nolo-cli 0.1.39 → 0.1.41

package/ai/context/retention.ts

@@ -0,0 +1,165 @@
+// 文件路径: packages/ai/context/retention.ts
+
+/**
+ * 将 contextRetention slider (1–100) 映射为实际保留比例 R ∈ (0, 1)。
+ *
+ * 分三档线性映射：
+ * - 低保留：  1–30  → 15% ~ 30%
+ * - 中保留： 31–70 → 30% ~ 60%
+ * - 高保留： 71–100 → 60% ~ 85%
+ */
+export const mapRetentionSliderToRatio = (slider: number): number => {
+  const s = Math.min(100, Math.max(1, slider));
+
+  if (s <= 30) {
+    // Low retention: 15% ~ 30%
+    return 0.15 + (0.30 - 0.15) * (s / 30);
+  }
+
+  if (s <= 70) {
+    // Medium retention: 30% ~ 60%
+    return 0.30 + (0.60 - 0.30) * ((s - 30) / 40);
+  }
+
+  // High retention: 60% ~ 85%
+  return 0.60 + (0.85 - 0.60) * ((s - 70) / 30);
+};
+
+/**
+ * 安全系数：只允许历史对话最多占用 contextWindow 的 SAFE_BUFFER_RATIO 部分，
+ * 预留剩余空间给 system prompt / 工具描述等固定开销。
+ */
+export const SAFE_BUFFER_RATIO = 0.9;
+
+/**
+ * 对话负载分档：
+ * - light  ：短句聊天为主
+ * - medium ：中等长度混合
+ * - heavy  ：长代码 / 长文档为主
+ */
+export type ConversationLoad = "light" | "medium" | "heavy";
+
+export interface ContextPlan {
+  /**
+   * 历史总预算（summary + 原始消息），单位：token
+   * 实际不会超过 contextWindow * SAFE_BUFFER_RATIO
+   */
+  historyBudget: number;
+  /**
+   * 历史中分配给「原始消息」的预算，单位：token
+   * 用于 trimMessagesWithSummary 之类的裁剪逻辑。
+   */
+  rawMessageBudget: number;
+  /**
+   * 希望至少保留的「最近尾部消息」的 token 数，用作二次兜底，
+   * 确保短句对话能多保几轮，重内容对话能保住最近几条大块内容。
+   */
+  minTailTokens: number;
+}
+
+const clamp = (v: number, min: number, max: number): number =>
+  Math.min(max, Math.max(min, v));
+
+/**
+ * 统一的上下文使用规划函数：
+ * - 输入：模型 window、用户 slider、现有 summary 长度、近期对话负载分档
+ * - 输出：历史总预算 + 原始消息预算 + 尾部最少保留 token 数
+ *
+ * 约束与策略：
+ * - 历史最多只能占用 contextWindow * SAFE_BUFFER_RATIO
+ * - slider 决定「整体偏好多记还是少记」
+ * - recentLoad 决定在该偏好下的微调：
+ *   - light ：适当放宽，给更多历史空间
+ *   - heavy ：适当收紧，为当前大块内容 / 工具描述预留空间
+ */
+export const planContextUsage = (params: {
+  contextWindow: number;
+  retentionSlider: number;
+  summaryTokens: number;
+  recentLoad: ConversationLoad;
+}): ContextPlan => {
+  const { contextWindow, retentionSlider, summaryTokens, recentLoad } = params;
+
+  const safeWindowLimit = Math.floor(contextWindow * SAFE_BUFFER_RATIO);
+
+  const s = Math.min(100, Math.max(1, retentionSlider));
+  const iq = s / 100; // 0~1，表示用户偏好的“记忆强度”
+
+  const isBigWindow = contextWindow >= 128000; // 例如 128k / 1M 等大模型
+  const isSmallWindow = contextWindow <= 32000; // 例如 8k / 16k 等小模型
+
+  // 基础历史比例（只考虑 slider，不考虑对话负载）
+  // - 小窗口：历史占 35% ~ 70% 的 safeWindow
+  // - 中窗口：历史占 40% ~ 75%
+  // - 大窗口：历史占 65% ~ 90%（尽量吃满大模型上下文）
+  let baseHistoryRatio: number;
+  if (isBigWindow) {
+    baseHistoryRatio = 0.65 + 0.25 * iq; // 0.65 ~ 0.90
+  } else if (isSmallWindow) {
+    baseHistoryRatio = 0.35 + 0.35 * iq; // 0.35 ~ 0.70
+  } else {
+    baseHistoryRatio = 0.40 + 0.35 * iq; // 0.40 ~ 0.75
+  }
+
+  // 1. 历史比例：在 baseHistoryRatio 基础上，根据对话负载做微调
+  let historyRatio = baseHistoryRatio;
+
+  switch (recentLoad) {
+    case "light": {
+      // 短句对话：可以多给一点历史空间，但仍保留少量余量给其他上下文
+      const maxRatio = isBigWindow ? 0.95 : 0.85;
+      historyRatio = clamp(historyRatio * 1.05, 0.3, maxRatio);
+      break;
+    }
+    case "heavy": {
+      // 重内容对话：适当收紧历史比例，为当前大块内容 / 工具描述等预留更多窗口
+      const maxRatio = isBigWindow ? 0.90 : 0.80;
+      historyRatio = clamp(historyRatio * 0.9, 0.3, maxRatio);
+      break;
+    }
+    case "medium":
+    default: {
+      const maxRatio = isBigWindow ? 0.90 : 0.85;
+      historyRatio = clamp(historyRatio, 0.3, maxRatio);
+      break;
+    }
+  }
+
+  const historyBudget = Math.max(
+    0,
+    Math.floor(safeWindowLimit * historyRatio),
+  );
+
+  // 2. 在历史预算里分配给「原始消息」的预算
+  //    - summary 已经占用一部分
+  //    - 同时保证原始消息至少有一部分空间（heavy 用户更多一点）
+  const minRawRatio = recentLoad === "heavy" ? 0.4 : 0.2;
+
+  const rawMessageBudget = Math.max(
+    Math.floor(historyBudget * minRawRatio),
+    historyBudget - summaryTokens,
+    0,
+  );
+
+  // 3. 希望至少保留的尾部 token 数
+  //    - light ：多保几轮短句（不超过 30% window，且上限 8k）
+  //    - heavy ：至少给几条大块代码空间（>= 16k，或 25% window）
+  //    - medium：取中间值
+  let minTailTokens: number;
+
+  if (recentLoad === "light") {
+    minTailTokens = Math.min(
+      Math.floor(safeWindowLimit * 0.3),
+      8000,
+    );
+  } else if (recentLoad === "heavy") {
+    minTailTokens = Math.max(
+      Math.floor(safeWindowLimit * 0.25),
+      16000,
+    );
+  } else {
+    minTailTokens = Math.floor(safeWindowLimit * 0.2);
+  }
+
+  return { historyBudget, rawMessageBudget, minTailTokens };
+};

package/ai/context/tokenUtils.ts

@@ -0,0 +1,78 @@
+// ai/context/tokenUtils.ts
+
+/**
+ * 简单的 Token 估算工具。
+ * 
+ * 规则：
+ * - 中文字符：约 1.5 tokens/字
+ * - 英文及其他字符：约 0.25 tokens/字符（即 4 字符 ≈ 1 token）
+ * 
+ * 这是一个粗略估算，精确值需要使用 tiktoken 等库。
+ */
+
+/**
+ * 估算文本的 Token 数量
+ */
+export const estimateTokenCount = (text: string): number => {
+    if (!text) return 0;
+
+    // 匹配中文字符（包括常用汉字范围）
+    const chineseChars = (text.match(/[\u4e00-\u9fa5\u3400-\u4dbf]/g) || []).length;
+    const otherChars = text.length - chineseChars;
+
+    // 中文 1.5 token/字，其他 0.25 token/字符
+    return Math.ceil(chineseChars * 1.5 + otherChars * 0.25);
+};
+
+/**
+ * 估算多个文本的 Token 总数
+ */
+export const estimateTotalTokens = (texts: (string | undefined | null)[]): number => {
+    return texts.reduce((sum, text) => sum + estimateTokenCount(text || ""), 0);
+};
+
+/**
+ * 格式化 Token 数量为可读字符串
+ * 例如：1500 -> "1.5k", 150000 -> "150k"
+ */
+export const formatTokenCount = (count: number): string => {
+    if (count < 1000) return String(count);
+    if (count < 10000) return `${(count / 1000).toFixed(1)}k`;
+    return `${Math.round(count / 1000)}k`;
+};
+
+/**
+ * 计算 Token 使用百分比
+ */
+export const calcTokenUsagePercent = (used: number, total: number): number => {
+    if (total <= 0) return 0;
+    return Math.min(100, Math.round((used / total) * 100));
+};
+
+/**
+ * Context 分配建议阈值
+ * - References/Space Context 不应超过总量的 40%
+ * - 为对话历史和用户输入预留 40%
+ * - 为系统提示预留 10%
+ * - 为输出预留 10%
+ */
+export const CONTEXT_BUDGET = {
+    REFERENCES_MAX_PERCENT: 40, // References + Space Context 最大占比
+    HISTORY_RESERVE_PERCENT: 40, // 为对话历史预留
+    SYSTEM_RESERVE_PERCENT: 10, // 系统提示预留
+    OUTPUT_RESERVE_PERCENT: 10, // 输出预留
+};
+
+/**
+ * 判断 Token 使用是否超过警告阈值
+ */
+export const isTokenUsageWarning = (usedPercent: number): boolean => {
+    return usedPercent > CONTEXT_BUDGET.REFERENCES_MAX_PERCENT;
+};
+
+/**
+ * 判断 Token 使用是否严重超标
+ */
+export const isTokenUsageCritical = (usedPercent: number): boolean => {
+    return usedPercent > 60; // 超过 60% 可能影响对话历史
+};

package/ai/llm/agentCapabilities.ts

@@ -0,0 +1,74 @@
+import {
+  findModelConfig,
+  getProviderByModelName,
+} from "./providers";
+
+type AgentCapabilityConfig = {
+  apiSource?: string | null;
+  provider?: string | null;
+  model?: string | null;
+  hasVision?: boolean | null;
+  useServerProxy?: boolean | null;
+};
+
+const isCustomAgent = (agent: AgentCapabilityConfig): boolean => {
+  const apiSource = agent.apiSource?.toLowerCase();
+  const provider = agent.provider?.toLowerCase();
+  return apiSource === "custom" || provider === "custom";
+};
+
+const lookupKnownModelVision = (
+  provider: string | null,
+  model: string,
+): boolean | undefined => {
+  if (!model) return undefined;
+
+  if (provider) {
+    const direct = findModelConfig(provider, model);
+    if (direct) return direct.hasVision;
+  }
+
+  const detected = getProviderByModelName(model);
+  if (detected) {
+    return findModelConfig(detected, model)?.hasVision;
+  }
+
+  if (!provider && model.includes("/")) {
+    const slash = model.indexOf("/");
+    const modelProvider = model.slice(0, slash).toLowerCase();
+    const modelName = model.slice(slash + 1);
+    const nested = findModelConfig(modelProvider, modelName);
+    if (nested) return nested.hasVision;
+  }
+
+  return undefined;
+};
+
+export const resolveAgentImageInputSupport = (
+  agent: AgentCapabilityConfig | null | undefined,
+): boolean => {
+  if (!agent) return true;
+
+  const provider = agent.provider?.trim().toLowerCase() || null;
+  const model = agent.model?.trim() || "";
+  const custom = isCustomAgent(agent);
+  const catalogHasVision = lookupKnownModelVision(
+    custom ? null : provider,
+    model,
+  );
+
+  if (!custom && catalogHasVision !== undefined) {
+    return catalogHasVision;
+  }
+
+  if (custom) {
+    if (catalogHasVision === true) return true;
+    return true;
+  }
+
+  if (typeof agent.hasVision === "boolean") {
+    return agent.hasVision;
+  }
+
+  return true;
+};

package/ai/llm/deepinfra.ts

@@ -0,0 +1,79 @@
+// ai/llm/deepinfra.ts
+import { DEEPINFRA_KIMI_FALLBACK_MODEL } from "./kimi";
+
+export const deepinfraModels = [
+  {
+    name: DEEPINFRA_KIMI_FALLBACK_MODEL,
+    displayName: "MoonshotAI: Kimi K2.6 (DeepInfra)",
+    hasVision: true,
+    price: {
+      input: 0.6 * 7,
+      output: 2.5 * 7,
+      inputCacheHit: 0.06 * 7,
+    },
+    maxOutputTokens: 262144,
+    contextWindow: 262144,
+    supportsTool: true,
+  },
+  {
+    name: "moonshotai/Kimi-K2.5",
+    displayName: "MoonshotAI: Kimi K2.5 (DeepInfra)",
+    hasVision: true,
+    price: {
+      input: 0.45 * 7,
+      output: 2.25 * 7,
+      inputCacheHit: 0.07 * 7,
+    },
+    maxOutputTokens: 262144,
+    contextWindow: 262144,
+    supportsTool: true,
+  },
+  {
+    name: "zai-org/GLM-5.1",
+    displayName: "Z.AI: GLM 5.1 (DeepInfra)",
+    hasVision: false,
+    price: {
+      input: 1.26 * 7,
+      output: 3.96 * 7,
+      inputCacheHit: 0.26 * 7,
+    },
+    contextWindow: 202752,
+    supportsTool: true,
+  },
+  {
+    name: "anthropic/claude-haiku-4-5",
+    displayName: "Anthropic: Claude Haiku 4.5",
+    hasVision: true,
+    price: {
+      input: 1 * 9,
+      output: 5 * 9,
+    },
+    contextWindow: 195000,
+    maxOutputTokens: 4092,
+    supportsTool: false,
+  },
+  {
+    name: "anthropic/claude-sonnet-4-6",
+    displayName: "Anthropic: Claude Sonnet 4.6",
+    hasVision: true,
+    price: {
+      input: 3 * 9,
+      output: 15 * 9,
+    },
+    contextWindow: 976000,
+    maxOutputTokens: 4092,
+    supportsTool: false,
+  },
+  {
+    name: "anthropic/claude-opus-4-7",
+    displayName: "Anthropic: Claude Opus 4.7",
+    hasVision: true,
+    price: {
+      input: 5 * 9,
+      output: 25 * 9,
+    },
+    contextWindow: 976000,
+    maxOutputTokens: 4092,
+    supportsTool: false,
+  },
+];

package/ai/llm/fireworks.ts

@@ -0,0 +1,68 @@
+// ai/llm/fireworks.ts
+import {
+    FIREWORKS_KIMI_CURRENT_MODEL,
+    FIREWORKS_KIMI_LATEST_MODEL,
+} from "./kimi";
+
+export const fireworksModels = [
+    // --- MoonshotAI Models (Kimi) ---
+    {
+        name: FIREWORKS_KIMI_LATEST_MODEL,
+        displayName: "MoonshotAI: Kimi Latest",
+        hasVision: true,
+        price: {
+            input: 0.95 * 8,
+            output: 4.0 * 8,
+            cachingRead: 0.16 * 8,
+        },
+        maxOutputTokens: 262144,
+        contextWindow: 262144,
+        supportsTool: true,
+    },
+    {
+        name: FIREWORKS_KIMI_CURRENT_MODEL,
+        displayName: "MoonshotAI: Kimi K2.6",
+        hasVision: true,
+        price: {
+            input: 0.95 * 8,
+            output: 4.0 * 8,
+            cachingRead: 0.16 * 8,
+        },
+        maxOutputTokens: 262144,
+        contextWindow: 262144,
+        supportsTool: true,
+    },
+    {
+        name: "accounts/fireworks/models/minimax-m2p7",
+        displayName: "MiniMax: MiniMax M2.7",
+        hasVision: true,
+        price: {
+            input: 0.3 * 8,
+            output: 1.2 * 8,
+            cachingRead: 0.06 * 8,
+        },
+        supportsTool: true,
+    },
+    {
+        name: "accounts/fireworks/models/qwen3p6-plus",
+        displayName: "Qwen: Qwen 3.6 Plus",
+        hasVision: true,
+        price: {
+            input: 0.5 * 8,
+            output: 3.0 * 8,
+            cachingRead: 0.1 * 8,
+        },
+        supportsTool: true,
+    },
+    {
+        name: "accounts/fireworks/models/glm-5p1",
+        displayName: "Z.AI: GLM 5.1",
+        hasVision: true,
+        price: {
+            input: 1.0 * 8,
+            output: 3.2 * 8,
+            cachingRead: 0.2 * 8,
+        },
+        supportsTool: true,
+    },
+];

package/ai/llm/generateRequestBody.ts

@@ -0,0 +1,165 @@
+// ai/llm/generateRequestBody.ts
+import { Agent, Message } from "../../app/types";
+import { generateOpenAIRequestBody } from "../../integrations/openai/generateOpenAIRequestBody";
+import { generateResponseRequestBody } from "../../integrations/openai/generateResponseRequestBody";
+
+import { isResponseAPIModel } from "./isResponseAPIModel";
+import { Contexts } from "../types";
+import { getModelConfig } from "./providers";
+import type { Model } from "./types";
+
+export interface GenerateRequestBodyArgs {
+  agentConfig: Agent;
+  messages: Message[]; // 动态历史消息/本轮新增
+  userInput: string; // 本次用户输入
+  contexts?: Contexts; // 可选上下文
+  stableMessages?: Message[]; // 稳定的、已截断的历史快照（用于缓存）
+  prependSystemPrompt?: boolean;
+}
+
+const IMAGE_OUTPUT_EXECUTION_PROMPT = [
+  "When the conversation already includes an input image and the user asks to change any visual attribute such as hairstyle, glasses, outfit, makeup, background, composition, or style, default to returning the edited image in this turn.",
+  "When the user asks you to generate, edit, redraw, or return an image, you must return the image in this turn.",
+  "Do not answer with only a textual description.",
+  "Do not claim that you uploaded, attached, or sent an image unless the response actually includes image output.",
+  "Only return text without an image when the user explicitly asks for analysis, recommendations, or explanation without generating an image.",
+].join(" ");
+
+type ImageOutputSettings = {
+  shouldEnableImage: boolean;
+  modelConfig: Model | null;
+};
+
+const getImageOutputSettings = (
+  agentConfig: Agent,
+  provider: string
+): ImageOutputSettings => {
+  const modelName = agentConfig.model;
+  if (!modelName) {
+    return { shouldEnableImage: false, modelConfig: null };
+  }
+
+  try {
+    const modelConfig = getModelConfig(provider as any, modelName);
+    const hasImageOutput = !!modelConfig.hasImageOutput;
+    const isExplicitlyDisabled = agentConfig.imageConfig?.enabled === false;
+    return {
+      shouldEnableImage: hasImageOutput && !isExplicitlyDisabled,
+      modelConfig,
+    };
+  } catch {
+    return { shouldEnableImage: false, modelConfig: null };
+  }
+};
+
+const withImageOutputPromptGuard = (
+  agentConfig: Agent,
+  shouldEnableImage: boolean
+): Agent => {
+  if (!shouldEnableImage) return agentConfig;
+
+  const basePrompt = typeof agentConfig.prompt === "string" ? agentConfig.prompt.trim() : "";
+  if (basePrompt.includes(IMAGE_OUTPUT_EXECUTION_PROMPT)) {
+    return agentConfig;
+  }
+
+  return {
+    ...agentConfig,
+    prompt: basePrompt
+      ? `${basePrompt}\n\n${IMAGE_OUTPUT_EXECUTION_PROMPT}`
+      : IMAGE_OUTPUT_EXECUTION_PROMPT,
+  };
+};
+
+/**
+ * 根据模型元数据 + Agent 配置，为 chat/completions 请求体注入
+ * - modalities: ["image", "text"]
+ * - image_config: { aspect_ratio, image_size }
+ */
+const applyImageConfigIfNeeded = (
+  body: any,
+  agentConfig: Agent,
+  provider: string
+): any => {
+  const { shouldEnableImage, modelConfig } = getImageOutputSettings(
+    agentConfig,
+    provider
+  );
+  const agentImageCfg = agentConfig.imageConfig;
+
+  if (!shouldEnableImage || !modelConfig) {
+    return body;
+  }
+
+  // 1) 处理 modalities
+  // 优先级：Agent.forceModalities > Model.defaultModalities > requiresImageModalities ? ["image","text"] : 不强制
+  const modalities: Array<"text" | "image"> | undefined =
+    agentImageCfg?.forceModalities ??
+    modelConfig.defaultModalities ??
+    (modelConfig.requiresImageModalities
+      ? (["image", "text"] as Array<"text" | "image">)
+      : undefined);
+
+  if (modalities && !body.modalities) {
+    body.modalities = modalities;
+  }
+
+  // 2) 处理 image_config（仅在模型声明 supportsImageConfig 时生效）
+  if (!modelConfig.supportsImageConfig) {
+    return body;
+  }
+
+  const aspectRatio = agentImageCfg?.aspectRatio;
+  const imageSize = agentImageCfg?.imageSize;
+
+  if (!aspectRatio && !imageSize) {
+    // Agent 未指定任何 image_config 参数，则沿用模型默认
+    return body;
+  }
+
+  body.image_config = {
+    ...(body.image_config ?? {}),
+    ...(aspectRatio ? { aspect_ratio: aspectRatio } : {}),
+    ...(imageSize ? { image_size: imageSize } : {}),
+  };
+
+  return body;
+};
+
+export const generateRequestBody = ({
+  agentConfig,
+  messages,
+  contexts,
+  stableMessages,
+  prependSystemPrompt = true,
+}: GenerateRequestBodyArgs) => {
+  const provider = (agentConfig.provider || "").toLowerCase();
+  const imageSettings = getImageOutputSettings(agentConfig, provider);
+  const agentConfigForRequest = withImageOutputPromptGuard(
+    agentConfig,
+    imageSettings.shouldEnableImage
+  );
+
+  // 1) Response API 走新版 /v1/responses（目前主要是 OpenAI）
+  if (isResponseAPIModel(agentConfigForRequest)) {
+    return generateResponseRequestBody(
+      agentConfigForRequest,
+      [...(stableMessages ?? []), ...messages],
+      contexts,
+      prependSystemPrompt
+    );
+  }
+
+  // 2) 其余走老版 chat/completions
+  const baseBody = generateOpenAIRequestBody(
+    agentConfigForRequest,
+    provider,
+    messages,
+    contexts,
+    stableMessages,
+    prependSystemPrompt
+  );
+
+  // 在基础请求体上注入图像相关配置（如果模型 + Agent 都支持/需要）
+  return applyImageConfigIfNeeded(baseBody, agentConfigForRequest, provider);
+};