newmark-agent 0.5.3 → 0.5.7

package/dist/conversation-utility-host.bundle.cjs

@@ -328670,7 +328670,7 @@ function providerStreamTimeoutError(timeoutMs) {
   error.message = `Stream read timeout after ${timeoutMs}ms`;
   return error;
 }
-async function readProviderStreamChunk(reader, signal, timeoutMs = 3e4) {
+async function readProviderStreamChunk(reader, signal, timeoutMs = 0) {
   if (signal.aborted) throw providerAbortError(signal);
   let timer;
   let onAbort;
@@ -328678,9 +328678,9 @@ async function readProviderStreamChunk(reader, signal, timeoutMs = 3e4) {
     onAbort = () => reject(providerAbortError(signal));
     signal.addEventListener("abort", onAbort, { once: true });
   });
-  const timeoutPromise = new Promise((_3, reject) => {
+  const timeoutPromise = timeoutMs > 0 ? new Promise((_3, reject) => {
     timer = setTimeout(() => reject(providerStreamTimeoutError(timeoutMs)), timeoutMs);
-  });
+  }) : new Promise(() => void 0);
   try {
     return await Promise.race([reader.read(), abortPromise, timeoutPromise]);
   } catch (error) {
@@ -336933,6 +336933,10 @@ var ToolExecutor = class {
   async webSearch(query) {
     return this.wsearch(query);
   }
+  /** OCR entry point for the runtime's final visual fallback. */
+  async finalVisualFallbackOcr(dataUrl, signal) {
+    return await this.localOcr.recognizeDataUrl(dataUrl, signal, "sparse-ui");
+  }
   setHostProfile(profile) {
     this.hostProfile = { ...profile };
   }
@@ -340846,9 +340850,20 @@ async function runAgentKernel(agent) {
   try {
     const linkedPlanRevisionBeforeRun = agent.getLinkedPlan().revision;
     const modelBeforeKernelRun = agent.model;
-    let lastTurn = await runWithCompressionResume([], false);
+    const preflightVisualFallback = !agent.activeModelConfig()?.vision ? await agent.finalVisualFallback("vision input not supported by the selected model", processSignal) : null;
+    let lastTurn = preflightVisualFallback ? { text: preflightVisualFallback, stopReason: "stop", errorMessage: "" } : await runWithCompressionResume([], false);
+    if (preflightVisualFallback) {
+      tokens.push({ type: "text", text: preflightVisualFallback });
+      agent.recordWorkStatus("Final visual fallback used: local mini OCR plus conservative text correction.");
+    }
     if (modelBeforeKernelRun && modelBeforeKernelRun !== agent.model && !tokens.some((t3) => t3.text?.includes("[Model fallback]"))) {
-      tokens.unshift({ type: "text", text: `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.` });
+      const notice = `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.`;
+      tokens.unshift({ type: "text", text: notice });
+      agent.emitWorkEvent({
+        type: "status",
+        content: notice,
+        fallback: { from: modelBeforeKernelRun, to: agent.model, providerId: agent.activeDeployment()?.providerId }
+      });
     }
     let emptyResponseRetries = 0;
     while (providerTurnIsEmpty(lastTurn) && emptyResponseRetries < 2) {
@@ -340869,6 +340884,11 @@ async function runAgentKernel(agent) {
       const notice = routeTransitionNotice(agent, previous);
       tokens.push({ type: "text", text: notice });
       agent.recordWorkStatus(notice);
+      agent.emitWorkEvent({
+        type: "status",
+        content: notice,
+        fallback: { from: previous, to: agent.model, providerId: agent.activeDeployment()?.providerId }
+      });
       kernel2.state.model = toKernelModel(agent);
       const fallbackToolSurface = refreshToolSurface(true);
       kernel2.state.systemPrompt = [agent.buildSystemPrompt(), fallbackToolSurface.systemPromptNotice].filter(Boolean).join("\n\n");
@@ -340876,6 +340896,14 @@ async function runAgentKernel(agent) {
       await agent.waitForPlannedRouteRetry();
       lastTurn = await runWithCompressionResume([], false);
     }
+    if (kernelTurnFailed(agent, lastTurn)) {
+      const visualFallback = await agent.finalVisualFallback(lastTurn.errorMessage || lastTurn.text, processSignal);
+      if (visualFallback) {
+        tokens.push({ type: "text", text: visualFallback });
+        agent.recordWorkStatus("Final visual fallback used: local mini OCR plus conservative text correction.");
+        lastTurn = { ...lastTurn, text: visualFallback, errorMessage: "", stopReason: "stop" };
+      }
+    }
     if (kernelTurnFailed(agent, lastTurn)) {
       throw new ProviderRunError(normalizePublicProviderError(lastTurn.errorMessage || lastTurn.text, [agent.activeModelConfig()?.api_key]));
     }
@@ -343285,14 +343313,16 @@ var AutoRouter = class {
         retryDelayMs
       });
     }
+    if (failure.allowModelFallback === false) return attempts.slice(0, remainingAttempts);
     const selection = decision.requestedSelection;
     if (selection.kind === "fixed") return attempts.slice(0, remainingAttempts);
     const scope = selection.kind === "auto" ? selection.scope : { kind: "provider", providerId: current.providerId };
     const subset = selection.kind === "auto" ? selection.subset : void 0;
     const currentGroup = current.logicalModelGroupId;
+    const currentProviderId = current.providerId;
     const now2 = this.now();
     const attemptedDeployments = decision.attempts.map((attempt) => attempt.deployment);
-    const eligible = candidates.filter((candidate) => candidate.enabled && inScope(candidate.deployment, scope) && inSubset(candidate.deployment, subset) && !sameDeployment(candidate.deployment, current) && !attemptedDeployments.some((attempted) => sameDeployment(candidate.deployment, attempted)) && validationEligible(candidate, now2).length === 0 && this.passedInitialHardFilters(decision, candidate) && this.circuitState(candidate.deployment, now2, false) !== "open");
+    const eligible = candidates.filter((candidate) => candidate.enabled && candidate.deployment.providerId === currentProviderId && inScope(candidate.deployment, scope) && inSubset(candidate.deployment, subset) && !sameDeployment(candidate.deployment, current) && !attemptedDeployments.some((attempted) => sameDeployment(candidate.deployment, attempted)) && validationEligible(candidate, now2).length === 0 && this.passedInitialHardFilters(decision, candidate) && this.circuitState(candidate.deployment, now2, false) !== "open");
     const equivalent = currentGroup ? eligible.find((candidate) => candidate.deployment.logicalModelGroupId === currentGroup && !candidate.fallbackOnly) : void 0;
     const fallback = eligible.find((candidate) => candidate.fallbackOnly);
     const rankedAlternates = decision.rankedCandidates.map((ranked) => eligible.find((candidate) => sameDeployment(candidate.deployment, ranked.deployment))).filter((candidate) => !!candidate && candidate !== equivalent && candidate !== fallback && !candidate.fallbackOnly);
@@ -346730,7 +346760,8 @@ ${String(event.toolArgs || "")}`;
       sequence,
       status: input.status,
       guide: !isToolEvent && input.guide ? this.normalizeGuideReceipt(input.guide) : void 0,
-      displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : void 0
+      displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : void 0,
+      fallback: input.fallback
     };
     if (activeRun && this.isPersistablePublicWorkEvent(event)) {
       activeRun.sequence = Number(sequence || activeRun.sequence + 1);
@@ -349649,9 +349680,15 @@ ${summary}`, segment, "local-summarize", true);
   }
   compressionBuildBlockStart(messages) {
     const activeRunId = this.currentWorkRunId();
-    if (!activeRunId) return 0;
-    const index = messages.findIndex((message) => String(message.run_id || message.runId || "") === activeRunId);
-    return index >= 0 ? index : 0;
+    if (activeRunId) {
+      const index = messages.findIndex((message) => String(message.run_id || message.runId || "") === activeRunId);
+      if (index >= 0) return index;
+    }
+    let lastRunBoundary = -1;
+    for (let index = 0; index < messages.length; index += 1) {
+      if (String(messages[index]?.run_id || messages[index]?.runId || "")) lastRunBoundary = index + 1;
+    }
+    return lastRunBoundary >= 0 ? lastRunBoundary : messages.length;
   }
   recentContextSuffix(messages, maxMessages, tokenBudget) {
     if (!messages.length) return [];
@@ -350580,17 +350617,12 @@ ${msg.content}
         currentAttempt.sideEffectBoundary = this.routeSideEffectCommitted;
         if (this.routeAttemptStartedAt > 0) currentAttempt.durationMs = Math.max(0, Date.now() - this.routeAttemptStartedAt);
       }
-      if (!fallbackEnabled) {
-        this.lastRouteDecision.finalStatus = "failed";
-        this.routeAttemptStartedAt = 0;
-        this.persistRouteDecision(this.lastRouteDecision);
-        return null;
-      }
       if (!this.pendingAutoAttempts.length) {
         this.pendingAutoAttempts = this.autoRouter.planAttempts(this.lastRouteDecision, this.autoRouteCandidates(), {
           error: failure,
           streamCommitted: this.routeStreamCommitted,
-          sideEffectCommitted: this.routeSideEffectCommitted
+          sideEffectCommitted: this.routeSideEffectCommitted,
+          allowModelFallback: fallbackEnabled
         });
       }
       const next2 = this.pendingAutoAttempts.shift();
@@ -350625,7 +350657,6 @@ ${msg.content}
     if (!next?.name) return null;
     this.model = next.name;
     this.fixedDeployment = this.deploymentRef(next);
-    if (next.provider) this.config.set("models", "auto_switch_anchor_provider", next.provider);
     return current;
   }
   isLlmErrorText(text) {
@@ -350633,12 +350664,10 @@ ${msg.content}
   }
   scopedSwitchModels(currentModelName) {
     const all = this.config.allModels();
-    if (this.config.autoSwitchScope() !== "provider") return all;
-    const current = currentModelName === "auto" ? void 0 : this.config.findModel(currentModelName);
-    const providerId = current?.provider_id || this.config.autoSwitchAnchorProvider() || this.config.findModel(this.config.getStr("models", "default_model"))?.provider_id || all[0]?.provider_id || "";
-    if (!providerId) return all;
-    const provider = this.config.findProvider(providerId);
-    return all.filter((m2) => m2.provider_id === (provider?.id || providerId));
+    const current = currentModelName === "auto" ? this.activeModelConfig() : this.config.findModel(currentModelName);
+    const providerId = current?.provider_id || this.fixedDeployment?.providerId || (currentModelName !== "auto" ? this.activeDeployment()?.providerId : void 0);
+    if (!providerId) return [];
+    return all.filter((m2) => m2.provider_id === providerId);
   }
   async validateModels(selectedNames, options = {}) {
     if (this.modelValidationPromise) return this.modelValidationPromise;
@@ -350820,6 +350849,60 @@ ${msg.content}
     };
     return results;
   }
+  /**
+   * Final visual safety net: OCR each submitted image and ask a text-only
+   * request to conservatively repair the OCR. This is intentionally callable
+   * only after a visual-input refusal and after same-provider vision routing
+   * has been exhausted; the original image is never sent again.
+   */
+  async finalVisualFallback(errorText, signal) {
+    if (!/(?:vision|image|multimodal|image_url|input_image).*(?:not supported|unsupported|拒绝|不支持|failed|failure|invalid)|(?:not supported|unsupported|拒绝|不支持).*(?:vision|image|multimodal|image_url|input_image)/i.test(String(errorText || ""))) return null;
+    const current = this.activeModelConfig();
+    if (!current) return null;
+    const alternateVision = this.config.allModels().some(
+      (model) => model.enabled !== false && model.provider_id === current.provider_id && model.name !== current.name && !!model.vision && !!model.api_key && !!model.provider_url && !["unavailable", "auth_error", "invalid_config"].includes(String(model.evaluation?.status || model.validation?.status || "").toLowerCase())
+    );
+    if (alternateVision) return null;
+    const latest = [...this.history].reverse().find((item) => item?.role === "user");
+    const parts = latest?.content && Array.isArray(latest.content) ? latest.content : [];
+    const images = parts.map((part) => {
+      const image = part.image_url;
+      return image && typeof image === "object" ? String(image.url || "") : "";
+    }).filter((value) => /^data:image\/(?:png|jpeg);base64,/i.test(value)).slice(0, 4);
+    if (!images.length) return null;
+    const ocr = [];
+    for (const [index, image] of images.entries()) {
+      try {
+        const result = await this.tools.finalVisualFallbackOcr(image, signal);
+        if (result.ok && result.text.trim()) ocr.push({ index: index + 1, text: result.text.slice(0, 5e4), confidence: result.confidence });
+      } catch {
+      }
+    }
+    if (!ocr.length) return JSON.stringify({ ok: false, fallback: "mini_ocr_llm", error: "Local OCR returned no readable text; no visual content was fabricated." });
+    const task = typeof latest?.content === "string" ? latest.content : "";
+    const evidence = ocr.map((item) => `Image ${item.index} (OCR confidence ${item.confidence.toFixed(1)}):
+${item.text}`).join("\n\n");
+    const prompt = `The provider rejected image input. Answer the user's task using only this approximate OCR evidence. Correct obvious character, spacing, and line-break errors only when supported by context. Preserve [uncertain] markers for ambiguity and never invent missing visual content.
+User task:
+${task.slice(0, 12e3)}
+OCR evidence:
+${evidence}`;
+    let corrected = "";
+    try {
+      const provider = this.engineModel();
+      if (provider) corrected = String(await provider.chat(this.activeModelName(), [{ role: "user", content: prompt }], "You are a text-only OCR correction assistant. Be conservative and explicit about uncertainty.", 0.05, 3e3, signal) || "").trim();
+    } catch {
+    }
+    return JSON.stringify({
+      ok: !!(corrected || ocr.length),
+      fallback: "mini_ocr_llm",
+      approximate: true,
+      warning: "\u89C6\u89C9\u8F93\u5165\u88AB\u62D2\u7EDD\uFF1B\u4EE5\u4E0B\u5185\u5BB9\u6765\u81EA\u672C\u5730 OCR\uFF0C\u5E76\u7ECF\u6587\u672C\u6A21\u578B\u4FDD\u5B88\u6821\u6B63\uFF0C\u53EF\u80FD\u4E0D\u5B8C\u6574\u3002",
+      raw_ocr: ocr,
+      corrected: corrected || ocr.map((item) => item.text).join("\n\n"),
+      uncertainty: corrected ? "preserved" : "raw_ocr_only"
+    }, null, 2);
+  }
   engineModel() {
     if (this.forcedProvider) {
       const active = this.activeDeployment();
@@ -351202,8 +351285,10 @@ ${persisted.map((guide, index) => `Guide ${index + 1}: ${guide.text}`).join("\n"
       }
       const selectedModel = this.activeModelConfig();
       if (images.length && !selectedModel?.vision) {
-        this.status = "idle";
-        return [{ type: "text", text: `[Vision unavailable] ${this.activeModelName() || this.model} has not passed image-input validation. Select a validated vision model before asking about attachments.` }];
+        const hasSameProviderVision = selectedModel && this.config.allModels().some(
+          (model) => model.enabled !== false && model.provider_id === selectedModel.provider_id && model.name !== selectedModel.name && !!model.vision
+        );
+        if (hasSameProviderVision) this.switchToFallbackModel("vision input not supported by the selected model");
       }
       const now2 = this.nowLabel();
       const visibleUserInput = inputEnvelope?.visibleUserInput === void 0 ? text : String(inputEnvelope.visibleUserInput || "");
@@ -351964,7 +352049,10 @@ ${items.map((item) => this.formatAutomation(item)).join("\n")}`;
     if (!sa) return "[Subagent] Not found.";
     const requestedModel = this.normalizeSubagentModelSelection(sa.model && sa.model !== "default" ? sa.model : this.modelSelectionValue() || this.model);
     const requestedDeployment = parseDeploymentSelectionValue2(requestedModel);
-    const assignedModel = requestedModel === "auto" ? this.activeModelConfig() : requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel);
+    let assignedModel = requestedModel === "auto" ? this.activeModelConfig() : requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel);
+    if (!assignedModel && !requestedDeployment && requestedModel !== "auto") {
+      assignedModel = this.activeModelConfig();
+    }
     const model = assignedModel?.name || (requestedModel === "auto" ? this.activeModelName() : requestedModel);
     const activeModel = this.activeModelConfig();
     const activeProvider = this.engineModel();
@@ -353393,6 +353481,32 @@ var ConversationKernel = class {
       followUp: queued?.followUp.slice() || []
     };
   }
+  /**
+   * dev-0.5.6: 排队 followUp 消息 drain 时转成普通用户消息载荷。
+   *
+   * 入队时（enqueueSameSession）为去重/编辑保留了 clientMessageId，并把
+   * runId 固定为入队时运行的 runId；若原样传给 process，写入 chatMessages
+   * 的用户消息会携带旧 runId + clientMessageId，渲染端（PC renderTranscript /
+   * 移动端 projectRemoteConversationItems）会把「clientMessageId + runId 命中
+   * workRuns」的消息误判为 guide 消息而跳过气泡，导致排队自动发送的消息不
+   * 显示在用户输入时间线。清除这两个字段后 process 走普通用户消息分支
+   * （chatMessages.push，mode 取 visibleMode，runId 取当前 run）。
+   */
+  drainQueuedFollowUpMessage(message) {
+    if (typeof message === "string") return message;
+    const text = String(message.text || "");
+    const images = message.images;
+    const attachments = message.attachments;
+    const visible = message.visibleUserInput;
+    const visibleMode = message.visibleMode;
+    return {
+      text,
+      ...images?.length ? { images } : {},
+      ...attachments?.length ? { attachments } : {},
+      ...visible ? { visibleUserInput: visible } : {},
+      ...visibleMode ? { visibleMode } : {}
+    };
+  }
   queueItems(target) {
     const runtime = this.findRuntime(target);
     if (!runtime) return [];
@@ -354134,7 +354248,7 @@ ${batchText}`,
           };
           lastTokens = await this.runSingle(runtime, batchMessage, "steer");
         } else {
-          lastTokens = await this.runSingle(runtime, next.message, next.queueMode);
+          lastTokens = await this.runSingle(runtime, this.drainQueuedFollowUpMessage(next.message), next.queueMode);
         }
       }
       const rootMessage = runtime.runner.subagents.readRootInbox()[0];
@@ -354593,7 +354707,11 @@ ${text}`;
       type: "queue_update",
       content: "Conversation queue updated.",
       conversationId: runtime.id,
-      queue: this.queued(runtime.target)
+      queue: this.queued(runtime.target),
+      // Structured rows with stable kernel ids so every consumer (PC UI and
+      // the paired mobile client) can render/edit/delete the same items.
+      queueItems: this.queueItems(runtime.target),
+      queuePaused: runtime.queuePaused === true
     });
   }
   clearQueued(runtime) {

package/dist/core/agent.d.ts

@@ -990,6 +990,13 @@ export declare class Agent {
     isModelValidationRunning(): boolean;
     modelValidationStatus(): ModelValidationProgress;
     private runModelValidation;
+    /**
+     * Final visual safety net: OCR each submitted image and ask a text-only
+     * request to conservatively repair the OCR. This is intentionally callable
+     * only after a visual-input refusal and after same-provider vision routing
+     * has been exhausted; the original image is never sent again.
+     */
+    finalVisualFallback(errorText: string, signal?: AbortSignal): Promise<string | null>;
     engineModel(): LLMProvider | null;
     /**
      * dev-0.4.3 模型原生思考强度档位映射表（模型名 → thinking_tier_map）。

package/dist/core/agent.js

@@ -2298,6 +2298,7 @@ class Agent {
             status: input.status,
             guide: !isToolEvent && input.guide ? this.normalizeGuideReceipt(input.guide) : undefined,
             displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : undefined,
+            fallback: input.fallback,
         };
         if (activeRun && this.isPersistablePublicWorkEvent(event)) {
             activeRun.sequence = Number(sequence || activeRun.sequence + 1);
@@ -5587,10 +5588,22 @@ class Agent {
     }
     compressionBuildBlockStart(messages) {
         const activeRunId = this.currentWorkRunId();
-        if (!activeRunId)
-            return 0;
-        const index = messages.findIndex(message => String(message.run_id || message.runId || '') === activeRunId);
-        return index >= 0 ? index : 0;
+        if (activeRunId) {
+            const index = messages.findIndex(message => String(message.run_id || message.runId || '') === activeRunId);
+            if (index >= 0)
+                return index;
+        }
+        // dev-0.5.6: 空闲对话（无活动 run）时不得把全部历史算进当前 Build Block，
+        // 否则上下文显示窗口的长期历史恒为 0。回退语义：
+        // - 存在带 run_id 的消息：boundary 取最后一个 run 的起点之后（该 run 及其
+        //   之前的历史属于长期历史，其后无归属的消息属于当前未命名区块）；
+        // - 完全没有 run_id：全部历史都是长期历史（不存在当前 Build Block）。
+        let lastRunBoundary = -1;
+        for (let index = 0; index < messages.length; index += 1) {
+            if (String(messages[index]?.run_id || messages[index]?.runId || ''))
+                lastRunBoundary = index + 1;
+        }
+        return lastRunBoundary >= 0 ? lastRunBoundary : messages.length;
     }
     recentContextSuffix(messages, maxMessages, tokenBudget) {
         if (!messages.length)
@@ -6651,17 +6664,12 @@ class Agent {
                 if (this.routeAttemptStartedAt > 0)
                     currentAttempt.durationMs = Math.max(0, Date.now() - this.routeAttemptStartedAt);
             }
-            if (!fallbackEnabled) {
-                this.lastRouteDecision.finalStatus = 'failed';
-                this.routeAttemptStartedAt = 0;
-                this.persistRouteDecision(this.lastRouteDecision);
-                return null;
-            }
             if (!this.pendingAutoAttempts.length) {
                 this.pendingAutoAttempts = this.autoRouter.planAttempts(this.lastRouteDecision, this.autoRouteCandidates(), {
                     error: failure,
                     streamCommitted: this.routeStreamCommitted,
                     sideEffectCommitted: this.routeSideEffectCommitted,
+                    allowModelFallback: fallbackEnabled,
                 });
             }
             const next = this.pendingAutoAttempts.shift();
@@ -6702,8 +6710,11 @@ class Agent {
             return null;
         this.model = next.name;
         this.fixedDeployment = this.deploymentRef(next);
-        if (next.provider)
-            this.config.set('models', 'auto_switch_anchor_provider', next.provider);
+        // Do NOT rewrite auto_switch_anchor_provider here. The anchor describes
+        // the user's Auto-routing scope; a fixed-model recovery must never
+        // repoint it (a single transient failure would otherwise silently move
+        // every future Auto route — and every future fallback pool — to the
+        // recovery provider, which is the cross-provider leak users observed).
         return current;
     }
     isLlmErrorText(text) {
@@ -6711,18 +6722,23 @@ class Agent {
     }
     scopedSwitchModels(currentModelName) {
         const all = this.config.allModels();
-        if (this.config.autoSwitchScope() !== 'provider')
-            return all;
-        const current = currentModelName === 'auto' ? undefined : this.config.findModel(currentModelName);
-        const providerId = current?.provider_id ||
-            this.config.autoSwitchAnchorProvider() ||
-            this.config.findModel(this.config.getStr('models', 'default_model'))?.provider_id ||
-            all[0]?.provider_id ||
-            '';
+        // Fixed-model recovery is always provider-local. A same-named model on a
+        // different provider is a different deployment and must never receive the
+        // failed provider's credentials or become an implicit fallback.
+        // Resolution order is strictly deployment-first: the pinned deployment,
+        // then the exact active deployment. The global auto anchor is deliberately
+        // NOT consulted here — it belongs to Auto routing only and can point at a
+        // different provider (for example after a previous cross-provider switch),
+        // which is exactly the same-name cross-provider leak this guard exists to
+        // prevent. When no provider can be established unambiguously, return an
+        // empty pool: failing the switch is safe, crossing providers is not.
+        const current = currentModelName === 'auto' ? this.activeModelConfig() : this.config.findModel(currentModelName);
+        const providerId = current?.provider_id
+            || this.fixedDeployment?.providerId
+            || (currentModelName !== 'auto' ? this.activeDeployment()?.providerId : undefined);
         if (!providerId)
-            return all;
-        const provider = this.config.findProvider(providerId);
-        return all.filter(m => m.provider_id === (provider?.id || providerId));
+            return [];
+        return all.filter(m => m.provider_id === providerId);
     }
     async validateModels(selectedNames, options = {}) {
         if (this.modelValidationPromise)
@@ -6913,6 +6929,62 @@ class Agent {
         };
         return results;
     }
+    /**
+     * Final visual safety net: OCR each submitted image and ask a text-only
+     * request to conservatively repair the OCR. This is intentionally callable
+     * only after a visual-input refusal and after same-provider vision routing
+     * has been exhausted; the original image is never sent again.
+     */
+    async finalVisualFallback(errorText, signal) {
+        if (!/(?:vision|image|multimodal|image_url|input_image).*(?:not supported|unsupported|拒绝|不支持|failed|failure|invalid)|(?:not supported|unsupported|拒绝|不支持).*(?:vision|image|multimodal|image_url|input_image)/i.test(String(errorText || '')))
+            return null;
+        const current = this.activeModelConfig();
+        if (!current)
+            return null;
+        const alternateVision = this.config.allModels().some(model => model.enabled !== false && model.provider_id === current.provider_id &&
+            model.name !== current.name && !!model.vision && !!model.api_key && !!model.provider_url &&
+            !['unavailable', 'auth_error', 'invalid_config'].includes(String(model.evaluation?.status || model.validation?.status || '').toLowerCase()));
+        if (alternateVision)
+            return null;
+        const latest = [...this.history].reverse().find(item => item?.role === 'user');
+        const parts = latest?.content && Array.isArray(latest.content) ? latest.content : [];
+        const images = parts.map(part => {
+            const image = part.image_url;
+            return image && typeof image === 'object' ? String(image.url || '') : '';
+        }).filter(value => /^data:image\/(?:png|jpeg);base64,/i.test(value)).slice(0, 4);
+        if (!images.length)
+            return null;
+        const ocr = [];
+        for (const [index, image] of images.entries()) {
+            try {
+                const result = await this.tools.finalVisualFallbackOcr(image, signal);
+                if (result.ok && result.text.trim())
+                    ocr.push({ index: index + 1, text: result.text.slice(0, 50_000), confidence: result.confidence });
+            }
+            catch { }
+        }
+        if (!ocr.length)
+            return JSON.stringify({ ok: false, fallback: 'mini_ocr_llm', error: 'Local OCR returned no readable text; no visual content was fabricated.' });
+        const task = typeof latest?.content === 'string' ? latest.content : '';
+        const evidence = ocr.map(item => `Image ${item.index} (OCR confidence ${item.confidence.toFixed(1)}):\n${item.text}`).join('\n\n');
+        const prompt = `The provider rejected image input. Answer the user's task using only this approximate OCR evidence. Correct obvious character, spacing, and line-break errors only when supported by context. Preserve [uncertain] markers for ambiguity and never invent missing visual content.\nUser task:\n${task.slice(0, 12_000)}\nOCR evidence:\n${evidence}`;
+        let corrected = '';
+        try {
+            const provider = this.engineModel();
+            if (provider)
+                corrected = String(await provider.chat(this.activeModelName(), [{ role: 'user', content: prompt }], 'You are a text-only OCR correction assistant. Be conservative and explicit about uncertainty.', 0.05, 3000, signal) || '').trim();
+        }
+        catch { }
+        return JSON.stringify({
+            ok: !!(corrected || ocr.length),
+            fallback: 'mini_ocr_llm',
+            approximate: true,
+            warning: '视觉输入被拒绝；以下内容来自本地 OCR，并经文本模型保守校正，可能不完整。',
+            raw_ocr: ocr,
+            corrected: corrected || ocr.map(item => item.text).join('\n\n'),
+            uncertainty: corrected ? 'preserved' : 'raw_ocr_only',
+        }, null, 2);
+    }
     engineModel() {
         if (this.forcedProvider) {
             const active = this.activeDeployment();
@@ -7353,8 +7425,13 @@ class Agent {
             }
             const selectedModel = this.activeModelConfig();
             if (images.length && !selectedModel?.vision) {
-                this.status = 'idle';
-                return [{ type: 'text', text: `[Vision unavailable] ${this.activeModelName() || this.model} has not passed image-input validation. Select a validated vision model before asking about attachments.` }];
+                // Give the normal route planner first chance to select another
+                // same-provider vision deployment. If none is available, the kernel
+                // preflight invokes the final mini-OCR + text-only correction path.
+                const hasSameProviderVision = selectedModel && this.config.allModels().some(model => model.enabled !== false && model.provider_id === selectedModel.provider_id &&
+                    model.name !== selectedModel.name && !!model.vision);
+                if (hasSameProviderVision)
+                    this.switchToFallbackModel('vision input not supported by the selected model');
             }
             const now = this.nowLabel();
             const visibleUserInput = inputEnvelope?.visibleUserInput === undefined
@@ -8163,9 +8240,17 @@ class Agent {
             return '[Subagent] Not found.';
         const requestedModel = this.normalizeSubagentModelSelection(sa.model && sa.model !== 'default' ? sa.model : this.modelSelectionValue() || this.model);
         const requestedDeployment = parseDeploymentSelectionValue(requestedModel);
-        const assignedModel = requestedModel === 'auto'
+        let assignedModel = requestedModel === 'auto'
             ? this.activeModelConfig()
             : (requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel));
+        // Imported presets (and stale records) may reference a bare model name
+        // that no longer exists in the configured catalog. Resolving by name must
+        // never guess a provider; instead the peer inherits the parent deployment
+        // so the job can still run. Qualified deployment values keep failing
+        // closed — they are exact references, not guesses.
+        if (!assignedModel && !requestedDeployment && requestedModel !== 'auto') {
+            assignedModel = this.activeModelConfig();
+        }
         const model = assignedModel?.name || (requestedModel === 'auto' ? this.activeModelName() : requestedModel);
         const activeModel = this.activeModelConfig();
         const activeProvider = this.engineModel();

package/dist/core/agentKernelRunner.js

@@ -429,9 +429,24 @@ async function runAgentKernel(agent) {
     try {
         const linkedPlanRevisionBeforeRun = agent.getLinkedPlan().revision;
         const modelBeforeKernelRun = agent.model;
-        let lastTurn = await runWithCompressionResume([], false);
+        const preflightVisualFallback = !agent.activeModelConfig()?.vision
+            ? await agent.finalVisualFallback('vision input not supported by the selected model', processSignal)
+            : null;
+        let lastTurn = preflightVisualFallback
+            ? { text: preflightVisualFallback, stopReason: 'stop', errorMessage: '' }
+            : await runWithCompressionResume([], false);
+        if (preflightVisualFallback) {
+            tokens.push({ type: 'text', text: preflightVisualFallback });
+            agent.recordWorkStatus('Final visual fallback used: local mini OCR plus conservative text correction.');
+        }
         if (modelBeforeKernelRun && modelBeforeKernelRun !== agent.model && !tokens.some(t => t.text?.includes('[Model fallback]'))) {
-            tokens.unshift({ type: 'text', text: `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.` });
+            const notice = `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.`;
+            tokens.unshift({ type: 'text', text: notice });
+            agent.emitWorkEvent({
+                type: 'status',
+                content: notice,
+                fallback: { from: modelBeforeKernelRun, to: agent.model, providerId: agent.activeDeployment()?.providerId },
+            });
         }
         let emptyResponseRetries = 0;
         while (providerTurnIsEmpty(lastTurn) && emptyResponseRetries < 2) {
@@ -453,6 +468,11 @@ async function runAgentKernel(agent) {
             const notice = routeTransitionNotice(agent, previous);
             tokens.push({ type: 'text', text: notice });
             agent.recordWorkStatus(notice);
+            agent.emitWorkEvent({
+                type: 'status',
+                content: notice,
+                fallback: { from: previous, to: agent.model, providerId: agent.activeDeployment()?.providerId },
+            });
             kernel.state.model = toKernelModel(agent);
             const fallbackToolSurface = refreshToolSurface(true);
             kernel.state.systemPrompt = [agent.buildSystemPrompt(), fallbackToolSurface.systemPromptNotice].filter(Boolean).join('\n\n');
@@ -460,6 +480,14 @@ async function runAgentKernel(agent) {
             await agent.waitForPlannedRouteRetry();
             lastTurn = await runWithCompressionResume([], false);
         }
+        if (kernelTurnFailed(agent, lastTurn)) {
+            const visualFallback = await agent.finalVisualFallback(lastTurn.errorMessage || lastTurn.text, processSignal);
+            if (visualFallback) {
+                tokens.push({ type: 'text', text: visualFallback });
+                agent.recordWorkStatus('Final visual fallback used: local mini OCR plus conservative text correction.');
+                lastTurn = { ...lastTurn, text: visualFallback, errorMessage: '', stopReason: 'stop' };
+            }
+        }
         if (kernelTurnFailed(agent, lastTurn)) {
             throw new ProviderRunError(normalizePublicProviderError(lastTurn.errorMessage || lastTurn.text, [agent.activeModelConfig()?.api_key]));
         }

package/dist/core/autoRouter.d.ts

@@ -153,6 +153,7 @@ export declare class AutoRouter {
         error: RouteFailure;
         streamCommitted: boolean;
         sideEffectCommitted: boolean;
+        allowModelFallback?: boolean;
     }): PlannedRouteAttempt[];
     recordEndpointFailure(deployment: DeploymentRef, failure: RouteFailureType | RouteFailure): void;
     recordEndpointSuccess(deployment: DeploymentRef, latencyMs?: number, throughput?: number): void;

package/dist/core/autoRouter.js

@@ -290,15 +290,24 @@ class AutoRouter {
                 retryDelayMs,
             });
         }
+        if (failure.allowModelFallback === false)
+            return attempts.slice(0, remainingAttempts);
         const selection = decision.requestedSelection;
         if (selection.kind === 'fixed')
             return attempts.slice(0, remainingAttempts);
         const scope = selection.kind === 'auto' ? selection.scope : { kind: 'provider', providerId: current.providerId };
         const subset = selection.kind === 'auto' ? selection.subset : undefined;
         const currentGroup = current.logicalModelGroupId;
+        const currentProviderId = current.providerId;
         const now = this.now();
         const attemptedDeployments = decision.attempts.map(attempt => attempt.deployment);
+        // Fallback is a recovery operation, not a new global routing decision.
+        // It must never cross the provider boundary, even when the original Auto
+        // selection has global scope. This prevents equal model ids from silently
+        // switching credentials/endpoints (for example provider A/model X to
+        // provider B/model X).
         const eligible = candidates.filter(candidate => candidate.enabled
+            && candidate.deployment.providerId === currentProviderId
             && inScope(candidate.deployment, scope)
             && inSubset(candidate.deployment, subset)
             && !sameDeployment(candidate.deployment, current)

package/dist/core/conversationKernel.d.ts

@@ -194,6 +194,18 @@ export declare class ConversationKernel {
         steering: string[];
         followUp: string[];
     };
+    /**
+     * dev-0.5.6: 排队 followUp 消息 drain 时转成普通用户消息载荷。
+     *
+     * 入队时（enqueueSameSession）为去重/编辑保留了 clientMessageId，并把
+     * runId 固定为入队时运行的 runId；若原样传给 process，写入 chatMessages
+     * 的用户消息会携带旧 runId + clientMessageId，渲染端（PC renderTranscript /
+     * 移动端 projectRemoteConversationItems）会把「clientMessageId + runId 命中
+     * workRuns」的消息误判为 guide 消息而跳过气泡，导致排队自动发送的消息不
+     * 显示在用户输入时间线。清除这两个字段后 process 走普通用户消息分支
+     * （chatMessages.push，mode 取 visibleMode，runId 取当前 run）。
+     */
+    private drainQueuedFollowUpMessage;
     queueItems(target: ConversationTargetInput): ConversationQueueItemSnapshot[];
     enqueueNext(target: ConversationTargetInput, input: {
         id: string;