.existsSync(p)) throw new Error('iris: 图片不存在');
-  const { buffer, width, height, mime } = await cropImage({ input: p, left: args.left, top: args.top, width: args.width, height: args.height });
-  const b64 = buffer.toString('base64');
-  return { ok: true, text: `[iris] 裁剪完成：${width}x${height}（原区域 ${args.left},${args.top},${args.width},${args.height}）`, imageDataUrl: `data:${mime};base64,${b64}` };
+  const result = await cropForDsh({
+    imagePath: p, left: args.left, top: args.top, width: args.width, height: args.height, signal
+  });
+  const b64 = result.bytes.toString('base64');
+  return { ok: true, text: `[iris] 裁剪完成：${result.width}x${result.height}（原区域 ${args.left},${args.top},${args.width},${args.height}）`, imageDataUrl: `data:${result.mediaType};base64,${b64}` };
 });
 
 /* ---------- 📷 像素 diff ---------- */
 register('diff', async (host, args, { signal }) => {
   if (!args.image_a_path || !args.image_b_path) throw new Error('iris: 需要 image_a_path 和 image_b_path');
-  const { ratio, diffPixels, totalPixels, width, height, worstRegions, heatmap, mime } = await pixelDiff({
-    inputA: args.image_a_path, inputB: args.image_b_path, grid: args.grid || 8, topRegions: args.top_regions || 3
+  const result = await diffForDsh({
+    imageA: { imagePath: args.image_a_path }, imageB: { imagePath: args.image_b_path },
+    grid: args.grid, topRegions: args.top_regions, signal
   });
+  const { ratio, diffPixels, totalPixels, width, height, worstRegions } = result.metrics;
   const pct = (ratio * 100).toFixed(2);
   const regions = worstRegions.map((r) => `(col ${r.col},row ${r.row}) ${(r.score * 100).toFixed(1)}%`).join(' ');
   return {
     ok: true,
+    artifactIds: [result.artifact.id],
     text: `[iris] 像素差异 ${pct}%（${diffPixels}/${totalPixels}px，归一化 ${width}x${height}）\n最差区域（${worstRegions.length} 格）: ${regions}`,
-    imageDataUrl: `data:${mime};base64,${heatmap.toString('base64')}`
+    imageDataUrl: `data:${result.mediaType};base64,${result.bytes.toString('base64')}`
   };
 });
 
 /* ---------- 📍 定位 ---------- */
 register('locate', async (host, args, { signal }) => {
+  const source = normalizeVisionImageSource(args);
   const p = String(args.image_path || '').trim();
-  if (!p || !path.isAbsolute(p)) throw new Error('iris: image_path 必须是绝对路径');
-  if (!fs.existsSync(p)) throw new Error('iris: 图片不存在');
-  const buf = fs.readFileSync(p);
+  if (source === 'image_path' && (!p || !path.isAbsolute(p))) throw new Error('iris: image_path 必须是绝对路径');
   const mt = MEDIA_TYPES[path.extname(p).toLowerCase()] || 'image/png';
-  const dataUrl = `data:${mt};base64,${buf.toString('base64')}`;
-  const { width, height } = await imageDimensions(buf);
-  const backends = buildVisionBackendsFromHost(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION), model: args.model });
-  const r = await locateObject(backends, { target: args.target, imageDataUrl: dataUrl, width, height, signal });
-  if (!r.found) return { ok: true, text: `[iris] 在图片中未找到「${args.target}」` };
-  return { ok: true, text: `[iris] 定位「${args.target}」：bbox (${r.x1},${r.y1},${r.x2},${r.y2}) / ${width}x${height}（${r.via} · ${r.model}）` };
+  const r = await runLocateRequest(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION), onRateLimit: store.recordRateLimit, model: args.model,
+    target: args.target, prepareImage: operationSignal => source === 'artifact_id'
+      ? visionImageForDsh({ artifactId: args.artifact_id, signal: operationSignal }) : readVisionImageFile(p, mt, operationSignal), signal });
+  const { width, height } = r;
+  const provenance = args.artifact_id ? { artifactId: args.artifact_id } : {};
+  if (!r.found) return { ok: true, ...provenance, text: `[iris] 在图片中未找到「${args.target}」` };
+  return { ok: true, ...provenance, text: `[iris] 定位「${args.target}」：bbox (${r.x1},${r.y1},${r.x2},${r.y2}) / ${width}x${height}（${r.via} · ${r.model}）` };
 });
 /* ---------- 🖼️ HTML 截图 ---------- */
 register('html', async (host, args, { signal }) => {
   const html = String(args.html || '').trim();
   if (!html) throw new Error('iris: html 不能为空');
   const browser = requireHostPort(host, 'browser', 'HTML 截图');
-  const rendered = await browser.renderHtml({ html, width: args.width, height: args.height,
+  const rendered = await htmlForDsh({ browser, html, width: args.width, height: args.height,
     fullPage: args.fullPage !== false, signal });
   const png = Buffer.from(rendered.bytes);
-  return { ok: true, text: `[iris] HTML 截图完成（fullPage: ${args.fullPage !== false}）`, imageDataUrl: `data:${rendered.mediaType || 'image/png'};base64,${png.toString('base64')}` };
+  return { ok: true, storage: 'core', artifactIds: [rendered.artifact.id], text: `[iris] HTML 截图完成（fullPage: ${args.fullPage !== false}）`, imageDataUrl: `data:image/png;base64,${png.toString('base64')}` };
 });
 
 /* ---------- 📄 长截图 OCR ---------- */
 register('ocr', async (host, args, { signal }) => {
+  const source = normalizeVisionImageSource(args);
   const p = String(args.image_path || '').trim();
-  if (!p || !path.isAbsolute(p)) throw new Error('iris: image_path 必须是绝对路径');
-  if (!fs.existsSync(p)) throw new Error('iris: 图片不存在');
-  const backends = buildVisionBackendsFromHost(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION) });
-  const result = await longOcr({ input: p, backends, chunkHeight: args.chunk_height || 1200, overlap: args.overlap || 120, signal });
-  const errCount = result.chunks.filter((c) => c.error).length;
-  const head = `[iris] 长截图 OCR 完成：${result.width}x${result.height}px，${result.totalChunks} 块${errCount ? '（' + errCount + ' 块失败）' : ''}\n`;
-  return { ok: true, text: result.fullText ? head + result.fullText : head + '（未识别到文字）' };
+  const mediaType = MEDIA_TYPES[path.extname(p).toLowerCase()];
+  if (source === 'image_path') {
+    if (!p || !path.isAbsolute(p)) throw new Error('iris: image_path 必须是绝对路径');
+    if (!mediaType || mediaType === 'image/gif') throw new Error('iris: 不支持的图片格式（png/jpg/jpeg/webp）');
+  }
+  const result = await runOcrRequest(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION),
+    onRateLimit: store.recordRateLimit,
+    prepareImage: operationSignal => source === 'artifact_id'
+      ? visionImageForDsh({ artifactId: args.artifact_id, signal: operationSignal }) : readVisionImageFile(p, mediaType, operationSignal),
+    chunkHeight: args.chunk_height, overlap: args.overlap, signal });
+  return { ok: result.status !== 'failed', status: result.status,
+    totalChunks: result.totalChunks, successfulChunks: result.successfulChunks,
+    failedChunks: result.failedChunks, skippedChunks: result.skippedChunks,
+    ...(args.artifact_id ? { artifactId: args.artifact_id } : {}), text: formatOcrResult(result) };
 });
 
 /* ---------- 📋 任务查询 ---------- */
 register('status', async (host, args, { signal }) => {
-  if (args.task_id) {
-    const t = tasks.get(String(args.task_id).trim());
-    if (!t) throw new Error('iris: 没有这个任务：' + args.task_id);
-    const lines = [`[${t.cap}] ${t.id} · ${t.status}${t.progress ? ' · ' + t.progress : ''} · ${t.model}`];
-    if (t.prompt) lines.push('prompt: ' + String(t.prompt).slice(0, 120));
-    if (t.error) lines.push('error: ' + t.error);
-    if ((t.files || []).length) lines.push('files:\n' + t.files.map((f) => path.join(outputDir(), f)).join('\n'));
-    const links = mediaLinksOf(t);
-    if (links.length) lines.push('links:\n' + links.join('\n'));
-    return { ok: true, text: lines.join('\n') };
+  const requestedId = String(args.task_id || '').trim();
+  if (requestedId) {
+    const legacy = tasks.get(requestedId);
+    if (legacy) {
+      const lines = [`[${legacy.cap}] ${legacy.id} · ${legacy.status}${legacy.progress ? ' · ' + legacy.progress : ''} · ${legacy.model}`];
+      if (legacy.prompt) lines.push('prompt: ' + String(legacy.prompt).slice(0, 120));
+      if (legacy.error) lines.push('error: ' + legacy.error);
+      if ((legacy.files || []).length) lines.push('files:\n' + legacy.files.map((f) => path.join(outputDir(), f)).join('\n'));
+      const links = mediaLinksOf(legacy);
+      if (links.length) lines.push('links:\n' + links.join('\n'));
+      return { ok: true, storage: 'legacy', taskId: legacy.id, text: lines.join('\n') };
+    }
+    if (/^task_[a-f0-9]{24}$/.test(requestedId)) {
+      try {
+        const core = await inspectProviderTaskForDsh(requestedId, { signal });
+        const lines = [`[${core.capability}] ${core.id} · Core ${core.phase} · ${core.outcome} / ${core.deliveryState} · ${core.modelRef || '未选择模型'}`];
+        if (core.lastError) lines.push('error: ' + core.lastError.safeMessage);
+        if (core.remoteTaskId) lines.push('remoteTask: ' + core.remoteTaskId);
+        if (core.artifactIds.length) {
+          lines.push('artifacts:\n' + core.artifactIds.map((id) => id + ' · /iris/api/core/artifact/' + encodeURIComponent(id) + '/media').join('\n'));
+        }
+        return { ok: true, storage: 'core', taskId: core.id, artifactIds: core.artifactIds, text: lines.join('\n') };
+      } catch (_) { /* 对外统一为不存在，不暴露 Core 数据根 */ }
+    }
+    throw new Error('iris: 没有这个任务：' + args.task_id);
   }
-  const recent = tasks.list().slice(0, 5);
-  if (!recent.length) return { ok: true, text: '[iris] 还没有生成任务记录。' };
-  return { ok: true, text: recent.map((t) => `[${t.cap}] ${t.id} · ${t.status} · ${t.model}`).join('\n') };
+
+  const legacyRows = tasks.list().slice(0, 5).map((task) => ({
+    at: task.createdAt,
+    text: `[${task.cap}] ${task.id} · ${task.status} · ${task.model}`
+  }));
+  let coreRows = [];
+  try {
+    const snapshot = await coreSnapshotForDsh({ limit: 5, signal });
+    coreRows = (snapshot.tasks.recent || []).map((task) => ({
+      at: task.updatedAt,
+      text: `[${task.capability}] ${task.id} · Core ${task.phase} · ${task.outcome} / ${task.deliveryState} · ${task.modelRef || '未选择模型'}`
+    }));
+  } catch (_) { /* legacy 查询仍然可用 */ }
+  const rows = legacyRows.concat(coreRows)
+    .sort((a, b) => String(b.at || '').localeCompare(String(a.at || '')))
+    .slice(0, 5);
+  return { ok: true, text: rows.length ? rows.map((item) => item.text).join('\n') : '[iris] 还没有生成任务记录。' };
 });
 
 /* ---------- Task v2 人工接管 ---------- */
@@ -1064,22 +964,23 @@ register('relook', async (host, args, { signal }) => {
   if (!id) throw new Error('iris: attachment_id 不能为空');
   const question = String(args.question || '').trim();
   if (!question) throw new Error('iris: question 不能为空');
-  // GUI 无会话事件扫描能力：只在 iris 自己的任务产物里找（findOwnAttachment 逻辑）
-  let hit = null;
-  for (const t of tasks.all()) {
-    for (const a of t.attachments || []) {
-      if (a.attachmentId === id) {
-        const absPath = path.join(store.irisHome(), 'outputs', a.file);
-        if (fs.existsSync(absPath)) { hit = { absPath, mediaType: a.mediaType }; break; }
+  const { answer, via, model } = await askSingleVisionWithHealth(host, { question, model: args.model,
+    async prepareImage(operationSignal) {
+      // GUI 没有会话上下文；记录扫描也在预取消检查及整体预算之后执行。
+      let hit;
+      for (const task of tasks.all()) {
+        for (const attachment of task.attachments || []) {
+          if (attachment.attachmentId === id) {
+            const absPath = path.join(store.irisHome(), 'outputs', attachment.file);
+            if (fs.existsSync(absPath)) { hit = { absPath, mediaType: attachment.mediaType }; break; }
+          }
+        }
+        if (hit) break;
       }
-    }
-    if (hit) break;
-  }
-  if (!hit) throw new Error('iris: 该 attachment 不在 iris 生成的任务产物中（GUI 重看仅支持 iris 画图产物）');
-  const mt = hit.mediaType || MEDIA_TYPES[path.extname(hit.absPath).toLowerCase()] || 'image/png';
-  const dataUrl = `data:${mt};base64,${fs.readFileSync(hit.absPath).toString('base64')}`;
-  const backends = buildVisionBackendsFromHost(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION), model: args.model });
-  const { answer, via, model } = await askWithHealth(backends, { question, imageDataUrl: dataUrl, signal });
+      if (!hit) throw new Error('iris: 图片引用不可用');
+      const mt = hit.mediaType || MEDIA_TYPES[path.extname(hit.absPath).toLowerCase()] || 'image/png';
+      return readVisionImageFile(hit.absPath, mt, operationSignal);
+    }, signal });
   const viaLabel = via === 'selfstack' ? 'iris 自持栈' : 'DSH 全局视觉模型';
   return { ok: true, text: `[iris] 重看回答（${model} · ${viaLabel}）：\n${answer}` };
 });
@@ -1096,12 +997,19 @@ register('providers_list', async (host, args) => {
       name: p.name || p.id,
       type: p.type || 'openai',
       baseUrl: p.baseUrl,
+      mediaBaseUrl: p.mediaBaseUrl || '',
       enabled: p.enabled !== false,
-      mediaProtocol: p.mediaProtocol || inferMediaProtocol(p.baseUrl),
+      auth: p.auth === 'none' ? 'none' : 'bearer',
+      ...selectMediaProtocol(p),
+      visionInput: p.visionInput || null,
       apiKeyHint: String(p.apiKey || '').length <= 8 ? '****' : String(p.apiKey).slice(0, 3) + '****' + String(p.apiKey).slice(-4),
       models: models.providerModels(p).map((m) => ({
         id: m.id,
         capabilities: m.capabilities,
+        ...(m.imageProtocol !== undefined ? { imageProtocol: m.imageProtocol } : {}),
+        ...(m.visionInput !== undefined ? { visionInput: m.visionInput } : {}),
+        ...(m.capabilities.includes(cap.CAPABILITIES.VISION) ? { visionInputEffective: visionInputLimits(p, m.id) } : {}),
+        ...(m.capabilities.includes(cap.CAPABILITIES.IMAGE) ? { imageRouting: selectImageProtocol(p, m.id) } : {}),
         verified: m.verified || null,
         health: Object.fromEntries(m.capabilities.map((capability) => [
           capability, store.modelHealth(p.id, m.id, capability)
@@ -1117,8 +1025,8 @@ register('providers_list', async (host, args) => {
 register('providers_upsert', async (host, args) => {
   const id = args.id || undefined;
   const patch = { id };
-  const allowed = ['name', 'type', 'baseUrl', 'apiKey', 'enabled', 'mediaProtocol',
-    'imageModel', 'videoModel', 'ttsModel', 'transcribeModel', 'visionModel', 'models'];
+  const allowed = ['auth', 'name', 'type', 'baseUrl', 'mediaBaseUrl', 'apiKey', 'enabled', 'mediaProtocol',
+    'imageModel', 'videoModel', 'ttsModel', 'transcribeModel', 'visionModel', 'visionInput', 'models'];
   for (const key of allowed) {
     if (Object.prototype.hasOwnProperty.call(args, key)) patch[key] = args[key];
   }
@@ -1141,44 +1049,42 @@ register('providers_set_models', async (host, args) => {
 });
 
 /**
- * 模型发现（P2）：GET /models 拉全量 → 按规则过滤出媒体能力模型 → 写入显式池。
+ * 模型发现：读取供应商目录 → 官方能力元数据优先、名称规则兜底 → 写入显式池。
  * 免费只读；拉不到/不支持则原样保留现有池（优雅降级），报人话错误。
  */
 register('providers_discover', async (host, args, { signal }) => {
   const p = store.providerById(String(args.id || ''));
   if (!p) throw new Error('iris: 供应商不存在或未启用');
-  let ids;
+  let discoveredModels;
   try {
     const providerAdapter = createConfiguredProviderAdapter(p);
     const discovered = await invokeProviderOperation(providerAdapter, 'discover', { signal });
-    ids = discovered.models.map((model) => model.id);
+    discoveredModels = discovered.models;
   } catch (err) {
     throw new Error('iris: 模型发现失败（保留现有池）：' + String(err.message || err));
   }
   const media = [];
-  for (const id of ids) {
-    const caps = models.capabilitiesOfModel(id);
-    if (caps.length) media.push({ id, capabilities: caps });
+  for (const model of discoveredModels) {
+    const caps = models.capabilitiesOfDiscoveredModel(model);
+    if (caps.length) media.push({ id: model.id, capabilities: caps });
   }
   if (!media.length) {
-    return { ok: true, total: ids.length, media: 0, text: `[iris] ${p.name}：拉到 ${ids.length} 个模型，无媒体能力匹配——可在模型池区手动添加并标能力` };
+    return { ok: true, total: discoveredModels.length, media: 0, text: `[iris] ${p.name}：拉到 ${discoveredModels.length} 个模型，无媒体能力匹配——可在模型池区手动添加并标能力` };
   }
-  store.setProviderModels(p.id, media);
-  return { ok: true, total: ids.length, media: media.length, text: `[iris] ${p.name}：发现 ${ids.length} 个模型，${media.length} 个媒体能力模型已入池（按能力标注）` };
+  store.mergeDiscoveredModels(p.id, media);
+  return { ok: true, total: discoveredModels.length, media: media.length, text: `[iris] ${p.name}：发现 ${discoveredModels.length} 个模型，${media.length} 个媒体能力模型已入池（官方元数据优先）` };
 });
 
 /** 能力实测（红色测试图）：对某供应商的 vision 模型跑 testVisionCapability */
 register('providers_test_vision', async (host, args, { signal }) => {
   const p = store.providerById(String(args.id || ''));
   if (!p) throw new Error('iris: 供应商不存在');
-  const visionProviders = store.pickAllFor(cap.CAPABILITIES.VISION).filter((x) => x.id === p.id);
+  const visionProviders = store.pickAllFor(cap.CAPABILITIES.VISION, { includeRateLimited: true }).filter((x) => x.id === p.id);
   if (!visionProviders.length) return { ok: true, tested: false, text: '该供应商未配置 vision 模型，跳过' };
-  const backends = buildVisionBackendsFromHost(host, { providers: visionProviders });
-  const backend = backends[0];
-  if (!backend) return { ok: true, tested: false, text: '该供应商的 vision 配置不可用（类型非 openai）' };
+  const model = configuredModel(visionProviders[0], cap.CAPABILITIES.VISION);
   if (args.confirm_paid !== true) throw new Error('iris: 视觉实测会发起真实供应商请求，请先明确确认');
-  const r = await testVisionCapability(backend, { timeoutMs: 15000, signal });
-  store.setModelVerified(p.id, backend.model, cap.CAPABILITIES.VISION, {
+  const r = await probeVisionModel(visionProviders[0], model, visionImageFromDataUrl(RED_TEST_IMAGE), { timeoutMs: 15000, signal });
+  store.setModelVerified(p.id, model, cap.CAPABILITIES.VISION, {
     ok: r.ok,
     note: r.ok ? '认出红色测试图' : (r.error || '未识别出红色'),
     category: r.category,
@@ -1189,7 +1095,7 @@ register('providers_test_vision', async (host, args, { signal }) => {
     tested: true,
     passed: r.ok,
     text: r.ok
-      ? `[iris] 视觉能力实测通过：模型认出红色测试图（${backend.model}）`
+      ? `[iris] 视觉能力实测通过：模型认出红色测试图（${model}）`
       : `[iris] 视觉能力实测未通过：${r.error || '未识别出红色'}`
   };
 });
@@ -1199,9 +1105,7 @@ register('providers_test_vision', async (host, args, { signal }) => {
 /** 按能力实测单个模型；返回 {ok, note}。真实调用仅在 UI 明确确认后执行。 */
 async function probeModel(host, p, modelId, capability, signal) {
   if (capability === cap.CAPABILITIES.VISION) {
-    const backend = buildVisionBackendsFromHost(host, { providers: [p], model: modelId })[0];
-    if (!backend) return { ok: false, note: '非 openai 类型，无法直测' };
-    const r = await testVisionCapability(backend, { timeoutMs: 15000, signal });
+    const r = await probeVisionModel(p, modelId, visionImageFromDataUrl(RED_TEST_IMAGE), { timeoutMs: 15000, signal });
     return {
       ok: r.ok,
       note: r.ok ? '认出红色测试图' : (r.error || '未识别出红色'),
@@ -1210,24 +1114,28 @@ async function probeModel(host, p, modelId, capability, signal) {
     };
   }
   if (capability === cap.CAPABILITIES.TRANSCRIBE) {
-    return { skipped: true, ok: false, note: '需真实音频，请通过“音频转写”动作验证' };
+    return { skipped: true, ok: false, note: '需真实音频；可用 CLI models test --capability transcribe --input 提供素材进行实测恢复' };
   }
   if (capability === cap.CAPABILITIES.TTS) {
-    const providerAdapter = createConfiguredProviderAdapter(p, { protocol: 'dashscope' });
+    const providerAdapter = createConfiguredProviderAdapter(p, { allowRateLimited: true });
     const submitted = await invokeProviderOperation(providerAdapter, 'submit', {
-      capability: 'tts', model: modelId, input: { text: '你好', voice: 'Cherry' }, signal
+      capability: 'tts', model: modelId, input: { text: '你好' }, signal
     });
     const value = submitted.kind === 'completed' ? submitted.value : null;
     const ok = Boolean(value && (value.audioUrl || value.audioB64));
-    return { ok, note: ok ? '合成返回音频' : (submitted.error?.safeMessage || '无音频返回') };
+    return { ok, note: ok ? '合成返回音频' : (submitted.error?.safeMessage || '无音频返回'),
+      category: submitted.error?.category, httpStatus: submitted.error?.httpStatus };
   }
   if (capability === cap.CAPABILITIES.IMAGE) {
-    const providerAdapter = createConfiguredProviderAdapter(p);
+    p = providerForImageModel(p, modelId);
+    const providerAdapter = createConfiguredProviderAdapter(p, { allowRateLimited: true, modelId });
     const submitted = await invokeProviderOperation(providerAdapter, 'submit', {
       capability: 'image', model: modelId,
       input: {
         prompt: providerAdapter.protocol === 'openai-images' ? 'red circle' : '红色圆形',
-        size: providerAdapter.protocol === 'openai-images' ? '256x256' : '512*512',
+        ...(['openai-chat-images', 'openai-responses-images'].includes(providerAdapter.protocol) ? {} : {
+          size: providerAdapter.protocol === 'openai-images' ? '256x256' : '512*512'
+        }),
         n: 1
       },
       signal
@@ -1264,7 +1172,7 @@ async function probeModel(host, p, modelId, capability, signal) {
     return { ok: false, note: '生成超时（>60s）' };
   }
   if (capability === cap.CAPABILITIES.VIDEO) {
-    return { skipped: true, ok: false, note: '视频提交即可能计费，请通过“视频生成”动作验证' };
+    return { skipped: true, ok: false, note: '视频提交即可能计费；可用 CLI models test --capability video-gen --input 提供素材进行实测恢复' };
   }
   return { ok: false, note: '未知能力' };
 }
@@ -1297,6 +1205,19 @@ register('providers_set_model_caps', async (host, args) => {
   return { ok: true, text: `[iris] ${args.model_id} 能力已设为：${caps.join(',') || '（空）'}` };
 });
 
+register('providers_set_model_image_protocol', async (host, args) => {
+  const ok = store.setModelImageProtocol(String(args.id || ''), String(args.model_id || ''), args.imageProtocol);
+  if (!ok) throw new Error('iris: 模型不在池中');
+  return { ok: true, text: '[iris] 模型图片协议已更新' };
+});
+
+register('providers_set_model_vision_input', async (host, args) => {
+  if (!Object.hasOwn(args, 'visionInput')) throw new Error('iris: 需要 visionInput 对象或 null');
+  const ok = store.setModelVisionInput(String(args.id || ''), String(args.model_id || ''), args.visionInput);
+  if (!ok) throw new Error('iris: 模型不在池中');
+  return { ok: true, text: '[iris] 模型看图输入预算已更新' };
+});
+
 /** 逐模型实测某能力 → 存 verified（"test it, don't guess it"；用户手动触发，发现不自动跑） */
 register('providers_test_model', async (host, args, { signal }) => {
   const p = store.providerById(String(args.id || ''));
@@ -1436,41 +1357,45 @@ register('attachment_export', async (host, args, { signal }) => {
 /* ============ 音频转写（阶段 7.2） ============ */
 
 register('transcribe', async (host, args, { signal }) => {
-  const audioPath = String(args.audio_path || '').trim();
-  if (!audioPath || !path.isAbsolute(audioPath)) throw new Error('iris: audio_path 必须是绝对路径');
-  if (!fs.existsSync(audioPath)) throw new Error('iris: 音频文件不存在');
+  const normalized = normalizeGenerationInput('transcribe', {
+    audio_url: args.audio_url, audio_path: args.audio_path
+  }, { allowAudioPath: true });
+  if (normalized.audioPath && !fs.existsSync(normalized.audioPath)) throw new Error('iris: 音频文件不存在');
   const providers = selectProviders(cap.CAPABILITIES.TRANSCRIBE, '转写', args.model);
   const routes = providers.map((provider) =>
-    simpleModelRoute(provider, args, 'transcribeModel', 'qwen-audio-3.0-asr-flash-filetrans')
+    simpleModelRoute(provider, args, cap.CAPABILITIES.TRANSCRIBE)
   );
-  const submitted = await submitTranscriptionV2(routes, audioPath, args, signal);
+  const submitted = await submitCoreTranscribe(routes, normalized, signal);
+  const firstArtifactId = submitted.task.artifactIds[0];
+  const ready = submitted.task.outcome === 'succeeded'
+    && submitted.task.deliveryState === 'ready' && Boolean(firstArtifactId);
   return {
-    ok: true, taskId: submitted.task.id, model: submitted.model, providerId: submitted.provider.id,
-    remoteTaskId: submitted.remoteTaskId,
-    text: '[iris] 音频转写任务已提交（' + submitted.model + '，task: ' + submitted.task.id + '）'
+    ok: true, storage: 'core', taskId: submitted.task.id, model: submitted.model,
+    providerId: submitted.provider.id, remoteTaskId: submitted.task.remoteTaskId || null,
+    artifactIds: submitted.task.artifactIds,
+    ...(ready ? { transcriptUrl: '/iris/api/core/artifact/' + encodeURIComponent(firstArtifactId) + '/media' } : {}),
+    text: ready
+      ? '[iris] 音频转写已完成（' + submitted.model + '，Core task: ' + submitted.task.id + '）'
+      : '[iris] 音频转写任务已提交并由 Core 后台观察（' + submitted.model + '，Core task: ' + submitted.task.id + '）'
   };
 });
 
 /* ============ 视频抽帧（阶段 7.1） ============ */
 
 register('video_frames', async (host, args, { signal }) => {
-  const p = String(args.video_path || '').trim();
-  if (!p || !path.isAbsolute(p)) throw new Error('iris: video_path 必须是绝对路径');
-  if (!fs.existsSync(p)) throw new Error('iris: 视频文件不存在');
-  if (!ffmpegAvailable()) throw new Error('iris: ffmpeg/ffprobe 不可用，无法抽帧（可选安装 ffmpeg 后启用）');
-  const meta = probeVideo(p);
-  const frames = await extractFrames({
-    inputPath: p, maxFrames: args.max_frames, targetWidth: args.target_width,
-    format: args.format || 'jpeg', quality: args.quality, signal
+  const { media: meta, frames } = await framesForDsh({
+    videoPath: args.video_path, maxFrames: args.max_frames, targetWidth: args.target_width,
+    format: optionalActionText(args.format), quality: args.quality, signal
   });
-  const mime = args.format === 'png' ? 'image/png' : 'image/jpeg';
+  const mime = frames[0].mediaType;
   const summary = frames.map((f, i) =>
-    `帧${i + 1} @${f.atSec.toFixed(1)}s  ${f.width}x${f.height}  ${(f.buffer.length / 1024).toFixed(0)}KB`
+    `帧${i + 1} @${f.atSec.toFixed(1)}s  ${f.width}x${f.height}  ${(f.bytes.length / 1024).toFixed(0)}KB`
   ).join('\n');
   // 首帧预览 base64
-  const firstFrame = `data:${mime};base64,${frames[0].buffer.toString('base64')}`;
+  const firstFrame = `data:${mime};base64,${frames[0].bytes.toString('base64')}`;
   return {
     ok: true,
+    artifactIds: frames.map((frame) => frame.artifact.id),
     text: `[iris] 视频抽帧完成：${frames.length} 帧（${meta.durationSec.toFixed(1)}s ${meta.width}x${meta.height}，${mime === 'image/png' ? 'PNG' : 'JPEG'}）\n${summary}\n\n（预览为首帧；全部帧请用对话工具 iris_video_frames 获取 attachment）`,
     imageDataUrl: firstFrame
   };
@@ -1479,28 +1404,32 @@ register('video_frames', async (host, args, { signal }) => {
 /* ============ 多模态视频摘要（阶段 7.3，GUI 同步快速版） ============ */
 
 register('media_summarize', async (host, args, { signal }) => {
+  const source = normalizeSummarySource(args);
   const p = String(args.video_path || '').trim();
-  if (!p || !path.isAbsolute(p)) throw new Error('iris: video_path 必须是绝对路径');
-  if (!fs.existsSync(p)) throw new Error('iris: 视频文件不存在');
-  if (!ffmpegAvailable()) throw new Error('iris: ffmpeg/ffprobe 不可用，无法分析视频（可选安装 ffmpeg 后启用）');
-  const meta = probeVideo(p);
-  const frames = await extractFrames({
-    inputPath: p, maxFrames: Math.min(Number(args.max_frames) || 8, 12), targetWidth: args.target_width,
-    format: 'jpeg', quality: 70, signal
-  });
-  const sheet = await buildContactSheet({ frames, cols: 3 });
-  const backends = buildVisionBackendsFromHost(host, { providers: store.pickAllFor(cap.CAPABILITIES.VISION), model: args.model });
-  const { answer, via, model } = await summarizeMedia({
-    backends, question: args.question,
-    transcript: args.transcribe_text, // GUI 同步版：不自动转写，可粘贴已有转写文本
-    imageDataUrl: `data:image/png;base64,${sheet.buffer.toString('base64')}`,
-    signal
+  if (source === 'video') {
+    if (!p || !path.isAbsolute(p)) throw new Error('iris: video_path 必须是绝对路径');
+    if (!fs.existsSync(p)) throw new Error('iris: 视频文件不存在');
+    if (!ffmpegAvailable()) throw new Error('iris: ffmpeg/ffprobe 不可用，无法分析视频（可选安装 ffmpeg 后启用）');
+  }
+  const { answer, via, model, sheet, meta, frames } = await runSummaryRequest(host, {
+    onRateLimit: store.recordRateLimit,
+    providers: store.pickAllFor(cap.CAPABILITIES.VISION), model: args.model, question: args.question, signal,
+    async prepareMedia(operationSignal) {
+      if (source === 'core-artifacts') {
+        return { ...await summaryFramesForDsh({ frameArtifactIds: args.frame_artifact_ids, signal: operationSignal }), transcript: args.transcribe_text };
+      }
+      const meta = probeVideo(p);
+      const frames = await extractFrames({ inputPath: p, maxFrames: Math.min(Number(args.max_frames) || 8, 12),
+        targetWidth: args.target_width, format: 'jpeg', quality: 70, signal: operationSignal });
+      return { meta, frames, transcript: args.transcribe_text };
+    }
   });
   const viaLabel = via === 'selfstack' ? 'iris 自持栈' : 'DSH 全局视觉模型';
   const hasTrans = args.transcribe_text && String(args.transcribe_text).trim();
   return {
     ok: true,
-    text: `[iris] 视频摘要完成（${model} · ${viaLabel}，${meta.durationSec.toFixed(1)}s / ${frames.length} 帧联系表${hasTrans ? ' + 转写文本' : ''}）：\n${answer}`,
+    frames: summaryFrameRecords(frames),
+    text: `[iris] 视频摘要完成（${model} · ${viaLabel}，${summaryTimelineLabel(meta)} / ${frames.length} 帧联系表${hasTrans ? ' + 转写文本' : ''}）：\n${answer}`,
     imageDataUrl: `data:image/png;base64,${sheet.buffer.toString('base64')}`
   };
 });