@witqq/agent-sdk 0.6.1 → 0.8.0

package/dist/chat/context.d.ts

@@ -0,0 +1,191 @@
+import { c as ChatMessage } from '../types-ajANVzf7.js';
+import '../agent-CW9XbmG_.js';
+import 'zod';
+import '../errors-C-so0M4t.js';
+import '../types-BxggH0Yh.js';
+import '../errors-CmVvczxZ.js';
+
+/**
+ * @witqq/agent-sdk/chat/context
+ *
+ * Context window manager for selecting which messages fit within a token budget.
+ * Stateless: takes messages in, returns trimmed messages out.
+ * Three overflow strategies: truncate-oldest, sliding-window, summarize-placeholder.
+ */
+
+/**
+ * Options for token estimation.
+ */
+interface TokenEstimationOptions {
+    /**
+     * Characters per token ratio.
+     * Lower = more conservative (fewer messages fit).
+     * @default 4
+     */
+    charsPerToken?: number;
+}
+/**
+ * Estimate token count for a single chat message.
+ * Uses character-based heuristic: `Math.ceil(charCount / charsPerToken)`.
+ *
+ * Counts:
+ * - Text content (string or text parts)
+ * - Serialized tool calls and tool results
+ * - Thinking blocks
+ * - Role overhead (~4 tokens)
+ *
+ * @param message - Chat message to estimate
+ * @param options - Estimation options
+ * @returns Estimated token count
+ *
+ * @example
+ * ```typescript
+ * const tokens = estimateTokens(message);
+ * const conservative = estimateTokens(message, { charsPerToken: 3 });
+ * ```
+ */
+declare function estimateTokens(message: ChatMessage, options?: TokenEstimationOptions): number;
+/** Overflow strategy type */
+type OverflowStrategy = "truncate-oldest" | "sliding-window" | "summarize-placeholder";
+/**
+ * Async summarizer function for the summarize-placeholder strategy.
+ * Receives removed messages and returns a summary string.
+ * When configured, replaces the static placeholder text with actual summary.
+ */
+type ContextSummarizer = (removedMessages: readonly ChatMessage[]) => Promise<string>;
+/**
+ * Configuration for the context window manager.
+ */
+interface ContextWindowConfig {
+    /** Maximum token budget for the context window */
+    maxTokens: number;
+    /**
+     * Tokens reserved for system prompt and response generation.
+     * Subtracted from maxTokens to get available budget.
+     * @default 0
+     */
+    reservedTokens?: number;
+    /**
+     * Strategy for handling overflow when messages exceed budget.
+     * @default "truncate-oldest"
+     */
+    strategy?: OverflowStrategy;
+    /**
+     * Token estimation options.
+     */
+    estimation?: TokenEstimationOptions;
+    /**
+     * Optional async summarizer for the summarize-placeholder strategy.
+     * When provided, replaces the static placeholder with a generated summary.
+     * Falls back to static placeholder if summarizer throws.
+     */
+    summarizer?: ContextSummarizer;
+}
+/**
+ * Result of context window trimming.
+ */
+interface ContextWindowResult {
+    /** Messages that fit within the budget */
+    messages: ChatMessage[];
+    /** Total estimated tokens for included messages */
+    totalTokens: number;
+    /** Number of messages removed */
+    removedCount: number;
+    /** Whether any messages were truncated */
+    wasTruncated: boolean;
+}
+/**
+ * Context usage statistics for a session.
+ * Returned by `IChatRuntime.getContextStats()`.
+ *
+ * When real usage data is available (after the first API response),
+ * `realPromptTokens` and `realCompletionTokens` contain actual token counts.
+ * `modelContextWindow` is the model's context window from `listModels()`.
+ */
+interface ContextStats {
+    /** Estimated total tokens in the trimmed context (heuristic, kept for backward compat) */
+    totalTokens: number;
+    /** Number of messages removed by trimming */
+    removedCount: number;
+    /** Whether context was truncated */
+    wasTruncated: boolean;
+    /** Available token budget (maxTokens − reservedTokens) */
+    availableBudget: number;
+    /** Real prompt tokens from the last API response (undefined before first response) */
+    realPromptTokens?: number;
+    /** Real completion tokens from the last API response (undefined before first response) */
+    realCompletionTokens?: number;
+    /** Model's context window in tokens from listModels() (undefined if not available) */
+    modelContextWindow?: number;
+}
+/**
+ * Stateless context window manager.
+ * Takes messages and returns the subset that fits within a token budget.
+ *
+ * @example
+ * ```typescript
+ * const manager = new ContextWindowManager({
+ *   maxTokens: 4096,
+ *   reservedTokens: 500,
+ *   strategy: "sliding-window",
+ * });
+ *
+ * const result = manager.fitMessages(messages);
+ * // result.messages — trimmed to fit budget
+ * // result.totalTokens — estimated token usage
+ * // result.wasTruncated — whether messages were removed
+ * ```
+ */
+declare class ContextWindowManager {
+    private readonly config;
+    constructor(config: ContextWindowConfig);
+    /** Available token budget after reserving tokens */
+    get availableBudget(): number;
+    /**
+     * Estimate tokens for a single message.
+     * @param message - Message to estimate
+     * @returns Estimated token count
+     */
+    estimateMessageTokens(message: ChatMessage): number;
+    /**
+     * Fit messages within the token budget using the configured strategy.
+     * @param messages - All messages to consider
+     * @returns Result with fitted messages and metadata
+     */
+    fitMessages(messages: readonly ChatMessage[]): ContextWindowResult;
+    /**
+     * Async variant of fitMessages that supports async summarization.
+     * When strategy is "summarize-placeholder" and a summarizer is configured,
+     * calls the summarizer with removed messages and replaces the placeholder text.
+     * Falls back to static placeholder if summarizer throws.
+     * For other strategies, behaves identically to fitMessages().
+     */
+    fitMessagesAsync(messages: readonly ChatMessage[]): Promise<ContextWindowResult>;
+    /**
+     * Trim messages using real token usage data from the previous API call.
+     * Uses average-based algorithm: `avgTokensPerMessage = lastPromptTokens / messageCount`.
+     * Removes oldest non-system messages until freed budget brings usage under modelContextWindow.
+     *
+     * @param messages - All messages in the session
+     * @param lastPromptTokens - Real prompt tokens from the last API response
+     * @param modelContextWindow - Model's total context window size in tokens
+     * @returns Result with fitted messages and metadata
+     */
+    fitMessagesWithUsage(messages: readonly ChatMessage[], lastPromptTokens: number, modelContextWindow: number): ContextWindowResult;
+    /**
+     * Truncate oldest: keeps system messages, removes oldest non-system messages first.
+     * Always keeps the most recent user message.
+     */
+    private truncateOldest;
+    /**
+     * Sliding window: keeps the most recent messages that fit within budget.
+     */
+    private slidingWindow;
+    /**
+     * Summarize placeholder: replaces truncated messages with a placeholder,
+     * preserving system messages and recent context.
+     */
+    private summarizePlaceholder;
+}
+
+export { type ContextStats, type ContextSummarizer, type ContextWindowConfig, ContextWindowManager, type ContextWindowResult, type OverflowStrategy, type TokenEstimationOptions, estimateTokens };

package/dist/chat/context.js

@@ -0,0 +1,277 @@
+// src/chat/context.ts
+function estimateTokens(message, options) {
+  const ratio = options?.charsPerToken ?? 4;
+  let charCount = 0;
+  charCount += message.role.length + 4;
+  for (const part of message.parts) {
+    charCount += estimatePartChars(part);
+  }
+  return Math.ceil(charCount / ratio);
+}
+function estimatePartChars(part) {
+  switch (part.type) {
+    case "text":
+      return part.text.length;
+    case "reasoning":
+      return part.text.length;
+    case "tool_call":
+      return JSON.stringify(part.args).length + part.name.length + 20 + (part.result !== void 0 ? JSON.stringify(part.result).length : 0);
+    case "source":
+      return (part.title?.length ?? 0) + part.url.length + 10;
+    case "file":
+      return part.name.length + part.data.length + 20;
+  }
+}
+var ContextWindowManager = class {
+  config;
+  constructor(config) {
+    this.config = {
+      maxTokens: config.maxTokens,
+      reservedTokens: config.reservedTokens ?? 0,
+      strategy: config.strategy ?? "truncate-oldest",
+      estimation: config.estimation,
+      summarizer: config.summarizer
+    };
+  }
+  /** Available token budget after reserving tokens */
+  get availableBudget() {
+    return Math.max(0, this.config.maxTokens - this.config.reservedTokens);
+  }
+  /**
+   * Estimate tokens for a single message.
+   * @param message - Message to estimate
+   * @returns Estimated token count
+   */
+  estimateMessageTokens(message) {
+    return estimateTokens(message, this.config.estimation);
+  }
+  /**
+   * Fit messages within the token budget using the configured strategy.
+   * @param messages - All messages to consider
+   * @returns Result with fitted messages and metadata
+   */
+  fitMessages(messages) {
+    if (messages.length === 0) {
+      return { messages: [], totalTokens: 0, removedCount: 0, wasTruncated: false };
+    }
+    const budget = this.availableBudget;
+    const tokenCounts = messages.map((m) => this.estimateMessageTokens(m));
+    const totalTokens = tokenCounts.reduce((a, b) => a + b, 0);
+    if (totalTokens <= budget) {
+      return {
+        messages: [...messages],
+        totalTokens,
+        removedCount: 0,
+        wasTruncated: false
+      };
+    }
+    switch (this.config.strategy) {
+      case "truncate-oldest":
+        return this.truncateOldest(messages, tokenCounts, budget);
+      case "sliding-window":
+        return this.slidingWindow(messages, tokenCounts, budget);
+      case "summarize-placeholder":
+        return this.summarizePlaceholder(messages, tokenCounts, budget);
+    }
+  }
+  /**
+   * Async variant of fitMessages that supports async summarization.
+   * When strategy is "summarize-placeholder" and a summarizer is configured,
+   * calls the summarizer with removed messages and replaces the placeholder text.
+   * Falls back to static placeholder if summarizer throws.
+   * For other strategies, behaves identically to fitMessages().
+   */
+  async fitMessagesAsync(messages) {
+    const result = this.fitMessages(messages);
+    if (this.config.strategy !== "summarize-placeholder" || !result.wasTruncated || !this.config.summarizer) {
+      return result;
+    }
+    const keptIds = new Set(result.messages.map((m) => m.id));
+    const removed = messages.filter((m) => !keptIds.has(m.id));
+    if (removed.length === 0) return result;
+    let summaryText;
+    try {
+      summaryText = await this.config.summarizer(removed);
+    } catch {
+      return result;
+    }
+    const updatedMessages = result.messages.map((m) => {
+      if (m.metadata?.isSummary === true) {
+        return {
+          ...m,
+          parts: [{ type: "text", text: summaryText, status: "complete" }]
+        };
+      }
+      return m;
+    });
+    return { ...result, messages: updatedMessages };
+  }
+  /**
+   * Trim messages using real token usage data from the previous API call.
+   * Uses average-based algorithm: `avgTokensPerMessage = lastPromptTokens / messageCount`.
+   * Removes oldest non-system messages until freed budget brings usage under modelContextWindow.
+   *
+   * @param messages - All messages in the session
+   * @param lastPromptTokens - Real prompt tokens from the last API response
+   * @param modelContextWindow - Model's total context window size in tokens
+   * @returns Result with fitted messages and metadata
+   */
+  fitMessagesWithUsage(messages, lastPromptTokens, modelContextWindow) {
+    if (messages.length === 0) {
+      return { messages: [], totalTokens: 0, removedCount: 0, wasTruncated: false };
+    }
+    const budget = modelContextWindow - this.config.reservedTokens;
+    if (budget <= 0 || lastPromptTokens <= budget) {
+      return {
+        messages: [...messages],
+        totalTokens: lastPromptTokens,
+        removedCount: 0,
+        wasTruncated: false
+      };
+    }
+    const avgTokensPerMessage = lastPromptTokens / messages.length;
+    const tokensToFree = lastPromptTokens - budget;
+    const messagesToRemove = Math.ceil(tokensToFree / avgTokensPerMessage);
+    const nonSystemIndices = [];
+    for (let i = 0; i < messages.length; i++) {
+      if (messages[i].role === "system") ; else {
+        nonSystemIndices.push(i);
+      }
+    }
+    const removableCount = Math.min(messagesToRemove, nonSystemIndices.length);
+    const removedIndices = new Set(nonSystemIndices.slice(0, removableCount));
+    const result = [];
+    for (let i = 0; i < messages.length; i++) {
+      if (!removedIndices.has(i)) {
+        result.push(messages[i]);
+      }
+    }
+    const estimatedTokens = Math.round(
+      lastPromptTokens * (result.length / messages.length)
+    );
+    return {
+      messages: result,
+      totalTokens: estimatedTokens,
+      removedCount: removableCount,
+      wasTruncated: removableCount > 0
+    };
+  }
+  /**
+   * Truncate oldest: keeps system messages, removes oldest non-system messages first.
+   * Always keeps the most recent user message.
+   */
+  truncateOldest(messages, tokenCounts, budget) {
+    const systemIndices = [];
+    const nonSystemIndices = [];
+    for (let i = 0; i < messages.length; i++) {
+      if (messages[i].role === "system") {
+        systemIndices.push(i);
+      } else {
+        nonSystemIndices.push(i);
+      }
+    }
+    let usedTokens = systemIndices.reduce(
+      (sum, i) => sum + tokenCounts[i],
+      0
+    );
+    const includedNonSystem = [];
+    for (let i = nonSystemIndices.length - 1; i >= 0; i--) {
+      const idx = nonSystemIndices[i];
+      if (usedTokens + tokenCounts[idx] <= budget) {
+        includedNonSystem.unshift(idx);
+        usedTokens += tokenCounts[idx];
+      }
+    }
+    const includedSet = /* @__PURE__ */ new Set([...systemIndices, ...includedNonSystem]);
+    const result = [];
+    let resultTokens = 0;
+    for (let i = 0; i < messages.length; i++) {
+      if (includedSet.has(i)) {
+        result.push(messages[i]);
+        resultTokens += tokenCounts[i];
+      }
+    }
+    return {
+      messages: result,
+      totalTokens: resultTokens,
+      removedCount: messages.length - result.length,
+      wasTruncated: true
+    };
+  }
+  /**
+   * Sliding window: keeps the most recent messages that fit within budget.
+   */
+  slidingWindow(messages, tokenCounts, budget) {
+    const result = [];
+    let usedTokens = 0;
+    for (let i = messages.length - 1; i >= 0; i--) {
+      if (usedTokens + tokenCounts[i] <= budget) {
+        result.unshift(messages[i]);
+        usedTokens += tokenCounts[i];
+      } else {
+        break;
+      }
+    }
+    return {
+      messages: result,
+      totalTokens: usedTokens,
+      removedCount: messages.length - result.length,
+      wasTruncated: true
+    };
+  }
+  /**
+   * Summarize placeholder: replaces truncated messages with a placeholder,
+   * preserving system messages and recent context.
+   */
+  summarizePlaceholder(messages, tokenCounts, budget) {
+    const systemMessages = [];
+    const nonSystem = [];
+    for (let i = 0; i < messages.length; i++) {
+      if (messages[i].role === "system") {
+        systemMessages.push({ msg: messages[i], tokens: tokenCounts[i] });
+      } else {
+        nonSystem.push({ msg: messages[i], tokens: tokenCounts[i], idx: i });
+      }
+    }
+    let usedTokens = systemMessages.reduce((s, m) => s + m.tokens, 0);
+    const placeholderTokens = 20;
+    usedTokens += placeholderTokens;
+    const recentKept = [];
+    for (let i = nonSystem.length - 1; i >= 0; i--) {
+      if (usedTokens + nonSystem[i].tokens <= budget) {
+        recentKept.unshift(nonSystem[i]);
+        usedTokens += nonSystem[i].tokens;
+      } else {
+        break;
+      }
+    }
+    const removedCount = messages.length - systemMessages.length - recentKept.length;
+    const result = [];
+    for (const sm of systemMessages) {
+      result.push(sm.msg);
+    }
+    if (removedCount > 0) {
+      result.push({
+        id: "context-placeholder",
+        role: "system",
+        parts: [{ type: "text", text: `[${removedCount} earlier message${removedCount === 1 ? "" : "s"} omitted for context window]`, status: "complete" }],
+        metadata: { isSummary: true },
+        createdAt: (/* @__PURE__ */ new Date()).toISOString(),
+        status: "complete"
+      });
+    }
+    for (const m of recentKept) {
+      result.push(m.msg);
+    }
+    return {
+      messages: result,
+      totalTokens: usedTokens,
+      removedCount,
+      wasTruncated: true
+    };
+  }
+};
+
+export { ContextWindowManager, estimateTokens };
+//# sourceMappingURL=context.js.map
+//# sourceMappingURL=context.js.map

package/dist/chat/context.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/chat/context.ts"],"names":[],"mappings":";AA4CO,SAAS,cAAA,CACd,SACA,OAAA,EACQ;AACR,EAAA,MAAM,KAAA,GAAQ,SAAS,aAAA,IAAiB,CAAA;AACxC,EAAA,IAAI,SAAA,GAAY,CAAA;AAGhB,EAAA,SAAA,IAAa,OAAA,CAAQ,KAAK,MAAA,GAAS,CAAA;AAGnC,EAAA,KAAA,MAAW,IAAA,IAAQ,QAAQ,KAAA,EAAO;AAChC,IAAA,SAAA,IAAa,kBAAkB,IAAI,CAAA;AAAA,EACrC;AAEA,EAAA,OAAO,IAAA,CAAK,IAAA,CAAK,SAAA,GAAY,KAAK,CAAA;AACpC;AAEA,SAAS,kBAAkB,IAAA,EAA2B;AACpD,EAAA,QAAQ,KAAK,IAAA;AAAM,IACjB,KAAK,MAAA;AACH,MAAA,OAAO,KAAK,IAAA,CAAK,MAAA;AAAA,IACnB,KAAK,WAAA;AACH,MAAA,OAAO,KAAK,IAAA,CAAK,MAAA;AAAA,IACnB,KAAK,WAAA;AACH,MAAA,OAAO,KAAK,SAAA,CAAU,IAAA,CAAK,IAAI,CAAA,CAAE,MAAA,GAAS,KAAK,IAAA,CAAK,MAAA,GAAS,EAAA,IAC1D,IAAA,CAAK,WAAW,MAAA,GAAY,IAAA,CAAK,UAAU,IAAA,CAAK,MAAM,EAAE,MAAA,GAAS,CAAA,CAAA;AAAA,IACtE,KAAK,QAAA;AACH,MAAA,OAAA,CAAQ,KAAK,KAAA,EAAO,MAAA,IAAU,CAAA,IAAK,IAAA,CAAK,IAAI,MAAA,GAAS,EAAA;AAAA,IACvD,KAAK,MAAA;AACH,MAAA,OAAO,IAAA,CAAK,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,KAAK,MAAA,GAAS,EAAA;AAAA;AAEnD;AAmHO,IAAM,uBAAN,MAA2B;AAAA,EACf,MAAA;AAAA,EAKjB,YAAY,MAAA,EAA6B;AACvC,IAAA,IAAA,CAAK,MAAA,GAAS;AAAA,MACZ,WAAW,MAAA,CAAO,SAAA;AAAA,MAClB,cAAA,EAAgB,OAAO,cAAA,IAAkB,CAAA;AAAA,MACzC,QAAA,EAAU,OAAO,QAAA,IAAY,iBAAA;AAAA,MAC7B,YAAY,MAAA,CAAO,UAAA;AAAA,MACnB,YAAY,MAAA,CAAO;AAAA,KACrB;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,eAAA,GAA0B;AAC5B,IAAA,OAAO,IAAA,CAAK,IAAI,CAAA,EAAG,IAAA,CAAK,OAAO,SAAA,GAAY,IAAA,CAAK,OAAO,cAAc,CAAA;AAAA,EACvE;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,sBAAsB,OAAA,EAA8B;AAClD,IAAA,OAAO,cAAA,CAAe,OAAA,EAAS,IAAA,CAAK,MAAA,CAAO,UAAU,CAAA;AAAA,EACvD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,YAAY,QAAA,EAAuD;AACjE,IAAA,IAAI,QAAA,CAAS,WAAW,CAAA,EAAG;AACzB,MAAA,OAAO,EAAE,UAAU,EAAC,EAAG,aAAa,CAAA,EAAG,YAAA,EAAc,CAAA,EAAG,YAAA,EAAc,KAAA,EAAM;AAAA,IAC9E;AAEA,IAAA,MAAM,SAAS,IAAA,CAAK,eAAA;AAGpB,IAAA,MAAM,WAAA,GAAc,SAAS,GAAA,CAAI,CAAC,MAAM,IAAA,CAAK,qBAAA,CAAsB,CAAC,CAAC,CAAA;AACrE,IAAA,MAAM,WAAA,GAAc,YAAY,MAAA,CAAO,CAAC,GAAG,CAAA,KAAM,CAAA,GAAI,GAAG,CAAC,CAAA;AAGzD,IAAA,IAAI,eAAe,MAAA,EAAQ;AACzB,MAAA,OAAO;AAAA,QACL,QAAA,EAAU,CAAC,GAAG,QAAQ,CAAA;AAAA,QACtB,WAAA;AAAA,QACA,YAAA,EAAc,CAAA;AAAA,QACd,YAAA,EAAc;AAAA,OAChB;AAAA,IACF;AAEA,IAAA,QAAQ,IAAA,CAAK,OAAO,QAAA;AAAU,MAC5B,KAAK,iBAAA;AACH,QAAA,OAAO,IAAA,CAAK,cAAA,CAAe,QAAA,EAAU,WAAA,EAAa,MAAM,CAAA;AAAA,MAC1D,KAAK,gBAAA;AACH,QAAA,OAAO,IAAA,CAAK,aAAA,CAAc,QAAA,EAAU,WAAA,EAAa,MAAM,CAAA;AAAA,MACzD,KAAK,uBAAA;AACH,QAAA,OAAO,IAAA,CAAK,oBAAA,CAAqB,QAAA,EAAU,WAAA,EAAa,MAAM,CAAA;AAAA;AAClE,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,iBAAiB,QAAA,EAAgE;AACrF,IAAA,MAAM,MAAA,GAAS,IAAA,CAAK,WAAA,CAAY,QAAQ,CAAA;AAGxC,IAAA,IACE,IAAA,CAAK,MAAA,CAAO,QAAA,KAAa,uBAAA,IACzB,CAAC,OAAO,YAAA,IACR,CAAC,IAAA,CAAK,MAAA,CAAO,UAAA,EACb;AACA,MAAA,OAAO,MAAA;AAAA,IACT;AAGA,IAAA,MAAM,OAAA,GAAU,IAAI,GAAA,CAAI,MAAA,CAAO,SAAS,GAAA,CAAI,CAAA,CAAA,KAAK,CAAA,CAAE,EAAE,CAAC,CAAA;AACtD,IAAA,MAAM,OAAA,GAAU,SAAS,MAAA,CAAO,CAAA,CAAA,KAAK,CAAC,OAAA,CAAQ,GAAA,CAAI,CAAA,CAAE,EAAE,CAAC,CAAA;AACvD,IAAA,IAAI,OAAA,CAAQ,MAAA,KAAW,CAAA,EAAG,OAAO,MAAA;AAGjC,IAAA,IAAI,WAAA;AACJ,IAAA,IAAI;AACF,MAAA,WAAA,GAAc,MAAM,IAAA,CAAK,MAAA,CAAO,UAAA,CAAW,OAAO,CAAA;AAAA,IACpD,CAAA,CAAA,MAAQ;AACN,MAAA,OAAO,MAAA;AAAA,IACT;AAGA,IAAA,MAAM,eAAA,GAAkB,MAAA,CAAO,QAAA,CAAS,GAAA,CAAI,CAAA,CAAA,KAAK;AAC/C,MAAA,IAAK,CAAA,CAAE,QAAA,EAAsC,SAAA,KAAc,IAAA,EAAM;AAC/D,QAAA,OAAO;AAAA,UACL,GAAG,CAAA;AAAA,UACH,KAAA,EAAO,CAAC,EAAE,IAAA,EAAM,QAAiB,IAAA,EAAM,WAAA,EAAa,MAAA,EAAQ,UAAA,EAAqB;AAAA,SACnF;AAAA,MACF;AACA,MAAA,OAAO,CAAA;AAAA,IACT,CAAC,CAAA;AAED,IAAA,OAAO,EAAE,GAAG,MAAA,EAAQ,QAAA,EAAU,eAAA,EAAgB;AAAA,EAChD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAYA,oBAAA,CACE,QAAA,EACA,gBAAA,EACA,kBAAA,EACqB;AACrB,IAAA,IAAI,QAAA,CAAS,WAAW,CAAA,EAAG;AACzB,MAAA,OAAO,EAAE,UAAU,EAAC,EAAG,aAAa,CAAA,EAAG,YAAA,EAAc,CAAA,EAAG,YAAA,EAAc,KAAA,EAAM;AAAA,IAC9E;AAEA,IAAA,MAAM,MAAA,GAAS,kBAAA,GAAqB,IAAA,CAAK,MAAA,CAAO,cAAA;AAChD,IAAA,IAAI,MAAA,IAAU,CAAA,IAAK,gBAAA,IAAoB,MAAA,EAAQ;AAC7C,MAAA,OAAO;AAAA,QACL,QAAA,EAAU,CAAC,GAAG,QAAQ,CAAA;AAAA,QACtB,WAAA,EAAa,gBAAA;AAAA,QACb,YAAA,EAAc,CAAA;AAAA,QACd,YAAA,EAAc;AAAA,OAChB;AAAA,IACF;AAGA,IAAA,MAAM,mBAAA,GAAsB,mBAAmB,QAAA,CAAS,MAAA;AAGxD,IAAA,MAAM,eAAe,gBAAA,GAAmB,MAAA;AAExC,IAAA,MAAM,gBAAA,GAAmB,IAAA,CAAK,IAAA,CAAK,YAAA,GAAe,mBAAmB,CAAA;AAIrE,IAAA,MAAM,mBAA6B,EAAC;AACpC,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,QAAA,CAAS,QAAQ,CAAA,EAAA,EAAK;AACxC,MAAA,IAAI,QAAA,CAAS,CAAC,CAAA,CAAE,IAAA,KAAS,QAAA,EAAU,CAEnC,MAAO;AACL,QAAA,gBAAA,CAAiB,KAAK,CAAC,CAAA;AAAA,MACzB;AAAA,IACF;AAGA,IAAA,MAAM,cAAA,GAAiB,IAAA,CAAK,GAAA,CAAI,gBAAA,EAAkB,iBAAiB,MAAM,CAAA;AACzE,IAAA,MAAM,iBAAiB,IAAI,GAAA,CAAI,iBAAiB,KAAA,CAAM,CAAA,EAAG,cAAc,CAAC,CAAA;AAExE,IAAA,MAAM,SAAwB,EAAC;AAC/B,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,QAAA,CAAS,QAAQ,CAAA,EAAA,EAAK;AACxC,MAAA,IAAI,CAAC,cAAA,CAAe,GAAA,CAAI,CAAC,CAAA,EAAG;AAC1B,QAAA,MAAA,CAAO,IAAA,CAAK,QAAA,CAAS,CAAC,CAAC,CAAA;AAAA,MACzB;AAAA,IACF;AAGA,IAAA,MAAM,kBAAkB,IAAA,CAAK,KAAA;AAAA,MAC3B,gBAAA,IAAoB,MAAA,CAAO,MAAA,GAAS,QAAA,CAAS,MAAA;AAAA,KAC/C;AAEA,IAAA,OAAO;AAAA,MACL,QAAA,EAAU,MAAA;AAAA,MACV,WAAA,EAAa,eAAA;AAAA,MACb,YAAA,EAAc,cAAA;AAAA,MACd,cAAc,cAAA,GAAiB;AAAA,KACjC;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,cAAA,CACN,QAAA,EACA,WAAA,EACA,MAAA,EACqB;AAErB,IAAA,MAAM,gBAA0B,EAAC;AACjC,IAAA,MAAM,mBAA6B,EAAC;AAEpC,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,QAAA,CAAS,QAAQ,CAAA,EAAA,EAAK;AACxC,MAAA,IAAI,QAAA,CAAS,CAAC,CAAA,CAAE,IAAA,KAAS,QAAA,EAAU;AACjC,QAAA,aAAA,CAAc,KAAK,CAAC,CAAA;AAAA,MACtB,CAAA,MAAO;AACL,QAAA,gBAAA,CAAiB,KAAK,CAAC,CAAA;AAAA,MACzB;AAAA,IACF;AAGA,IAAA,IAAI,aAAa,aAAA,CAAc,MAAA;AAAA,MAC7B,CAAC,GAAA,EAAK,CAAA,KAAM,GAAA,GAAM,YAAY,CAAC,CAAA;AAAA,MAC/B;AAAA,KACF;AAMA,IAAA,MAAM,oBAA8B,EAAC;AACrC,IAAA,KAAA,IAAS,IAAI,gBAAA,CAAiB,MAAA,GAAS,CAAA,EAAG,CAAA,IAAK,GAAG,CAAA,EAAA,EAAK;AACrD,MAAA,MAAM,GAAA,GAAM,iBAAiB,CAAC,CAAA;AAC9B,MAAA,IAAI,UAAA,GAAa,WAAA,CAAY,GAAG,CAAA,IAAK,MAAA,EAAQ;AAC3C,QAAA,iBAAA,CAAkB,QAAQ,GAAG,CAAA;AAC7B,QAAA,UAAA,IAAc,YAAY,GAAG,CAAA;AAAA,MAC/B;AAAA,IACF;AAGA,IAAA,MAAM,WAAA,uBAAkB,GAAA,CAAI,CAAC,GAAG,aAAA,EAAe,GAAG,iBAAiB,CAAC,CAAA;AACpE,IAAA,MAAM,SAAwB,EAAC;AAC/B,IAAA,IAAI,YAAA,GAAe,CAAA;AACnB,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,QAAA,CAAS,QAAQ,CAAA,EAAA,EAAK;AACxC,MAAA,IAAI,WAAA,CAAY,GAAA,CAAI,CAAC,CAAA,EAAG;AACtB,QAAA,MAAA,CAAO,IAAA,CAAK,QAAA,CAAS,CAAC,CAAC,CAAA;AACvB,QAAA,YAAA,IAAgB,YAAY,CAAC,CAAA;AAAA,MAC/B;AAAA,IACF;AAEA,IAAA,OAAO;AAAA,MACL,QAAA,EAAU,MAAA;AAAA,MACV,WAAA,EAAa,YAAA;AAAA,MACb,YAAA,EAAc,QAAA,CAAS,MAAA,GAAS,MAAA,CAAO,MAAA;AAAA,MACvC,YAAA,EAAc;AAAA,KAChB;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKQ,aAAA,CACN,QAAA,EACA,WAAA,EACA,MAAA,EACqB;AACrB,IAAA,MAAM,SAAwB,EAAC;AAC/B,IAAA,IAAI,UAAA,GAAa,CAAA;AAGjB,IAAA,KAAA,IAAS,IAAI,QAAA,CAAS,MAAA,GAAS,CAAA,EAAG,CAAA,IAAK,GAAG,CAAA,EAAA,EAAK;AAC7C,MAAA,IAAI,UAAA,GAAa,WAAA,CAAY,CAAC,CAAA,IAAK,MAAA,EAAQ;AACzC,QAAA,MAAA,CAAO,OAAA,CAAQ,QAAA,CAAS,CAAC,CAAC,CAAA;AAC1B,QAAA,UAAA,IAAc,YAAY,CAAC,CAAA;AAAA,MAC7B,CAAA,MAAO;AACL,QAAA;AAAA,MACF;AAAA,IACF;AAEA,IAAA,OAAO;AAAA,MACL,QAAA,EAAU,MAAA;AAAA,MACV,WAAA,EAAa,UAAA;AAAA,MACb,YAAA,EAAc,QAAA,CAAS,MAAA,GAAS,MAAA,CAAO,MAAA;AAAA,MACvC,YAAA,EAAc;AAAA,KAChB;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,oBAAA,CACN,QAAA,EACA,WAAA,EACA,MAAA,EACqB;AAErB,IAAA,MAAM,iBAAyD,EAAC;AAChE,IAAA,MAAM,YAAiE,EAAC;AAExE,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,QAAA,CAAS,QAAQ,CAAA,EAAA,EAAK;AACxC,MAAA,IAAI,QAAA,CAAS,CAAC,CAAA,CAAE,IAAA,KAAS,QAAA,EAAU;AACjC,QAAA,cAAA,CAAe,IAAA,CAAK,EAAE,GAAA,EAAK,QAAA,CAAS,CAAC,GAAG,MAAA,EAAQ,WAAA,CAAY,CAAC,CAAA,EAAG,CAAA;AAAA,MAClE,CAAA,MAAO;AACL,QAAA,SAAA,CAAU,IAAA,CAAK,EAAE,GAAA,EAAK,QAAA,CAAS,CAAC,CAAA,EAAG,MAAA,EAAQ,WAAA,CAAY,CAAC,CAAA,EAAG,GAAA,EAAK,CAAA,EAAG,CAAA;AAAA,MACrE;AAAA,IACF;AAGA,IAAA,IAAI,UAAA,GAAa,eAAe,MAAA,CAAO,CAAC,GAAG,CAAA,KAAM,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAC,CAAA;AAGhE,IAAA,MAAM,iBAAA,GAAoB,EAAA;AAC1B,IAAA,UAAA,IAAc,iBAAA;AAGd,IAAA,MAAM,aAA+B,EAAC;AACtC,IAAA,KAAA,IAAS,IAAI,SAAA,CAAU,MAAA,GAAS,CAAA,EAAG,CAAA,IAAK,GAAG,CAAA,EAAA,EAAK;AAC9C,MAAA,IAAI,UAAA,GAAa,SAAA,CAAU,CAAC,CAAA,CAAE,UAAU,MAAA,EAAQ;AAC9C,QAAA,UAAA,CAAW,OAAA,CAAQ,SAAA,CAAU,CAAC,CAAC,CAAA;AAC/B,QAAA,UAAA,IAAc,SAAA,CAAU,CAAC,CAAA,CAAE,MAAA;AAAA,MAC7B,CAAA,MAAO;AACL,QAAA;AAAA,MACF;AAAA,IACF;AAEA,IAAA,MAAM,YAAA,GACJ,QAAA,CAAS,MAAA,GACT,cAAA,CAAe,SACf,UAAA,CAAW,MAAA;AAGb,IAAA,MAAM,SAAwB,EAAC;AAG/B,IAAA,KAAA,MAAW,MAAM,cAAA,EAAgB;AAC/B,MAAA,MAAA,CAAO,IAAA,CAAK,GAAG,GAAG,CAAA;AAAA,IACpB;AAGA,IAAA,IAAI,eAAe,CAAA,EAAG;AACpB,MAAA,MAAA,CAAO,IAAA,CAAK;AAAA,QACV,EAAA,EAAI,qBAAA;AAAA,QACJ,IAAA,EAAM,QAAA;AAAA,QACN,OAAO,CAAC,EAAE,IAAA,EAAM,MAAA,EAAQ,MAAM,CAAA,CAAA,EAAI,YAAY,CAAA,gBAAA,EAAmB,YAAA,KAAiB,IAAI,EAAA,GAAK,GAAG,CAAA,4BAAA,CAAA,EAAgC,MAAA,EAAQ,YAAqB,CAAA;AAAA,QAC3J,QAAA,EAAU,EAAE,SAAA,EAAW,IAAA,EAAK;AAAA,QAC5B,SAAA,EAAA,iBAAW,IAAI,IAAA,EAAK,EAAE,WAAA,EAAY;AAAA,QAClC,MAAA,EAAQ;AAAA,OACT,CAAA;AAAA,IACH;AAGA,IAAA,KAAA,MAAW,KAAK,UAAA,EAAY;AAC1B,MAAA,MAAA,CAAO,IAAA,CAAK,EAAE,GAAG,CAAA;AAAA,IACnB;AAEA,IAAA,OAAO;AAAA,MACL,QAAA,EAAU,MAAA;AAAA,MACV,WAAA,EAAa,UAAA;AAAA,MACb,YAAA;AAAA,MACA,YAAA,EAAc;AAAA,KAChB;AAAA,EACF;AACF","file":"context.js","sourcesContent":["/**\n * @witqq/agent-sdk/chat/context\n *\n * Context window manager for selecting which messages fit within a token budget.\n * Stateless: takes messages in, returns trimmed messages out.\n * Three overflow strategies: truncate-oldest, sliding-window, summarize-placeholder.\n */\n\nimport type { ChatMessage, MessagePart } from \"./core.js\";\n\n// ─── Token Estimation ──────────────────────────────────────────\n\n/**\n * Options for token estimation.\n */\nexport interface TokenEstimationOptions {\n  /**\n   * Characters per token ratio.\n   * Lower = more conservative (fewer messages fit).\n   * @default 4\n   */\n  charsPerToken?: number;\n}\n\n/**\n * Estimate token count for a single chat message.\n * Uses character-based heuristic: `Math.ceil(charCount / charsPerToken)`.\n *\n * Counts:\n * - Text content (string or text parts)\n * - Serialized tool calls and tool results\n * - Thinking blocks\n * - Role overhead (~4 tokens)\n *\n * @param message - Chat message to estimate\n * @param options - Estimation options\n * @returns Estimated token count\n *\n * @example\n * ```typescript\n * const tokens = estimateTokens(message);\n * const conservative = estimateTokens(message, { charsPerToken: 3 });\n * ```\n */\nexport function estimateTokens(\n  message: ChatMessage,\n  options?: TokenEstimationOptions,\n): number {\n  const ratio = options?.charsPerToken ?? 4;\n  let charCount = 0;\n\n  // Role overhead\n  charCount += message.role.length + 4;\n\n  // Parts\n  for (const part of message.parts) {\n    charCount += estimatePartChars(part);\n  }\n\n  return Math.ceil(charCount / ratio);\n}\n\nfunction estimatePartChars(part: MessagePart): number {\n  switch (part.type) {\n    case \"text\":\n      return part.text.length;\n    case \"reasoning\":\n      return part.text.length;\n    case \"tool_call\":\n      return JSON.stringify(part.args).length + part.name.length + 20 +\n        (part.result !== undefined ? JSON.stringify(part.result).length : 0);\n    case \"source\":\n      return (part.title?.length ?? 0) + part.url.length + 10;\n    case \"file\":\n      return part.name.length + part.data.length + 20;\n  }\n}\n\n// ─── Overflow Strategies ───────────────────────────────────────\n\n/** Overflow strategy type */\nexport type OverflowStrategy =\n  | \"truncate-oldest\"\n  | \"sliding-window\"\n  | \"summarize-placeholder\";\n\n// ─── Context Window Configuration ──────────────────────────────\n\n/**\n * Async summarizer function for the summarize-placeholder strategy.\n * Receives removed messages and returns a summary string.\n * When configured, replaces the static placeholder text with actual summary.\n */\nexport type ContextSummarizer = (removedMessages: readonly ChatMessage[]) => Promise<string>;\n\n/**\n * Configuration for the context window manager.\n */\nexport interface ContextWindowConfig {\n  /** Maximum token budget for the context window */\n  maxTokens: number;\n\n  /**\n   * Tokens reserved for system prompt and response generation.\n   * Subtracted from maxTokens to get available budget.\n   * @default 0\n   */\n  reservedTokens?: number;\n\n  /**\n   * Strategy for handling overflow when messages exceed budget.\n   * @default \"truncate-oldest\"\n   */\n  strategy?: OverflowStrategy;\n\n  /**\n   * Token estimation options.\n   */\n  estimation?: TokenEstimationOptions;\n\n  /**\n   * Optional async summarizer for the summarize-placeholder strategy.\n   * When provided, replaces the static placeholder with a generated summary.\n   * Falls back to static placeholder if summarizer throws.\n   */\n  summarizer?: ContextSummarizer;\n}\n\n// ─── Context Window Result ─────────────────────────────────────\n\n/**\n * Result of context window trimming.\n */\nexport interface ContextWindowResult {\n  /** Messages that fit within the budget */\n  messages: ChatMessage[];\n  /** Total estimated tokens for included messages */\n  totalTokens: number;\n  /** Number of messages removed */\n  removedCount: number;\n  /** Whether any messages were truncated */\n  wasTruncated: boolean;\n}\n\n// ─── Context Stats ─────────────────────────────────────────────\n\n/**\n * Context usage statistics for a session.\n * Returned by `IChatRuntime.getContextStats()`.\n *\n * When real usage data is available (after the first API response),\n * `realPromptTokens` and `realCompletionTokens` contain actual token counts.\n * `modelContextWindow` is the model's context window from `listModels()`.\n */\nexport interface ContextStats {\n  /** Estimated total tokens in the trimmed context (heuristic, kept for backward compat) */\n  totalTokens: number;\n  /** Number of messages removed by trimming */\n  removedCount: number;\n  /** Whether context was truncated */\n  wasTruncated: boolean;\n  /** Available token budget (maxTokens − reservedTokens) */\n  availableBudget: number;\n  /** Real prompt tokens from the last API response (undefined before first response) */\n  realPromptTokens?: number;\n  /** Real completion tokens from the last API response (undefined before first response) */\n  realCompletionTokens?: number;\n  /** Model's context window in tokens from listModels() (undefined if not available) */\n  modelContextWindow?: number;\n}\n\n// ─── Context Window Manager ────────────────────────────────────\n\n/**\n * Stateless context window manager.\n * Takes messages and returns the subset that fits within a token budget.\n *\n * @example\n * ```typescript\n * const manager = new ContextWindowManager({\n *   maxTokens: 4096,\n *   reservedTokens: 500,\n *   strategy: \"sliding-window\",\n * });\n *\n * const result = manager.fitMessages(messages);\n * // result.messages — trimmed to fit budget\n * // result.totalTokens — estimated token usage\n * // result.wasTruncated — whether messages were removed\n * ```\n */\nexport class ContextWindowManager {\n  private readonly config: Required<\n    Pick<ContextWindowConfig, \"maxTokens\" | \"reservedTokens\" | \"strategy\">\n  > &\n    Pick<ContextWindowConfig, \"estimation\" | \"summarizer\">;\n\n  constructor(config: ContextWindowConfig) {\n    this.config = {\n      maxTokens: config.maxTokens,\n      reservedTokens: config.reservedTokens ?? 0,\n      strategy: config.strategy ?? \"truncate-oldest\",\n      estimation: config.estimation,\n      summarizer: config.summarizer,\n    };\n  }\n\n  /** Available token budget after reserving tokens */\n  get availableBudget(): number {\n    return Math.max(0, this.config.maxTokens - this.config.reservedTokens);\n  }\n\n  /**\n   * Estimate tokens for a single message.\n   * @param message - Message to estimate\n   * @returns Estimated token count\n   */\n  estimateMessageTokens(message: ChatMessage): number {\n    return estimateTokens(message, this.config.estimation);\n  }\n\n  /**\n   * Fit messages within the token budget using the configured strategy.\n   * @param messages - All messages to consider\n   * @returns Result with fitted messages and metadata\n   */\n  fitMessages(messages: readonly ChatMessage[]): ContextWindowResult {\n    if (messages.length === 0) {\n      return { messages: [], totalTokens: 0, removedCount: 0, wasTruncated: false };\n    }\n\n    const budget = this.availableBudget;\n\n    // Calculate tokens for each message\n    const tokenCounts = messages.map((m) => this.estimateMessageTokens(m));\n    const totalTokens = tokenCounts.reduce((a, b) => a + b, 0);\n\n    // All messages fit\n    if (totalTokens <= budget) {\n      return {\n        messages: [...messages],\n        totalTokens,\n        removedCount: 0,\n        wasTruncated: false,\n      };\n    }\n\n    switch (this.config.strategy) {\n      case \"truncate-oldest\":\n        return this.truncateOldest(messages, tokenCounts, budget);\n      case \"sliding-window\":\n        return this.slidingWindow(messages, tokenCounts, budget);\n      case \"summarize-placeholder\":\n        return this.summarizePlaceholder(messages, tokenCounts, budget);\n    }\n  }\n\n  /**\n   * Async variant of fitMessages that supports async summarization.\n   * When strategy is \"summarize-placeholder\" and a summarizer is configured,\n   * calls the summarizer with removed messages and replaces the placeholder text.\n   * Falls back to static placeholder if summarizer throws.\n   * For other strategies, behaves identically to fitMessages().\n   */\n  async fitMessagesAsync(messages: readonly ChatMessage[]): Promise<ContextWindowResult> {\n    const result = this.fitMessages(messages);\n\n    // Only enhance if summarize-placeholder strategy, messages were removed, and summarizer is configured\n    if (\n      this.config.strategy !== \"summarize-placeholder\" ||\n      !result.wasTruncated ||\n      !this.config.summarizer\n    ) {\n      return result;\n    }\n\n    // Find removed messages (those in original but not in result)\n    const keptIds = new Set(result.messages.map(m => m.id));\n    const removed = messages.filter(m => !keptIds.has(m.id));\n    if (removed.length === 0) return result;\n\n    // Call async summarizer, fall back to static placeholder on error\n    let summaryText: string;\n    try {\n      summaryText = await this.config.summarizer(removed);\n    } catch {\n      return result; // Keep static placeholder on summarizer failure\n    }\n\n    // Replace placeholder text with summarizer output\n    const updatedMessages = result.messages.map(m => {\n      if ((m.metadata as Record<string, unknown>)?.isSummary === true) {\n        return {\n          ...m,\n          parts: [{ type: \"text\" as const, text: summaryText, status: \"complete\" as const }],\n        };\n      }\n      return m;\n    });\n\n    return { ...result, messages: updatedMessages };\n  }\n\n  /**\n   * Trim messages using real token usage data from the previous API call.\n   * Uses average-based algorithm: `avgTokensPerMessage = lastPromptTokens / messageCount`.\n   * Removes oldest non-system messages until freed budget brings usage under modelContextWindow.\n   *\n   * @param messages - All messages in the session\n   * @param lastPromptTokens - Real prompt tokens from the last API response\n   * @param modelContextWindow - Model's total context window size in tokens\n   * @returns Result with fitted messages and metadata\n   */\n  fitMessagesWithUsage(\n    messages: readonly ChatMessage[],\n    lastPromptTokens: number,\n    modelContextWindow: number,\n  ): ContextWindowResult {\n    if (messages.length === 0) {\n      return { messages: [], totalTokens: 0, removedCount: 0, wasTruncated: false };\n    }\n\n    const budget = modelContextWindow - this.config.reservedTokens;\n    if (budget <= 0 || lastPromptTokens <= budget) {\n      return {\n        messages: [...messages],\n        totalTokens: lastPromptTokens,\n        removedCount: 0,\n        wasTruncated: false,\n      };\n    }\n\n    // Average tokens per message from real data\n    const avgTokensPerMessage = lastPromptTokens / messages.length;\n\n    // How many tokens we need to free\n    const tokensToFree = lastPromptTokens - budget;\n    // How many messages to remove (ceil to be safe)\n    const messagesToRemove = Math.ceil(tokensToFree / avgTokensPerMessage);\n\n    // Separate system and non-system messages\n    const systemIndices: number[] = [];\n    const nonSystemIndices: number[] = [];\n    for (let i = 0; i < messages.length; i++) {\n      if (messages[i].role === \"system\") {\n        systemIndices.push(i);\n      } else {\n        nonSystemIndices.push(i);\n      }\n    }\n\n    // Remove oldest non-system messages (from the beginning of conversation)\n    const removableCount = Math.min(messagesToRemove, nonSystemIndices.length);\n    const removedIndices = new Set(nonSystemIndices.slice(0, removableCount));\n\n    const result: ChatMessage[] = [];\n    for (let i = 0; i < messages.length; i++) {\n      if (!removedIndices.has(i)) {\n        result.push(messages[i]);\n      }\n    }\n\n    // Estimate new total: proportional reduction\n    const estimatedTokens = Math.round(\n      lastPromptTokens * (result.length / messages.length),\n    );\n\n    return {\n      messages: result,\n      totalTokens: estimatedTokens,\n      removedCount: removableCount,\n      wasTruncated: removableCount > 0,\n    };\n  }\n\n  /**\n   * Truncate oldest: keeps system messages, removes oldest non-system messages first.\n   * Always keeps the most recent user message.\n   */\n  private truncateOldest(\n    messages: readonly ChatMessage[],\n    tokenCounts: number[],\n    budget: number,\n  ): ContextWindowResult {\n    // Separate system messages (always kept) and non-system\n    const systemIndices: number[] = [];\n    const nonSystemIndices: number[] = [];\n\n    for (let i = 0; i < messages.length; i++) {\n      if (messages[i].role === \"system\") {\n        systemIndices.push(i);\n      } else {\n        nonSystemIndices.push(i);\n      }\n    }\n\n    // System messages cost\n    let usedTokens = systemIndices.reduce(\n      (sum, i) => sum + tokenCounts[i],\n      0,\n    );\n\n    // If system messages alone exceed budget, still include them\n    // (caller should configure reservedTokens properly)\n\n    // Try to fit non-system from newest to oldest\n    const includedNonSystem: number[] = [];\n    for (let i = nonSystemIndices.length - 1; i >= 0; i--) {\n      const idx = nonSystemIndices[i];\n      if (usedTokens + tokenCounts[idx] <= budget) {\n        includedNonSystem.unshift(idx);\n        usedTokens += tokenCounts[idx];\n      }\n    }\n\n    // Build result preserving original order\n    const includedSet = new Set([...systemIndices, ...includedNonSystem]);\n    const result: ChatMessage[] = [];\n    let resultTokens = 0;\n    for (let i = 0; i < messages.length; i++) {\n      if (includedSet.has(i)) {\n        result.push(messages[i]);\n        resultTokens += tokenCounts[i];\n      }\n    }\n\n    return {\n      messages: result,\n      totalTokens: resultTokens,\n      removedCount: messages.length - result.length,\n      wasTruncated: true,\n    };\n  }\n\n  /**\n   * Sliding window: keeps the most recent messages that fit within budget.\n   */\n  private slidingWindow(\n    messages: readonly ChatMessage[],\n    tokenCounts: number[],\n    budget: number,\n  ): ContextWindowResult {\n    const result: ChatMessage[] = [];\n    let usedTokens = 0;\n\n    // Walk from newest to oldest\n    for (let i = messages.length - 1; i >= 0; i--) {\n      if (usedTokens + tokenCounts[i] <= budget) {\n        result.unshift(messages[i]);\n        usedTokens += tokenCounts[i];\n      } else {\n        break;\n      }\n    }\n\n    return {\n      messages: result,\n      totalTokens: usedTokens,\n      removedCount: messages.length - result.length,\n      wasTruncated: true,\n    };\n  }\n\n  /**\n   * Summarize placeholder: replaces truncated messages with a placeholder,\n   * preserving system messages and recent context.\n   */\n  private summarizePlaceholder(\n    messages: readonly ChatMessage[],\n    tokenCounts: number[],\n    budget: number,\n  ): ContextWindowResult {\n    // First, identify system messages and recent messages\n    const systemMessages: { msg: ChatMessage; tokens: number }[] = [];\n    const nonSystem: { msg: ChatMessage; tokens: number; idx: number }[] = [];\n\n    for (let i = 0; i < messages.length; i++) {\n      if (messages[i].role === \"system\") {\n        systemMessages.push({ msg: messages[i], tokens: tokenCounts[i] });\n      } else {\n        nonSystem.push({ msg: messages[i], tokens: tokenCounts[i], idx: i });\n      }\n    }\n\n    // System message cost\n    let usedTokens = systemMessages.reduce((s, m) => s + m.tokens, 0);\n\n    // Placeholder costs ~20 tokens\n    const placeholderTokens = 20;\n    usedTokens += placeholderTokens;\n\n    // Fit recent non-system messages from newest\n    const recentKept: typeof nonSystem = [];\n    for (let i = nonSystem.length - 1; i >= 0; i--) {\n      if (usedTokens + nonSystem[i].tokens <= budget) {\n        recentKept.unshift(nonSystem[i]);\n        usedTokens += nonSystem[i].tokens;\n      } else {\n        break;\n      }\n    }\n\n    const removedCount =\n      messages.length -\n      systemMessages.length -\n      recentKept.length;\n\n    // Build result: system messages, placeholder, recent messages\n    const result: ChatMessage[] = [];\n\n    // System messages first\n    for (const sm of systemMessages) {\n      result.push(sm.msg);\n    }\n\n    // Placeholder if messages were removed\n    if (removedCount > 0) {\n      result.push({\n        id: \"context-placeholder\" as ChatMessage[\"id\"],\n        role: \"system\",\n        parts: [{ type: \"text\", text: `[${removedCount} earlier message${removedCount === 1 ? \"\" : \"s\"} omitted for context window]`, status: \"complete\" as const }],\n        metadata: { isSummary: true },\n        createdAt: new Date().toISOString(),\n        status: \"complete\",\n      });\n    }\n\n    // Recent messages\n    for (const m of recentKept) {\n      result.push(m.msg);\n    }\n\n    return {\n      messages: result,\n      totalTokens: usedTokens,\n      removedCount,\n      wasTruncated: true,\n    };\n  }\n}\n"]}