@pipedog/core 0.0.0 → 0.2.0

package/dist/processors/aggregators/llm_context_summarizer.d.ts

@@ -0,0 +1,219 @@
+import { type Frame, LLMContextSummaryRequestFrame, LLMContextSummaryResultFrame, LLMFullResponseStartFrame, LLMSummarizeContextFrame } from "../../frames/frames.ts";
+import type { LLMService } from "../../services/llm_service.ts";
+import { BaseObject } from "../../utils/base_object.ts";
+import { LLMAutoContextSummarizationConfig, type LLMContextSummaryConfig } from "../../utils/context/llm_context_summarization.ts";
+import type { FrameProcessorSetup } from "../frame_processor.ts";
+import { type LLMContext, type LLMContextMessage } from "./llm_context.ts";
+export interface SummaryAppliedEventOptions {
+    /** Number of messages before summarization. */
+    originalMessageCount: number;
+    /** Number of messages after summarization. */
+    newMessageCount: number;
+    /** Number of messages that were compressed into the summary. */
+    summarizedMessageCount: number;
+    /** Number of recent messages preserved uncompressed. */
+    preservedMessageCount: number;
+}
+/**
+ * Event data emitted when context summarization completes successfully.
+ *
+ * - `originalMessageCount`: Number of messages before summarization.
+ * - `newMessageCount`: Number of messages after summarization.
+ * - `summarizedMessageCount`: Number of messages that were compressed into the summary.
+ * - `preservedMessageCount`: Number of recent messages preserved uncompressed.
+ */
+export declare class SummaryAppliedEvent {
+    originalMessageCount: number;
+    newMessageCount: number;
+    summarizedMessageCount: number;
+    preservedMessageCount: number;
+    constructor(options: SummaryAppliedEventOptions);
+}
+export interface LLMContextSummarizerOptions {
+    /** The LLM context to monitor and summarize. */
+    context: LLMContext;
+    /**
+     * Auto-summarization configuration controlling both trigger thresholds and
+     * default summary generation parameters. If null, uses default
+     * `LLMAutoContextSummarizationConfig` values.
+     */
+    config?: LLMAutoContextSummarizationConfig | null;
+    /**
+     * Whether to automatically trigger summarization when thresholds are
+     * reached. When false, summarization only happens when an
+     * `LLMSummarizeContextFrame` is pushed into the pipeline. Defaults to true.
+     */
+    autoTrigger?: boolean;
+}
+/**
+ * Summarizer for managing LLM context summarization.
+ *
+ * This class manages context summarization, either automatically when token
+ * or message limits are reached, or on-demand when an
+ * `LLMSummarizeContextFrame` is received. It monitors the LLM context size,
+ * triggers summarization requests, and applies the results to compress
+ * conversation history.
+ *
+ * When `autoTrigger` is true (the default), summarization is triggered
+ * automatically based on the configured thresholds in
+ * `LLMAutoContextSummarizationConfig`. When `autoTrigger` is false, threshold
+ * checks are skipped and summarization only happens when an
+ * `LLMSummarizeContextFrame` is explicitly pushed into the pipeline.
+ *
+ * Both modes can coexist: set `autoTrigger: true` and also push
+ * `LLMSummarizeContextFrame` at any time to force an immediate summarization
+ * (subject to the `_summarizationInProgress` guard).
+ *
+ * Event handlers available:
+ *
+ * - on_request_summarization: Emitted when summarization should be triggered.
+ *   The aggregator should broadcast this frame to the LLM service.
+ * - on_summary_applied: Emitted after a summary has been successfully applied
+ *   to the context. Receives a SummaryAppliedEvent with metrics about the
+ *   compression.
+ *
+ * @example
+ * summarizer.eventHandler("on_request_summarization", async (summarizer, frame: LLMContextSummaryRequestFrame) => {
+ *   await aggregator.broadcastFrame(LLMContextSummaryRequestFrame, {
+ *     requestId: frame.requestId,
+ *     context: frame.context,
+ *     // ...
+ *   });
+ * });
+ *
+ * summarizer.eventHandler("on_summary_applied", async (summarizer, event: SummaryAppliedEvent) => {
+ *   logger.info(`Compressed ${event.originalMessageCount} -> ${event.newMessageCount} messages`);
+ * });
+ */
+export declare class LLMContextSummarizer extends BaseObject {
+    protected _context: LLMContext;
+    protected _autoConfig: LLMAutoContextSummarizationConfig;
+    protected _autoTrigger: boolean;
+    protected _summarizationInProgress: boolean;
+    protected _pendingSummaryRequestId: string | null;
+    protected _summaryMessage: LLMContextMessage | null;
+    /** Initialize the context summarizer. */
+    constructor(options: LLMContextSummarizerOptions);
+    /**
+     * Return 1 if the first message is a system message, else 0.
+     *
+     * Mirrors the messages[0] system-message check used elsewhere when
+     * selecting messages to summarize and when reconstructing the context after
+     * a summary, so the message count stays consistent with what those steps
+     * actually preserve.
+     *
+     * @param messages The context messages to check.
+     * @returns 1 if `messages[0]` is a system message, otherwise 0.
+     */
+    protected static _leadingSystemMessageCount(messages: LLMContextMessage[]): number;
+    /**
+     * Count the conversation messages added since the last applied summary.
+     *
+     * Excludes the leading system message, if present, and the summary message
+     * injected by the last summary while it still sits directly after it. Both
+     * are computed from the current context so that replacing the messages (for
+     * example, resetting the conversation) resets the count as well.
+     *
+     * @returns The number of messages that count toward `maxUnsummarizedMessages`.
+     */
+    protected _messagesSinceSummary(): number;
+    /**
+     * Set up the summarizer.
+     *
+     * @param setup Configuration object containing setup parameters.
+     */
+    setup(setup: FrameProcessorSetup | unknown): Promise<void>;
+    /** Cleanup the summarizer. */
+    cleanup(): Promise<void>;
+    /**
+     * Process an incoming frame to detect when summarization is needed.
+     *
+     * @param frame The frame to be processed.
+     */
+    processFrame(frame: Frame): Promise<void>;
+    /**
+     * Handle LLM response start to check if summarization is needed.
+     *
+     * @param frame The LLM response start frame.
+     */
+    protected _handleLlmResponseStart(_frame: LLMFullResponseStartFrame): Promise<void>;
+    /**
+     * Handle an explicit on-demand summarization request.
+     *
+     * Reuses the same `_requestSummarization()` code path as auto mode, so
+     * bookkeeping (`_summarizationInProgress`, `_pendingSummaryRequestId`) is
+     * always updated correctly.
+     *
+     * @param frame The manual summarization request frame, optionally carrying
+     *   a per-request LLMContextSummaryConfig.
+     */
+    protected _handleManualSummarizationRequest(frame: LLMSummarizeContextFrame): Promise<void>;
+    /** Handle interruption by canceling summarization in progress. */
+    protected _handleInterruption(): Promise<void>;
+    /** Cancel pending summarization. */
+    protected _clearSummarizationState(): Promise<void>;
+    /**
+     * Determine if context summarization should be triggered.
+     *
+     * Evaluates whether the current context has reached either the token
+     * threshold or message count threshold that warrants compression. Either
+     * threshold can be null to disable that check; at least one must be set
+     * (enforced at config construction time).
+     *
+     * @returns True when `autoTrigger` is enabled, no summarization is in
+     *   progress, and either the token count exceeds `maxContextTokens` or the
+     *   message count since the last summary exceeds `maxUnsummarizedMessages`
+     *   (whichever of the two is set).
+     */
+    protected _shouldSummarize(): boolean;
+    /**
+     * Request context summarization from LLM service.
+     *
+     * Creates a summarization request frame and either handles it directly
+     * using a dedicated LLM (if configured) or emits it via event handler for
+     * the pipeline's primary LLM. Tracks the request ID to match async
+     * responses and prevent race conditions.
+     *
+     * @param configOverride Optional per-request summary configuration. If
+     *   provided, overrides the default summary generation settings from
+     *   `this._autoConfig.summaryConfig`.
+     */
+    protected _requestSummarization(configOverride?: LLMContextSummaryConfig | null): Promise<void>;
+    /**
+     * Generate summary using a dedicated LLM service.
+     *
+     * Calls the dedicated LLM's `_generateSummary` directly and feeds the
+     * result back through `_handleSummaryResult`, bypassing the pipeline.
+     *
+     * @param llm The dedicated LLM service to use for summarization.
+     * @param frame The summarization request frame.
+     */
+    protected _generateSummaryWithDedicatedLlm(llm: LLMService, frame: LLMContextSummaryRequestFrame): Promise<void>;
+    /**
+     * Handle context summarization result from LLM service.
+     *
+     * Processes the summary result by validating the request ID, checking for
+     * errors, validating context state, and applying the summary.
+     *
+     * @param frame The summary result frame containing the generated summary.
+     */
+    protected _handleSummaryResult(frame: LLMContextSummaryResultFrame): Promise<void>;
+    /**
+     * Validate that context state is still valid for applying summary.
+     *
+     * @param lastSummarizedIndex The index of the last summarized message.
+     * @returns True if the context state is still consistent with the summary.
+     */
+    protected _validateSummaryContext(lastSummarizedIndex: number): boolean;
+    /**
+     * Apply summary to compress the conversation context.
+     *
+     * Reconstructs the context with:
+     * [first_system_message] + [summary_message] + [recent_messages]
+     *
+     * @param summary The generated summary text.
+     * @param lastSummarizedIndex Index of the last message that was summarized.
+     */
+    protected _applySummary(summary: string, lastSummarizedIndex: number): Promise<void>;
+}
+//# sourceMappingURL=llm_context_summarizer.d.ts.map

package/dist/processors/aggregators/llm_context_summarizer.js

@@ -0,0 +1,447 @@
+// Copyright (c) 2024-2026, Daily
+// Copyright (c) 2026, pipedog contributors
+// SPDX-License-Identifier: BSD-2-Clause
+//
+// Ported from pipecat: src/pipecat/processors/aggregators/llm_context_summarizer.py
+/** This module defines a summarizer for managing LLM context summarization. */
+import { rethrowIfCancelled, TimeoutError, waitFor } from "@pipedog/asyncio";
+import { InterruptionFrame, LLMContextSummaryRequestFrame, LLMContextSummaryResultFrame, LLMFullResponseStartFrame, LLMSummarizeContextFrame, } from "../../frames/frames.js";
+import { BaseObject } from "../../utils/base_object.js";
+import { DEFAULT_SUMMARIZATION_TIMEOUT, LLMAutoContextSummarizationConfig, LLMContextSummarizationUtil, } from "../../utils/context/llm_context_summarization.js";
+import { logger } from "../../utils/logger.js";
+import { LLMSpecificMessage } from "./llm_context.js";
+/**
+ * Event data emitted when context summarization completes successfully.
+ *
+ * - `originalMessageCount`: Number of messages before summarization.
+ * - `newMessageCount`: Number of messages after summarization.
+ * - `summarizedMessageCount`: Number of messages that were compressed into the summary.
+ * - `preservedMessageCount`: Number of recent messages preserved uncompressed.
+ */
+export class SummaryAppliedEvent {
+    originalMessageCount;
+    newMessageCount;
+    summarizedMessageCount;
+    preservedMessageCount;
+    constructor(options) {
+        this.originalMessageCount = options.originalMessageCount;
+        this.newMessageCount = options.newMessageCount;
+        this.summarizedMessageCount = options.summarizedMessageCount;
+        this.preservedMessageCount = options.preservedMessageCount;
+    }
+}
+/**
+ * Summarizer for managing LLM context summarization.
+ *
+ * This class manages context summarization, either automatically when token
+ * or message limits are reached, or on-demand when an
+ * `LLMSummarizeContextFrame` is received. It monitors the LLM context size,
+ * triggers summarization requests, and applies the results to compress
+ * conversation history.
+ *
+ * When `autoTrigger` is true (the default), summarization is triggered
+ * automatically based on the configured thresholds in
+ * `LLMAutoContextSummarizationConfig`. When `autoTrigger` is false, threshold
+ * checks are skipped and summarization only happens when an
+ * `LLMSummarizeContextFrame` is explicitly pushed into the pipeline.
+ *
+ * Both modes can coexist: set `autoTrigger: true` and also push
+ * `LLMSummarizeContextFrame` at any time to force an immediate summarization
+ * (subject to the `_summarizationInProgress` guard).
+ *
+ * Event handlers available:
+ *
+ * - on_request_summarization: Emitted when summarization should be triggered.
+ *   The aggregator should broadcast this frame to the LLM service.
+ * - on_summary_applied: Emitted after a summary has been successfully applied
+ *   to the context. Receives a SummaryAppliedEvent with metrics about the
+ *   compression.
+ *
+ * @example
+ * summarizer.eventHandler("on_request_summarization", async (summarizer, frame: LLMContextSummaryRequestFrame) => {
+ *   await aggregator.broadcastFrame(LLMContextSummaryRequestFrame, {
+ *     requestId: frame.requestId,
+ *     context: frame.context,
+ *     // ...
+ *   });
+ * });
+ *
+ * summarizer.eventHandler("on_summary_applied", async (summarizer, event: SummaryAppliedEvent) => {
+ *   logger.info(`Compressed ${event.originalMessageCount} -> ${event.newMessageCount} messages`);
+ * });
+ */
+export class LLMContextSummarizer extends BaseObject {
+    _context;
+    _autoConfig;
+    _autoTrigger;
+    _summarizationInProgress = false;
+    _pendingSummaryRequestId = null;
+    // The summary message injected by the last applied summary, tracked by
+    // identity so the message-count trigger can exclude it only while it is
+    // still at the head of the context.
+    _summaryMessage = null;
+    /** Initialize the context summarizer. */
+    constructor(options) {
+        super();
+        this._context = options.context;
+        this._autoConfig = options.config || new LLMAutoContextSummarizationConfig();
+        this._autoTrigger = options.autoTrigger ?? true;
+        this._registerEventHandler("on_request_summarization", true);
+        this._registerEventHandler("on_summary_applied");
+    }
+    /**
+     * Return 1 if the first message is a system message, else 0.
+     *
+     * Mirrors the messages[0] system-message check used elsewhere when
+     * selecting messages to summarize and when reconstructing the context after
+     * a summary, so the message count stays consistent with what those steps
+     * actually preserve.
+     *
+     * @param messages The context messages to check.
+     * @returns 1 if `messages[0]` is a system message, otherwise 0.
+     */
+    static _leadingSystemMessageCount(messages) {
+        const first = messages[0];
+        if (first !== undefined && !(first instanceof LLMSpecificMessage) && first.role === "system") {
+            return 1;
+        }
+        return 0;
+    }
+    /**
+     * Count the conversation messages added since the last applied summary.
+     *
+     * Excludes the leading system message, if present, and the summary message
+     * injected by the last summary while it still sits directly after it. Both
+     * are computed from the current context so that replacing the messages (for
+     * example, resetting the conversation) resets the count as well.
+     *
+     * @returns The number of messages that count toward `maxUnsummarizedMessages`.
+     */
+    _messagesSinceSummary() {
+        const messages = this._context.messages;
+        let excluded = LLMContextSummarizer._leadingSystemMessageCount(messages);
+        if (this._summaryMessage !== null &&
+            messages.length > excluded &&
+            messages[excluded] === this._summaryMessage) {
+            excluded += 1;
+        }
+        return messages.length - excluded;
+    }
+    /**
+     * Set up the summarizer.
+     *
+     * @param setup Configuration object containing setup parameters.
+     */
+    async setup(setup) {
+        await super.setup(setup.taskManager);
+    }
+    /** Cleanup the summarizer. */
+    async cleanup() {
+        await super.cleanup();
+        await this._clearSummarizationState();
+    }
+    /**
+     * Process an incoming frame to detect when summarization is needed.
+     *
+     * @param frame The frame to be processed.
+     */
+    async processFrame(frame) {
+        if (frame instanceof LLMFullResponseStartFrame) {
+            await this._handleLlmResponseStart(frame);
+        }
+        else if (frame instanceof LLMSummarizeContextFrame) {
+            await this._handleManualSummarizationRequest(frame);
+        }
+        else if (frame instanceof LLMContextSummaryResultFrame) {
+            await this._handleSummaryResult(frame);
+        }
+        else if (frame instanceof InterruptionFrame) {
+            await this._handleInterruption();
+        }
+    }
+    /**
+     * Handle LLM response start to check if summarization is needed.
+     *
+     * @param frame The LLM response start frame.
+     */
+    async _handleLlmResponseStart(_frame) {
+        if (this._shouldSummarize()) {
+            await this._requestSummarization();
+        }
+    }
+    /**
+     * Handle an explicit on-demand summarization request.
+     *
+     * Reuses the same `_requestSummarization()` code path as auto mode, so
+     * bookkeeping (`_summarizationInProgress`, `_pendingSummaryRequestId`) is
+     * always updated correctly.
+     *
+     * @param frame The manual summarization request frame, optionally carrying
+     *   a per-request LLMContextSummaryConfig.
+     */
+    async _handleManualSummarizationRequest(frame) {
+        if (this._summarizationInProgress) {
+            logger.debug(`${this}: Summarization already in progress, ignoring manual request`);
+            return;
+        }
+        await this._requestSummarization(frame.config);
+    }
+    /** Handle interruption by canceling summarization in progress. */
+    async _handleInterruption() {
+        // Reset summarization state to allow new requests. This is necessary
+        // because the request frame (LLMContextSummaryRequestFrame) may have been
+        // cancelled during interruption. We preserve _pendingSummaryRequestId to
+        // handle the response frame (LLMContextSummaryResultFrame), which is
+        // uninterruptible and will still be delivered.
+        this._summarizationInProgress = false;
+    }
+    /** Cancel pending summarization. */
+    async _clearSummarizationState() {
+        if (this._summarizationInProgress) {
+            logger.debug(`${this}: Clearing pending summarization`);
+            this._summarizationInProgress = false;
+            this._pendingSummaryRequestId = null;
+        }
+    }
+    /**
+     * Determine if context summarization should be triggered.
+     *
+     * Evaluates whether the current context has reached either the token
+     * threshold or message count threshold that warrants compression. Either
+     * threshold can be null to disable that check; at least one must be set
+     * (enforced at config construction time).
+     *
+     * @returns True when `autoTrigger` is enabled, no summarization is in
+     *   progress, and either the token count exceeds `maxContextTokens` or the
+     *   message count since the last summary exceeds `maxUnsummarizedMessages`
+     *   (whichever of the two is set).
+     */
+    _shouldSummarize() {
+        logger.trace(`${this}: Checking if context summarization is needed`);
+        if (!this._autoTrigger) {
+            return false;
+        }
+        if (this._summarizationInProgress) {
+            logger.debug(`${this}: Summarization already in progress`);
+            return false;
+        }
+        // Estimate tokens in context
+        const totalTokens = LLMContextSummarizationUtil.estimateContextTokens(this._context);
+        const numMessages = this._context.messages.length;
+        // Check if we've reached the token limit
+        const tokenLimit = this._autoConfig.maxContextTokens;
+        const tokenLimitExceeded = tokenLimit !== null && totalTokens >= tokenLimit;
+        // Check if we've exceeded max unsummarized messages
+        const messagesSinceSummary = this._messagesSinceSummary();
+        const messageThreshold = this._autoConfig.maxUnsummarizedMessages;
+        const messageThresholdExceeded = messageThreshold !== null && messagesSinceSummary >= messageThreshold;
+        logger.trace(`${this}: Context has ${numMessages} messages, ` +
+            `~${totalTokens} tokens (limit: ${tokenLimit !== null ? tokenLimit : "disabled"}), ` +
+            `${messagesSinceSummary} messages since last summary ` +
+            `(message threshold: ${messageThreshold !== null ? messageThreshold : "disabled"})`);
+        // Trigger if either limit is exceeded
+        if (!tokenLimitExceeded && !messageThresholdExceeded) {
+            logger.trace(`${this}: Neither token limit nor message threshold exceeded, skipping summarization`);
+            return false;
+        }
+        const reason = [];
+        if (tokenLimitExceeded) {
+            reason.push(`~${totalTokens} tokens (>=${tokenLimit} limit)`);
+        }
+        if (messageThresholdExceeded) {
+            reason.push(`${messagesSinceSummary} messages (>=${messageThreshold} threshold)`);
+        }
+        logger.debug(`${this}: ✓ Summarization needed - ${reason.join(", ")}`);
+        return true;
+    }
+    /**
+     * Request context summarization from LLM service.
+     *
+     * Creates a summarization request frame and either handles it directly
+     * using a dedicated LLM (if configured) or emits it via event handler for
+     * the pipeline's primary LLM. Tracks the request ID to match async
+     * responses and prevent race conditions.
+     *
+     * @param configOverride Optional per-request summary configuration. If
+     *   provided, overrides the default summary generation settings from
+     *   `this._autoConfig.summaryConfig`.
+     */
+    async _requestSummarization(configOverride = null) {
+        // Generate unique request ID
+        const requestId = crypto.randomUUID();
+        const summaryConfig = configOverride || this._autoConfig.summaryConfig;
+        // Mark summarization in progress
+        this._summarizationInProgress = true;
+        this._pendingSummaryRequestId = requestId;
+        logger.debug(`${this}: Sending summarization request (request_id=${requestId})`);
+        // Create the request frame
+        const requestFrame = new LLMContextSummaryRequestFrame({
+            requestId,
+            context: this._context,
+            minMessagesToKeep: summaryConfig.minMessagesAfterSummary,
+            targetContextTokens: summaryConfig.targetContextTokens,
+            summarizationPrompt: summaryConfig.summaryPrompt,
+            summarizationTimeout: summaryConfig.summarizationTimeout,
+        });
+        if (summaryConfig.llm) {
+            // Use dedicated LLM directly — no need to involve the pipeline
+            const llm = summaryConfig.llm;
+            this.taskManager.createTask(() => this._generateSummaryWithDedicatedLlm(llm, requestFrame), `${this}-dedicated-llm-summary`);
+        }
+        else {
+            // Emit event for aggregator to broadcast to the pipeline LLM
+            await this._callEventHandler("on_request_summarization", requestFrame);
+        }
+    }
+    /**
+     * Generate summary using a dedicated LLM service.
+     *
+     * Calls the dedicated LLM's `_generateSummary` directly and feeds the
+     * result back through `_handleSummaryResult`, bypassing the pipeline.
+     *
+     * @param llm The dedicated LLM service to use for summarization.
+     * @param frame The summarization request frame.
+     */
+    async _generateSummaryWithDedicatedLlm(llm, frame) {
+        // Python `or`: a 0 timeout also falls back to the default.
+        const timeout = frame.summarizationTimeout || DEFAULT_SUMMARIZATION_TIMEOUT;
+        let resultFrame;
+        try {
+            const generator = llm;
+            const [summary, lastIndex] = await waitFor(() => generator._generateSummary(frame), timeout);
+            resultFrame = new LLMContextSummaryResultFrame({
+                requestId: frame.requestId,
+                summary,
+                lastSummarizedIndex: lastIndex,
+            });
+        }
+        catch (e) {
+            rethrowIfCancelled(e);
+            let error;
+            if (e instanceof TimeoutError) {
+                error = `Context summarization timed out after ${timeout}s`;
+            }
+            else {
+                error = `Error generating context summary: ${e instanceof Error ? e.message : String(e)}`;
+            }
+            logger.error(`${this}: ${error}`);
+            resultFrame = new LLMContextSummaryResultFrame({
+                requestId: frame.requestId,
+                summary: "",
+                lastSummarizedIndex: -1,
+                error,
+            });
+        }
+        await this._handleSummaryResult(resultFrame);
+    }
+    /**
+     * Handle context summarization result from LLM service.
+     *
+     * Processes the summary result by validating the request ID, checking for
+     * errors, validating context state, and applying the summary.
+     *
+     * @param frame The summary result frame containing the generated summary.
+     */
+    async _handleSummaryResult(frame) {
+        logger.debug(`${this}: Received summary result (request_id=${frame.requestId})`);
+        // Check if this is the result we're waiting for. Both auto and manual
+        // summarization set _pendingSummaryRequestId via _requestSummarization(),
+        // so this check always applies.
+        if (frame.requestId !== this._pendingSummaryRequestId) {
+            logger.debug(`${this}: Ignoring stale summary result (request_id=${frame.requestId})`);
+            return;
+        }
+        // Clear pending state
+        await this._clearSummarizationState();
+        // Check for errors
+        if (frame.error) {
+            logger.error(`${this}: Context summarization failed: ${frame.error}`);
+            return;
+        }
+        // Validate context state
+        if (!this._validateSummaryContext(frame.lastSummarizedIndex)) {
+            logger.warning(`${this}: Context state changed, skipping summary application`);
+            return;
+        }
+        // Apply summary
+        await this._applySummary(frame.summary, frame.lastSummarizedIndex);
+    }
+    /**
+     * Validate that context state is still valid for applying summary.
+     *
+     * @param lastSummarizedIndex The index of the last summarized message.
+     * @returns True if the context state is still consistent with the summary.
+     */
+    _validateSummaryContext(lastSummarizedIndex) {
+        if (lastSummarizedIndex < 0) {
+            return false;
+        }
+        // Check if we still have enough messages
+        if (lastSummarizedIndex >= this._context.messages.length) {
+            return false;
+        }
+        const minKeep = this._autoConfig.summaryConfig.minMessagesAfterSummary;
+        const remaining = this._context.messages.length - 1 - lastSummarizedIndex;
+        if (remaining < minKeep) {
+            return false;
+        }
+        return true;
+    }
+    /**
+     * Apply summary to compress the conversation context.
+     *
+     * Reconstructs the context with:
+     * [first_system_message] + [summary_message] + [recent_messages]
+     *
+     * @param summary The generated summary text.
+     * @param lastSummarizedIndex Index of the last message that was summarized.
+     */
+    async _applySummary(summary, lastSummarizedIndex) {
+        const config = this._autoConfig.summaryConfig;
+        const messages = this._context.messages;
+        // Preserve the first message if it is a system message (initial system
+        // prompt). Only messages[0] is treated as the system preamble — system
+        // messages at other positions are mid-conversation injections and are not
+        // preserved separately (they will be part of the summary or the recent
+        // messages).
+        const numSystemPreserved = LLMContextSummarizer._leadingSystemMessageCount(messages);
+        const firstSystemMsg = numSystemPreserved ? messages[0] : null;
+        // Get recent messages to keep
+        const recentMessages = messages.slice(lastSummarizedIndex + 1);
+        // Create summary message as a user message (the summary is context
+        // provided *to* the assistant, not something the assistant said)
+        const summaryContent = formatSummaryTemplate(config.summaryMessageTemplate, summary);
+        const summaryMessage = { role: "user", content: summaryContent };
+        // Reconstruct context
+        const newMessages = [];
+        if (firstSystemMsg) {
+            newMessages.push(firstSystemMsg);
+        }
+        newMessages.push(summaryMessage);
+        newMessages.push(...recentMessages);
+        // Update context
+        const originalMessageCount = messages.length;
+        this._context.setMessages(newMessages);
+        this._summaryMessage = summaryMessage;
+        // Messages actually summarized = index range minus the preserved system message
+        const summarizedCount = lastSummarizedIndex + 1 - numSystemPreserved;
+        logger.info(`${this}: Applied context summary, compressed ${summarizedCount} messages ` +
+            `into summary. Context now has ${newMessages.length} messages (was ${originalMessageCount})`);
+        // Emit event for observability
+        const event = new SummaryAppliedEvent({
+            originalMessageCount,
+            newMessageCount: newMessages.length,
+            summarizedMessageCount: summarizedCount,
+            preservedMessageCount: recentMessages.length + numSystemPreserved,
+        });
+        await this._callEventHandler("on_summary_applied", event);
+    }
+}
+/**
+ * Python `template.format(summary=summary)`: replaces `{summary}` and
+ * unescapes doubled braces.
+ */
+function formatSummaryTemplate(template, summary) {
+    return template.replace(/\{\{|\}\}|\{summary\}/g, (m) => m === "{{" ? "{" : m === "}}" ? "}" : summary);
+}
+//# sourceMappingURL=llm_context_summarizer.js.map