import axios from "axios";

import { buildApiUrl, isSystemProxyBaseUrl, resolveBackendApiUrl, resolveModelRequestConfig, type AiConfig, type ModelChannel } from "@/stores/use-config-store";
import { sanitizeChannelModelCatalogItem, type ChannelModelCatalogItem } from "@/lib/channel-model-catalog";
import { nanoid } from "nanoid";
import { dataUrlToFile } from "@/lib/image-utils";
import { buildImageReferencePromptText } from "@/lib/image-reference-prompt";
import { createClientId } from "@/lib/client-id";
import { projectDesktopLocalChannelRuntime } from "@/lib/desktop-local-channel";
import { channelRequest } from "@/services/api/custom-channel-relay";
import { imageToDataUrl } from "@/services/image-storage";
import type { ReferenceImage } from "@/types/image";
import { withOpenAIPromptCacheKey } from "@/lib/openai-prompt-cache";
import { imageSizeRequest, modelCapabilityConfigFor, normalizeImageValue, type ImageCapabilityConfig } from "@/lib/model-capabilities";
import { buildGeminiImageGenerationConfig, parseGeminiImageDataUrl, type GeminiImageGenerationConfig } from "@/lib/gemini-image";

export type AiTextMessage = {
    role: "system" | "user" | "assistant";
    content: string | AiTextContentPart[];
};

export type AiTextContentPart = { type: "text"; text: string } | { type: "image_url"; image_url: { url: string } } | { type: "file_url"; file_url: { url: string; name: string; mimeType: string } };

export type ResponseToolCall = {
    id: string;
    type: "function";
    function: { name: string; arguments: string };
    thoughtSignature?: string;
};

export type ResponseInputMessage = AiTextMessage | { type: "function_call"; call_id: string; name: string; arguments: string; thoughtSignature?: string } | { role: "tool"; tool_call_id: string; content: string };

export type ResponseFunctionTool = {
    type: "function";
    function: {
        name: string;
        description?: string;
        parameters: Record<string, unknown>;
        strict?: boolean;
    };
};

export type ToolResponseResult = {
    content: string;
    toolCalls: ResponseToolCall[];
    reasoning?: string;
};

type ToolChoice = "auto" | "required" | { type: "function"; name: string };
type ResponseMessageContent = AiTextMessage["content"] | string;
type ResponseInputContent = { type: "input_text"; text: string } | { type: "input_image"; image_url: string } | { type: "input_file"; filename: string; file_data?: string; file_url?: string };
type ResponseInputItem = { role: "system" | "user" | "assistant"; content: string | ResponseInputContent[] } | { type: "function_call"; call_id: string; name: string; arguments: string } | { type: "function_call_output"; call_id: string; output: string };
type ResponseApiToolDefinition = {
    type: "function";
    name: string;
    description?: string;
    parameters: Record<string, unknown>;
    strict?: boolean;
};
type ResponseApiOutputItem = { type?: "message"; content?: Array<{ type?: string; text?: string }> } | { type?: "function_call"; id?: string; call_id?: string; name?: string; arguments?: string };
type ResponseApiPayload = {
    id?: string;
    output?: ResponseApiOutputItem[];
    output_text?: string;
    error?: { message?: string };
    code?: number;
    msg?: string;
};
type ResponseStreamState = { buffer: string; text: string; reasoning: string; payload?: ResponseApiPayload; error?: string };
type ChatCompletionToolCall = { id?: string; type?: "function"; function?: { name?: string; arguments?: string } };
type ChatCompletionPayload = {
    choices?: Array<{ message?: { content?: string | null; reasoning_content?: string | null; tool_calls?: ChatCompletionToolCall[] } }>;
    error?: { message?: string };
    code?: number;
    msg?: string;
};
type ChatCompletionStreamToolCall = { id: string; name: string; arguments: string };
type ChatCompletionStreamState = { buffer: string; text: string; reasoning: string; toolCalls: Map<number, ChatCompletionStreamToolCall>; error?: string };

type ImageApiResponse = {
    data?: Array<Record<string, unknown>>;
    error?: { message?: string };
    code?: number;
    msg?: string;
};
type GeminiPart = {
    text?: string;
    thought?: boolean;
    inlineData?: { mimeType?: string; data?: string };
    inline_data?: { mime_type?: string; mimeType?: string; data?: string };
    fileData?: { mimeType?: string; fileUri?: string };
    functionCall?: { id?: string; name?: string; args?: Record<string, unknown> };
    functionResponse?: { id?: string; name?: string; response?: Record<string, unknown> };
    thoughtSignature?: string;
    thought_signature?: string;
};
type GeminiContent = { role?: "user" | "model"; parts: GeminiPart[] };
type GeminiPayload = {
    candidates?: Array<{ content?: { parts?: GeminiPart[] }; finishReason?: string }>;
    models?: Array<{ name?: string }>;
    error?: { message?: string };
    promptFeedback?: { blockReason?: string };
};
type GeminiStreamState = { buffer: string; text: string; reasoning: string; toolCalls: ResponseToolCall[]; error?: string };
type RequestOptions = { signal?: AbortSignal; promptCacheKey?: string; onReasoning?: (text: string) => void };

const QUALITY_BASE: Record<string, number> = {
    low: 1024,
    medium: 2048,
    high: 2880,
    standard: 1024,
    hd: 2048,
};
const QUALITY_ALIASES: Record<string, string> = {
    "1k": "low",
    "2k": "medium",
    "4k": "high",
};
const DEFAULT_IMAGE_SHORT_SIDE = 1024;
const IMAGE_SIZE_STEP = 16;
const IMAGE_MIN_PIXELS = 655360;
const IMAGE_MAX_PIXELS = 8294400;
const IMAGE_MAX_EDGE = 3840;
const IMAGE_MAX_RATIO = 3;
const IMAGE_OUTPUT_FORMAT = "png";
const VOLCENGINE_ARK_IMAGE_MAX_PIXELS = 4624220;

function normalizeQuality(quality: string) {
    const value = quality.trim().toLowerCase();
    const normalized = QUALITY_ALIASES[value] || value;
    return QUALITY_BASE[normalized] ? normalized : undefined;
}

/** grok2api / xAI Imagine：画布 quality 映射为 resolution（1k/2k）。 */
function normalizeGrokImageResolution(quality: string | undefined) {
    const value = (quality || "").trim().toLowerCase();
    if (!value || value === "auto") return undefined;
    if (value === "1k" || value === "low" || value === "standard") return "1k";
    if (value === "2k" || value === "medium" || value === "hd" || value === "high" || value === "4k") return "2k";
    return undefined;
}

/** Map "quality + ratio" to an explicit pixel dimension like "3840x2160". */
function resolveSize(quality: string | undefined, ratio: string): string {
    const parsedRatio = parseImageRatio(ratio);
    const basePixels = quality ? QUALITY_BASE[quality] : undefined;
    const isLandscape = parsedRatio.width >= parsedRatio.height;
    const longRatio = isLandscape ? parsedRatio.width / parsedRatio.height : parsedRatio.height / parsedRatio.width;
    let longSide: number;
    let shortSide: number;

    if (basePixels) {
        const targetPixels = basePixels * basePixels;
        const longSideRaw = Math.sqrt(targetPixels * longRatio);
        longSide = Math.floor(longSideRaw / IMAGE_SIZE_STEP) * IMAGE_SIZE_STEP;
        shortSide = Math.round(longSide / longRatio / IMAGE_SIZE_STEP) * IMAGE_SIZE_STEP;
    } else {
        shortSide = DEFAULT_IMAGE_SHORT_SIDE;
        longSide = Math.round((shortSide * longRatio) / IMAGE_SIZE_STEP) * IMAGE_SIZE_STEP;
    }

    const width = isLandscape ? longSide : shortSide;
    const height = isLandscape ? shortSide : longSide;
    validateImageSize(width, height);
    return `${width}x${height}`;
}

function parseImageRatio(value: string) {
    const parts = value.split(":");
    if (parts.length !== 2) throw new Error("图像尺寸格式不支持，请使用 auto、9:16 或 1024x1024");
    const w = Number(parts[0]);
    const h = Number(parts[1]);
    if (!Number.isFinite(w) || !Number.isFinite(h) || w <= 0 || h <= 0) throw new Error("图像比例必须是正数，例如 9:16");
    if (Math.max(w, h) / Math.min(w, h) > IMAGE_MAX_RATIO) throw new Error("图像宽高比不能超过 3:1，请调整尺寸");
    return { width: w, height: h };
}

function parseImageDimensions(value: string) {
    const match = value.match(/^(\d+)x(\d+)$/i);
    if (!match) return null;
    return { width: Number(match[1]), height: Number(match[2]) };
}

function validateImageSize(width: number, height: number) {
    if (!Number.isInteger(width) || !Number.isInteger(height) || width <= 0 || height <= 0) throw new Error("图像尺寸必须是正整数，例如 1024x1024");
    if (width % IMAGE_SIZE_STEP !== 0 || height % IMAGE_SIZE_STEP !== 0) throw new Error("图像尺寸的宽高必须是 16 的倍数，请调整尺寸");
    if (Math.max(width, height) > IMAGE_MAX_EDGE) throw new Error("图像尺寸最长边不能超过 3840px，请调整尺寸");
    if (Math.max(width, height) / Math.min(width, height) > IMAGE_MAX_RATIO) throw new Error("图像宽高比不能超过 3:1，请调整尺寸");
    const pixels = width * height;
    if (pixels < IMAGE_MIN_PIXELS || pixels > IMAGE_MAX_PIXELS) throw new Error("图像总像素需在 655360 到 8294400 之间，请调整尺寸");
}

function resolveRequestSize(quality: string | undefined, size: string) {
    const value = size.trim();
    if (!value || value.toLowerCase() === "auto") return undefined;
    const dimensions = parseImageDimensions(value);
    if (dimensions) {
        validateImageSize(dimensions.width, dimensions.height);
        return `${dimensions.width}x${dimensions.height}`;
    }
    if (value.includes(":")) return resolveSize(quality, value);
    throw new Error("图像尺寸格式不支持，请使用 auto、9:16 或 1024x1024");
}

function resolveAspectRatio(value: string) {
    const normalized = value.trim().toLowerCase().replace("×", "x");
    if (normalized.includes(":")) return normalized;
    const dimensions = parseImageDimensions(normalized);
    if (!dimensions) throw new Error("图像比例格式不支持，请使用 3:4 或 1024x1360");
    const divisor = dimensionGCD(dimensions.width, dimensions.height);
    return `${dimensions.width / divisor}:${dimensions.height / divisor}`;
}

function dimensionGCD(left: number, right: number) {
    while (right) [left, right] = [right, left % right];
    return Math.max(1, left);
}

function resolveImageRequestSize(profile: ImageCapabilityConfig, quality: string | undefined, size: string) {
    const request = imageSizeRequest(profile, size);
    if (!request) return undefined;
    const value = request.parameter === "size" ? resolveRequestSize(quality, request.value) : resolveAspectRatio(request.value);
    return value ? { parameter: request.parameter, value } : undefined;
}

function validateImageCapability(profile: ImageCapabilityConfig, references: ReferenceImage[], mask?: ReferenceImage) {
    if (references.length > profile.references.maxImages) throw new Error(`当前图片模型最多支持 ${profile.references.maxImages} 张参考图`);
    if (mask && !profile.references.maskSupported) throw new Error("当前图片模型不支持蒙版编辑");
    if (profile.references.maxImageBytes > 0 && references.some((image) => (image.bytes || 0) > profile.references.maxImageBytes)) throw new Error("参考图片文件超过当前模型大小限制");
}

function normalizeVolcengineArkImageSize(size: string | undefined) {
    if (!size) return undefined;
    const dimensions = parseImageDimensions(size);
    if (!dimensions || dimensions.width * dimensions.height <= VOLCENGINE_ARK_IMAGE_MAX_PIXELS) return size;
    const scale = Math.sqrt(VOLCENGINE_ARK_IMAGE_MAX_PIXELS / (dimensions.width * dimensions.height));
    let width = Math.floor((dimensions.width * scale) / 2) * 2;
    let height = Math.floor((dimensions.height * scale) / 2) * 2;
    while (width > 2 && height > 2 && width * height > VOLCENGINE_ARK_IMAGE_MAX_PIXELS) {
        if (width >= height) width -= 2;
        else height -= 2;
    }
    return `${width}x${height}`;
}

function resolveImageDataUrl(item: Record<string, unknown>) {
    if (typeof item.b64_json === "string" && item.b64_json) {
        return `data:image/png;base64,${item.b64_json}`;
    }
    if (typeof item.url === "string" && item.url) {
        return item.url;
    }
    return null;
}

function parseImagePayload(payload: ImageApiResponse) {
    if (typeof payload.code === "number" && payload.code !== 0) {
        throw new Error(payload.msg || "请求失败");
    }
    const images =
        payload.data
            ?.map(resolveImageDataUrl)
            .filter((value): value is string => Boolean(value))
            .map((dataUrl) => ({ id: nanoid(), dataUrl })) || [];

    if (images.length === 0) {
        throw new Error("接口没有返回图片");
    }

    return images;
}

function readAxiosError(error: unknown, fallback: string) {
    if (axios.isCancel(error)) return "请求已取消";
    if (axios.isAxiosError<{ error?: { message?: string }; msg?: string; code?: number }>(error)) {
        const responseData = error.response?.data;
        return responseData?.msg || responseData?.error?.message || readStatusError(error.response?.status, fallback);
    }
    if (error instanceof DOMException && error.name === "AbortError") return "请求已取消";
    return error instanceof Error ? error.message : fallback;
}

function readStatusError(status: number | undefined, fallback: string) {
    if (status === 401 || status === 403) return "鉴权失败，请检查 API Key、套餐权限或模型权限";
    if (status === 429) return "请求被限流或额度不足，请稍后重试";
    return status ? `${fallback}：${status}` : fallback;
}

function withSystemPrompt(config: AiConfig, prompt: string) {
    const systemPrompt = config.systemPrompt.trim();
    return systemPrompt ? `${systemPrompt}\n\n${prompt}` : prompt;
}

function aiApiUrl(config: AiConfig, path: string) {
    return buildApiUrl(config.baseUrl, path);
}

function aiHeaders(config: AiConfig, contentType?: string) {
    return {
        Authorization: `Bearer ${config.apiKey}`,
        ...(contentType ? { "Content-Type": contentType } : {}),
        ...(isSystemProxyBaseUrl(config.baseUrl) ? { "X-Canvas-Scene": "image", "X-Idempotency-Key": createClientId() } : {}),
    };
}

async function postVolcengineArkImage(config: ReturnType<typeof resolveModelRequestConfig>, payload: Record<string, unknown>, options?: RequestOptions) {
    const upstreamUrl = aiApiUrl(config, "/images/generations");
    const request = channelRequest(config, upstreamUrl, aiHeaders(config, "application/json"));
    return (
        await axios.post<ImageApiResponse>(request.url, payload, {
            headers: request.headers,
            withCredentials: request.credentials === "include",
            signal: options?.signal,
        })
    ).data;
}

function geminiBaseUrl(config: Pick<AiConfig, "baseUrl">) {
    const normalizedBaseUrl = resolveBackendApiUrl(config.baseUrl).replace(/\/+$/, "");
    const lowerBaseUrl = normalizedBaseUrl.toLowerCase();
    return isSystemProxyBaseUrl(normalizedBaseUrl) || lowerBaseUrl.endsWith("/v1") || lowerBaseUrl.endsWith("/v1beta") ? normalizedBaseUrl : `${normalizedBaseUrl}/v1beta`;
}

function geminiModelName(model: string) {
    return model.trim().replace(/^models\//, "");
}

function geminiApiUrl(config: Pick<AiConfig, "baseUrl" | "model">, action?: "generateContent" | "streamGenerateContent") {
    const baseUrl = geminiBaseUrl(config);
    if (!action) return `${baseUrl}/models`;
    return `${baseUrl}/models/${encodeURIComponent(geminiModelName(config.model))}:${action}`;
}

function geminiHeaders(config: Pick<AiConfig, "apiKey">) {
    return {
        "x-goog-api-key": config.apiKey,
        "Content-Type": "application/json",
    };
}

function withSystemMessage<T extends ResponseInputMessage>(config: AiConfig, messages: T[]): ResponseInputMessage[] {
    const systemPrompt = config.systemPrompt.trim();
    return systemPrompt ? [{ role: "system" as const, content: systemPrompt }, ...messages] : messages;
}

function toResponseInput(messages: ResponseInputMessage[]): ResponseInputItem[] {
    return messages.flatMap((message): ResponseInputItem[] => {
        if ("type" in message) return [message];
        if (message.role === "tool") return [{ type: "function_call_output", call_id: message.tool_call_id, output: message.content }];
        return [{ role: message.role, content: toResponseContent(message.content || "") }];
    });
}

function toResponseContent(content: ResponseMessageContent): string | ResponseInputContent[] {
    if (!Array.isArray(content)) return String(content || "");
    return content.map((item) => {
        if (item.type === "text") return { type: "input_text" as const, text: item.text };
        if (item.type === "image_url") return { type: "input_image" as const, image_url: item.image_url.url };
        return item.file_url.url.startsWith("data:") ? { type: "input_file" as const, filename: item.file_url.name, file_data: item.file_url.url } : { type: "input_file" as const, filename: item.file_url.name, file_url: item.file_url.url };
    });
}

function toResponseTool(tool: ResponseFunctionTool): ResponseApiToolDefinition {
    return {
        type: "function",
        name: tool.function.name,
        description: tool.function.description,
        parameters: tool.function.parameters,
        strict: tool.function.strict,
    };
}

function toChatCompletionMessages(messages: ResponseInputMessage[]) {
    const result: Array<Record<string, unknown>> = [];
    for (let index = 0; index < messages.length;) {
        const message = messages[index];
        if ("type" in message) {
            const toolCalls: Array<Record<string, unknown>> = [];
            while (index < messages.length && "type" in messages[index]) {
                const call = messages[index] as Extract<ResponseInputMessage, { type: "function_call" }>;
                toolCalls.push({ id: call.call_id, type: "function", function: { name: call.name, arguments: call.arguments } });
                index += 1;
            }
            result.push({ role: "assistant", content: null, tool_calls: toolCalls });
            continue;
        }
        if (message.role === "tool") {
            result.push({ role: "tool", tool_call_id: message.tool_call_id, content: message.content });
        } else {
            result.push({ role: message.role, content: toChatCompletionContent(message.content) });
        }
        index += 1;
    }
    return result;
}

function toChatCompletionContent(content: ResponseMessageContent) {
    if (!Array.isArray(content)) return content;
    return content.map((item) => {
        if (item.type !== "file_url") return item;
        const file = item.file_url.url.startsWith("data:") ? { filename: item.file_url.name, file_data: item.file_url.url } : { filename: item.file_url.name, file_url: item.file_url.url };
        return { type: "file", file };
    });
}

function toChatCompletionToolChoice(toolChoice: ToolChoice) {
    return typeof toolChoice === "object" ? { type: "function", function: { name: toolChoice.name } } : toolChoice;
}

function isToolChoiceCompatibilityError(error: unknown) {
    const message = error instanceof Error ? error.message : "";
    return /tool[_\s-]?choice|thinking\s+mode/i.test(message);
}

function parseChatCompletionPayload(payload: ChatCompletionPayload): ToolResponseResult {
    if (typeof payload.code === "number" && payload.code !== 0) throw new Error(payload.msg || "请求失败");
    if (payload.error?.message) throw new Error(payload.error.message);
    const message = payload.choices?.[0]?.message;
    const toolCalls = (message?.tool_calls || [])
        .map((call) => ({
            id: call.id || "",
            type: "function" as const,
            function: { name: call.function?.name || "", arguments: call.function?.arguments || "{}" },
        }))
        .filter((call) => call.id && call.function.name);
    return { content: message?.content || "", toolCalls, ...(message?.reasoning_content ? { reasoning: message.reasoning_content } : {}) };
}

function parseToolResponse(payload: ResponseApiPayload): ToolResponseResult {
    const output = payload.output || [];
    const content =
        payload.output_text ||
        output
            .flatMap((item) => (item.type === "message" ? item.content || [] : []))
            .map((item) => item.text || "")
            .join("");
    const toolCalls = output
        .filter((item): item is Extract<ResponseApiOutputItem, { type?: "function_call" }> => item.type === "function_call")
        .map((item) => ({
            id: item.call_id || item.id || "",
            type: "function" as const,
            function: { name: item.name || "", arguments: item.arguments || "{}" },
        }))
        .filter((item) => item.id && item.function.name);
    return { content, toolCalls };
}

function isRecord(value: unknown): value is Record<string, unknown> {
    return Boolean(value && typeof value === "object" && !Array.isArray(value));
}

function responseErrorMessage(value: unknown) {
    if (!isRecord(value)) return "";
    const error = isRecord(value.error) ? value.error : undefined;
    const response = isRecord(value.response) ? value.response : undefined;
    const responseError = response && isRecord(response.error) ? response.error : undefined;
    return stringValue(value.msg) || stringValue(error?.message) || stringValue(responseError?.message);
}

function stringValue(value: unknown) {
    return typeof value === "string" ? value : "";
}

function validateResponsePayload(payload: ResponseApiPayload) {
    if (typeof payload.code === "number" && payload.code !== 0) throw new Error(payload.msg || "请求失败");
    if (payload.error?.message) throw new Error(payload.error.message);
}

function validateGeminiPayload(payload: GeminiPayload) {
    if (payload.error?.message) throw new Error(payload.error.message);
    if (payload.promptFeedback?.blockReason) throw new Error(`Gemini 拒绝了本次请求：${payload.promptFeedback.blockReason}`);
}

async function readFetchError(response: Response, fallback: string) {
    const text = await response.text();
    if (!text) return readStatusError(response.status, fallback);
    try {
        return responseErrorMessage(JSON.parse(text)) || readStatusError(response.status, fallback);
    } catch {
        return text.slice(0, 300) || readStatusError(response.status, fallback);
    }
}

async function readJsonPayload<T>(response: Response, fallback: string): Promise<T> {
    const text = await response.text();
    try {
        return JSON.parse(text) as T;
    } catch {
        if (/^\s*(?:<!doctype|<html)/i.test(text)) throw new Error("后端代理返回了前端网页，请检查 VITE_CANVAS_BACKEND_URL 和反向代理配置");
        throw new Error(`${fallback}：接口没有返回有效 JSON`);
    }
}

function consumeResponseStreamBlock(block: string, state: ResponseStreamState, onDelta?: (text: string) => void, onReasoning?: (text: string) => void) {
    const data = block
        .split(/\r?\n/)
        .filter((line) => line.startsWith("data:"))
        .map((line) => line.slice(5).replace(/^ /, ""))
        .join("\n")
        .trim();
    if (!data || data === "[DONE]") return;
    const event = JSON.parse(data) as Record<string, unknown>;
    const type = stringValue(event.type);
    const errorMessage = responseErrorMessage(event);
    if (errorMessage) state.error = errorMessage;
    if (type === "response.output_text.delta" && typeof event.delta === "string") {
        state.text += event.delta;
        onDelta?.(state.text);
    }
    if (type === "response.output_text.done" && !state.text && typeof event.text === "string") {
        state.text = event.text;
        onDelta?.(state.text);
    }
    if (["response.reasoning.delta", "response.reasoning_text.delta", "response.reasoning_summary_text.delta"].includes(type) && typeof event.delta === "string") {
        state.reasoning += event.delta;
        onReasoning?.(state.reasoning);
    }
    if (["response.reasoning.done", "response.reasoning_text.done", "response.reasoning_summary_text.done"].includes(type) && !state.reasoning && typeof event.text === "string") {
        state.reasoning = event.text;
        onReasoning?.(state.reasoning);
    }
    if (type === "response.completed" && isRecord(event.response)) {
        state.payload = event.response as ResponseApiPayload;
    } else if (Array.isArray(event.output)) {
        state.payload = event as ResponseApiPayload;
    }
}

function consumeResponseStreamText(state: ResponseStreamState, text: string, onDelta?: (text: string) => void, onReasoning?: (text: string) => void, flush = false) {
    state.buffer += text;
    for (;;) {
        const match = state.buffer.match(/\r?\n\r?\n/);
        if (!match) break;
        const index = match.index ?? 0;
        consumeResponseStreamBlock(state.buffer.slice(0, index), state, onDelta, onReasoning);
        state.buffer = state.buffer.slice(index + match[0].length);
    }
    if (flush && state.buffer.trim()) {
        consumeResponseStreamBlock(state.buffer, state, onDelta, onReasoning);
        state.buffer = "";
    }
}

async function requestStreamingResponse(config: AiConfig, body: Record<string, unknown>, onDelta?: (text: string) => void, options?: RequestOptions): Promise<ToolResponseResult> {
    const request = channelRequest(config, aiApiUrl(config, "/responses"), { ...aiHeaders(config, "application/json"), Accept: "text/event-stream" });
    const response = await fetch(request.url, {
        method: "POST",
        headers: request.headers,
        body: JSON.stringify({ ...body, stream: true }),
        signal: options?.signal,
        credentials: request.credentials,
    });
    if (!response.ok) throw new Error(await readFetchError(response, "请求失败"));
    if (!response.body) {
        const payload = (await response.json()) as ResponseApiPayload;
        validateResponsePayload(payload);
        return parseToolResponse(payload);
    }

    const reader = response.body.getReader();
    const decoder = new TextDecoder();
    const state: ResponseStreamState = { buffer: "", text: "", reasoning: "" };
    for (;;) {
        const { done, value } = await reader.read();
        if (done) break;
        consumeResponseStreamText(state, decoder.decode(value, { stream: true }), onDelta, options?.onReasoning);
        if (state.error) throw new Error(state.error);
    }
    consumeResponseStreamText(state, decoder.decode(), onDelta, options?.onReasoning, true);
    if (state.error) throw new Error(state.error);
    if (!state.payload) return { content: state.text, toolCalls: [], ...(state.reasoning ? { reasoning: state.reasoning } : {}) };
    validateResponsePayload(state.payload);
    const result = parseToolResponse(state.payload);
    return { ...result, content: state.text || result.content, ...(state.reasoning ? { reasoning: state.reasoning } : {}) };
}

function consumeChatCompletionStreamBlock(block: string, state: ChatCompletionStreamState, onDelta?: (text: string) => void, onReasoning?: (text: string) => void) {
    const data = block
        .split(/\r?\n/)
        .filter((line) => line.startsWith("data:"))
        .map((line) => line.slice(5).replace(/^ /, ""))
        .join("\n")
        .trim();
    if (!data || data === "[DONE]") return;
    const event = JSON.parse(data) as Record<string, unknown>;
    const errorMessage = responseErrorMessage(event);
    if (errorMessage) state.error = errorMessage;
    const choices = Array.isArray(event.choices) ? event.choices : [];
    const choice = isRecord(choices[0]) ? choices[0] : undefined;
    const delta = choice && isRecord(choice.delta) ? choice.delta : undefined;
    if (!delta) return;
    if (typeof delta.content === "string") {
        state.text += delta.content;
        onDelta?.(state.text);
    }
    const reasoningDelta = stringValue(delta.reasoning_content) || stringValue(delta.reasoning) || stringValue(delta.reasoning_text);
    if (reasoningDelta) {
        state.reasoning += reasoningDelta;
        onReasoning?.(state.reasoning);
    }
    const chunks = Array.isArray(delta.tool_calls) ? delta.tool_calls : [];
    chunks.forEach((value, fallbackIndex) => {
        if (!isRecord(value)) return;
        const callIndex = typeof value.index === "number" ? value.index : fallbackIndex;
        const current = state.toolCalls.get(callIndex) || { id: "", name: "", arguments: "" };
        const fn = isRecord(value.function) ? value.function : undefined;
        state.toolCalls.set(callIndex, {
            id: stringValue(value.id) || current.id,
            name: stringValue(fn?.name) || current.name,
            arguments: current.arguments + stringValue(fn?.arguments),
        });
    });
}

function consumeChatCompletionStreamText(state: ChatCompletionStreamState, text: string, onDelta?: (text: string) => void, onReasoning?: (text: string) => void, flush = false) {
    state.buffer += text;
    for (;;) {
        const match = state.buffer.match(/\r?\n\r?\n/);
        if (!match) break;
        const index = match.index ?? 0;
        consumeChatCompletionStreamBlock(state.buffer.slice(0, index), state, onDelta, onReasoning);
        state.buffer = state.buffer.slice(index + match[0].length);
    }
    if (flush && state.buffer.trim()) {
        consumeChatCompletionStreamBlock(state.buffer, state, onDelta, onReasoning);
        state.buffer = "";
    }
}

async function requestStreamingChatCompletion(config: AiConfig, body: Record<string, unknown>, onDelta?: (text: string) => void, options?: RequestOptions): Promise<ToolResponseResult> {
    const request = channelRequest(config, aiApiUrl(config, "/chat/completions"), { ...aiHeaders(config, "application/json"), Accept: "text/event-stream" });
    const response = await fetch(request.url, {
        method: "POST",
        headers: request.headers,
        body: JSON.stringify({ ...body, stream: true }),
        signal: options?.signal,
        credentials: request.credentials,
    });
    if (!response.ok) throw new Error(await readFetchError(response, "请求失败"));
    const contentType = response.headers.get("content-type") || "";
    if (!response.body || !contentType.includes("text/event-stream")) {
        const result = parseChatCompletionPayload(await readJsonPayload<ChatCompletionPayload>(response, "请求失败"));
        if (result.reasoning) options?.onReasoning?.(result.reasoning);
        return result;
    }

    const reader = response.body.getReader();
    const decoder = new TextDecoder();
    const state: ChatCompletionStreamState = { buffer: "", text: "", reasoning: "", toolCalls: new Map() };
    for (;;) {
        const { done, value } = await reader.read();
        if (done) break;
        consumeChatCompletionStreamText(state, decoder.decode(value, { stream: true }), onDelta, options?.onReasoning);
        if (state.error) throw new Error(state.error);
    }
    consumeChatCompletionStreamText(state, decoder.decode(), onDelta, options?.onReasoning, true);
    if (state.error) throw new Error(state.error);
    const toolCalls = Array.from(state.toolCalls.entries())
        .sort(([left], [right]) => left - right)
        .map(([, call]) => ({ id: call.id, type: "function" as const, function: { name: call.name, arguments: call.arguments || "{}" } }))
        .filter((call) => call.id && call.function.name);
    return { content: state.text, toolCalls, ...(state.reasoning ? { reasoning: state.reasoning } : {}) };
}

function toGeminiBody(config: AiConfig, messages: ResponseInputMessage[], extra?: Record<string, unknown>) {
    const systemText = [config.systemPrompt.trim(), ...messages.flatMap((message) => (!("type" in message) && message.role === "system" ? [geminiTextContent(message.content)] : []))].filter(Boolean).join("\n\n");
    const contents = toGeminiContents(messages.filter((message) => ("type" in message ? true : message.role !== "system")));
    return {
        contents,
        ...(systemText ? { systemInstruction: { parts: [{ text: systemText }] } } : {}),
        ...extra,
    };
}

function toGeminiContents(messages: ResponseInputMessage[]): GeminiContent[] {
    const callNameById = new Map<string, string>();
    return messages.flatMap((message): GeminiContent[] => {
        if ("type" in message) {
            callNameById.set(message.call_id, message.name);
            return [{ role: "model", parts: [{ functionCall: { id: message.call_id, name: message.name, args: jsonObject(message.arguments) }, ...(message.thoughtSignature ? { thoughtSignature: message.thoughtSignature } : {}) }] }];
        }
        if (message.role === "tool") {
            const name = callNameById.get(message.tool_call_id) || "tool_result";
            return [{ role: "user", parts: [{ functionResponse: { id: message.tool_call_id, name, response: { result: jsonValue(message.content) } } }] }];
        }
        return [{ role: message.role === "assistant" ? "model" : "user", parts: toGeminiParts(message.content) }];
    });
}

function toGeminiParts(content: ResponseMessageContent): GeminiPart[] {
    if (!Array.isArray(content)) return [{ text: String(content || "") }];
    return content.map((item) => {
        if (item.type === "text") return { text: item.text };
        if (item.type === "image_url") return toGeminiFilePart(item.image_url.url, "image/png");
        return toGeminiFilePart(item.file_url.url, item.file_url.mimeType);
    });
}

function toGeminiFilePart(url: string, mimeType: string): GeminiPart {
    const match = url.match(/^data:([^;,]+);base64,(.+)$/);
    if (match) return { inlineData: { mimeType: match[1], data: match[2] } };
    return { fileData: { fileUri: url, mimeType } };
}

function geminiTextContent(content: ResponseMessageContent) {
    if (!Array.isArray(content)) return String(content || "");
    return content.map((item) => (item.type === "text" ? item.text : item.type === "image_url" ? item.image_url.url : `${item.file_url.name}: ${item.file_url.url}`)).join("\n");
}

function jsonObject(value: string): Record<string, unknown> {
    const parsed = jsonValue(value);
    return isRecord(parsed) ? parsed : {};
}

function jsonValue(value: string): unknown {
    try {
        return JSON.parse(value);
    } catch {
        return value;
    }
}

function toGeminiToolOptions(tools: ResponseFunctionTool[], toolChoice: ToolChoice) {
    if (!tools.length) return {};
    const functionDeclarations = tools.map((tool) => ({
        name: tool.function.name,
        description: tool.function.description,
        parameters: tool.function.parameters,
    }));
    const functionCallingConfig = typeof toolChoice === "object" ? { mode: "ANY", allowedFunctionNames: [toolChoice.name] } : { mode: toolChoice === "required" ? "ANY" : "AUTO" };
    return {
        tools: [{ functionDeclarations }],
        toolConfig: { functionCallingConfig },
    };
}

async function requestGeminiStreamingResponse(config: AiConfig, body: Record<string, unknown>, onDelta?: (text: string) => void, options?: RequestOptions): Promise<ToolResponseResult> {
    const request = channelRequest(config, `${geminiApiUrl(config, "streamGenerateContent")}?alt=sse`, geminiHeaders(config));
    const response = await fetch(request.url, {
        method: "POST",
        headers: request.headers,
        body: JSON.stringify(body),
        signal: options?.signal,
        credentials: request.credentials,
    });
    if (!response.ok) throw new Error(await readFetchError(response, "请求失败"));
    if (!response.body) {
        const payload = (await response.json()) as GeminiPayload;
        const result = parseGeminiToolResponse(payload);
        if (result.reasoning) options?.onReasoning?.(result.reasoning);
        return result;
    }

    const reader = response.body.getReader();
    const decoder = new TextDecoder();
    const state: GeminiStreamState = { buffer: "", text: "", reasoning: "", toolCalls: [] };
    for (;;) {
        const { done, value } = await reader.read();
        if (done) break;
        consumeGeminiStreamText(state, decoder.decode(value, { stream: true }), onDelta, options?.onReasoning);
        if (state.error) throw new Error(state.error);
    }
    consumeGeminiStreamText(state, decoder.decode(), onDelta, options?.onReasoning, true);
    if (state.error) throw new Error(state.error);
    return { content: state.text, toolCalls: state.toolCalls, ...(state.reasoning ? { reasoning: state.reasoning } : {}) };
}

function consumeGeminiStreamText(state: GeminiStreamState, text: string, onDelta?: (text: string) => void, onReasoning?: (text: string) => void, flush = false) {
    state.buffer += text;
    for (;;) {
        const match = state.buffer.match(/\r?\n\r?\n/);
        if (!match) break;
        const index = match.index ?? 0;
        consumeGeminiStreamBlock(state.buffer.slice(0, index), state, onDelta, onReasoning);
        state.buffer = state.buffer.slice(index + match[0].length);
    }
    if (flush && state.buffer.trim()) {
        consumeGeminiStreamBlock(state.buffer, state, onDelta, onReasoning);
        state.buffer = "";
    }
}

function consumeGeminiStreamBlock(block: string, state: GeminiStreamState, onDelta?: (text: string) => void, onReasoning?: (text: string) => void) {
    const data = block
        .split(/\r?\n/)
        .filter((line) => line.startsWith("data:"))
        .map((line) => line.slice(5).replace(/^ /, ""))
        .join("\n")
        .trim();
    if (!data || data === "[DONE]") return;
    const result = parseGeminiToolResponse(JSON.parse(data) as GeminiPayload);
    if (result.reasoning) {
        state.reasoning += result.reasoning;
        onReasoning?.(state.reasoning);
    }
    if (result.content) {
        state.text += result.content;
        onDelta?.(state.text);
    }
    state.toolCalls.push(...result.toolCalls);
}

function parseGeminiToolResponse(payload: GeminiPayload): ToolResponseResult {
    validateGeminiPayload(payload);
    const parts = payload.candidates?.flatMap((candidate) => candidate.content?.parts || []) || [];
    const content = parts
        .filter((part) => !part.thought)
        .map((part) => part.text || "")
        .join("");
    const reasoning = parts
        .filter((part) => part.thought)
        .map((part) => part.text || "")
        .join("");
    const toolCalls = parts
        .map((part) => part.functionCall)
        .filter((call): call is NonNullable<GeminiPart["functionCall"]> => Boolean(call?.name))
        .map((call) => {
            const part = parts.find((item) => item.functionCall === call);
            const thoughtSignature = part?.thoughtSignature || part?.thought_signature;
            return {
                id: call.id || nanoid(),
                type: "function" as const,
                function: { name: call.name || "", arguments: JSON.stringify(call.args || {}) },
                ...(thoughtSignature ? { thoughtSignature } : {}),
            };
        });
    return { content, toolCalls, ...(reasoning ? { reasoning } : {}) };
}

async function requestGeminiImages(config: AiConfig, prompt: string, references: ReferenceImage[], count: number, generationConfig: GeminiImageGenerationConfig, options?: RequestOptions) {
    // 参考图先完整读取并校验一次，再复用已解析的 inlineData；不能让每个输出重新读取 storageKey，
    // 否则资源缓存瞬时未命中时会出现“偶发缺参考图”或把 storageKey 当 URL 请求的问题。
    const referenceParts = await Promise.all(
        references.map(async (image) => {
            const dataUrl = await imageToDataUrl(image);
            const inlineData = parseGeminiImageDataUrl(dataUrl);
            return { inlineData };
        }),
    );
    const requests = Array.from({ length: count }, () => requestGeminiImagesOnce(config, prompt, referenceParts, generationConfig, options));
    return (await Promise.all(requests)).flat();
}

async function requestGeminiImagesOnce(config: AiConfig, prompt: string, referenceParts: GeminiPart[], generationConfig: GeminiImageGenerationConfig, options?: RequestOptions) {
    const parts: GeminiPart[] = [{ text: prompt }, ...referenceParts];
    const request = channelRequest(config, geminiApiUrl(config, "generateContent"), geminiHeaders(config));
    const response = await axios.post<GeminiPayload>(
        request.url,
        {
            ...toGeminiBody(config, [{ role: "user", content: prompt }], { generationConfig }),
            contents: [{ role: "user", parts }],
        },
        { headers: request.headers, withCredentials: request.credentials === "include", signal: options?.signal },
    );
    return parseGeminiImagePayload(response.data);
}

function parseGeminiImagePayload(payload: GeminiPayload) {
    validateGeminiPayload(payload);
    const images =
        payload.candidates
            ?.flatMap((candidate) => candidate.content?.parts || [])
            .map((part) => {
                const inlineData = part.inlineData || (part.inline_data ? { mimeType: part.inline_data.mimeType || part.inline_data.mime_type, data: part.inline_data.data } : undefined);
                if (inlineData?.data) return `data:${inlineData.mimeType || "image/png"};base64,${inlineData.data}`;
                return part.fileData?.fileUri || null;
            })
            .filter((value): value is string => Boolean(value))
            .map((dataUrl) => ({ id: nanoid(), dataUrl })) || [];
    if (!images.length) throw new Error("Gemini 接口没有返回图片");
    return images;
}

export async function requestGeneration(config: AiConfig, prompt: string, options?: RequestOptions) {
    const selectedModel = config.model || config.imageModel;
    const requestConfig = resolveModelRequestConfig(config, selectedModel);
    const imageProfile = modelCapabilityConfigFor(config, selectedModel).image!;
    validateImageCapability(imageProfile, []);
    const normalizedImage = normalizeImageValue(imageProfile, config);
    const n = Number(normalizedImage.count);
    if (requestConfig.interfaceType === "gemini-image") {
        try {
            return await requestGeminiImages(requestConfig, prompt, [], n, buildGeminiImageGenerationConfig(normalizedImage.size, normalizedImage.quality), options);
        } catch (error) {
            throw new Error(readAxiosError(error, "请求失败"));
        }
    }
    if (requestConfig.interfaceType === "grok-image") {
        try {
            const size = normalizedImage.size && normalizedImage.size !== "auto" ? normalizedImage.size : undefined;
            const aspectRatio = size?.includes(":") ? size : undefined;
            const resolution = normalizeGrokImageResolution(normalizedImage.quality);
            const responseData = await postChannelJSON<ImageApiResponse>(
                requestConfig,
                aiApiUrl(requestConfig, "/images/generations"),
                {
                    model: requestConfig.model,
                    prompt: withSystemPrompt(requestConfig, prompt),
                    n,
                    response_format: "url",
                    ...(size ? { size } : {}),
                    ...(aspectRatio ? { aspect_ratio: aspectRatio } : {}),
                    ...(resolution ? { resolution } : {}),
                },
                options,
            );
            return parseImagePayload(responseData);
        } catch (error) {
            throw new Error(readAxiosError(error, "Grok 图片生成失败"));
        }
    }
    const quality = imageProfile.quality.supported && normalizedImage.quality !== "auto" ? normalizeQuality(normalizedImage.quality) || normalizedImage.quality : undefined;
    const requestSize = resolveImageRequestSize(imageProfile, quality, normalizedImage.size);
    const isVolcengineArk = requestConfig.interfaceType === "volcengine-ark-image";
    const normalizedRequestSize = requestSize?.parameter === "size" && isVolcengineArk ? { ...requestSize, value: normalizeVolcengineArkImageSize(requestSize.value)! } : requestSize;
    try {
        const payload = isVolcengineArk
            ? {
                  model: requestConfig.model,
                  prompt: withSystemPrompt(requestConfig, prompt),
                  n,
                  watermark: false,
                  ...(normalizedRequestSize ? { [normalizedRequestSize.parameter]: normalizedRequestSize.value } : {}),
              }
            : {
                  model: requestConfig.model,
                  prompt: withSystemPrompt(requestConfig, prompt),
                  n,
                  ...(quality ? { quality } : {}),
                  ...(requestSize ? { [requestSize.parameter]: requestSize.value } : {}),
                  ...(imageProfile.responseFormat.supported ? { response_format: "b64_json" } : {}),
                  ...(imageProfile.outputFormat.supported ? { output_format: IMAGE_OUTPUT_FORMAT } : {}),
                  ...(imageProfile.transparentBackground.supported && normalizedImage.transparentBackground === "true" ? { background: "transparent" } : {}),
              };
        const responseData = isVolcengineArk ? await postVolcengineArkImage(requestConfig, payload, options) : await postChannelJSON<ImageApiResponse>(requestConfig, aiApiUrl(requestConfig, "/images/generations"), payload, options);
        const images = parseImagePayload(responseData);
        return images;
    } catch (error) {
        throw new Error(readAxiosError(error, "请求失败"));
    }
}

async function postChannelJSON<T>(config: ReturnType<typeof resolveModelRequestConfig>, upstreamUrl: string, body: unknown, options?: RequestOptions) {
    const request = channelRequest(config, upstreamUrl, aiHeaders(config, "application/json"));
    return (
        await axios.post<T>(request.url, body, {
            headers: request.headers,
            withCredentials: request.credentials === "include",
            signal: options?.signal,
        })
    ).data;
}

async function grokImageInputURL(image: ReferenceImage) {
    const candidate = image.url?.trim() || "";
    if (/^https?:\/\//i.test(candidate)) return candidate;
    return imageToDataUrl(image);
}

export async function requestEdit(config: AiConfig, prompt: string, references: ReferenceImage[], mask?: ReferenceImage, options?: RequestOptions) {
    const selectedModel = config.model || config.imageModel;
    const requestConfig = resolveModelRequestConfig(config, selectedModel);
    const imageProfile = modelCapabilityConfigFor(config, selectedModel).image!;
    validateImageCapability(imageProfile, references, mask);
    const normalizedImage = normalizeImageValue(imageProfile, config);
    const n = Number(normalizedImage.count);
    const requestPrompt = buildImageReferencePromptText(prompt, references);
    if (requestConfig.interfaceType === "gemini-image") {
        if (mask) throw new Error("Gemini 调用格式暂不支持蒙版编辑");
        try {
            return await requestGeminiImages(requestConfig, requestPrompt, references, n, buildGeminiImageGenerationConfig(normalizedImage.size, normalizedImage.quality), options);
        } catch (error) {
            throw new Error(readAxiosError(error, "请求失败"));
        }
    }
    if (requestConfig.interfaceType === "grok-image") {
        if (mask) throw new Error("Grok 图片协议不支持蒙版编辑，请移除蒙版后重试");
        if (references.length !== 1) throw new Error("Grok 图片编辑必须提供且仅支持 1 张参考图");
        try {
            const imageUrl = await grokImageInputURL(references[0]);
            const size = normalizedImage.size && normalizedImage.size !== "auto" ? normalizedImage.size : undefined;
            const aspectRatio = size?.includes(":") ? size : undefined;
            const resolution = normalizeGrokImageResolution(normalizedImage.quality);
            const response = await postChannelJSON<ImageApiResponse>(
                requestConfig,
                aiApiUrl(requestConfig, "/images/edits"),
                {
                    model: requestConfig.model,
                    prompt: withSystemPrompt(requestConfig, requestPrompt),
                    image: { url: imageUrl },
                    n,
                    response_format: "url",
                    ...(size ? { size } : {}),
                    ...(aspectRatio ? { aspect_ratio: aspectRatio } : {}),
                    ...(resolution ? { resolution } : {}),
                },
                options,
            );
            return parseImagePayload(response);
        } catch (error) {
            throw new Error(readAxiosError(error, "Grok 图片编辑失败"));
        }
    }
    if (requestConfig.interfaceType === "volcengine-ark-image") {
        if (mask) throw new Error("火山方舟图片协议不支持蒙版编辑，请移除蒙版后重试");
        const quality = imageProfile.quality.supported && normalizedImage.quality !== "auto" ? normalizeQuality(normalizedImage.quality) || normalizedImage.quality : undefined;
        const sizeRequest = resolveImageRequestSize(imageProfile, quality, normalizedImage.size);
        const requestSize = sizeRequest?.parameter === "size" ? { ...sizeRequest, value: normalizeVolcengineArkImageSize(sizeRequest.value)! } : sizeRequest;
        try {
            const images = await Promise.all(references.map((image) => imageToDataUrl(image)));
            const response = await postVolcengineArkImage(
                requestConfig,
                {
                    model: requestConfig.model,
                    prompt: withSystemPrompt(requestConfig, requestPrompt),
                    n,
                    watermark: false,
                    ...(requestSize ? { [requestSize.parameter]: requestSize.value } : {}),
                    ...(images.length === 1 ? { image: images[0] } : images.length > 1 ? { image: images } : {}),
                },
                options,
            );
            return parseImagePayload(response);
        } catch (error) {
            throw new Error(readAxiosError(error, "火山方舟图片生成失败"));
        }
    }
    const quality = imageProfile.quality.supported && normalizedImage.quality !== "auto" ? normalizeQuality(normalizedImage.quality) || normalizedImage.quality : undefined;
    const requestSize = resolveImageRequestSize(imageProfile, quality, normalizedImage.size);
    const formData = new FormData();
    formData.set("model", requestConfig.model);
    formData.set("prompt", withSystemPrompt(requestConfig, requestPrompt));
    formData.set("n", String(n));
    if (imageProfile.responseFormat.supported) formData.set("response_format", "b64_json");
    if (imageProfile.outputFormat.supported) formData.set("output_format", IMAGE_OUTPUT_FORMAT);
    if (imageProfile.transparentBackground.supported && normalizedImage.transparentBackground === "true") {
        formData.set("background", "transparent");
    }
    if (quality) {
        formData.set("quality", quality);
    }
    if (requestSize) {
        formData.set(requestSize.parameter, requestSize.value);
    }
    const files = await Promise.all(references.map(async (image) => dataUrlToFile({ ...image, dataUrl: await imageToDataUrl(image) })));
    files.forEach((file) => formData.append("image", file));
    if (mask) formData.set("mask", dataUrlToFile(mask));

    try {
        const request = channelRequest(requestConfig, aiApiUrl(requestConfig, "/images/edits"), aiHeaders(requestConfig));
        const response = await axios.post<ImageApiResponse>(request.url, formData, { headers: request.headers, withCredentials: request.credentials === "include", signal: options?.signal });
        const images = parseImagePayload(response.data);
        return images;
    } catch (error) {
        throw new Error(readAxiosError(error, "请求失败"));
    }
}

export async function requestImageQuestion(config: AiConfig, messages: AiTextMessage[], onDelta: (text: string) => void, options?: RequestOptions) {
    const requestConfig = resolveModelRequestConfig(config, config.model || config.textModel);
    try {
        if (requestConfig.apiFormat === "gemini") {
            const answer = (await requestGeminiStreamingResponse(requestConfig, toGeminiBody(requestConfig, messages), onDelta, options)).content || "没有返回内容";
            if (answer === "没有返回内容") onDelta(answer);
            return answer;
        }
        if (requestConfig.interfaceType === "chat-completion" || !requestConfig.interfaceType) {
            const answer =
                (
                    await requestStreamingChatCompletion(
                        requestConfig,
                        {
                            model: requestConfig.model,
                            messages: toChatCompletionMessages(withSystemMessage(requestConfig, messages)),
                        },
                        onDelta,
                        options,
                    )
                ).content || "没有返回内容";
            if (answer === "没有返回内容") onDelta(answer);
            return answer;
        }
        const answer =
            (
                await requestStreamingResponse(
                    requestConfig,
                    {
                        model: requestConfig.model,
                        input: toResponseInput(withSystemMessage(requestConfig, messages)),
                    },
                    onDelta,
                    options,
                )
            ).content || "没有返回内容";
        if (answer === "没有返回内容") onDelta(answer);
        return answer;
    } catch (error) {
        throw new Error(readAxiosError(error, "请求失败"));
    }
}

export async function requestToolResponse(config: AiConfig, messages: ResponseInputMessage[], tools: ResponseFunctionTool[], toolChoice: ToolChoice = "auto", onDelta?: (text: string) => void, options?: RequestOptions): Promise<ToolResponseResult> {
    const requestConfig = resolveModelRequestConfig(config, config.model || config.textModel);
    try {
        if (requestConfig.apiFormat === "gemini") {
            return await requestGeminiStreamingResponse(requestConfig, toGeminiBody(requestConfig, messages, toGeminiToolOptions(tools, toolChoice)), onDelta, options);
        }
        if (requestConfig.interfaceType === "chat-completion" || !requestConfig.interfaceType) {
            const chatPayload: Record<string, unknown> = {
                model: requestConfig.model,
                messages: toChatCompletionMessages(withSystemMessage(requestConfig, messages)),
                tools,
                tool_choice: toChatCompletionToolChoice(toolChoice),
                parallel_tool_calls: false,
            };
            try {
                return await requestStreamingChatCompletion(requestConfig, chatPayload, onDelta, options);
            } catch (error) {
                if (!isToolChoiceCompatibilityError(error)) throw error;

                // 部分 OpenAI 兼容上游仅支持 auto，思考模式则可能要求完全省略该字段。
                if (toolChoice !== "auto") {
                    try {
                        return await requestStreamingChatCompletion(requestConfig, { ...chatPayload, tool_choice: toChatCompletionToolChoice("auto") }, onDelta, options);
                    } catch (autoError) {
                        if (!isToolChoiceCompatibilityError(autoError)) throw autoError;
                    }
                }
                const { tool_choice: _ignored, ...withoutToolChoice } = chatPayload;
                void _ignored;
                return await requestStreamingChatCompletion(requestConfig, withoutToolChoice, onDelta, options);
            }
        }
        return await requestStreamingResponse(
            requestConfig,
            withOpenAIPromptCacheKey(
                {
                    model: requestConfig.model,
                    input: toResponseInput(withSystemMessage(requestConfig, messages)),
                    tools: tools.map(toResponseTool),
                    tool_choice: toolChoice,
                    parallel_tool_calls: false,
                },
                options?.promptCacheKey,
            ),
            onDelta,
            options,
        );
    } catch (error) {
        throw new Error(readAxiosError(error, "请求失败"));
    }
}

export async function fetchImageModels(config: Pick<AiConfig, "baseUrl" | "apiKey" | "apiFormat"> & { allowLocalChannel?: boolean }) {
    try {
        if (config.apiFormat === "gemini") {
            const requestConfig = { ...defaultGeminiConfig, ...config };
            const request = channelRequest(requestConfig, geminiApiUrl(requestConfig), geminiHeaders(requestConfig));
            const response = await axios.get<GeminiPayload>(request.url, { headers: request.headers, withCredentials: request.credentials === "include" });
            validateGeminiPayload(response.data);
            return (response.data.models || [])
                .map((model) => model.name?.replace(/^models\//, ""))
                .filter((id): id is string => Boolean(id))
                .sort((a, b) => a.localeCompare(b));
        }
        const request = channelRequest(config, buildApiUrl(config.baseUrl, "/models"), { Authorization: `Bearer ${config.apiKey}` });
        const response = await axios.get<{ data?: Array<{ id?: string }>; error?: { message?: string } }>(request.url, { headers: request.headers, withCredentials: request.credentials === "include" });
        return (response.data.data || [])
            .map((model) => model.id)
            .filter((id): id is string => Boolean(id))
            .sort((a, b) => a.localeCompare(b));
    } catch (error) {
        throw new Error(readAxiosError(error, "读取模型失败"));
    }
}

export type ChannelModelFetchResult = { models: string[]; catalog: ChannelModelCatalogItem[] };

export async function fetchChannelModels(channel: ModelChannel, viaBackend = false): Promise<ChannelModelFetchResult> {
    const runtimeChannel = projectDesktopLocalChannelRuntime(channel);
    if (!viaBackend) {
        const models = await fetchImageModels({ baseUrl: runtimeChannel.baseUrl, allowLocalChannel: runtimeChannel.allowLocalChannel === true, apiKey: runtimeChannel.apiKey, apiFormat: runtimeChannel.apiFormat });
        return { models, catalog: models.map((id) => ({ id })) };
    }
    try {
        // 登录态由同源后端代取模型目录，避免每个 OpenAI 兼容服务分别维护浏览器 CORS 白名单。
        const response = await axios.post<{ code?: number; data?: { models?: Array<string | ChannelModelCatalogItem> }; msg?: string }>(
            resolveBackendApiUrl("/api/ai/models"),
            {
                baseUrl: runtimeChannel.baseUrl,
                allowLocalChannel: runtimeChannel.allowLocalChannel === true,
                apiKey: runtimeChannel.apiKey,
                apiFormat: runtimeChannel.apiFormat,
                headers: runtimeChannel.headers,
            },
            { withCredentials: true },
        );
        if (typeof response.data.code === "number" && response.data.code !== 0) {
            throw new Error(response.data.msg || "读取模型失败");
        }
        const catalog = new Map<string, ChannelModelCatalogItem>();
        for (const item of response.data.data?.models || []) {
            const entry = typeof item === "string" ? sanitizeChannelModelCatalogItem({ id: item }) : sanitizeChannelModelCatalogItem(item);
            if (!entry) continue;
            const existing = catalog.get(entry.id);
            catalog.set(entry.id, existing || entry);
        }
        const models = Array.from(catalog.keys()).sort((a, b) => a.localeCompare(b));
        const sortedCatalog = Array.from(catalog.values()).sort((a, b) => a.id.localeCompare(b.id));
        return { models, catalog: sortedCatalog };
    } catch (error) {
        throw new Error(readAxiosError(error, "读取模型失败"));
    }
}

const defaultGeminiConfig: Pick<AiConfig, "baseUrl" | "apiKey" | "apiFormat" | "model" | "systemPrompt"> = {
    baseUrl: "https://generativelanguage.googleapis.com",
    apiKey: "",
    apiFormat: "gemini",
    model: "",
    systemPrompt: "",
};