feat(channel): add customizable JS model-call scripts with split-panel editor

- Run user-authored async scripts per model with flat locals (prompt,
  images, messages, params, model, baseUrl, apiKey, http, request, poll,
  sleep, signal, onDelta); empty script falls back to system default.
- Wire image/video/audio/text request entrypoints through the plugin
  runtime, keeping default handlers unchanged.
- Split-panel script editor: variable/return docs on the left, CodeMirror
  JS editor on the right; title shows capability - model.
- Provide OpenAI and Gemini template buttons per capability; text template
  uses the /responses API; image template branches text2img vs img2img.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

web/src/services/api/audio.ts

@@ -32,7 +32,8 @@ export async function requestAudioGeneration(config: AiConfig, prompt: string, o
                 capability: "audio",
                 script,
                 config: requestConfig,
-                input: { prompt, params: { voice: normalizeAudioVoiceValue(config.audioVoice), format, speed: normalizeAudioSpeedValue(config.audioSpeed), instructions: config.audioInstructions.trim() } },
+                prompt,
+                params: { voice: normalizeAudioVoiceValue(config.audioVoice), format, speed: normalizeAudioSpeedValue(config.audioSpeed), instructions: config.audioInstructions.trim() },
                 signal: options?.signal,
             });
             return await audioPluginBlob(result, format);

web/src/services/api/image.ts

@@ -666,12 +666,9 @@ export async function requestGeneration(config: AiConfig, prompt: string, option
                 capability: "image",
                 script,
                 config: requestConfig,
-                input: {
-                    prompt: withSystemPrompt(requestConfig, prompt),
-                    references: [],
-                    params: { size: requestSize, quality, count: n },
-                    body: { model: requestConfig.model, n, ...(quality ? { quality } : {}), ...(requestSize ? { size: requestSize } : {}), response_format: "b64_json", output_format: IMAGE_OUTPUT_FORMAT },
-                },
+                prompt: withSystemPrompt(requestConfig, prompt),
+                images: [],
+                params: { size: requestSize, quality, count: n },
                 signal: options?.signal,
             });
             return normalizePluginImages(result).map((dataUrl) => ({ id: nanoid(), dataUrl }));
@@ -726,12 +723,9 @@ export async function requestEdit(config: AiConfig, prompt: string, references:
                 capability: "image",
                 script,
                 config: requestConfig,
-                input: {
-                    prompt: withSystemPrompt(requestConfig, requestPrompt),
-                    references: refs,
-                    params: { size: requestSize, quality, count: n },
-                    body: { model: requestConfig.model, n, ...(quality ? { quality } : {}), ...(requestSize ? { size: requestSize } : {}), response_format: "b64_json", output_format: IMAGE_OUTPUT_FORMAT },
-                },
+                prompt: withSystemPrompt(requestConfig, requestPrompt),
+                images: refs,
+                params: { size: requestSize, quality, count: n },
                 signal: options?.signal,
             });
             return normalizePluginImages(result).map((dataUrl) => ({ id: nanoid(), dataUrl }));
@@ -783,7 +777,7 @@ export async function requestImageQuestion(config: AiConfig, messages: AiTextMes
                 capability: "text",
                 script,
                 config: requestConfig,
-                input: { messages: withSystemMessage(requestConfig, messages), body: { model: requestConfig.model } },
+                messages: withSystemMessage(requestConfig, messages),
                 signal: options?.signal,
                 onDelta,
             });

web/src/services/api/model-plugin.ts

@@ -1,4 +1,4 @@
-import axios from "axios";
+import axios, { type AxiosRequestConfig } from "axios";
 
 import { buildApiUrl, type AiConfig, type ModelCapability } from "@/stores/use-config-store";
 
@@ -18,27 +18,20 @@ export type PluginHttp = {
 
 export type PluginPollOptions = { intervalMs?: number; timeoutMs?: number };
 
-export type PluginConfigView = {
-    baseUrl: string;
-    apiKey: string;
-    model: string;
-    apiFormat: string;
-    systemPrompt: string;
-};
-
 export type RunPluginArgs = {
     capability: ModelCapability;
     script: string;
     config: AiConfig;
-    input: Record<string, unknown>;
+    prompt?: string;
+    images?: string[];
+    messages?: unknown[];
+    params?: Record<string, unknown>;
     signal?: AbortSignal;
     onDelta?: (text: string) => void;
 };
 
-function pluginHeaders(config: AiConfig, extra?: Record<string, string>, hasJsonBody = false): Record<string, string> {
+function pluginHeaders(extra?: Record<string, string>, hasJsonBody = false): Record<string, string> {
     const headers: Record<string, string> = {};
-    if (config.apiFormat === "gemini") headers["x-goog-api-key"] = config.apiKey;
-    else headers.Authorization = `Bearer ${config.apiKey}`;
     if (hasJsonBody) headers["Content-Type"] = "application/json";
     return { ...headers, ...extra };
 }
@@ -49,14 +42,14 @@ function pluginUrl(config: AiConfig, path: string) {
 }
 
 function createPluginHttp(config: AiConfig, options?: RequestOptions): PluginHttp {
-    const request = async (method: "get" | "post", path: string, body: unknown, opts?: PluginHttpOptions) => {
+    const run = async (method: "get" | "post", path: string, body: unknown, opts?: PluginHttpOptions) => {
         const isForm = typeof FormData !== "undefined" && body instanceof FormData;
         const response = await axios.request({
             method,
             url: pluginUrl(config, path),
             data: method === "post" ? body : undefined,
             params: opts?.params,
-            headers: pluginHeaders(config, opts?.headers, method === "post" && !isForm && body !== undefined),
+            headers: pluginHeaders({ Authorization: `Bearer ${config.apiKey}`, ...opts?.headers }, method === "post" && !isForm && body !== undefined),
             responseType: opts?.responseType || "json",
             signal: options?.signal,
         });
@@ -64,8 +57,16 @@ function createPluginHttp(config: AiConfig, options?: RequestOptions): PluginHtt
     };
     return {
         url: (path) => pluginUrl(config, path),
-        post: (path, body, opts) => request("post", path, body, opts),
-        get: (path, opts) => request("get", path, undefined, opts),
+        post: (path, body, opts) => run("post", path, body, opts),
+        get: (path, opts) => run("get", path, undefined, opts),
+    };
+}
+
+/** Raw request with no automatic auth header — the script controls method, url, headers, body entirely. */
+function createPluginRequest(config: AiConfig, options?: RequestOptions) {
+    return async (requestConfig: AxiosRequestConfig & { url: string }) => {
+        const response = await axios.request({ ...requestConfig, url: pluginUrl(config, requestConfig.url), signal: options?.signal });
+        return response.data;
     };
 }
 
@@ -103,38 +104,51 @@ function createPoll(signal?: AbortSignal) {
 }
 
 /**
- * Run a user-authored model call script. The script body runs as an async function with these locals:
- *   input   —— normalized request input for this capability (prompt / references / messages / params / body)
- *   config  —— { baseUrl, apiKey, model, apiFormat, systemPrompt }
- *   http    —— { url(path), post(path, body, opts), get(path, opts) } bound to the model's channel
- *   poll    —— poll(request, extract, { intervalMs, timeoutMs }) resolves with the first truthy extract result
- *   sleep   —— sleep(ms)
- *   signal  —— AbortSignal for cancellation
- *   onDelta —— (text) => void, push streaming text (text capability only)
+ * Run a user-authored model call script as an async function body with flat locals (see PLUGIN_VARIABLES):
+ *   prompt / images / messages / params        —— 本次请求的输入
+ *   model / baseUrl / apiKey / systemPrompt     —— 当前渠道信息
+ *   http / request / poll / sleep / signal / onDelta    —— 调用辅助
  * The script must `return` the result; each caller normalizes it to its capability's shape.
  */
 export async function runModelPlugin<T = unknown>(args: RunPluginArgs): Promise<T> {
-    const configView: PluginConfigView = {
-        baseUrl: args.config.baseUrl,
-        apiKey: args.config.apiKey,
-        model: args.config.model,
-        apiFormat: args.config.apiFormat,
-        systemPrompt: args.config.systemPrompt,
-    };
-    const http = createPluginHttp(args.config, { signal: args.signal });
+    const { config } = args;
+    const http = createPluginHttp(config, { signal: args.signal });
+    const request = createPluginRequest(config, { signal: args.signal });
     const poll = createPoll(args.signal);
     const runner = new Function(
-        "input",
-        "config",
+        "prompt",
+        "images",
+        "messages",
+        "params",
+        "model",
+        "baseUrl",
+        "apiKey",
+        "systemPrompt",
         "http",
+        "request",
         "poll",
         "sleep",
         "signal",
         "onDelta",
         `"use strict"; return (async () => {\n${args.script}\n})();`,
-    ) as (input: unknown, config: PluginConfigView, http: PluginHttp, poll: ReturnType<typeof createPoll>, sleep: (ms: number) => Promise<void>, signal: AbortSignal | undefined, onDelta?: (text: string) => void) => Promise<T>;
+    ) as (...fnArgs: unknown[]) => Promise<T>;
     try {
-        return await runner(args.input, configView, http, poll, (ms: number) => sleep(ms, args.signal), args.signal, args.onDelta);
+        return await runner(
+            args.prompt || "",
+            args.images || [],
+            args.messages || [],
+            args.params || {},
+            config.model,
+            config.baseUrl,
+            config.apiKey,
+            config.systemPrompt || "",
+            http,
+            request,
+            poll,
+            (ms: number) => sleep(ms, args.signal),
+            args.signal,
+            args.onDelta,
+        );
     } catch (error) {
         if (error instanceof DOMException && error.name === "AbortError") throw error;
         if (axios.isCancel(error)) throw error;
@@ -143,45 +157,201 @@ export async function runModelPlugin<T = unknown>(args: RunPluginArgs): Promise<
     }
 }
 
-export const PLUGIN_TEMPLATES: Record<ModelCapability, string> = {
-    image: `// 输入：input.prompt / input.references(dataURL[]) / input.body(默认请求体) / input.params
-// 返回：dataURL 或 URL 字符串，或它们的数组，或 [{ dataUrl }]
-const data = await http.post("/images/generations", {
-  ...input.body,
-  model: config.model,
-  prompt: input.prompt,
+export type PluginVariable = { name: string; type: string; desc: string; capabilities?: ModelCapability[] };
+
+/** Documentation surface shown in the script editor. */
+export const PLUGIN_VARIABLES: PluginVariable[] = [
+    { name: "prompt", type: "string", desc: "用户输入的提示词（已拼接系统提示词）", capabilities: ["image", "video", "audio"] },
+    { name: "images", type: "string[]", desc: "参考图，dataURL 数组（改图 / 图生视频时有值）", capabilities: ["image", "video"] },
+    { name: "messages", type: "{ role, content }[]", desc: "对话消息数组，含系统消息", capabilities: ["text"] },
+    { name: "params", type: "object", desc: "生成参数：生图 {size,quality,count}、视频 {seconds,size,resolution,ratio,generateAudio,watermark}、音频 {voice,format,speed,instructions}" },
+    { name: "model", type: "string", desc: "模型名称（不含渠道前缀）" },
+    { name: "baseUrl", type: "string", desc: "渠道接口地址（原样，未拼 /v1）" },
+    { name: "apiKey", type: "string", desc: "渠道 API Key，请求头里自己带上" },
+    { name: "systemPrompt", type: "string", desc: "系统提示词原文" },
+    { name: "http", type: "object", desc: "便捷请求：http.post(path, body, {headers,params,responseType})、http.get(path, opts)、http.url(path)；默认带 Authorization: Bearer apiKey，可用 headers 覆盖；path 相对时按 baseUrl 拼 /v1" },
+    { name: "request", type: "function", desc: "原始请求 request({ method, url, headers, params, data, responseType })，不加任何默认头，鉴权头自己写；url 相对时按 baseUrl 拼接（不加 /v1）" },
+    { name: "poll", type: "function", desc: "轮询 poll(request, extract, {intervalMs,timeoutMs})，extract 返回真值即结束" },
+    { name: "sleep", type: "function", desc: "sleep(ms) 延时" },
+    { name: "signal", type: "AbortSignal", desc: "取消信号，可透传给 http/request" },
+    { name: "onDelta", type: "function", desc: "onDelta(text) 推送流式文本（文本模型）", capabilities: ["text"] },
+];
+
+export const PLUGIN_RETURNS: Record<ModelCapability, string> = {
+    image: "文生图（images 为空）和图生图（images 有参考图）接口不同，脚本需自行区分；返回图片 URL 或 dataURL 字符串，也可返回它们的数组，或 [{ dataUrl }] / [{ url }] / [{ b64_json }]",
+    video: "脚本内部完成轮询，返回 { url } 或 { blob } 或视频 URL 字符串",
+    audio: "返回 Blob，或 base64 / dataURL 字符串，或 { b64_json } / { data } / { url }",
+    text: "用 onDelta(text) 推送流式，最终 return 完整文本字符串",
+};
+
+export type PluginTemplate = { label: string; script: string };
+
+export const PLUGIN_TEMPLATES: Record<ModelCapability, PluginTemplate[]> = {
+    image: [
+        {
+            label: "OpenAI 规范",
+            script: `// 生图 / 改图：两者接口不同，用 images 是否为空来区分。
+// 可用：prompt、images(dataURL[])、params{size,quality,count}、model、baseUrl、apiKey
+if (images.length === 0) {
+  // 文生图：/images/generations（JSON）
+  const data = await request({
+    method: "post",
+    url: \`\${baseUrl}/v1/images/generations\`,
+    headers: { "Content-Type": "application/json", Authorization: \`Bearer \${apiKey}\` },
+    data: { model, prompt, n: params.count, size: params.size, response_format: "b64_json" },
+  });
+  return (data.data || []).map((item) => item.b64_json ? \`data:image/png;base64,\${item.b64_json}\` : item.url);
+}
+
+// 图生图：/images/edits（multipart/form-data，参考图作为文件上传）
+const form = new FormData();
+form.set("model", model);
+form.set("prompt", prompt);
+form.set("n", String(params.count));
+form.set("response_format", "b64_json");
+for (const dataUrl of images) {
+  form.append("image", await (await fetch(dataUrl)).blob(), "ref.png");
+}
+const edited = await request({
+  method: "post",
+  url: \`\${baseUrl}/v1/images/edits\`,
+  headers: { Authorization: \`Bearer \${apiKey}\` }, // 不要手动设 Content-Type，交给浏览器带 boundary
+  data: form,
 });
-return (data.data || []).map((item) => item.b64_json ? \`data:image/png;base64,\${item.b64_json}\` : item.url);`,
-    video: `// 输入：input.prompt / input.references(dataURL[]) / input.params
-// 返回：{ url } 或 { blob } 或视频 URL 字符串
-const task = await http.post("/videos", {
-  model: config.model,
-  prompt: input.prompt,
-  seconds: input.params.seconds,
+return (edited.data || []).map((item) => item.b64_json ? \`data:image/png;base64,\${item.b64_json}\` : item.url);`,
+        },
+        {
+            label: "Gemini 规范",
+            script: `// Gemini 文生图 / 图生图：都走 generateContent，参考图放进 parts 的 inline_data。
+// 可用：prompt、images(dataURL[])、model、baseUrl、apiKey
+const parts = [{ text: prompt }];
+for (const dataUrl of images) {
+  const match = dataUrl.match(/^data:([^;]+);base64,(.*)$/);
+  if (match) parts.push({ inline_data: { mime_type: match[1], data: match[2] } });
+}
+const data = await request({
+  method: "post",
+  url: \`\${baseUrl}/v1beta/models/\${model}:generateContent\`,
+  headers: { "Content-Type": "application/json", "x-goog-api-key": apiKey },
+  data: { contents: [{ role: "user", parts }], generationConfig: { responseModalities: ["IMAGE"] } },
+});
+return (data.candidates || [])
+  .flatMap((c) => c.content?.parts || [])
+  .map((p) => p.inlineData || p.inline_data)
+  .filter(Boolean)
+  .map((img) => \`data:\${img.mimeType || img.mime_type || "image/png"};base64,\${img.data}\`);`,
+        },
+    ],
+    video: [
+        {
+            label: "OpenAI 规范",
+            script: `// 视频（脚本内部自行轮询）。可用：prompt、images(dataURL[])、params{seconds,size,resolution,ratio}
+const headers = { "Content-Type": "application/json", Authorization: \`Bearer \${apiKey}\` };
+const task = await request({
+  method: "post",
+  url: \`\${baseUrl}/v1/videos\`,
+  headers,
+  data: { model, prompt, seconds: params.seconds },
 });
 return await poll(
-  () => http.get(\`/videos/\${task.id}\`),
+  () => request({ method: "get", url: \`\${baseUrl}/v1/videos/\${task.id}\`, headers }),
   (state) => state.status === "completed" ? { url: state.video_url || state.url } : null,
   { intervalMs: 2500, timeoutMs: 300000 },
 );`,
-    audio: `// 输入：input.prompt / input.params(voice/format/speed/instructions)
-// 返回：Blob，或 base64/dataURL 字符串
-return await http.post("/audio/speech", {
-  model: config.model,
-  input: input.prompt,
-  voice: input.params.voice,
-  response_format: input.params.format,
-  speed: Number(input.params.speed),
-}, { responseType: "blob" });`,
-    text: `// 输入：input.messages([{role,content}]) / input.body
-// 用 onDelta(text) 推送流式文本；返回最终完整文本
-const data = await http.post("/chat/completions", {
-  model: config.model,
-  messages: input.messages,
+        },
+        {
+            label: "Gemini 规范",
+            script: `// Gemini(Veo) 视频：predictLongRunning 提交，轮询 operation 拿视频 URI。
+// 可用：prompt、images(dataURL[])、params、model、baseUrl、apiKey
+const headers = { "Content-Type": "application/json", "x-goog-api-key": apiKey };
+const instance = { prompt };
+const first = images[0] && images[0].match(/^data:([^;]+);base64,(.*)$/);
+if (first) instance.image = { bytesBase64Encoded: first[2], mimeType: first[1] };
+const op = await request({
+  method: "post",
+  url: \`\${baseUrl}/v1beta/models/\${model}:predictLongRunning\`,
+  headers,
+  data: { instances: [instance], parameters: { aspectRatio: params.ratio } },
 });
-const text = data.choices?.[0]?.message?.content || "";
+return await poll(
+  () => request({ method: "get", url: \`\${baseUrl}/v1beta/\${op.name}\`, headers }),
+  (state) => {
+    if (!state.done) return null;
+    const uri = state.response?.generateVideoResponse?.generatedSamples?.[0]?.video?.uri;
+    if (!uri) throw new Error("Gemini 未返回视频 URI");
+    return { url: uri.includes("key=") ? uri : \`\${uri}\${uri.includes("?") ? "&" : "?"}key=\${apiKey}\` };
+  },
+  { intervalMs: 5000, timeoutMs: 300000 },
+);`,
+        },
+    ],
+    audio: [
+        {
+            label: "OpenAI 规范",
+            script: `// 音频 TTS。可用：prompt、params{voice,format,speed,instructions}、model
+return await request({
+  method: "post",
+  url: \`\${baseUrl}/v1/audio/speech\`,
+  headers: { "Content-Type": "application/json", Authorization: \`Bearer \${apiKey}\` },
+  responseType: "blob",
+  data: { model, input: prompt, voice: params.voice, response_format: params.format, speed: Number(params.speed) },
+});`,
+        },
+        {
+            label: "Gemini 规范",
+            script: `// Gemini TTS：generateContent + AUDIO 模态，返回 base64 PCM（音频数据在 inlineData.data）。
+// 可用：prompt、params{voice}、model、baseUrl、apiKey
+const data = await request({
+  method: "post",
+  url: \`\${baseUrl}/v1beta/models/\${model}:generateContent\`,
+  headers: { "Content-Type": "application/json", "x-goog-api-key": apiKey },
+  data: {
+    contents: [{ role: "user", parts: [{ text: prompt }] }],
+    generationConfig: {
+      responseModalities: ["AUDIO"],
+      speechConfig: { voiceConfig: { prebuiltVoiceConfig: { voiceName: params.voice } } },
+    },
+  },
+});
+const audio = data.candidates?.[0]?.content?.parts?.map((p) => p.inlineData || p.inline_data).find(Boolean);
+if (!audio?.data) throw new Error("Gemini 未返回音频");
+return { data: audio.data };`,
+        },
+    ],
+    text: [
+        {
+            label: "OpenAI 规范",
+            script: `// 文本对话（OpenAI Responses 接口）。可用：messages([{role,content}])、systemPrompt、model
+const data = await request({
+  method: "post",
+  url: \`\${baseUrl}/v1/responses\`,
+  headers: { "Content-Type": "application/json", Authorization: \`Bearer \${apiKey}\` },
+  data: { model, input: messages },
+});
+const text = data.output_text
+  || (data.output || []).flatMap((o) => o.content || []).map((c) => c.text || "").join("")
+  || "";
 onDelta(text);
 return text;`,
+        },
+        {
+            label: "Gemini 规范",
+            script: `// Gemini 文本：generateContent，system 消息放 systemInstruction。
+// 可用：messages([{role,content}])、systemPrompt、model、baseUrl、apiKey
+const contents = messages
+  .filter((m) => m.role !== "system")
+  .map((m) => ({ role: m.role === "assistant" ? "model" : "user", parts: [{ text: m.content }] }));
+const data = await request({
+  method: "post",
+  url: \`\${baseUrl}/v1beta/models/\${model}:generateContent\`,
+  headers: { "Content-Type": "application/json", "x-goog-api-key": apiKey },
+  data: { contents, ...(systemPrompt ? { systemInstruction: { parts: [{ text: systemPrompt }] } } : {}) },
+});
+const text = data.candidates?.[0]?.content?.parts?.map((p) => p.text || "").join("") || "";
+onDelta(text);
+return text;`,
+        },
+    ],
 };
 
 /** Normalize whatever an image script returns into the app's generated-image shape. */

web/src/services/api/video.ts

@@ -90,17 +90,15 @@ async function createPluginVideoTask(config: AiConfig, model: string, script: st
             capability: "video",
             script,
             config,
-            input: {
-                prompt,
-                references: refs,
-                params: {
-                    seconds: normalizeVideoSeconds(config.videoSeconds),
-                    size: normalizeVideoSize(config.size),
-                    resolution: normalizeVideoResolution(config.vquality),
-                    ratio: config.size,
-                    generateAudio: boolConfig(config.videoGenerateAudio, true),
-                    watermark: boolConfig(config.videoWatermark, false),
-                },
+            prompt,
+            images: refs,
+            params: {
+                seconds: normalizeVideoSeconds(config.videoSeconds),
+                size: normalizeVideoSize(config.size),
+                resolution: normalizeVideoResolution(config.vquality),
+                ratio: config.size,
+                generateAudio: boolConfig(config.videoGenerateAudio, true),
+                watermark: boolConfig(config.videoWatermark, false),
             },
             signal: options?.signal,
         }),