diff --git a/src/agent/context-window/guard.ts b/src/agent/context-window/guard.ts
new file mode 100644
index 00000000..20fdd4f2
--- /dev/null
+++ b/src/agent/context-window/guard.ts
@@ -0,0 +1,110 @@
+/**
+ * Context Window Guard - 上下文窗口验证
+ *
+ * 在 agent 运行前验证 context window 是否足够，防止 token 溢出
+ */
+
+import type { ContextWindowInfo, ContextWindowGuardResult, ContextWindowSource } from "./types.js";
+
+/** 硬性最小 token 数，低于此值将阻止运行 */
+export const CONTEXT_WINDOW_HARD_MIN_TOKENS = 16_000;
+
+/** 警告阈值，低于此值会发出警告 */
+export const CONTEXT_WINDOW_WARN_BELOW_TOKENS = 32_000;
+
+/** 默认 context window（当无法获取时） */
+export const DEFAULT_CONTEXT_TOKENS = 200_000;
+
+/**
+ * 标准化为正整数
+ */
+function normalizePositiveInt(value: unknown): number | null {
+  if (typeof value !== "number" || !Number.isFinite(value)) return null;
+  const int = Math.floor(value);
+  return int > 0 ? int : null;
+}
+
+/**
+ * 解析 context window 信息
+ *
+ * 优先级：model > config > default
+ */
+export function resolveContextWindowInfo(params: {
+  /** Model 的 contextWindow 属性 */
+  modelContextWindow?: number | undefined;
+  /** 配置中指定的 context tokens */
+  configContextTokens?: number | undefined;
+  /** 默认值 */
+  defaultTokens?: number | undefined;
+}): ContextWindowInfo {
+  // 1. 尝试从 model 获取
+  const fromModel = normalizePositiveInt(params.modelContextWindow);
+  if (fromModel) {
+    return { tokens: fromModel, source: "model" };
+  }
+
+  // 2. 尝试从配置获取
+  const fromConfig = normalizePositiveInt(params.configContextTokens);
+  if (fromConfig) {
+    return { tokens: fromConfig, source: "config" };
+  }
+
+  // 3. 使用默认值
+  return {
+    tokens: Math.floor(params.defaultTokens ?? DEFAULT_CONTEXT_TOKENS),
+    source: "default",
+  };
+}
+
+/**
+ * 评估 context window guard
+ *
+ * 返回是否需要警告或阻止运行
+ */
+export function evaluateContextWindowGuard(params: {
+  info: ContextWindowInfo;
+  warnBelowTokens?: number | undefined;
+  hardMinTokens?: number | undefined;
+}): ContextWindowGuardResult {
+  const warnBelow = Math.max(
+    1,
+    Math.floor(params.warnBelowTokens ?? CONTEXT_WINDOW_WARN_BELOW_TOKENS),
+  );
+  const hardMin = Math.max(
+    1,
+    Math.floor(params.hardMinTokens ?? CONTEXT_WINDOW_HARD_MIN_TOKENS),
+  );
+  const tokens = Math.max(0, Math.floor(params.info.tokens));
+
+  return {
+    ...params.info,
+    tokens,
+    shouldWarn: tokens > 0 && tokens < warnBelow,
+    shouldBlock: tokens > 0 && tokens < hardMin,
+  };
+}
+
+/**
+ * 完整的 context window guard 流程
+ *
+ * 解析 + 评估一步完成
+ */
+export function checkContextWindow(params: {
+  modelContextWindow?: number | undefined;
+  configContextTokens?: number | undefined;
+  defaultTokens?: number | undefined;
+  warnBelowTokens?: number | undefined;
+  hardMinTokens?: number | undefined;
+}): ContextWindowGuardResult {
+  const info = resolveContextWindowInfo({
+    modelContextWindow: params.modelContextWindow,
+    configContextTokens: params.configContextTokens,
+    defaultTokens: params.defaultTokens,
+  });
+
+  return evaluateContextWindowGuard({
+    info,
+    warnBelowTokens: params.warnBelowTokens,
+    hardMinTokens: params.hardMinTokens,
+  });
+}
diff --git a/src/agent/context-window/index.ts b/src/agent/context-window/index.ts
new file mode 100644
index 00000000..ec57f696
--- /dev/null
+++ b/src/agent/context-window/index.ts
@@ -0,0 +1,46 @@
+/**
+ * Context Window Guard
+ *
+ * 提供上下文窗口管理功能，防止 token 溢出
+ */
+
+// Types
+export type {
+  ContextWindowSource,
+  ContextWindowInfo,
+  ContextWindowGuardResult,
+  TokenEstimation,
+  TokenAwareCompactionResult,
+} from "./types.js";
+
+// Guard
+export {
+  CONTEXT_WINDOW_HARD_MIN_TOKENS,
+  CONTEXT_WINDOW_WARN_BELOW_TOKENS,
+  DEFAULT_CONTEXT_TOKENS,
+  resolveContextWindowInfo,
+  evaluateContextWindowGuard,
+  checkContextWindow,
+} from "./guard.js";
+
+// Token estimation
+export {
+  ESTIMATION_SAFETY_MARGIN,
+  COMPACTION_TRIGGER_RATIO,
+  COMPACTION_TARGET_RATIO,
+  MIN_KEEP_MESSAGES,
+  estimateMessagesTokens,
+  estimateSystemPromptTokens,
+  estimateTokenUsage,
+  shouldCompact,
+  compactMessagesTokenAware,
+  isMessageOversized,
+} from "./token-estimation.js";
+
+// Summarization
+export type { SummaryCompactionResult, SummaryCompactionParams } from "./summarization.js";
+export {
+  splitMessagesForSummary,
+  compactMessagesWithSummary,
+  compactMessagesWithChunkedSummary,
+} from "./summarization.js";
diff --git a/src/agent/context-window/summarization.ts b/src/agent/context-window/summarization.ts
new file mode 100644
index 00000000..e6808e51
--- /dev/null
+++ b/src/agent/context-window/summarization.ts
@@ -0,0 +1,299 @@
+/**
+ * 摘要式 Compaction
+ *
+ * 使用 LLM 生成历史消息的摘要，而不是简单截断
+ */
+
+import type { AgentMessage } from "@mariozechner/pi-agent-core";
+import { generateSummary, estimateTokens } from "@mariozechner/pi-coding-agent";
+import type { Model } from "@mariozechner/pi-ai";
+import { estimateMessagesTokens } from "./token-estimation.js";
+
+/** 摘要 Compaction 结果 */
+export type SummaryCompactionResult = {
+  /** 保留的消息（包含摘要消息） */
+  kept: AgentMessage[];
+  /** 移除的消息数量 */
+  removedCount: number;
+  /** 移除的 token 数 */
+  tokensRemoved: number;
+  /** 保留的 token 数 */
+  tokensKept: number;
+  /** 生成的摘要 */
+  summary: string;
+  /** compaction 原因 */
+  reason: "summary";
+};
+
+/** 摘要 Compaction 参数 */
+export type SummaryCompactionParams = {
+  /** 消息列表 */
+  messages: AgentMessage[];
+  /** LLM Model（用于生成摘要） */
+  model: Model<any>;
+  /** API Key */
+  apiKey: string;
+  /** 可用 token 数 */
+  availableTokens: number;
+  /** 目标利用率 (0-1)，默认 0.5 */
+  targetRatio?: number | undefined;
+  /** 最少保留消息数，默认 10 */
+  minKeepMessages?: number | undefined;
+  /** 预留给摘要生成的 token 数，默认 2048 */
+  reserveTokens?: number | undefined;
+  /** 自定义摘要指令 */
+  customInstructions?: string | undefined;
+  /** 之前的摘要（用于增量更新） */
+  previousSummary?: string | undefined;
+  /** AbortSignal */
+  signal?: AbortSignal | undefined;
+};
+
+/** 默认摘要提示词 */
+const DEFAULT_SUMMARY_INSTRUCTIONS = `Summarize the conversation history concisely, focusing on:
+- Key decisions made
+- Important context and constraints
+- Open questions or TODOs
+- Technical details that may be needed later
+
+Keep the summary concise but complete. Use bullet points for clarity.`;
+
+/**
+ * 将消息分割为需要摘要的部分和保留的部分
+ */
+export function splitMessagesForSummary(
+  messages: AgentMessage[],
+  availableTokens: number,
+  options?: {
+    targetRatio?: number | undefined;
+    minKeepMessages?: number | undefined;
+  },
+): { toSummarize: AgentMessage[]; toKeep: AgentMessage[] } | null {
+  const targetRatio = options?.targetRatio ?? 0.5;
+  const minKeep = options?.minKeepMessages ?? 10;
+
+  if (messages.length <= minKeep) {
+    return null; // 消息太少，不需要压缩
+  }
+
+  const totalTokens = estimateMessagesTokens(messages);
+  const targetTokens = Math.floor(availableTokens * targetRatio);
+
+  // 如果当前已经在目标内，不需要压缩
+  if (totalTokens <= targetTokens) {
+    return null;
+  }
+
+  // 从后往前保留消息
+  const toKeep: AgentMessage[] = [];
+  let keptTokens = 0;
+
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i]!;
+    const msgTokens = estimateTokens(msg);
+
+    // 检查是否可以添加这条消息
+    if (keptTokens + msgTokens <= targetTokens || toKeep.length < minKeep) {
+      toKeep.unshift(msg);
+      keptTokens += msgTokens;
+    }
+
+    // 如果已经达到最小保留数且超过目标，停止
+    if (toKeep.length >= minKeep && keptTokens >= targetTokens) {
+      break;
+    }
+  }
+
+  // 需要摘要的消息
+  const toSummarize = messages.slice(0, messages.length - toKeep.length);
+
+  if (toSummarize.length === 0) {
+    return null;
+  }
+
+  return { toSummarize, toKeep };
+}
+
+/**
+ * 创建摘要消息
+ */
+function createSummaryMessage(summary: string, previousSummary?: string): AgentMessage {
+  const content = previousSummary
+    ? `## Previous Context Summary\n${previousSummary}\n\n## Recent Context Summary\n${summary}`
+    : `## Conversation Summary\n${summary}`;
+
+  return {
+    role: "user",
+    content: `[System Note: The following is a summary of the earlier conversation history that has been compacted to save context space.]\n\n${content}\n\n[End of Summary]`,
+    timestamp: Date.now(),
+  };
+}
+
+/**
+ * 执行摘要式 Compaction
+ *
+ * 使用 LLM 生成历史消息的摘要，然后将摘要和最近的消息组合
+ */
+export async function compactMessagesWithSummary(
+  params: SummaryCompactionParams,
+): Promise<SummaryCompactionResult | null> {
+  const {
+    messages,
+    model,
+    apiKey,
+    availableTokens,
+    targetRatio,
+    minKeepMessages,
+    reserveTokens = 2048,
+    customInstructions,
+    previousSummary,
+    signal,
+  } = params;
+
+  // 分割消息
+  const split = splitMessagesForSummary(messages, availableTokens, {
+    targetRatio,
+    minKeepMessages,
+  });
+
+  if (!split) {
+    return null;
+  }
+
+  const { toSummarize, toKeep } = split;
+
+  // 生成摘要
+  const instructions = customInstructions || DEFAULT_SUMMARY_INSTRUCTIONS;
+  const summary = await generateSummary(
+    toSummarize,
+    model,
+    reserveTokens,
+    apiKey,
+    signal,
+    instructions,
+    previousSummary,
+  );
+
+  // 创建摘要消息
+  const summaryMessage = createSummaryMessage(summary, previousSummary);
+
+  // 组合结果
+  const kept = [summaryMessage, ...toKeep];
+
+  const tokensRemoved = estimateMessagesTokens(toSummarize);
+  const tokensKept = estimateMessagesTokens(kept);
+
+  return {
+    kept,
+    removedCount: toSummarize.length,
+    tokensRemoved,
+    tokensKept,
+    summary,
+    reason: "summary",
+  };
+}
+
+/**
+ * 分块生成摘要（用于超大历史）
+ *
+ * 当历史太大时，分块生成摘要然后合并
+ */
+export async function compactMessagesWithChunkedSummary(
+  params: SummaryCompactionParams & {
+    maxChunkTokens?: number | undefined;
+  },
+): Promise<SummaryCompactionResult | null> {
+  const {
+    messages,
+    model,
+    apiKey,
+    availableTokens,
+    targetRatio,
+    minKeepMessages,
+    reserveTokens = 2048,
+    customInstructions,
+    previousSummary,
+    signal,
+    maxChunkTokens = 50000,
+  } = params;
+
+  // 分割消息
+  const split = splitMessagesForSummary(messages, availableTokens, {
+    targetRatio,
+    minKeepMessages,
+  });
+
+  if (!split) {
+    return null;
+  }
+
+  const { toSummarize, toKeep } = split;
+
+  // 如果需要摘要的消息不多，直接摘要
+  const toSummarizeTokens = estimateMessagesTokens(toSummarize);
+  if (toSummarizeTokens <= maxChunkTokens) {
+    return compactMessagesWithSummary(params);
+  }
+
+  // 分块处理
+  const chunks: AgentMessage[][] = [];
+  let currentChunk: AgentMessage[] = [];
+  let currentTokens = 0;
+
+  for (const msg of toSummarize) {
+    const msgTokens = estimateTokens(msg);
+
+    if (currentTokens + msgTokens > maxChunkTokens && currentChunk.length > 0) {
+      chunks.push(currentChunk);
+      currentChunk = [];
+      currentTokens = 0;
+    }
+
+    currentChunk.push(msg);
+    currentTokens += msgTokens;
+  }
+
+  if (currentChunk.length > 0) {
+    chunks.push(currentChunk);
+  }
+
+  // 为每个块生成摘要
+  const instructions = customInstructions || DEFAULT_SUMMARY_INSTRUCTIONS;
+  const chunkSummaries: string[] = [];
+
+  let runningContext = previousSummary;
+  for (const chunk of chunks) {
+    const chunkSummary = await generateSummary(
+      chunk,
+      model,
+      reserveTokens,
+      apiKey,
+      signal,
+      instructions,
+      runningContext,
+    );
+    chunkSummaries.push(chunkSummary);
+    runningContext = chunkSummary;
+  }
+
+  // 最终摘要就是最后一个块的摘要（已经包含了之前的上下文）
+  const finalSummary = chunkSummaries[chunkSummaries.length - 1] ?? "";
+
+  // 创建摘要消息
+  const summaryMessage = createSummaryMessage(finalSummary);
+
+  // 组合结果
+  const kept = [summaryMessage, ...toKeep];
+
+  const tokensRemoved = estimateMessagesTokens(toSummarize);
+  const tokensKept = estimateMessagesTokens(kept);
+
+  return {
+    kept,
+    removedCount: toSummarize.length,
+    tokensRemoved,
+    tokensKept,
+    summary: finalSummary,
+    reason: "summary",
+  };
+}
diff --git a/src/agent/context-window/token-estimation.ts b/src/agent/context-window/token-estimation.ts
new file mode 100644
index 00000000..e0a67018
--- /dev/null
+++ b/src/agent/context-window/token-estimation.ts
@@ -0,0 +1,156 @@
+/**
+ * Token 估算工具
+ *
+ * 提供消息和系统提示词的 token 计数功能
+ */
+
+import type { AgentMessage } from "@mariozechner/pi-agent-core";
+import { estimateTokens } from "@mariozechner/pi-coding-agent";
+import type { TokenEstimation, TokenAwareCompactionResult } from "./types.js";
+
+/** 安全边界系数，用于补偿估算不准确 */
+export const ESTIMATION_SAFETY_MARGIN = 1.2; // 20% buffer
+
+/** 触发 compaction 的利用率阈值 */
+export const COMPACTION_TRIGGER_RATIO = 0.8; // 80%
+
+/** Compaction 目标利用率 */
+export const COMPACTION_TARGET_RATIO = 0.5; // 50%
+
+/** 最小保留消息数 */
+export const MIN_KEEP_MESSAGES = 10;
+
+/**
+ * 估算消息数组的总 token 数
+ */
+export function estimateMessagesTokens(messages: AgentMessage[]): number {
+  return messages.reduce((sum, message) => sum + estimateTokens(message), 0);
+}
+
+/**
+ * 估算系统提示词的 token 数
+ */
+export function estimateSystemPromptTokens(systemPrompt: string | undefined): number {
+  if (!systemPrompt) return 0;
+  // 简单估算：约 4 字符 = 1 token（适用于英文/代码混合文本）
+  // 中文约 2 字符 = 1 token
+  // 取平均值 3
+  return Math.ceil(systemPrompt.length / 3);
+}
+
+/**
+ * 计算完整的 token 使用情况
+ */
+export function estimateTokenUsage(params: {
+  messages: AgentMessage[];
+  systemPrompt?: string | undefined;
+  contextWindowTokens: number;
+  reserveTokens?: number | undefined;
+}): TokenEstimation {
+  const messageTokens = estimateMessagesTokens(params.messages);
+  const systemPromptTokens = estimateSystemPromptTokens(params.systemPrompt);
+  const reserve = params.reserveTokens ?? 1024; // 预留给响应生成
+
+  // 可用 token = 总窗口 - 系统提示 - 预留
+  const availableTokens = Math.max(
+    0,
+    params.contextWindowTokens - systemPromptTokens - reserve,
+  );
+
+  // 计算利用率（带安全边界）
+  const safeMessageTokens = messageTokens * ESTIMATION_SAFETY_MARGIN;
+  const utilizationRatio = availableTokens > 0 ? safeMessageTokens / availableTokens : 1;
+
+  return {
+    messageTokens,
+    systemPromptTokens,
+    availableTokens,
+    utilizationRatio,
+  };
+}
+
+/**
+ * 判断是否需要 compaction
+ */
+export function shouldCompact(estimation: TokenEstimation): boolean {
+  return estimation.utilizationRatio >= COMPACTION_TRIGGER_RATIO;
+}
+
+/**
+ * Token 感知的消息压缩
+ *
+ * 策略：从最旧的消息开始移除，直到达到目标利用率
+ */
+export function compactMessagesTokenAware(
+  messages: AgentMessage[],
+  availableTokens: number,
+  options?: {
+    targetRatio?: number;
+    minKeepMessages?: number;
+  },
+): TokenAwareCompactionResult | null {
+  const targetRatio = options?.targetRatio ?? COMPACTION_TARGET_RATIO;
+  const minKeep = options?.minKeepMessages ?? MIN_KEEP_MESSAGES;
+
+  if (messages.length <= minKeep) {
+    return null; // 消息太少，不压缩
+  }
+
+  const currentTokens = estimateMessagesTokens(messages);
+  const targetTokens = Math.floor(availableTokens * targetRatio);
+
+  // 如果当前已经在目标内，不需要压缩
+  if (currentTokens <= targetTokens) {
+    return null;
+  }
+
+  // 从后往前保留消息，直到达到目标 token 数
+  const kept: AgentMessage[] = [];
+  let keptTokens = 0;
+
+  // 反向遍历，保留最新的消息
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i]!;
+    const msgTokens = estimateTokens(msg);
+
+    // 检查是否可以添加这条消息
+    if (keptTokens + msgTokens <= targetTokens || kept.length < minKeep) {
+      kept.unshift(msg);
+      keptTokens += msgTokens;
+    }
+
+    // 如果已经达到最小保留数且超过目标，停止
+    if (kept.length >= minKeep && keptTokens >= targetTokens) {
+      break;
+    }
+  }
+
+  // 如果保留的消息数量不变，说明没有压缩
+  if (kept.length >= messages.length) {
+    return null;
+  }
+
+  const removedCount = messages.length - kept.length;
+  const tokensRemoved = currentTokens - keptTokens;
+
+  return {
+    kept,
+    removedCount,
+    tokensRemoved,
+    tokensKept: keptTokens,
+  };
+}
+
+/**
+ * 检查单条消息是否过大
+ *
+ * 如果单条消息超过 context window 的一定比例，可能需要特殊处理
+ */
+export function isMessageOversized(
+  message: AgentMessage,
+  contextWindowTokens: number,
+  maxRatio: number = 0.5,
+): boolean {
+  const tokens = estimateTokens(message) * ESTIMATION_SAFETY_MARGIN;
+  return tokens > contextWindowTokens * maxRatio;
+}
diff --git a/src/agent/context-window/types.ts b/src/agent/context-window/types.ts
new file mode 100644
index 00000000..11cbb8ee
--- /dev/null
+++ b/src/agent/context-window/types.ts
@@ -0,0 +1,48 @@
+/**
+ * Context Window Guard - 类型定义
+ *
+ * 用于管理和验证 LLM 上下文窗口限制
+ */
+
+/** Context window 信息来源 */
+export type ContextWindowSource = "model" | "config" | "default";
+
+/** Context window 信息 */
+export type ContextWindowInfo = {
+  /** Token 数量 */
+  tokens: number;
+  /** 来源 */
+  source: ContextWindowSource;
+};
+
+/** Context window guard 验证结果 */
+export type ContextWindowGuardResult = ContextWindowInfo & {
+  /** 是否需要警告（窗口较小） */
+  shouldWarn: boolean;
+  /** 是否应该阻止运行（窗口太小） */
+  shouldBlock: boolean;
+};
+
+/** Token 估算结果 */
+export type TokenEstimation = {
+  /** 消息总 token 数 */
+  messageTokens: number;
+  /** 系统提示词 token 数 */
+  systemPromptTokens: number;
+  /** 可用 token 数 */
+  availableTokens: number;
+  /** 使用率 (0-1) */
+  utilizationRatio: number;
+};
+
+/** Compaction 结果（带 token 信息） */
+export type TokenAwareCompactionResult = {
+  /** 保留的消息 */
+  kept: import("@mariozechner/pi-agent-core").AgentMessage[];
+  /** 移除的消息数量 */
+  removedCount: number;
+  /** 移除的 token 数 */
+  tokensRemoved: number;
+  /** 保留的 token 数 */
+  tokensKept: number;
+};
diff --git a/src/agent/index.ts b/src/agent/index.ts
index 4f53d09f..6f569f3f 100644
--- a/src/agent/index.ts
+++ b/src/agent/index.ts
@@ -1,3 +1,4 @@
 export * from "./runner.js";
 export * from "./types.js";
 export * from "./profile/index.js";
+export * from "./context-window/index.js";
diff --git a/src/agent/runner.ts b/src/agent/runner.ts
index 4efcb1b7..da7da224 100644
--- a/src/agent/runner.ts
+++ b/src/agent/runner.ts
@@ -5,12 +5,45 @@ import { createAgentOutput } from "./output.js";
 import { resolveModel, resolveTools } from "./tools.js";
 import { SessionManager } from "./session/session-manager.js";
 import { ProfileManager } from "./profile/index.js";
+import {
+  checkContextWindow,
+  DEFAULT_CONTEXT_TOKENS,
+  type ContextWindowGuardResult,
+} from "./context-window/index.js";
+
+/**
+ * 根据 provider 获取 API Key
+ */
+function resolveApiKey(provider: string): string | undefined {
+  const providerEnvMap: Record<string, string> = {
+    openai: "OPENAI_API_KEY",
+    anthropic: "ANTHROPIC_API_KEY",
+    google: "GOOGLE_API_KEY",
+    "google-genai": "GOOGLE_API_KEY",
+    kimi: "MOONSHOT_API_KEY",
+    "kimi-coding": "MOONSHOT_API_KEY",
+    deepseek: "DEEPSEEK_API_KEY",
+    groq: "GROQ_API_KEY",
+    mistral: "MISTRAL_API_KEY",
+    together: "TOGETHER_API_KEY",
+  };
+
+  const envVar = providerEnvMap[provider];
+  if (envVar) {
+    return process.env[envVar];
+  }
+
+  // 尝试通用格式: PROVIDER_API_KEY
+  const normalizedProvider = provider.toUpperCase().replace(/-/g, "_");
+  return process.env[`${normalizedProvider}_API_KEY`];
+}
 
 export class Agent {
   private readonly agent: PiAgentCore;
   private readonly output;
   private readonly session: SessionManager;
   private readonly profile?: ProfileManager;
+  private readonly contextWindowGuard: ContextWindowGuardResult;
 
   /** 当前会话 ID */
   readonly sessionId: string;
@@ -23,34 +56,87 @@ export class Agent {
     this.agent = new PiAgentCore();
 
     // 加载 Agent Profile（如果指定了 profileId）
+    let systemPrompt: string | undefined;
     if (options.profileId) {
       this.profile = new ProfileManager({
         profileId: options.profileId,
         baseDir: options.profileBaseDir,
       });
-      const systemPrompt = this.profile.buildSystemPrompt();
+      systemPrompt = this.profile.buildSystemPrompt();
       if (systemPrompt) {
         this.agent.setSystemPrompt(systemPrompt);
       }
     } else if (options.systemPrompt) {
       // 直接使用传入的 systemPrompt
+      systemPrompt = options.systemPrompt;
       this.agent.setSystemPrompt(options.systemPrompt);
     }
 
     this.sessionId = options.sessionId ?? uuidv7();
-    this.session = new SessionManager({ sessionId: this.sessionId });
-    const storedMeta = this.session.getMeta();
-    if (!options.thinkingLevel && storedMeta?.thinkingLevel) {
-      this.agent.setThinkingLevel(storedMeta.thinkingLevel as any);
-    } else if (options.thinkingLevel) {
-      this.agent.setThinkingLevel(options.thinkingLevel);
-    }
+
+    // 解析 model（用于获取 context window）
+    const storedMeta = (() => {
+      // 临时创建 session 获取 meta，避免循环依赖
+      const tempSession = new SessionManager({ sessionId: this.sessionId });
+      return tempSession.getMeta();
+    })();
 
     const model = options.provider && options.model ? resolveModel(options) : resolveModel({
       ...options,
       provider: storedMeta?.provider,
       model: storedMeta?.model,
     });
+
+    // === Context Window Guard ===
+    this.contextWindowGuard = checkContextWindow({
+      modelContextWindow: model.contextWindow,
+      configContextTokens: options.contextWindowTokens,
+      defaultTokens: DEFAULT_CONTEXT_TOKENS,
+    });
+
+    // 警告：context window 较小
+    if (this.contextWindowGuard.shouldWarn) {
+      stderr.write(
+        `[Context Window Guard] WARNING: Low context window: ${this.contextWindowGuard.tokens} tokens (source: ${this.contextWindowGuard.source})\n`,
+      );
+    }
+
+    // 阻止：context window 太小
+    if (this.contextWindowGuard.shouldBlock) {
+      throw new Error(
+        `[Context Window Guard] Context window too small: ${this.contextWindowGuard.tokens} tokens. ` +
+          `Minimum required: 16,000 tokens. Please use a model with a larger context window.`,
+      );
+    }
+
+    // 确定 compaction 模式
+    const compactionMode = options.compactionMode ?? "tokens"; // 默认使用 token 模式
+
+    // 获取 API Key（用于 summary 模式）
+    const apiKey = compactionMode === "summary" ? resolveApiKey(model.provider) : undefined;
+
+    // 创建 SessionManager（带 context window 配置）
+    this.session = new SessionManager({
+      sessionId: this.sessionId,
+      compactionMode,
+      // Token 模式参数
+      contextWindowTokens: this.contextWindowGuard.tokens,
+      systemPrompt,
+      reserveTokens: options.reserveTokens,
+      targetRatio: options.compactionTargetRatio,
+      minKeepMessages: options.minKeepMessages,
+      // Summary 模式参数
+      model: compactionMode === "summary" ? model : undefined,
+      apiKey,
+      customInstructions: options.summaryInstructions,
+    });
+
+    if (!options.thinkingLevel && storedMeta?.thinkingLevel) {
+      this.agent.setThinkingLevel(storedMeta.thinkingLevel as any);
+    } else if (options.thinkingLevel) {
+      this.agent.setThinkingLevel(options.thinkingLevel);
+    }
+
     this.agent.setModel(model);
     this.agent.setTools(resolveTools(options));
 
@@ -63,6 +149,7 @@ export class Agent {
       provider: this.agent.state.model?.provider,
       model: this.agent.state.model?.id,
       thinkingLevel: this.agent.state.thinkingLevel,
+      contextWindowTokens: this.contextWindowGuard.tokens,
     });
 
     this.agent.subscribe((event: AgentEvent) => {
diff --git a/src/agent/session/compaction.ts b/src/agent/session/compaction.ts
index dec56363..636c36c0 100644
--- a/src/agent/session/compaction.ts
+++ b/src/agent/session/compaction.ts
@@ -1,15 +1,186 @@
 import type { AgentMessage } from "@mariozechner/pi-agent-core";
+import type { Model } from "@mariozechner/pi-ai";
+import {
+  estimateMessagesTokens,
+  compactMessagesTokenAware,
+  estimateTokenUsage,
+  shouldCompact as shouldCompactTokens,
+  compactMessagesWithSummary,
+  compactMessagesWithChunkedSummary,
+  COMPACTION_TARGET_RATIO,
+  MIN_KEEP_MESSAGES,
+} from "../context-window/index.js";
 
 export type CompactionResult = {
   kept: AgentMessage[];
   removedCount: number;
+  /** Token 感知模式下的额外信息 */
+  tokensRemoved?: number | undefined;
+  tokensKept?: number | undefined;
+  /** 摘要模式下生成的摘要 */
+  summary?: string | undefined;
+  reason: "count" | "tokens" | "summary";
 };
 
-export function compactMessages(messages: AgentMessage[], maxMessages: number, keepLast: number) {
+/**
+ * 基于消息数量的简单压缩（旧逻辑，保持向后兼容）
+ */
+export function compactMessagesByCount(
+  messages: AgentMessage[],
+  maxMessages: number,
+  keepLast: number,
+): CompactionResult | null {
   if (messages.length <= maxMessages) return null;
   const kept = messages.slice(-keepLast);
   return {
     kept,
     removedCount: messages.length - kept.length,
-  } satisfies CompactionResult;
+    reason: "count",
+  };
+}
+
+/**
+ * 基于 token 的智能压缩
+ */
+export function compactMessagesByTokens(
+  messages: AgentMessage[],
+  availableTokens: number,
+  options?: {
+    targetRatio?: number;
+    minKeepMessages?: number;
+  },
+): CompactionResult | null {
+  const result = compactMessagesTokenAware(messages, availableTokens, options);
+  if (!result) return null;
+
+  return {
+    kept: result.kept,
+    removedCount: result.removedCount,
+    tokensRemoved: result.tokensRemoved,
+    tokensKept: result.tokensKept,
+    reason: "tokens",
+  };
+}
+
+/** 同步压缩选项（count/tokens 模式） */
+export type SyncCompactionOptions = {
+  mode: "count" | "tokens";
+  // count 模式参数
+  maxMessages?: number | undefined;
+  keepLast?: number | undefined;
+  // tokens 模式参数
+  contextWindowTokens?: number | undefined;
+  systemPrompt?: string | undefined;
+  reserveTokens?: number | undefined;
+  targetRatio?: number | undefined;
+  minKeepMessages?: number | undefined;
+};
+
+/** 摘要压缩选项（summary 模式） */
+export type SummaryCompactionOptions = {
+  mode: "summary";
+  // 必需参数
+  model: Model<any>;
+  apiKey: string;
+  // tokens 模式参数（复用）
+  contextWindowTokens?: number | undefined;
+  systemPrompt?: string | undefined;
+  reserveTokens?: number | undefined;
+  targetRatio?: number | undefined;
+  minKeepMessages?: number | undefined;
+  // summary 特有参数
+  customInstructions?: string | undefined;
+  previousSummary?: string | undefined;
+  signal?: AbortSignal | undefined;
+  maxChunkTokens?: number | undefined;
+};
+
+export type CompactionOptions = SyncCompactionOptions | SummaryCompactionOptions;
+
+/**
+ * 统一的压缩入口（同步版本，用于 count/tokens 模式）
+ *
+ * 根据模式选择压缩策略
+ */
+export function compactMessages(
+  messages: AgentMessage[],
+  options: SyncCompactionOptions,
+): CompactionResult | null {
+  if (options.mode === "count") {
+    return compactMessagesByCount(
+      messages,
+      options.maxMessages ?? 80,
+      options.keepLast ?? 60,
+    );
+  }
+
+  // Token 模式
+  const contextWindowTokens = options.contextWindowTokens ?? 200_000;
+  const estimation = estimateTokenUsage({
+    messages,
+    systemPrompt: options.systemPrompt,
+    contextWindowTokens,
+    reserveTokens: options.reserveTokens,
+  });
+
+  // 检查是否需要压缩
+  if (!shouldCompactTokens(estimation)) {
+    return null;
+  }
+
+  return compactMessagesByTokens(messages, estimation.availableTokens, {
+    targetRatio: options.targetRatio ?? COMPACTION_TARGET_RATIO,
+    minKeepMessages: options.minKeepMessages ?? MIN_KEEP_MESSAGES,
+  });
+}
+
+/**
+ * 摘要式压缩（异步版本）
+ *
+ * 使用 LLM 生成历史消息的摘要
+ */
+export async function compactMessagesAsync(
+  messages: AgentMessage[],
+  options: SummaryCompactionOptions,
+): Promise<CompactionResult | null> {
+  const contextWindowTokens = options.contextWindowTokens ?? 200_000;
+  const estimation = estimateTokenUsage({
+    messages,
+    systemPrompt: options.systemPrompt,
+    contextWindowTokens,
+    reserveTokens: options.reserveTokens,
+  });
+
+  // 检查是否需要压缩
+  if (!shouldCompactTokens(estimation)) {
+    return null;
+  }
+
+  // 使用分块摘要处理超大历史
+  const result = await compactMessagesWithChunkedSummary({
+    messages,
+    model: options.model,
+    apiKey: options.apiKey,
+    availableTokens: estimation.availableTokens,
+    targetRatio: options.targetRatio ?? COMPACTION_TARGET_RATIO,
+    minKeepMessages: options.minKeepMessages ?? MIN_KEEP_MESSAGES,
+    reserveTokens: options.reserveTokens ?? 2048,
+    customInstructions: options.customInstructions,
+    previousSummary: options.previousSummary,
+    signal: options.signal,
+    maxChunkTokens: options.maxChunkTokens,
+  });
+
+  if (!result) {
+    return null;
+  }
+
+  return {
+    kept: result.kept,
+    removedCount: result.removedCount,
+    tokensRemoved: result.tokensRemoved,
+    tokensKept: result.tokensKept,
+    summary: result.summary,
+    reason: "summary",
+  };
 }
diff --git a/src/agent/session/session-manager.ts b/src/agent/session/session-manager.ts
index b7c6d8bb..17f13fcd 100644
--- a/src/agent/session/session-manager.ts
+++ b/src/agent/session/session-manager.ts
@@ -1,31 +1,125 @@
 import type { AgentMessage } from "@mariozechner/pi-agent-core";
+import type { Model } from "@mariozechner/pi-ai";
 import type { SessionEntry, SessionMeta } from "./types.js";
 import { appendEntry, readEntries, writeEntries } from "./storage.js";
-import { compactMessages } from "./compaction.js";
+import { compactMessages, compactMessagesAsync } from "./compaction.js";
 
 export type SessionManagerOptions = {
   sessionId: string;
-  baseDir?: string;
-  maxMessages?: number;
-  keepLast?: number;
+  baseDir?: string | undefined;
+
+  // Compaction 模式配置
+  /** Compaction 模式: "count" 使用消息计数, "tokens" 使用 token 感知, "summary" 使用 LLM 摘要 */
+  compactionMode?: "count" | "tokens" | "summary" | undefined;
+
+  // Count 模式参数
+  maxMessages?: number | undefined;
+  keepLast?: number | undefined;
+
+  // Token 模式参数
+  /** Context window token 数 */
+  contextWindowTokens?: number | undefined;
+  /** 系统提示词（用于计算可用 token） */
+  systemPrompt?: string | undefined;
+  /** 预留给响应的 token 数 */
+  reserveTokens?: number | undefined;
+  /** Compaction 目标利用率 (0-1) */
+  targetRatio?: number | undefined;
+  /** 最小保留消息数 */
+  minKeepMessages?: number | undefined;
+
+  // Summary 模式参数
+  /** LLM Model（用于生成摘要） */
+  model?: Model<any> | undefined;
+  /** API Key */
+  apiKey?: string | undefined;
+  /** 自定义摘要指令 */
+  customInstructions?: string | undefined;
 };
 
 export class SessionManager {
   private readonly sessionId: string;
   private readonly baseDir: string | undefined;
+  private readonly compactionMode: "count" | "tokens" | "summary";
+  // Count 模式
   private readonly maxMessages: number;
   private readonly keepLast: number;
+  // Token 模式
+  private readonly contextWindowTokens: number;
+  private systemPrompt: string | undefined;
+  private readonly reserveTokens: number;
+  private readonly targetRatio: number;
+  private readonly minKeepMessages: number;
+  // Summary 模式
+  private model: Model<any> | undefined;
+  private apiKey: string | undefined;
+  private readonly customInstructions: string | undefined;
+  private previousSummary: string | undefined;
+
   private queue: Promise<void> = Promise.resolve();
   private meta: SessionMeta | undefined;
 
   constructor(options: SessionManagerOptions) {
     this.sessionId = options.sessionId;
     this.baseDir = options.baseDir;
+
+    // Compaction 模式
+    this.compactionMode = options.compactionMode ?? "count";
+
+    // Count 模式参数
     this.maxMessages = options.maxMessages ?? 80;
     this.keepLast = options.keepLast ?? 60;
+
+    // Token 模式参数
+    this.contextWindowTokens = options.contextWindowTokens ?? 200_000;
+    this.systemPrompt = options.systemPrompt;
+    this.reserveTokens = options.reserveTokens ?? 1024;
+    this.targetRatio = options.targetRatio ?? 0.5;
+    this.minKeepMessages = options.minKeepMessages ?? 10;
+
+    // Summary 模式参数
+    this.model = options.model;
+    this.apiKey = options.apiKey;
+    this.customInstructions = options.customInstructions;
+
     this.meta = this.loadMeta();
   }
 
+  /**
+   * 更新系统提示词（用于 token 模式计算）
+   */
+  setSystemPrompt(systemPrompt: string | undefined) {
+    this.systemPrompt = systemPrompt;
+  }
+
+  /**
+   * 获取当前 context window token 数
+   */
+  getContextWindowTokens(): number {
+    return this.contextWindowTokens;
+  }
+
+  /**
+   * 设置 LLM Model（用于 summary 模式）
+   */
+  setModel(model: Model<any> | undefined) {
+    this.model = model;
+  }
+
+  /**
+   * 设置 API Key（用于 summary 模式）
+   */
+  setApiKey(apiKey: string | undefined) {
+    this.apiKey = apiKey;
+  }
+
+  /**
+   * 获取当前 compaction 模式
+   */
+  getCompactionMode(): "count" | "tokens" | "summary" {
+    return this.compactionMode;
+  }
+
   loadEntries(): SessionEntry[] {
     return readEntries(this.sessionId, { baseDir: this.baseDir });
   }
@@ -74,8 +168,56 @@ export class SessionManager {
   }
 
   async maybeCompact(messages: AgentMessage[]) {
-    const result = compactMessages(messages, this.maxMessages, this.keepLast);
+    let result;
+
+    if (this.compactionMode === "summary") {
+      // Summary 模式需要 model 和 apiKey
+      if (!this.model || !this.apiKey) {
+        // 降级到 tokens 模式
+        result = compactMessages(messages, {
+          mode: "tokens",
+          contextWindowTokens: this.contextWindowTokens,
+          systemPrompt: this.systemPrompt,
+          reserveTokens: this.reserveTokens,
+          targetRatio: this.targetRatio,
+          minKeepMessages: this.minKeepMessages,
+        });
+      } else {
+        result = await compactMessagesAsync(messages, {
+          mode: "summary",
+          model: this.model,
+          apiKey: this.apiKey,
+          contextWindowTokens: this.contextWindowTokens,
+          systemPrompt: this.systemPrompt,
+          reserveTokens: this.reserveTokens,
+          targetRatio: this.targetRatio,
+          minKeepMessages: this.minKeepMessages,
+          customInstructions: this.customInstructions,
+          previousSummary: this.previousSummary,
+        });
+
+        // 保存摘要用于下次增量更新
+        if (result?.summary) {
+          this.previousSummary = result.summary;
+        }
+      }
+    } else {
+      result = compactMessages(messages, {
+        mode: this.compactionMode,
+        // Count 模式参数
+        maxMessages: this.maxMessages,
+        keepLast: this.keepLast,
+        // Token 模式参数
+        contextWindowTokens: this.contextWindowTokens,
+        systemPrompt: this.systemPrompt,
+        reserveTokens: this.reserveTokens,
+        targetRatio: this.targetRatio,
+        minKeepMessages: this.minKeepMessages,
+      });
+    }
+
     if (!result) return null;
+
     const entries: SessionEntry[] = [];
     if (this.meta) {
       entries.push({ type: "meta", meta: this.meta, timestamp: Date.now() });
@@ -88,7 +230,13 @@ export class SessionManager {
       removed: result.removedCount,
       kept: result.kept.length,
       timestamp: Date.now(),
+      // Token/Summary 模式下的额外信息
+      tokensRemoved: result.tokensRemoved,
+      tokensKept: result.tokensKept,
+      summary: result.summary,
+      reason: result.reason,
     });
+
     await this.enqueue(() =>
       writeEntries(this.sessionId, entries, { baseDir: this.baseDir }),
     );
diff --git a/src/agent/session/types.ts b/src/agent/session/types.ts
index cbc82685..c9086b74 100644
--- a/src/agent/session/types.ts
+++ b/src/agent/session/types.ts
@@ -4,9 +4,22 @@ export type SessionMeta = {
   provider?: string;
   model?: string;
   thinkingLevel?: string;
+  /** Context window token 数 */
+  contextWindowTokens?: number;
 };
 
 export type SessionEntry =
   | { type: "message"; message: AgentMessage; timestamp: number }
   | { type: "meta"; meta: SessionMeta; timestamp: number }
-  | { type: "compaction"; removed: number; kept: number; timestamp: number };
+  | {
+      type: "compaction";
+      removed: number;
+      kept: number;
+      timestamp: number;
+      /** Token 感知 compaction 信息（可选，向后兼容） */
+      tokensRemoved?: number | undefined;
+      tokensKept?: number | undefined;
+      /** 摘要模式生成的摘要 */
+      summary?: string | undefined;
+      reason?: "count" | "tokens" | "summary" | undefined;
+    };
diff --git a/src/agent/types.ts b/src/agent/types.ts
index 9773b6ce..fc7ad9bd 100644
--- a/src/agent/types.ts
+++ b/src/agent/types.ts
@@ -24,4 +24,25 @@ export type AgentOptions = {
   cwd?: string | undefined;
   sessionId?: string | undefined;
   logger?: AgentLogger | undefined;
+
+  // === Context Window Guard 配置 ===
+  /** 手动指定 context window token 数（覆盖 model 的值） */
+  contextWindowTokens?: number | undefined;
+  /** 预留给响应生成的 token 数，默认 1024 */
+  reserveTokens?: number | undefined;
+  /**
+   * Compaction 模式:
+   * - "count": 使用旧的消息计数
+   * - "tokens": 使用 token 感知（默认）
+   * - "summary": 使用 LLM 生成摘要
+   */
+  compactionMode?: "count" | "tokens" | "summary" | undefined;
+  /** Compaction 目标利用率 (0-1)，默认 0.5 */
+  compactionTargetRatio?: number | undefined;
+  /** 最小保留消息数，默认 10 */
+  minKeepMessages?: number | undefined;
+
+  // === Summary Compaction 配置 ===
+  /** 自定义摘要生成指令 */
+  summaryInstructions?: string | undefined;
 };