@dickpy/dsh-imagegen 1.2.1 → 1.2.3

package/lib/index.js

@@ -6,7 +6,6 @@ import { homedir } from "node:os";
 import path from "node:path";
 import { fileURLToPath } from "node:url";
 import { spawn } from "node:child_process";
-import { createUserMessage } from "@deepseek-ai/dsh-llm/message";
 import { defineTool } from "@deepseek-ai/dsh-tools";
 //#region src/protocol.ts
 /**
@@ -17,7 +16,7 @@ import { defineTool } from "@deepseek-ai/dsh-tools";
 /** Settings namespace this plugin owns (host settings seam + bridge). */
 const IMAGEGEN_SETTINGS_NAMESPACE = "dsh-imagegen";
 /** Published package version shared by the host updater and the client UI. */
-const PLUGIN_VERSION = "1.2.1";
+const PLUGIN_VERSION = "1.2.3";
 /** Same-origin route family (loopback-only, mirroring the dsh-ssh fence). */
 const SETTINGS_API = {
 	describe: "/api/dsh-imagegen/settings/describe",
@@ -141,7 +140,14 @@ async function enhancePrompt(config, prompt) {
 * `/models` exposes candidates only: the configured list is the explicit
 * allow-list because OpenAI-compatible gateways rarely advertise modalities.
 */
-const DEFAULT_IMAGE_MODELS = ["gpt-image-2", "grok-imagine-image"];
+const DEFAULT_IMAGE_MODELS = [
+	"gpt-image-2",
+	"grok-imagine-image",
+	"nanobanana2",
+	"nanobanana2-lite",
+	"nanobanana-pro",
+	"seedream-5.0-pro"
+];
 /** Normalize user-entered model identifiers and retain a usable legacy default. */
 function normalizeImageModels(value) {
 	const candidates = Array.isArray(value) ? value : [];
@@ -184,6 +190,24 @@ const DALLE3_SIZES = /* @__PURE__ */ new Set([
 function isGrokImagine(model) {
 	return /^grok-imagine(?:-|$)/.test(model);
 }
+/** Whether the model belongs to the Google Nano Banana family (nanobanana2 /
+*  nanobanana2-lite / nanobanana-pro, plus the official Gemini image IDs the
+*  gateways expose). OpenAI-compatible gateways serve these with their own
+*  aspect_ratio / image_size vocabulary instead of the OpenAI size/quality
+*  passthrough. */
+function isNanoBanana(model) {
+	if (/^nanobanana/i.test(model)) return true;
+	return model === "gemini-3-pro-image" || model === "gemini-3-pro-image-preview" || model === "gemini-3.1-flash-image" || model === "gemini-3.1-flash-image-preview" || model === "gemini-3.1-flash-lite-image" || model === "gemini-2.5-flash-image";
+}
+/** Whether the model belongs to the ByteDance Seedream family (seedream-5.0-pro,
+*  seedream-5.0, seedream-4.x, doubao-seedream-…). OpenAI-compatible gateways
+*  serve Seedream through a unified generate-and-edit architecture:
+*  generation AND editing both go to /images/generations, reference images are
+*  a JSON URL / data-URL array, and the clarity tier is `resolution` while
+*  `size` carries the aspect ratio (or exact pixels). */
+function isSeedream(model) {
+	return /^(?:doubao-)?seedream/i.test(model);
+}
 /** The panel's aspect ratios mapped to the closest OpenAI pixel size
 *  (gpt-image-2 / generic OpenAI-compatible endpoints). */
 const OPENAI_SIZE_BY_RATIO = {
@@ -274,6 +298,18 @@ function effectiveParams(request) {
 		...request.quality !== "" && request.quality !== "auto" ? { resolution: request.quality === "4k" ? "2k" : request.quality } : {},
 		response_format: "b64_json"
 	};
+	if (isNanoBanana(model)) return {
+		model,
+		...request.size !== "" && request.size !== "auto" ? { aspect_ratio: request.size } : {},
+		...request.quality !== "" && request.quality !== "auto" ? { image_size: request.quality.toUpperCase() } : {},
+		response_format: "b64_json"
+	};
+	if (isSeedream(model)) return {
+		model,
+		...request.size !== "" && request.size !== "auto" ? { size: request.size } : {},
+		...request.quality !== "" && request.quality !== "auto" ? { resolution: request.quality === "4k" ? "2K" : request.quality.toUpperCase() } : {},
+		response_format: "b64_json"
+	};
 	return {
 		model,
 		...request.size !== "" && request.size !== "auto" && OPENAI_SIZE_BY_RATIO[request.size] !== void 0 ? { size: OPENAI_SIZE_BY_RATIO[request.size] } : {},
@@ -359,6 +395,24 @@ async function requestOneImage(baseUrl, upstream, request, params, signal) {
 				...params.aspect_ratio !== void 0 ? { aspect_ratio: params.aspect_ratio } : {},
 				response_format: "b64_json"
 			});
+		} else if (isNanoBanana(params.model)) {
+			const form = new FormData();
+			form.append("image", new Blob([bytes], { type: parsed.mime }), `reference.${extensionOf$2(parsed.mime)}`);
+			form.append("prompt", request.prompt);
+			form.append("model", params.model);
+			if (params.aspect_ratio !== void 0) form.append("aspect_ratio", params.aspect_ratio);
+			if (params.image_size !== void 0) form.append("image_size", params.image_size);
+			body = form;
+		} else if (isSeedream(params.model)) {
+			headers["content-type"] = "application/json";
+			body = JSON.stringify({
+				model: params.model,
+				prompt: request.prompt,
+				image: [request.image],
+				...params.size !== void 0 ? { size: params.size } : {},
+				...params.resolution !== void 0 ? { resolution: params.resolution } : {},
+				response_format: "b64_json"
+			});
 		} else {
 			const form = new FormData();
 			form.append("image", new Blob([bytes], { type: parsed.mime }), `reference.${extensionOf$2(parsed.mime)}`);
@@ -379,7 +433,8 @@ async function requestOneImage(baseUrl, upstream, request, params, signal) {
 	const budget = requestSignal(signal, UPSTREAM_TIMEOUT_MS);
 	let response;
 	try {
-		response = await fetch(`${baseUrl}/images/${request.mode === "edit" ? "edits" : "generations"}`, {
+		const endpoint = request.mode === "edit" && !isSeedream(params.model) ? "/images/edits" : "/images/generations";
+		response = await fetch(`${baseUrl}${endpoint}`, {
 			method: "POST",
 			headers,
 			body,
@@ -2297,6 +2352,8 @@ const taskResultSchema = {
 		}
 	}
 };
+/** Agent calls stay pending until the provider and history write settle. */
+const AGENT_GENERATION_TIMEOUT_MS = 3e5;
 function acceptedMediaType(value) {
 	return value === "image/png" || value === "image/jpeg" || value === "image/webp" || value === "image/gif";
 }
@@ -2337,7 +2394,6 @@ function renderTaskResult(value) {
 /** Register the global Agent tools and unregister them with the plugin lifecycle. */
 function registerAgentImageTools(ctx, runtime, resolve) {
 	const attachmentRefs = /* @__PURE__ */ new Map();
-	const taskSubscriptions = /* @__PURE__ */ new Set();
 	const ensureConfigured = () => {
 		const config = resolve();
 		if (!config.enabled) throw new ImageGenError("AI image generation is disabled. Open Settings > Plugins > AI Image and enable it.", "plugin-disabled");
@@ -2366,7 +2422,7 @@ function registerAgentImageTools(ctx, runtime, resolve) {
 		return {
 			task_id: task.id,
 			status: task.status,
-			message: task.status === "completed" ? "Generation completed. The images are attached below and can be reused as source_image in edit_image." : task.status === "failed" ? "Generation failed." : task.status === "cancelled" ? "Generation was cancelled." : "Generation is still running. Completion will be delivered to the conversation automatically with image attachments.",
+			message: task.status === "completed" ? "Generation completed. The images are attached below and can be reused as source_image in edit_image." : task.status === "failed" ? "Generation failed." : task.status === "cancelled" ? "Generation was cancelled." : "Generation is still running. Query the task again when you need its current status.",
 			...task.error === void 0 ? {} : { error: task.error },
 			images
 		};
@@ -2376,39 +2432,66 @@ function registerAgentImageTools(ctx, runtime, resolve) {
 		if (task === void 0) throw new ImageGenError(`Image generation task ${id} was not found.`, "task-not-found");
 		return task;
 	};
-	const notifyCompletion = async (agent, task) => {
-		const result = await taskResult(task);
-		const completed = task.status === "completed";
-		const text = completed ? `图像生成任务已完成（${task.id}）。图片已附在这条消息中，可以直接查看、下载或作为后续图生图的参考。` : task.status === "cancelled" ? `图像生成任务已取消（${task.id}）。` : `图像生成任务失败（${task.id}）：${task.error ?? "未知错误"}`;
-		agent.send(createUserMessage({
-			content: [{
-				type: "text",
-				text
-			}, ...completed ? result.images.map((image) => ({
-				type: "image",
-				attachment: restoreRef(image)
-			})) : []],
-			source: { kind: "user" }
-		}), "next-turn", true);
-	};
-	const watchTask = (task, agent) => {
-		if (agent === void 0) return;
-		let dispose;
+	const waitForTask = (id, signal) => new Promise((resolveTask, rejectTask) => {
+		let settled = false;
+		let dispose = () => {};
+		let timer;
+		let abort = () => {};
+		const cleanup = () => {
+			dispose();
+			if (timer !== void 0) clearTimeout(timer);
+			signal?.removeEventListener("abort", abort);
+		};
+		const resolve = (task) => {
+			if (settled) return;
+			settled = true;
+			cleanup();
+			resolveTask(task);
+		};
+		const reject = (error) => {
+			if (settled) return;
+			settled = true;
+			cleanup();
+			rejectTask(error);
+		};
+		abort = () => {
+			if (settled) return;
+			const reason = signal?.reason instanceof Error ? signal.reason : /* @__PURE__ */ new Error("Image generation was cancelled.");
+			settled = true;
+			cleanup();
+			runtime.queue.cancel(id);
+			rejectTask(reason);
+		};
 		const onChange = (updated) => {
-			if (updated.id !== task.id || !isFinalTask(updated)) return;
-			dispose?.();
-			if (dispose !== void 0) taskSubscriptions.delete(dispose);
-			notifyCompletion(agent, updated).catch(() => {});
+			if (updated.id === id && isFinalTask(updated)) resolve(updated);
 		};
+		if (signal?.aborted === true) {
+			abort();
+			return;
+		}
 		dispose = runtime.queue.subscribe(onChange);
-		taskSubscriptions.add(dispose);
-		const current = findTask(task.id);
-		if (isFinalTask(current)) onChange(current);
-	};
+		signal?.addEventListener("abort", abort, { once: true });
+		timer = setTimeout(() => {
+			if (settled) return;
+			const timeout = new ImageGenError(`Image generation task ${id} timed out after ${AGENT_GENERATION_TIMEOUT_MS / 1e3} seconds.`, "generation-timeout");
+			settled = true;
+			cleanup();
+			runtime.queue.cancel(id);
+			rejectTask(timeout);
+		}, AGENT_GENERATION_TIMEOUT_MS);
+		let current;
+		try {
+			current = findTask(id);
+		} catch (error) {
+			reject(error);
+			return;
+		}
+		if (isFinalTask(current)) resolve(current);
+	});
 	const disposers = [
 		ctx.tools.register(defineTool({
 			name: "generate_image",
-			description: "Queue a text-to-image generation request. This returns immediately with a task id. When it finishes, the conversation automatically receives a visible image-attachment notification; do not repeatedly poll. Only use models configured for this plugin; omit model to use the first configured image model. Use get_image_generation_task only for an explicit status check or recovery.",
+			description: "Generate an image. By default this tool call stays pending until the task reaches a final state, and completed images are returned directly as tool-result attachments without creating a user message. Set wait_for_completion to false for background mode, then use get_image_generation_task explicitly. Only use models configured for this plugin; omit model to use the first configured image model.",
 			parameters: {
 				prompt: {
 					type: "string",
@@ -2434,6 +2517,10 @@ function registerAgentImageTools(ctx, runtime, resolve) {
 				detail: {
 					type: "string",
 					description: "Optional provider detail value, for example standard or high."
+				},
+				wait_for_completion: {
+					type: "boolean",
+					description: "Wait for images and return them in this tool result. Defaults to true; set false for background mode."
 				}
 			},
 			output: {
@@ -2451,13 +2538,12 @@ function registerAgentImageTools(ctx, runtime, resolve) {
 					n: Math.min(4, Math.max(1, args.count ?? 1)),
 					detail: args.detail ?? ""
 				});
-				watchTask(task, exec.agent);
-				return taskResult(task);
+				return taskResult(args.wait_for_completion === false ? task : await waitForTask(task.id, exec.signal));
 			}
 		})),
 		ctx.tools.register(defineTool({
 			name: "edit_image",
-			description: "Queue an image-to-image edit. source_image must be an image reference returned by a completed generation notification or get_image_generation_task; pass that entire object unchanged. Only configured image models are allowed; omit model to use the first configured model. Completion is automatically delivered to the conversation with visible image attachments; do not repeatedly poll.",
+			description: "Edit an image. By default this tool call stays pending until the task reaches a final state, and completed images are returned directly as tool-result attachments without creating a user message. Set wait_for_completion to false for background mode, then use get_image_generation_task explicitly. source_image must be an image reference returned by a completed generation or get_image_generation_task; pass that entire object unchanged. Only configured image models are allowed; omit model to use the first configured model.",
 			parameters: {
 				prompt: {
 					type: "string",
@@ -2488,6 +2574,10 @@ function registerAgentImageTools(ctx, runtime, resolve) {
 				detail: {
 					type: "string",
 					description: "Optional provider detail value."
+				},
+				wait_for_completion: {
+					type: "boolean",
+					description: "Wait for images and return them in this tool result. Defaults to true; set false for background mode."
 				}
 			},
 			output: {
@@ -2508,13 +2598,12 @@ function registerAgentImageTools(ctx, runtime, resolve) {
 					image: imageDataUrl(reference),
 					...reference.ref.name === void 0 ? {} : { refName: reference.ref.name }
 				});
-				watchTask(task, exec.agent);
-				return taskResult(task);
+				return taskResult(args.wait_for_completion === false ? task : await waitForTask(task.id, exec.signal));
 			}
 		})),
 		ctx.tools.register(defineTool({
 			name: "get_image_generation_task",
-			description: "Optionally check an image-generation task status. Completed tasks return image references for edit_image, but the conversation already receives a visible completion notification automatically; do not poll repeatedly.",
+			description: "Check an image-generation task status. Completed tasks return image references and image attachments for edit_image. Generation tools normally wait for completion, so use this for explicit recovery or status checks.",
 			parameters: { task_id: {
 				type: "string",
 				required: true,
@@ -2550,8 +2639,6 @@ function registerAgentImageTools(ctx, runtime, resolve) {
 		}))
 	];
 	return () => {
-		for (const dispose of taskSubscriptions) dispose();
-		taskSubscriptions.clear();
 		for (const dispose of disposers) dispose();
 	};
 }
@@ -2592,7 +2679,7 @@ const DEFAULT_ALLOW_AGENT_IMAGE_GENERATION = true;
 /** Order of the announcement section within the tool-guidance band. */
 const SECTION_ORDER = 150;
 /** Model-facing announcement: plugin presence, capabilities, and limits. */
-const IMAGEGEN_GUIDANCE = "本机已安装 dsh-imagegen 插件（DSH AI 生图）：侧边栏「AI 生图」入口。能力：对接 OpenAI 兼容图像生成 API，模型由用户在「设置 → 插件 → AI 生图」中检测或手动配置的生图模型列表决定；支持文生图（/images/generations）与图生图（/images/edits，上传参考图，grok-imagine 模型按官方 JSON image_url 协议发送）。API 地址与密钥在 GUI 设置中配置，密钥仅存于本机设置文档；生成请求由本地宿主代理转发，结果以 base64 返回面板，可预览与下载。模型只能使用已配置的生图模型；模型出现在 /models 中不等于其网关原生支持生图协议，遇到 Qwen、Gemini 等非 OpenAI 生图协议时应如实说明上游兼容性。可一键把满意的图片加入「画廊」。内置「提示词模板库」（面板提示词框左下角「模板库」按钮）：打包 awesome-gpt-image-2 的数百条提示词案例，可搜索、筛选与复用。Agent 可直接调用 `generate_image` 提交文生图，也可用 `edit_image` 图生图；任务后台异步执行，完成后插件会自动唤醒原对话，并以可直接查看和复用的图片附件回贴结果，因此不要反复轮询。仅在用户明确要求进度或需要恢复任务时，才使用 `get_image_generation_task` 查询状态。限制：生成消耗上游 API 额度；图片内容由上游模型生成，可能不符合预期或包含不适宜内容；api_key 以明文存储在设置文档中；参考图会发送至所配置的 API 服务；模板库在线刷新与参考图首次加载需要访问 vibeui.top。用户提到「生图 / 绘画 / 生成图片 / 文生图 / 图生图 / 画廊 / 提示词模板」时即指本插件，请据此协作。";
+const IMAGEGEN_GUIDANCE = "本机已安装 dsh-imagegen 插件（DSH AI 生图）：侧边栏「AI 生图」入口。能力：对接 OpenAI 兼容图像生成 API，模型由用户在「设置 → 插件 → AI 生图」中检测或手动配置的生图模型列表决定；支持文生图（/images/generations）与图生图（/images/edits，上传参考图，grok-imagine 模型按官方 JSON image_url 协议发送，nanobanana 系列按 aspect_ratio / image_size 参数协议发送；seedream 系列统一走 /images/generations，参考图以 JSON image 数组发送）。API 地址与密钥在 GUI 设置中配置，密钥仅存于本机设置文档；生成请求由本地宿主代理转发，结果以 base64 返回面板，可预览与下载。模型只能使用已配置的生图模型；模型出现在 /models 中不等于其网关原生支持生图协议，遇到 Qwen、Gemini 等非 OpenAI 生图协议时应如实说明上游兼容性。可一键把满意的图片加入「画廊」。内置「提示词模板库」（面板提示词框左下角「模板库」按钮）：打包 awesome-gpt-image-2 的数百条提示词案例，可搜索、筛选与复用。Agent 可直接调用 `generate_image` 提交文生图，也可用 `edit_image` 图生图；默认保持工具调用等待直到任务完成，完成图片直接作为工具结果附件返回，不会额外伪造用户消息。若明确需要后台执行，可传 `wait_for_completion: false`，之后再用 `get_image_generation_task` 查询；不要反复轮询。限制：生成消耗上游 API 额度；图片内容由上游模型生成，可能不符合预期或包含不适宜内容；api_key 以明文存储在设置文档中；参考图会发送至所配置的 API 服务；模板库在线刷新与参考图首次加载需要访问 vibeui.top。用户提到「生图 / 绘画 / 生成图片 / 文生图 / 图生图 / 画廊 / 提示词模板」时即指本插件，请据此协作。";
 /** Add the live allow-list so an Agent can honor a user's model choice. */
 function guidanceFor(imageModels) {
 	return `${IMAGEGEN_GUIDANCE} 当前允许调用的生图模型：${imageModels.join("、")}。用户指定其中某个模型时，工具参数 model 必须使用该精确名称；未指定时使用列表中的第一个。`;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@dickpy/dsh-imagegen",
-  "description": "AI 鐢熷浘 (image generation) plugin for the dsh web GUI: text-to-image and image-to-image through a configurable OpenAI-compatible endpoint (gpt-image-2 / grok-imagine-image / dall-e-3, with native xAI Grok Imagine request shaping), with a settings card for api_url / api_key and a sidebar entry opening a split-pane generation studio.",
-  "version": "1.2.1",
+  "description": "AI 鐢熷浘 (image generation) plugin for the dsh web GUI: text-to-image and image-to-image through a configurable OpenAI-compatible endpoint (gpt-image-2 / grok-imagine-image / nanobanana series / seedream-5.0-pro / dall-e-3, with native xAI Grok Imagine, Google Nano Banana and ByteDance Seedream request shaping), with a settings card for api_url / api_key and a sidebar entry opening a split-pane generation studio.",
+  "version": "1.2.3",
   "type": "module",
   "main": "lib/index.js",
   "exports": {
@@ -17,7 +17,8 @@
       "inject": [
         "@deepseek-ai/dsh-client-runtime",
         "@deepseek-ai/dsh-client-connection",
-        "@deepseek-ai/dsh-client-ui-settings"
+        "@deepseek-ai/dsh-client-ui-settings",
+        "@deepseek-ai/dsh-client-ui-tool"
       ],
       "platform": "web"
     }

package/src/agent-image-tools.ts

@@ -1,7 +1,6 @@
 /** Agent-facing image-generation tools backed by the shared host queue. */
 
 import type { Context } from '@deepseek-ai/cordis'
-import { createUserMessage } from '@deepseek-ai/dsh-llm/message'
 import { defineTool } from '@deepseek-ai/dsh-tools'
 import type { ImageAttachmentRef, ImageMediaType } from '@deepseek-ai/dsh-attachment'
 import type {} from '@deepseek-ai/dsh-attachment'
@@ -61,6 +60,9 @@ const taskResultSchema = {
   },
 } as const
 
+/** Agent calls stay pending until the provider and history write settle. */
+const AGENT_GENERATION_TIMEOUT_MS = 300_000
+
 function acceptedMediaType(value: string): value is ImageMediaType {
   return value === 'image/png' || value === 'image/jpeg' || value === 'image/webp' || value === 'image/gif'
 }
@@ -106,7 +108,6 @@ function renderTaskResult(value: AgentTaskResult): Array<{ type: 'text'; text: s
 /** Register the global Agent tools and unregister them with the plugin lifecycle. */
 export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRuntime, resolve: () => AgentImageToolConfig): () => void {
   const attachmentRefs = new Map<string, Promise<AgentImageRef[]>>()
-  const taskSubscriptions = new Set<() => void>()
   const ensureConfigured = (): void => {
     const config = resolve()
     if (!config.enabled) throw new ImageGenError('AI image generation is disabled. Open Settings > Plugins > AI Image and enable it.', 'plugin-disabled')
@@ -144,7 +145,7 @@ export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRu
           ? 'Generation failed.'
           : task.status === 'cancelled'
             ? 'Generation was cancelled.'
-            : 'Generation is still running. Completion will be delivered to the conversation automatically with image attachments.',
+            : 'Generation is still running. Query the task again when you need its current status.',
       ...task.error === undefined ? {} : { error: task.error },
       images,
     }
@@ -154,42 +155,70 @@ export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRu
     if (task === undefined) throw new ImageGenError(`Image generation task ${id} was not found.`, 'task-not-found')
     return task
   }
-  const notifyCompletion = async (agent: { send: (message: ReturnType<typeof createUserMessage>, target: 'next-turn', wakeup: true) => void }, task: GenerationTask): Promise<void> => {
-    const result = await taskResult(task)
-    const completed = task.status === 'completed'
-    const text = completed
-      ? `图像生成任务已完成（${task.id}）。图片已附在这条消息中，可以直接查看、下载或作为后续图生图的参考。`
-      : task.status === 'cancelled'
-        ? `图像生成任务已取消（${task.id}）。`
-        : `图像生成任务失败（${task.id}）：${task.error ?? '未知错误'}`
-    agent.send(createUserMessage({
-      content: [
-        { type: 'text', text },
-        ...completed ? result.images.map(image => ({ type: 'image' as const, attachment: restoreRef(image) })) : [],
-      ],
-      // The stock conversation's plugin-context row renders text only; a
-      // user-role message is the supported path that renders image attachments.
-      source: { kind: 'user' },
-    }), 'next-turn', true)
-  }
-  const watchTask = (task: GenerationTask, agent: { send: (message: ReturnType<typeof createUserMessage>, target: 'next-turn', wakeup: true) => void } | undefined): void => {
-    if (agent === undefined) return
-    let dispose: (() => void) | undefined
+  const waitForTask = (id: string, signal: AbortSignal | undefined): Promise<GenerationTask> => new Promise((resolveTask, rejectTask) => {
+    let settled = false
+    let dispose = (): void => {}
+    let timer: ReturnType<typeof setTimeout> | undefined
+    let abort = (): void => {}
+
+    const cleanup = (): void => {
+      dispose()
+      if (timer !== undefined) clearTimeout(timer)
+      signal?.removeEventListener('abort', abort)
+    }
+    const resolve = (task: GenerationTask): void => {
+      if (settled) return
+      settled = true
+      cleanup()
+      resolveTask(task)
+    }
+    const reject = (error: unknown): void => {
+      if (settled) return
+      settled = true
+      cleanup()
+      rejectTask(error)
+    }
+    abort = (): void => {
+      if (settled) return
+      const reason = signal?.reason instanceof Error ? signal.reason : new Error('Image generation was cancelled.')
+      // Remove the listener before publishing cancellation so the queue event
+      // cannot turn an execution abort into a successful cancelled result.
+      settled = true
+      cleanup()
+      runtime.queue.cancel(id)
+      rejectTask(reason)
+    }
     const onChange = (updated: GenerationTask): void => {
-      if (updated.id !== task.id || !isFinalTask(updated)) return
-      dispose?.()
-      if (dispose !== undefined) taskSubscriptions.delete(dispose)
-      void notifyCompletion(agent, updated).catch(() => {})
+      if (updated.id === id && isFinalTask(updated)) resolve(updated)
+    }
+
+    if (signal?.aborted === true) {
+      abort()
+      return
     }
     dispose = runtime.queue.subscribe(onChange)
-    taskSubscriptions.add(dispose)
-    const current = findTask(task.id)
-    if (isFinalTask(current)) onChange(current)
-  }
+    signal?.addEventListener('abort', abort, { once: true })
+    timer = setTimeout(() => {
+      if (settled) return
+      const timeout = new ImageGenError(`Image generation task ${id} timed out after ${AGENT_GENERATION_TIMEOUT_MS / 1000} seconds.`, 'generation-timeout')
+      settled = true
+      cleanup()
+      runtime.queue.cancel(id)
+      rejectTask(timeout)
+    }, AGENT_GENERATION_TIMEOUT_MS)
+    let current: GenerationTask
+    try {
+      current = findTask(id)
+    } catch (error) {
+      reject(error)
+      return
+    }
+    if (isFinalTask(current)) resolve(current)
+  })
   const disposers = [
     ctx.tools.register(defineTool({
       name: 'generate_image',
-      description: 'Queue a text-to-image generation request. This returns immediately with a task id. When it finishes, the conversation automatically receives a visible image-attachment notification; do not repeatedly poll. Only use models configured for this plugin; omit model to use the first configured image model. Use get_image_generation_task only for an explicit status check or recovery.',
+      description: 'Generate an image. By default this tool call stays pending until the task reaches a final state, and completed images are returned directly as tool-result attachments without creating a user message. Set wait_for_completion to false for background mode, then use get_image_generation_task explicitly. Only use models configured for this plugin; omit model to use the first configured image model.',
       parameters: {
         prompt: { type: 'string', required: true, description: 'Detailed image-generation prompt.' },
         model: { type: 'string', description: 'One of the configured image models. Defaults to the first configured model.' },
@@ -197,6 +226,7 @@ export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRu
         quality: { type: 'string', description: 'auto, 1k, 2k, or 4k.' },
         count: { type: 'integer', description: 'Number of images, 1 to 4. Defaults to 1.' },
         detail: { type: 'string', description: 'Optional provider detail value, for example standard or high.' },
+        wait_for_completion: { type: 'boolean', description: 'Wait for images and return them in this tool result. Defaults to true; set false for background mode.' },
       },
       output: { schema: taskResultSchema, render: (_args, value) => renderTaskResult(value) },
       async execute(args, exec) {
@@ -210,13 +240,12 @@ export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRu
           n: Math.min(4, Math.max(1, args.count ?? 1)),
           detail: args.detail ?? '',
         })
-        watchTask(task, exec.agent)
-        return taskResult(task)
+        return taskResult(args.wait_for_completion === false ? task : await waitForTask(task.id, exec.signal))
       },
     })),
     ctx.tools.register(defineTool({
       name: 'edit_image',
-      description: 'Queue an image-to-image edit. source_image must be an image reference returned by a completed generation notification or get_image_generation_task; pass that entire object unchanged. Only configured image models are allowed; omit model to use the first configured model. Completion is automatically delivered to the conversation with visible image attachments; do not repeatedly poll.',
+      description: 'Edit an image. By default this tool call stays pending until the task reaches a final state, and completed images are returned directly as tool-result attachments without creating a user message. Set wait_for_completion to false for background mode, then use get_image_generation_task explicitly. source_image must be an image reference returned by a completed generation or get_image_generation_task; pass that entire object unchanged. Only configured image models are allowed; omit model to use the first configured model.',
       parameters: {
         prompt: { type: 'string', required: true, description: 'How to transform the source image.' },
         source_image: { ...imageRefSchema, required: true, description: 'Image reference returned by get_image_generation_task.' },
@@ -225,6 +254,7 @@ export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRu
         quality: { type: 'string', description: 'auto, 1k, 2k, or 4k.' },
         count: { type: 'integer', description: 'Number of images, 1 to 4. Defaults to 1.' },
         detail: { type: 'string', description: 'Optional provider detail value.' },
+        wait_for_completion: { type: 'boolean', description: 'Wait for images and return them in this tool result. Defaults to true; set false for background mode.' },
       },
       output: { schema: taskResultSchema, render: (_args, value) => renderTaskResult(value) },
       async execute(args, exec) {
@@ -241,13 +271,12 @@ export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRu
           image: imageDataUrl(reference),
           ...reference.ref.name === undefined ? {} : { refName: reference.ref.name },
         })
-        watchTask(task, exec.agent)
-        return taskResult(task)
+        return taskResult(args.wait_for_completion === false ? task : await waitForTask(task.id, exec.signal))
       },
     })),
     ctx.tools.register(defineTool({
       name: 'get_image_generation_task',
-      description: 'Optionally check an image-generation task status. Completed tasks return image references for edit_image, but the conversation already receives a visible completion notification automatically; do not poll repeatedly.',
+      description: 'Check an image-generation task status. Completed tasks return image references and image attachments for edit_image. Generation tools normally wait for completion, so use this for explicit recovery or status checks.',
       parameters: { task_id: { type: 'string', required: true, description: 'Task id returned by generate_image or edit_image.' } },
       output: { schema: taskResultSchema, render: (_args, value) => renderTaskResult(value) },
       async execute(args) {
@@ -269,8 +298,6 @@ export function registerAgentImageTools(ctx: Context, runtime: ImageGenerationRu
     })),
   ]
   return () => {
-    for (const dispose of taskSubscriptions) dispose()
-    taskSubscriptions.clear()
     for (const dispose of disposers) dispose()
   }
 }

package/src/client/ImageGenPanel.tsx

@@ -1127,7 +1127,7 @@ export function ImageGenPanel(props: {
                           <span className={css.galleryBadge}>{entry.mode === 'edit' ? tt('mode.edit') : tt('mode.text')}</span>
                         </button>
                         <div className={css.galleryCardFooter}>
-                          <span className={css.galleryAvatar}>{entry.model.startsWith('grok') ? 'G' : 'D'}</span>
+                          <span className={css.galleryAvatar}>{entry.model.toLowerCase().startsWith('nanobanana') ? 'N' : entry.model.toLowerCase().startsWith('seedream') ? 'S' : entry.model.startsWith('grok') ? 'G' : 'D'}</span>
                           <span className={css.galleryCardInfo}>
                             <strong>{entry.prompt || tt('gallery.untitled')}</strong>
                             <small>{entry.model} · {normalizeSize(entry.size)} · {formatTime(entry.createdAt)}</small>

package/src/client/image-toolview.module.css

@@ -0,0 +1,73 @@
+.root {
+  display: flex;
+  flex-direction: column;
+  gap: 7px;
+  margin: 4px 0;
+  padding: 8px 10px 10px;
+  border: 1px solid var(--dsw-alias-border-l1);
+  border-radius: 8px;
+  background: var(--dsw-alias-bg-layer-1);
+}
+
+.header {
+  display: flex;
+  align-items: center;
+  gap: 7px;
+  min-height: 20px;
+  color: var(--dsw-alias-label-secondary);
+  font-size: 12px;
+}
+
+.icon {
+  color: var(--dsw-alias-brand-primary);
+  font-size: 15px;
+  line-height: 1;
+}
+
+.status {
+  margin-left: auto;
+  color: var(--dsw-alias-label-tertiary);
+  font-size: 11px;
+}
+
+.message,
+.loading,
+.error {
+  margin: 0;
+  color: var(--dsw-alias-label-tertiary);
+  font-size: 12px;
+  line-height: 1.5;
+  overflow-wrap: anywhere;
+}
+
+.images {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+}
+
+.imageLink {
+  display: block;
+  max-width: min(280px, 100%);
+  overflow: hidden;
+  border: 1px solid var(--dsw-alias-border-l1);
+  border-radius: 6px;
+  background: var(--dsw-alias-bg-base);
+}
+
+.imageLink:hover {
+  border-color: var(--dsw-alias-brand-primary);
+}
+
+.image {
+  display: block;
+  width: auto;
+  max-width: 280px;
+  height: auto;
+  max-height: 220px;
+  object-fit: contain;
+}
+
+.error {
+  color: var(--dsw-alias-state-error-primary);
+}