@juspay/neurolink 9.5.2 → 9.6.0

package/dist/cli/commands/setup-google-ai.d.ts

@@ -8,9 +8,9 @@
  *
  * Follows the same UX patterns as setup-openai and setup-gcp
  */
-interface GoogleAISetupArgv {
+type GoogleAISetupArgv = {
     check?: boolean;
     nonInteractive?: boolean;
-}
+};
 export declare function handleGoogleAISetup(argv: GoogleAISetupArgv): Promise<void>;
 export {};

package/dist/cli/commands/setup-huggingface.d.ts

@@ -1,8 +1,8 @@
 import type { Arguments, CommandBuilder } from "yargs";
-interface SetupHuggingFaceArgs {
+type SetupHuggingFaceArgs = {
     check?: boolean;
     "non-interactive"?: boolean;
-}
+};
 export declare const handleHuggingFaceSetup: (argv: Arguments<SetupHuggingFaceArgs>) => Promise<void>;
 export declare const setupHuggingFaceBuilder: CommandBuilder;
 export {};

package/dist/cli/commands/setup-mistral.d.ts

@@ -1,8 +1,8 @@
 import type { Arguments, CommandBuilder } from "yargs";
-interface SetupMistralArgs {
+type SetupMistralArgs = {
     check?: boolean;
     "non-interactive"?: boolean;
-}
+};
 export declare const handleMistralSetup: (argv: Arguments<SetupMistralArgs>) => Promise<void>;
 export declare const setupMistralBuilder: CommandBuilder;
 export {};

package/dist/cli/commands/setup-openai.d.ts

@@ -8,9 +8,9 @@
  *
  * Follows the same UX patterns as setup-gcp and setup-bedrock
  */
-interface OpenAISetupArgv {
+type OpenAISetupArgv = {
     check?: boolean;
     nonInteractive?: boolean;
-}
+};
 export declare function handleOpenAISetup(argv: OpenAISetupArgv): Promise<void>;
 export {};

package/dist/cli/commands/setup.d.ts

@@ -5,13 +5,13 @@
  * Provides a beautiful welcome experience for new users and guided
  * provider selection, while delegating to existing setup commands.
  */
-interface SetupArgs {
+type SetupArgs = {
     provider?: string;
     list?: boolean;
     status?: boolean;
     interactive?: boolean;
     help?: boolean;
-}
+};
 /**
  * Main setup command handler
  */

package/dist/cli/factories/commandFactory.js

@@ -6,7 +6,7 @@ import { ModelResolver } from "../../lib/models/modelResolver.js";
 import { globalSession } from "../../lib/session/globalSessionState.js";
 // Use TokenUsage from standard types - no local interface needed
 import { ContextFactory, } from "../../lib/types/contextTypes.js";
-import { checkRedisAvailability } from "../../lib/utils/conversationMemoryUtils.js";
+import { checkRedisAvailability } from "../../lib/utils/conversationMemory.js";
 import { normalizeEvaluationData } from "../../lib/utils/evaluationUtils.js";
 import { logger } from "../../lib/utils/logger.js";
 import { createThinkingConfigFromRecord } from "../../lib/utils/thinkingConfig.js";
@@ -1152,6 +1152,16 @@ export class CLICommandFactory {
                     description: "List available conversations and exit",
                     alias: "l",
                 },
+                "compact-threshold": {
+                    describe: "Context compaction trigger threshold (0.0-1.0)",
+                    type: "number",
+                    default: 0.8,
+                },
+                "disable-compaction": {
+                    describe: "Disable automatic context compaction",
+                    type: "boolean",
+                    default: false,
+                },
             })
                 .example("$0 loop", "Start interactive session with conversation selection")
                 .example("$0 loop --new", "Force start new conversation")
@@ -1165,7 +1175,7 @@ export class CLICommandFactory {
                     return;
                 }
                 let conversationMemoryConfig;
-                const { enableConversationMemory, maxSessions, maxTurnsPerSession, autoRedis, listConversations, } = argv;
+                const { enableConversationMemory, maxSessions, maxTurnsPerSession, autoRedis, listConversations, compactThreshold, disableCompaction, } = argv;
                 if (enableConversationMemory) {
                     let storageType = "memory";
                     if (autoRedis) {
@@ -1188,6 +1198,10 @@ export class CLICommandFactory {
                         enabled: true,
                         maxSessions: maxSessions,
                         maxTurnsPerSession: maxTurnsPerSession,
+                        contextCompaction: {
+                            enabled: !disableCompaction,
+                            threshold: compactThreshold,
+                        },
                     };
                 }
                 // Handle --list-conversations option

package/dist/cli/loop/optionsSchema.d.ts

@@ -1,8 +1,8 @@
-import type { TextGenerationOptions } from "../../lib/types/generateTypes.js";
 import type { OptionSchema } from "../../lib/types/cli.js";
+import type { TextGenerationOptions } from "../../lib/types/generateTypes.js";
 /**
  * Master schema for all text generation options.
  * This object provides metadata for validation and help text in the CLI loop.
  * It is derived from the main TextGenerationOptions interface to ensure consistency.
  */
-export declare const textGenerationOptionsSchema: Record<keyof Omit<TextGenerationOptions, "prompt" | "input" | "schema" | "tools" | "context" | "conversationHistory" | "conversationMessages" | "conversationMemoryConfig" | "originalPrompt" | "middleware" | "expectedOutcome" | "evaluationCriteria" | "region" | "csvOptions" | "tts" | "thinkingConfig">, OptionSchema>;
+export declare const textGenerationOptionsSchema: Record<keyof Omit<TextGenerationOptions, "prompt" | "input" | "schema" | "tools" | "context" | "conversationHistory" | "conversationMessages" | "conversationMemoryConfig" | "originalPrompt" | "middleware" | "expectedOutcome" | "evaluationCriteria" | "region" | "csvOptions" | "tts" | "thinkingConfig" | "fileRegistry">, OptionSchema>;

package/dist/cli/loop/session.d.ts

@@ -21,6 +21,10 @@ export declare class LoopSession {
      * Handle conversation selection logic when no direct resume is specified
      */
     private handleConversationSelection;
+    /**
+     * Check context budget and warn if approaching limits.
+     */
+    private checkContextBudgetWarning;
     /**
      * Clean up session resources and connections
      */

package/dist/cli/loop/session.js

@@ -1,12 +1,13 @@
 import chalk from "chalk";
 import readline from "readline";
-import { logger } from "../../lib/utils/logger.js";
+import { checkContextBudget } from "../../lib/context/budgetChecker.js";
+import { NeuroLink } from "../../lib/neurolink.js";
 import { globalSession } from "../../lib/session/globalSessionState.js";
-import { textGenerationOptionsSchema } from "./optionsSchema.js";
+import { logger } from "../../lib/utils/logger.js";
+import { displayConversationPreview, displaySessionMessage, getConversationPreview, loadCommandHistory, parseValue, restoreSessionVariables, saveCommandToHistory, verifyConversationContext, } from "../../lib/utils/loopUtils.js";
 import { handleError } from "../errorHandler.js";
 import { ConversationSelector } from "./conversationSelector.js";
-import { NeuroLink } from "../../lib/neurolink.js";
-import { displaySessionMessage, verifyConversationContext, getConversationPreview, loadCommandHistory, saveCommandToHistory, displayConversationPreview, parseValue, restoreSessionVariables, } from "../../lib/utils/loopUtils.js";
+import { textGenerationOptionsSchema } from "./optionsSchema.js";
 // Banner Art
 const NEUROLINK_BANNER = `
 ▗▖  ▗▖▗▄▄▄▖▗▖ ▗▖▗▄▄▖  ▗▄▖ ▗▖   ▗▄▄▄▖▗▖  ▗▖▗▖ ▗▖
@@ -120,6 +121,8 @@ export class LoopSession {
                 })
                     .exitProcess(false)
                     .parse(processedCommand);
+                // Check context budget after each generation command
+                await this.checkContextBudgetWarning();
             }
             catch (error) {
                 // Handle command execution errors gracefully
@@ -205,6 +208,48 @@ export class LoopSession {
             this.sessionId = globalSession.setLoopSession(this.conversationMemoryConfig);
         }
     }
+    /**
+     * Check context budget and warn if approaching limits.
+     */
+    async checkContextBudgetWarning() {
+        const compactionConfig = this.conversationMemoryConfig?.contextCompaction;
+        if (!compactionConfig?.enabled) {
+            return;
+        }
+        try {
+            const provider = globalSession.getSessionVariable("provider") || "openai";
+            const model = globalSession.getSessionVariable("model");
+            const neurolinkInstance = globalSession.getOrCreateNeuroLink();
+            if (!neurolinkInstance?.conversationMemory || !this.sessionId) {
+                return;
+            }
+            const messages = await neurolinkInstance.conversationMemory.buildContextMessages(this.sessionId);
+            if (!messages || messages.length === 0) {
+                return;
+            }
+            const budgetResult = checkContextBudget({
+                provider,
+                model,
+                conversationMessages: messages.map((m) => ({
+                    role: m.role,
+                    content: m.content,
+                })),
+            });
+            const usagePercent = budgetResult.usageRatio * 100;
+            if (budgetResult.shouldCompact) {
+                logger.always(chalk.yellow(`\n  Context usage: ${usagePercent.toFixed(0)}% of window (${budgetResult.estimatedInputTokens.toLocaleString()} / ${budgetResult.availableInputTokens.toLocaleString()} tokens)`));
+                logger.always(chalk.yellow(`  Auto-compaction will trigger to preserve conversation quality.\n`));
+            }
+            else if (usagePercent > 60) {
+                logger.always(chalk.gray(`  Context: ${usagePercent.toFixed(0)}% used`));
+            }
+        }
+        catch (error) {
+            logger.debug("Context budget check failed", {
+                error: error instanceof Error ? error.message : String(error),
+            });
+        }
+    }
     /**
      * Clean up session resources and connections
      */

package/dist/cli/utils/interactiveSetup.d.ts

@@ -5,7 +5,7 @@
  * credential collection, testing, and environment file management.
  */
 import { AIProviderName } from "../../lib/constants/enums.js";
-export interface ProviderConfig {
+export type ProviderConfig = {
     id: AIProviderName;
     name: string;
     description: string;
@@ -16,9 +16,9 @@ export interface ProviderConfig {
         default?: string;
         optional?: boolean;
     }>;
-}
+};
 export declare const PROVIDER_CONFIGS: ProviderConfig[];
-export interface SetupResult {
+export type SetupResult = {
     selectedProviders: AIProviderName[];
     credentials: Record<string, string>;
     envFileBackup?: string;
@@ -28,7 +28,7 @@ export interface SetupResult {
         error?: string;
         responseTime?: number;
     }>;
-}
+};
 /**
  * Run the interactive setup wizard
  */

package/dist/config/conversationMemory.d.ts

@@ -5,6 +5,8 @@
 import type { ConversationMemoryConfig } from "../types/conversation.js";
 /**
  * Default maximum number of turns per session
+ * @deprecated Use tokenThreshold-based memory management instead.
+ * This constant is kept for backward compatibility with legacy turn-based memory.
  */
 export declare const DEFAULT_MAX_TURNS_PER_SESSION = 50;
 /**

package/dist/config/conversationMemory.js

@@ -4,6 +4,8 @@
  */
 /**
  * Default maximum number of turns per session
+ * @deprecated Use tokenThreshold-based memory management instead.
+ * This constant is kept for backward compatibility with legacy turn-based memory.
  */
 export const DEFAULT_MAX_TURNS_PER_SESSION = 50;
 /**
@@ -77,10 +79,8 @@ export function getConversationMemoryDefaults() {
             : undefined,
         summarizationProvider: process.env.NEUROLINK_SUMMARIZATION_PROVIDER || "vertex",
         summarizationModel: process.env.NEUROLINK_SUMMARIZATION_MODEL || "gemini-2.5-flash",
-        // Deprecated (for backward compatibility)
-        maxTurnsPerSession: Number(process.env.NEUROLINK_MEMORY_MAX_TURNS_PER_SESSION) ||
-            DEFAULT_MAX_TURNS_PER_SESSION,
-        summarizationThresholdTurns: Number(process.env.NEUROLINK_SUMMARIZATION_THRESHOLD_TURNS) || 20,
-        summarizationTargetTurns: Number(process.env.NEUROLINK_SUMMARIZATION_TARGET_TURNS) || 10,
+        // @deprecated — Turn-based fields are superseded by tokenThreshold-based memory.
+        // Kept for backward compatibility; will be removed in a future major version.
+        maxTurnsPerSession: DEFAULT_MAX_TURNS_PER_SESSION,
     };
 }

package/dist/constants/contextWindows.d.ts

@@ -0,0 +1,46 @@
+/**
+ * Context Window Registry
+ *
+ * Accurate per-provider, per-model context window sizes (INPUT token limits).
+ * These are distinct from OUTPUT token limits in tokens.ts.
+ *
+ * Sources:
+ * - Anthropic: https://docs.anthropic.com/en/docs/about-claude/models
+ * - OpenAI: https://platform.openai.com/docs/models
+ * - Google: https://ai.google.dev/gemini-api/docs/models
+ * - Others: Provider documentation as of Feb 2026
+ */
+/** Default context window when provider/model is unknown */
+export declare const DEFAULT_CONTEXT_WINDOW = 128000;
+/** Maximum output reserve when maxTokens not specified */
+export declare const MAX_DEFAULT_OUTPUT_RESERVE = 64000;
+/** Default output reserve ratio (35% of context) */
+export declare const DEFAULT_OUTPUT_RESERVE_RATIO = 0.35;
+/**
+ * Per-provider, per-model context window sizes.
+ * The "_default" key is the fallback for unknown models within a provider.
+ */
+export declare const MODEL_CONTEXT_WINDOWS: Record<string, Record<string, number>>;
+/**
+ * Resolve context window size for a provider/model combination.
+ *
+ * Priority:
+ *  1. Exact model match under provider
+ *  2. Provider's _default
+ *  3. Global DEFAULT_CONTEXT_WINDOW
+ */
+export declare function getContextWindowSize(provider: string, model?: string): number;
+/**
+ * Calculate output token reserve for a given context window.
+ *
+ * @param contextWindow - Total context window size
+ * @param maxTokens - Explicit maxTokens from user config (if set)
+ * @returns Number of tokens reserved for output
+ */
+export declare function getOutputReserve(contextWindow: number, maxTokens?: number): number;
+/**
+ * Calculate available input tokens for a given provider/model.
+ *
+ * available = contextWindow - outputReserve
+ */
+export declare function getAvailableInputTokens(provider: string, model?: string, maxTokens?: number): number;

package/dist/constants/contextWindows.js

@@ -0,0 +1,156 @@
+/**
+ * Context Window Registry
+ *
+ * Accurate per-provider, per-model context window sizes (INPUT token limits).
+ * These are distinct from OUTPUT token limits in tokens.ts.
+ *
+ * Sources:
+ * - Anthropic: https://docs.anthropic.com/en/docs/about-claude/models
+ * - OpenAI: https://platform.openai.com/docs/models
+ * - Google: https://ai.google.dev/gemini-api/docs/models
+ * - Others: Provider documentation as of Feb 2026
+ */
+/** Default context window when provider/model is unknown */
+export const DEFAULT_CONTEXT_WINDOW = 128_000;
+/** Maximum output reserve when maxTokens not specified */
+export const MAX_DEFAULT_OUTPUT_RESERVE = 64_000;
+/** Default output reserve ratio (35% of context) */
+export const DEFAULT_OUTPUT_RESERVE_RATIO = 0.35;
+/**
+ * Per-provider, per-model context window sizes.
+ * The "_default" key is the fallback for unknown models within a provider.
+ */
+export const MODEL_CONTEXT_WINDOWS = {
+    anthropic: {
+        _default: 200_000,
+        "claude-opus-4-20250514": 200_000,
+        "claude-sonnet-4-20250514": 200_000,
+        "claude-3-7-sonnet-20250219": 200_000,
+        "claude-3-5-sonnet-20241022": 200_000,
+        "claude-3-5-haiku-20241022": 200_000,
+        "claude-3-opus-20240229": 200_000,
+        "claude-3-sonnet-20240229": 200_000,
+        "claude-3-haiku-20240307": 200_000,
+    },
+    openai: {
+        _default: 128_000,
+        "gpt-4o": 128_000,
+        "gpt-4o-mini": 128_000,
+        "gpt-4-turbo": 128_000,
+        "gpt-4": 8_192,
+        "gpt-3.5-turbo": 16_385,
+        o1: 200_000,
+        "o1-mini": 128_000,
+        "o1-pro": 200_000,
+        o3: 200_000,
+        "o3-mini": 200_000,
+        "o4-mini": 200_000,
+        "gpt-4.1": 1_047_576,
+        "gpt-4.1-mini": 1_047_576,
+        "gpt-4.1-nano": 1_047_576,
+        "gpt-5": 1_047_576,
+    },
+    "google-ai": {
+        _default: 1_048_576,
+        "gemini-2.5-pro": 1_048_576,
+        "gemini-2.5-flash": 1_048_576,
+        "gemini-2.0-flash": 1_048_576,
+        "gemini-1.5-pro": 2_097_152,
+        "gemini-1.5-flash": 1_048_576,
+        "gemini-3-flash-preview": 1_048_576,
+        "gemini-3-pro-preview": 1_048_576,
+    },
+    vertex: {
+        _default: 1_048_576,
+        "gemini-2.5-pro": 1_048_576,
+        "gemini-2.5-flash": 1_048_576,
+        "gemini-2.0-flash": 1_048_576,
+        "gemini-1.5-pro": 2_097_152,
+        "gemini-1.5-flash": 1_048_576,
+    },
+    bedrock: {
+        _default: 200_000,
+        "anthropic.claude-3-5-sonnet-20241022-v2:0": 200_000,
+        "anthropic.claude-3-5-haiku-20241022-v1:0": 200_000,
+        "anthropic.claude-3-opus-20240229-v1:0": 200_000,
+        "anthropic.claude-3-sonnet-20240229-v1:0": 200_000,
+        "anthropic.claude-3-haiku-20240307-v1:0": 200_000,
+        "amazon.nova-pro-v1:0": 300_000,
+        "amazon.nova-lite-v1:0": 300_000,
+    },
+    azure: {
+        _default: 128_000,
+        "gpt-4o": 128_000,
+        "gpt-4o-mini": 128_000,
+        "gpt-4-turbo": 128_000,
+        "gpt-4": 8_192,
+    },
+    mistral: {
+        _default: 128_000,
+        "mistral-large-latest": 128_000,
+        "mistral-medium-latest": 32_000,
+        "mistral-small-latest": 128_000,
+        "codestral-latest": 256_000,
+    },
+    ollama: {
+        _default: 128_000,
+    },
+    litellm: {
+        _default: 128_000,
+    },
+    huggingface: {
+        _default: 32_000,
+    },
+    sagemaker: {
+        _default: 128_000,
+    },
+};
+/**
+ * Resolve context window size for a provider/model combination.
+ *
+ * Priority:
+ *  1. Exact model match under provider
+ *  2. Provider's _default
+ *  3. Global DEFAULT_CONTEXT_WINDOW
+ */
+export function getContextWindowSize(provider, model) {
+    const providerWindows = MODEL_CONTEXT_WINDOWS[provider];
+    if (!providerWindows) {
+        return DEFAULT_CONTEXT_WINDOW;
+    }
+    if (model && providerWindows[model] !== undefined) {
+        return providerWindows[model];
+    }
+    // Try partial match (model name may be a prefix)
+    if (model) {
+        for (const [key, value] of Object.entries(providerWindows)) {
+            if (key !== "_default" && model.startsWith(key)) {
+                return value;
+            }
+        }
+    }
+    return providerWindows._default ?? DEFAULT_CONTEXT_WINDOW;
+}
+/**
+ * Calculate output token reserve for a given context window.
+ *
+ * @param contextWindow - Total context window size
+ * @param maxTokens - Explicit maxTokens from user config (if set)
+ * @returns Number of tokens reserved for output
+ */
+export function getOutputReserve(contextWindow, maxTokens) {
+    if (maxTokens !== undefined && maxTokens > 0) {
+        return maxTokens;
+    }
+    return Math.min(MAX_DEFAULT_OUTPUT_RESERVE, Math.ceil(contextWindow * DEFAULT_OUTPUT_RESERVE_RATIO));
+}
+/**
+ * Calculate available input tokens for a given provider/model.
+ *
+ * available = contextWindow - outputReserve
+ */
+export function getAvailableInputTokens(provider, model, maxTokens) {
+    const contextWindow = getContextWindowSize(provider, model);
+    const outputReserve = getOutputReserve(contextWindow, maxTokens);
+    return contextWindow - outputReserve;
+}

package/dist/context/budgetChecker.d.ts

@@ -0,0 +1,18 @@
+/**
+ * Context Budget Checker
+ *
+ * Pre-generation validation that estimates total input token cost
+ * and compares against the model's available input space.
+ *
+ * This runs BEFORE every LLM call to prevent context overflow.
+ */
+import type { BudgetCheckResult, BudgetCheckParams } from "../types/contextTypes.js";
+export type { BudgetCheckResult, BudgetCheckParams, } from "../types/contextTypes.js";
+/**
+ * Check whether a request fits within the model's context budget.
+ *
+ * Estimates total input tokens from: system prompt + tool definitions +
+ * conversation history + current prompt + file attachments, and compares
+ * against available input space.
+ */
+export declare function checkContextBudget(params: BudgetCheckParams): BudgetCheckResult;

package/dist/context/budgetChecker.js

@@ -0,0 +1,71 @@
+/**
+ * Context Budget Checker
+ *
+ * Pre-generation validation that estimates total input token cost
+ * and compares against the model's available input space.
+ *
+ * This runs BEFORE every LLM call to prevent context overflow.
+ */
+import { getAvailableInputTokens } from "../constants/contextWindows.js";
+import { estimateMessagesTokens, estimateTokens, TOKENS_PER_MESSAGE, } from "../utils/tokenEstimation.js";
+/** Default compaction threshold (80% of available input) */
+const DEFAULT_COMPACTION_THRESHOLD = 0.8;
+/** Estimated tokens per tool definition */
+const TOKENS_PER_TOOL_DEFINITION = 200;
+/**
+ * Check whether a request fits within the model's context budget.
+ *
+ * Estimates total input tokens from: system prompt + tool definitions +
+ * conversation history + current prompt + file attachments, and compares
+ * against available input space.
+ */
+export function checkContextBudget(params) {
+    const { provider, model, maxTokens, systemPrompt, conversationMessages, currentPrompt, toolDefinitions, fileAttachments, compactionThreshold = DEFAULT_COMPACTION_THRESHOLD, } = params;
+    const availableInputTokens = getAvailableInputTokens(provider, model, maxTokens);
+    // Estimate each category
+    const systemPromptTokens = systemPrompt
+        ? estimateTokens(systemPrompt, provider) + TOKENS_PER_MESSAGE
+        : 0;
+    const conversationHistoryTokens = conversationMessages?.length
+        ? estimateMessagesTokens(conversationMessages, provider)
+        : 0;
+    const currentPromptTokens = currentPrompt
+        ? estimateTokens(currentPrompt, provider) + TOKENS_PER_MESSAGE
+        : 0;
+    const toolDefinitionTokens = toolDefinitions?.length
+        ? toolDefinitions.reduce((sum, tool) => {
+            try {
+                const serialized = JSON.stringify(tool);
+                return sum + Math.ceil(serialized.length / 4);
+            }
+            catch {
+                return sum + TOKENS_PER_TOOL_DEFINITION;
+            }
+        }, 0)
+        : 0;
+    const fileAttachmentTokens = fileAttachments?.length
+        ? fileAttachments.reduce((sum, file) => sum + estimateTokens(file.content, provider), 0)
+        : 0;
+    const estimatedInputTokens = systemPromptTokens +
+        conversationHistoryTokens +
+        currentPromptTokens +
+        toolDefinitionTokens +
+        fileAttachmentTokens;
+    const usageRatio = availableInputTokens > 0 ? estimatedInputTokens / availableInputTokens : 1;
+    const withinBudget = estimatedInputTokens <= availableInputTokens;
+    const shouldCompact = usageRatio >= compactionThreshold;
+    return {
+        withinBudget,
+        estimatedInputTokens,
+        availableInputTokens,
+        usageRatio,
+        shouldCompact,
+        breakdown: {
+            systemPrompt: systemPromptTokens,
+            conversationHistory: conversationHistoryTokens,
+            currentPrompt: currentPromptTokens,
+            toolDefinitions: toolDefinitionTokens,
+            fileAttachments: fileAttachmentTokens,
+        },
+    };
+}

package/dist/context/contextCompactor.d.ts

@@ -0,0 +1,22 @@
+/**
+ * ContextCompactor
+ *
+ * Orchestrates multi-stage context reduction:
+ *
+ * Stage 1: Tool Output Pruning (cheapest -- no LLM call)
+ * Stage 2: File Read Deduplication (cheap -- no LLM call)
+ * Stage 3: LLM Summarization (expensive -- requires LLM call)
+ * Stage 4: Sliding Window Truncation (fallback -- no LLM call)
+ */
+import type { ChatMessage, ConversationMemoryConfig } from "../types/conversation.js";
+import type { CompactionResult, CompactionConfig } from "../types/contextTypes.js";
+export type { CompactionResult, CompactionConfig, } from "../types/contextTypes.js";
+export type CompactionStage = "prune" | "deduplicate" | "summarize" | "truncate";
+export declare class ContextCompactor {
+    private config;
+    constructor(config?: CompactionConfig);
+    /**
+     * Run the multi-stage compaction pipeline until messages fit within budget.
+     */
+    compact(messages: ChatMessage[], targetTokens: number, memoryConfig?: Partial<ConversationMemoryConfig>): Promise<CompactionResult>;
+}