llmist 1.2.0 → 1.3.1

package/dist/index.cjs

@@ -375,10 +375,21 @@ function resolveRulesTemplate(rules, context) {
   }
   return [resolved];
 }
-var DEFAULT_PROMPTS;
+function resolveHintTemplate(template, defaultValue, context) {
+  const resolved = template ?? defaultValue;
+  if (typeof resolved === "function") {
+    return resolved(context);
+  }
+  return resolved.replace(/\{iteration\}/g, String(context.iteration)).replace(/\{maxIterations\}/g, String(context.maxIterations)).replace(/\{remaining\}/g, String(context.remaining));
+}
+var DEFAULT_HINTS, DEFAULT_PROMPTS;
 var init_prompt_config = __esm({
   "src/core/prompt-config.ts"() {
     "use strict";
+    DEFAULT_HINTS = {
+      parallelGadgetsHint: "Tip: You can call multiple gadgets in a single response for efficiency.",
+      iterationProgressHint: "[Iteration {iteration}/{maxIterations}] Plan your actions accordingly."
+    };
     DEFAULT_PROMPTS = {
       mainInstruction: [
         "\u26A0\uFE0F CRITICAL: RESPOND ONLY WITH GADGET INVOCATIONS",
@@ -1129,6 +1140,417 @@ var init_output_viewer = __esm({
   }
 });
 
+// src/agent/compaction/config.ts
+function resolveCompactionConfig(config = {}) {
+  const trigger = config.triggerThresholdPercent ?? DEFAULT_COMPACTION_CONFIG.triggerThresholdPercent;
+  const target = config.targetPercent ?? DEFAULT_COMPACTION_CONFIG.targetPercent;
+  if (target >= trigger) {
+    console.warn(
+      `[llmist/compaction] targetPercent (${target}) should be less than triggerThresholdPercent (${trigger}) to be effective.`
+    );
+  }
+  const strategy = config.strategy ?? DEFAULT_COMPACTION_CONFIG.strategy;
+  const strategyName = typeof strategy === "object" && "name" in strategy ? strategy.name : strategy;
+  return {
+    enabled: config.enabled ?? DEFAULT_COMPACTION_CONFIG.enabled,
+    strategy: strategyName,
+    triggerThresholdPercent: trigger,
+    targetPercent: target,
+    preserveRecentTurns: config.preserveRecentTurns ?? DEFAULT_COMPACTION_CONFIG.preserveRecentTurns,
+    summarizationModel: config.summarizationModel,
+    summarizationPrompt: config.summarizationPrompt ?? DEFAULT_SUMMARIZATION_PROMPT,
+    onCompaction: config.onCompaction
+  };
+}
+var DEFAULT_COMPACTION_CONFIG, DEFAULT_SUMMARIZATION_PROMPT;
+var init_config = __esm({
+  "src/agent/compaction/config.ts"() {
+    "use strict";
+    DEFAULT_COMPACTION_CONFIG = {
+      enabled: true,
+      strategy: "hybrid",
+      triggerThresholdPercent: 80,
+      targetPercent: 50,
+      preserveRecentTurns: 5
+    };
+    DEFAULT_SUMMARIZATION_PROMPT = `Summarize this conversation history concisely, preserving:
+1. Key decisions made and their rationale
+2. Important facts and data discovered
+3. Errors encountered and how they were resolved
+4. Current task context and goals
+
+Format as a brief narrative paragraph, not bullet points.
+Previous conversation:`;
+  }
+});
+
+// src/agent/compaction/strategy.ts
+function groupIntoTurns(messages) {
+  const turns = [];
+  let currentTurn = [];
+  for (const msg of messages) {
+    if (msg.role === "user" && currentTurn.length > 0) {
+      turns.push({
+        messages: currentTurn,
+        tokenEstimate: estimateTurnTokens(currentTurn)
+      });
+      currentTurn = [msg];
+    } else {
+      currentTurn.push(msg);
+    }
+  }
+  if (currentTurn.length > 0) {
+    turns.push({
+      messages: currentTurn,
+      tokenEstimate: estimateTurnTokens(currentTurn)
+    });
+  }
+  return turns;
+}
+function estimateTurnTokens(messages) {
+  return Math.ceil(messages.reduce((sum, msg) => sum + (msg.content?.length ?? 0), 0) / 4);
+}
+function flattenTurns(turns) {
+  return turns.flatMap((turn) => turn.messages);
+}
+var init_strategy = __esm({
+  "src/agent/compaction/strategy.ts"() {
+    "use strict";
+  }
+});
+
+// src/agent/compaction/strategies/sliding-window.ts
+var TRUNCATION_MARKER_TEMPLATE, SlidingWindowStrategy;
+var init_sliding_window = __esm({
+  "src/agent/compaction/strategies/sliding-window.ts"() {
+    "use strict";
+    init_strategy();
+    TRUNCATION_MARKER_TEMPLATE = "[Previous conversation truncated. Removed {count} turn(s) to fit context window.]";
+    SlidingWindowStrategy = class {
+      name = "sliding-window";
+      async compact(messages, config, context) {
+        const turns = groupIntoTurns(messages);
+        const preserveCount = Math.min(config.preserveRecentTurns, turns.length);
+        if (turns.length <= preserveCount) {
+          return {
+            messages,
+            strategyName: this.name,
+            metadata: {
+              originalCount: messages.length,
+              compactedCount: messages.length,
+              tokensBefore: context.currentTokens,
+              tokensAfter: context.currentTokens
+            }
+          };
+        }
+        const turnsToKeep = turns.slice(-preserveCount);
+        const turnsRemoved = turns.length - preserveCount;
+        const truncationMarker = {
+          role: "user",
+          content: TRUNCATION_MARKER_TEMPLATE.replace("{count}", turnsRemoved.toString())
+        };
+        const compactedMessages = [truncationMarker, ...flattenTurns(turnsToKeep)];
+        const tokensAfter = Math.ceil(
+          compactedMessages.reduce((sum, msg) => sum + (msg.content?.length ?? 0), 0) / 4
+        );
+        return {
+          messages: compactedMessages,
+          strategyName: this.name,
+          metadata: {
+            originalCount: messages.length,
+            compactedCount: compactedMessages.length,
+            tokensBefore: context.currentTokens,
+            tokensAfter
+          }
+        };
+      }
+    };
+  }
+});
+
+// src/agent/compaction/strategies/summarization.ts
+var SummarizationStrategy;
+var init_summarization = __esm({
+  "src/agent/compaction/strategies/summarization.ts"() {
+    "use strict";
+    init_strategy();
+    SummarizationStrategy = class {
+      name = "summarization";
+      async compact(messages, config, context) {
+        const turns = groupIntoTurns(messages);
+        const preserveCount = Math.min(config.preserveRecentTurns, turns.length);
+        if (turns.length <= preserveCount) {
+          return {
+            messages,
+            strategyName: this.name,
+            metadata: {
+              originalCount: messages.length,
+              compactedCount: messages.length,
+              tokensBefore: context.currentTokens,
+              tokensAfter: context.currentTokens
+            }
+          };
+        }
+        const turnsToSummarize = turns.slice(0, -preserveCount);
+        const turnsToKeep = turns.slice(-preserveCount);
+        const conversationToSummarize = this.formatTurnsForSummary(flattenTurns(turnsToSummarize));
+        const summary = await this.generateSummary(conversationToSummarize, config, context);
+        const summaryMessage = {
+          role: "user",
+          content: `[Previous conversation summary]
+${summary}
+[End of summary - conversation continues below]`
+        };
+        const compactedMessages = [summaryMessage, ...flattenTurns(turnsToKeep)];
+        const tokensAfter = Math.ceil(
+          compactedMessages.reduce((sum, msg) => sum + (msg.content?.length ?? 0), 0) / 4
+        );
+        return {
+          messages: compactedMessages,
+          summary,
+          strategyName: this.name,
+          metadata: {
+            originalCount: messages.length,
+            compactedCount: compactedMessages.length,
+            tokensBefore: context.currentTokens,
+            tokensAfter
+          }
+        };
+      }
+      /**
+       * Formats messages into a readable conversation format for summarization.
+       */
+      formatTurnsForSummary(messages) {
+        return messages.map((msg) => {
+          const role = msg.role.charAt(0).toUpperCase() + msg.role.slice(1);
+          return `${role}: ${msg.content}`;
+        }).join("\n\n");
+      }
+      /**
+       * Generates a summary using the configured LLM.
+       */
+      async generateSummary(conversation, config, context) {
+        const model = config.summarizationModel ?? context.model;
+        const prompt = `${config.summarizationPrompt}
+
+${conversation}`;
+        const response = await context.client.complete(prompt, {
+          model,
+          temperature: 0.3
+          // Low temperature for factual summarization
+        });
+        return response.trim();
+      }
+    };
+  }
+});
+
+// src/agent/compaction/strategies/hybrid.ts
+var MIN_TURNS_FOR_SUMMARIZATION, HybridStrategy;
+var init_hybrid = __esm({
+  "src/agent/compaction/strategies/hybrid.ts"() {
+    "use strict";
+    init_strategy();
+    init_sliding_window();
+    init_summarization();
+    MIN_TURNS_FOR_SUMMARIZATION = 3;
+    HybridStrategy = class {
+      name = "hybrid";
+      slidingWindow = new SlidingWindowStrategy();
+      summarization = new SummarizationStrategy();
+      async compact(messages, config, context) {
+        const turns = groupIntoTurns(messages);
+        const preserveCount = Math.min(config.preserveRecentTurns, turns.length);
+        if (turns.length <= preserveCount) {
+          return {
+            messages,
+            strategyName: this.name,
+            metadata: {
+              originalCount: messages.length,
+              compactedCount: messages.length,
+              tokensBefore: context.currentTokens,
+              tokensAfter: context.currentTokens
+            }
+          };
+        }
+        const turnsToSummarize = turns.length - preserveCount;
+        if (turnsToSummarize < MIN_TURNS_FOR_SUMMARIZATION) {
+          return this.slidingWindow.compact(messages, config, context);
+        }
+        return this.summarization.compact(messages, config, context);
+      }
+    };
+  }
+});
+
+// src/agent/compaction/strategies/index.ts
+var init_strategies = __esm({
+  "src/agent/compaction/strategies/index.ts"() {
+    "use strict";
+    init_sliding_window();
+    init_summarization();
+    init_hybrid();
+  }
+});
+
+// src/agent/compaction/manager.ts
+function createStrategy(name) {
+  switch (name) {
+    case "sliding-window":
+      return new SlidingWindowStrategy();
+    case "summarization":
+      return new SummarizationStrategy();
+    case "hybrid":
+      return new HybridStrategy();
+    default:
+      throw new Error(`Unknown compaction strategy: ${name}`);
+  }
+}
+var CompactionManager;
+var init_manager = __esm({
+  "src/agent/compaction/manager.ts"() {
+    "use strict";
+    init_config();
+    init_strategies();
+    CompactionManager = class {
+      client;
+      model;
+      config;
+      strategy;
+      modelLimits;
+      // Statistics
+      totalCompactions = 0;
+      totalTokensSaved = 0;
+      lastTokenCount = 0;
+      constructor(client, model, config = {}) {
+        this.client = client;
+        this.model = model;
+        this.config = resolveCompactionConfig(config);
+        if (typeof config.strategy === "object" && "compact" in config.strategy) {
+          this.strategy = config.strategy;
+        } else {
+          this.strategy = createStrategy(this.config.strategy);
+        }
+      }
+      /**
+       * Check if compaction is needed and perform it if so.
+       *
+       * @param conversation - The conversation manager to compact
+       * @param iteration - Current agent iteration (for event metadata)
+       * @returns CompactionEvent if compaction was performed, null otherwise
+       */
+      async checkAndCompact(conversation, iteration) {
+        if (!this.config.enabled) {
+          return null;
+        }
+        if (!this.modelLimits) {
+          this.modelLimits = this.client.modelRegistry.getModelLimits(this.model);
+          if (!this.modelLimits) {
+            return null;
+          }
+        }
+        if (!this.client.countTokens) {
+          return null;
+        }
+        const messages = conversation.getMessages();
+        const currentTokens = await this.client.countTokens(this.model, messages);
+        this.lastTokenCount = currentTokens;
+        const usagePercent = currentTokens / this.modelLimits.contextWindow * 100;
+        if (usagePercent < this.config.triggerThresholdPercent) {
+          return null;
+        }
+        const historyMessages = conversation.getHistoryMessages();
+        const baseMessages = conversation.getBaseMessages();
+        const historyTokens = await this.client.countTokens(this.model, historyMessages);
+        const baseTokens = await this.client.countTokens(this.model, baseMessages);
+        return this.compact(conversation, iteration, {
+          historyMessages,
+          baseMessages,
+          historyTokens,
+          baseTokens,
+          currentTokens: historyTokens + baseTokens
+        });
+      }
+      /**
+       * Force compaction regardless of threshold.
+       *
+       * @param conversation - The conversation manager to compact
+       * @param iteration - Current agent iteration (for event metadata). Use -1 for manual compaction.
+       * @param precomputed - Optional pre-computed token counts (passed from checkAndCompact for efficiency)
+       * @returns CompactionEvent with compaction details
+       */
+      async compact(conversation, iteration, precomputed) {
+        if (!this.modelLimits) {
+          this.modelLimits = this.client.modelRegistry.getModelLimits(this.model);
+          if (!this.modelLimits) {
+            return null;
+          }
+        }
+        const historyMessages = precomputed?.historyMessages ?? conversation.getHistoryMessages();
+        const baseMessages = precomputed?.baseMessages ?? conversation.getBaseMessages();
+        const historyTokens = precomputed?.historyTokens ?? await this.client.countTokens(this.model, historyMessages);
+        const baseTokens = precomputed?.baseTokens ?? await this.client.countTokens(this.model, baseMessages);
+        const currentTokens = precomputed?.currentTokens ?? historyTokens + baseTokens;
+        const targetTotalTokens = Math.floor(
+          this.modelLimits.contextWindow * this.config.targetPercent / 100
+        );
+        const targetHistoryTokens = Math.max(0, targetTotalTokens - baseTokens);
+        const result = await this.strategy.compact(historyMessages, this.config, {
+          currentTokens: historyTokens,
+          targetTokens: targetHistoryTokens,
+          modelLimits: this.modelLimits,
+          client: this.client,
+          model: this.config.summarizationModel ?? this.model
+        });
+        conversation.replaceHistory(result.messages);
+        const afterTokens = await this.client.countTokens(this.model, conversation.getMessages());
+        const tokensSaved = currentTokens - afterTokens;
+        this.totalCompactions++;
+        this.totalTokensSaved += tokensSaved;
+        this.lastTokenCount = afterTokens;
+        const event = {
+          strategy: result.strategyName,
+          tokensBefore: currentTokens,
+          tokensAfter: afterTokens,
+          messagesBefore: historyMessages.length + baseMessages.length,
+          messagesAfter: result.messages.length + baseMessages.length,
+          summary: result.summary,
+          iteration
+        };
+        if (this.config.onCompaction) {
+          try {
+            this.config.onCompaction(event);
+          } catch (err) {
+            console.warn("[llmist/compaction] onCompaction callback error:", err);
+          }
+        }
+        return event;
+      }
+      /**
+       * Get compaction statistics.
+       */
+      getStats() {
+        const contextWindow = this.modelLimits?.contextWindow ?? 0;
+        return {
+          totalCompactions: this.totalCompactions,
+          totalTokensSaved: this.totalTokensSaved,
+          currentUsage: {
+            tokens: this.lastTokenCount,
+            percent: contextWindow > 0 ? this.lastTokenCount / contextWindow * 100 : 0
+          },
+          contextWindow
+        };
+      }
+      /**
+       * Check if compaction is enabled.
+       */
+      isEnabled() {
+        return this.config.enabled;
+      }
+    };
+  }
+});
+
 // src/agent/gadget-output-store.ts
 var import_node_crypto, GadgetOutputStore;
 var init_gadget_output_store = __esm({
@@ -1231,10 +1653,16 @@ var init_conversation_manager = __esm({
       baseMessages;
       initialMessages;
       historyBuilder;
+      startPrefix;
+      endPrefix;
+      argPrefix;
       constructor(baseMessages, initialMessages, options = {}) {
         this.baseMessages = baseMessages;
         this.initialMessages = initialMessages;
         this.historyBuilder = new LLMMessageBuilder();
+        this.startPrefix = options.startPrefix;
+        this.endPrefix = options.endPrefix;
+        this.argPrefix = options.argPrefix;
         if (options.startPrefix && options.endPrefix) {
           this.historyBuilder.withPrefixes(options.startPrefix, options.endPrefix, options.argPrefix);
         }
@@ -1251,6 +1679,25 @@ var init_conversation_manager = __esm({
       getMessages() {
         return [...this.baseMessages, ...this.initialMessages, ...this.historyBuilder.build()];
       }
+      getHistoryMessages() {
+        return this.historyBuilder.build();
+      }
+      getBaseMessages() {
+        return [...this.baseMessages, ...this.initialMessages];
+      }
+      replaceHistory(newHistory) {
+        this.historyBuilder = new LLMMessageBuilder();
+        if (this.startPrefix && this.endPrefix) {
+          this.historyBuilder.withPrefixes(this.startPrefix, this.endPrefix, this.argPrefix);
+        }
+        for (const msg of newHistory) {
+          if (msg.role === "user") {
+            this.historyBuilder.addUser(msg.content);
+          } else if (msg.role === "assistant") {
+            this.historyBuilder.addAssistant(msg.content);
+          }
+        }
+      }
     };
   }
 });
@@ -2827,6 +3274,7 @@ var init_agent = __esm({
     init_model_shortcuts();
     init_output_viewer();
     init_logger();
+    init_manager();
     init_gadget_output_store();
     init_agent_internal_key();
     init_conversation_manager();
@@ -2857,6 +3305,8 @@ var init_agent = __esm({
       outputStore;
       outputLimitEnabled;
       outputLimitCharLimit;
+      // Context compaction
+      compactionManager;
       /**
        * Creates a new Agent instance.
        * @internal This constructor is private. Use LLMist.createAgent() or AgentBuilder instead.
@@ -2916,6 +3366,14 @@ var init_agent = __esm({
         if (options.userPrompt) {
           this.conversation.addUserMessage(options.userPrompt);
         }
+        const compactionEnabled = options.compactionConfig?.enabled ?? true;
+        if (compactionEnabled) {
+          this.compactionManager = new CompactionManager(
+            this.client,
+            this.model,
+            options.compactionConfig
+          );
+        }
       }
       /**
        * Get the gadget registry for this agent.
@@ -2938,6 +3396,53 @@ var init_agent = __esm({
       getRegistry() {
         return this.registry;
       }
+      /**
+       * Manually trigger context compaction.
+       *
+       * Forces compaction regardless of threshold. Useful for:
+       * - Pre-emptive context management before expected long operations
+       * - Testing compaction behavior
+       *
+       * @returns CompactionEvent if compaction was performed, null if not configured or no history
+       *
+       * @example
+       * ```typescript
+       * const agent = await LLMist.createAgent()
+       *   .withModel('sonnet')
+       *   .withCompaction()
+       *   .ask('...');
+       *
+       * // Manually compact before a long operation
+       * const event = await agent.compact();
+       * if (event) {
+       *   console.log(`Saved ${event.tokensBefore - event.tokensAfter} tokens`);
+       * }
+       * ```
+       */
+      async compact() {
+        if (!this.compactionManager) {
+          return null;
+        }
+        return this.compactionManager.compact(this.conversation, -1);
+      }
+      /**
+       * Get compaction statistics.
+       *
+       * @returns CompactionStats if compaction is enabled, null otherwise
+       *
+       * @example
+       * ```typescript
+       * const stats = agent.getCompactionStats();
+       * if (stats) {
+       *   console.log(`Total compactions: ${stats.totalCompactions}`);
+       *   console.log(`Tokens saved: ${stats.totalTokensSaved}`);
+       *   console.log(`Current usage: ${stats.currentUsage.percent.toFixed(1)}%`);
+       * }
+       * ```
+       */
+      getCompactionStats() {
+        return this.compactionManager?.getStats() ?? null;
+      }
       /**
        * Run the agent loop.
        * Clean, simple orchestration - all complexity is in StreamProcessor.
@@ -2958,6 +3463,30 @@ var init_agent = __esm({
         while (currentIteration < this.maxIterations) {
           this.logger.debug("Starting iteration", { iteration: currentIteration });
           try {
+            if (this.compactionManager) {
+              const compactionEvent = await this.compactionManager.checkAndCompact(
+                this.conversation,
+                currentIteration
+              );
+              if (compactionEvent) {
+                this.logger.info("Context compacted", {
+                  strategy: compactionEvent.strategy,
+                  tokensBefore: compactionEvent.tokensBefore,
+                  tokensAfter: compactionEvent.tokensAfter
+                });
+                yield { type: "compaction", event: compactionEvent };
+                await this.safeObserve(async () => {
+                  if (this.hooks.observers?.onCompaction) {
+                    await this.hooks.observers.onCompaction({
+                      iteration: currentIteration,
+                      event: compactionEvent,
+                      stats: this.compactionManager.getStats(),
+                      logger: this.logger
+                    });
+                  }
+                });
+              }
+            }
             let llmOptions = {
               model: this.model,
               messages: this.conversation.getMessages(),
@@ -2977,6 +3506,7 @@ var init_agent = __esm({
             if (this.hooks.controllers?.beforeLLMCall) {
               const context = {
                 iteration: currentIteration,
+                maxIterations: this.maxIterations,
                 options: llmOptions,
                 logger: this.logger
               };
@@ -3041,12 +3571,17 @@ var init_agent = __esm({
             });
             let finalMessage = result.finalMessage;
             if (this.hooks.controllers?.afterLLMCall) {
+              const gadgetCallCount = result.outputs.filter(
+                (output) => output.type === "gadget_result"
+              ).length;
               const context = {
                 iteration: currentIteration,
+                maxIterations: this.maxIterations,
                 options: llmOptions,
                 finishReason: result.finishReason,
                 usage: result.usage,
                 finalMessage: result.finalMessage,
+                gadgetCallCount,
                 logger: this.logger
               };
               const action = await this.hooks.controllers.afterLLMCall(context);
@@ -5281,6 +5816,7 @@ var init_builder = __esm({
       defaultGadgetTimeoutMs;
       gadgetOutputLimit;
       gadgetOutputLimitPercent;
+      compactionConfig;
       constructor(client) {
         this.client = client;
       }
@@ -5676,6 +6212,57 @@ var init_builder = __esm({
         this.gadgetOutputLimitPercent = percent;
         return this;
       }
+      /**
+       * Configure context compaction.
+       *
+       * Context compaction automatically manages conversation history to prevent
+       * context window overflow in long-running agent conversations.
+       *
+       * @param config - Compaction configuration options
+       * @returns This builder for chaining
+       *
+       * @example
+       * ```typescript
+       * // Custom thresholds
+       * .withCompaction({
+       *   triggerThresholdPercent: 70,
+       *   targetPercent: 40,
+       *   preserveRecentTurns: 10,
+       * })
+       *
+       * // Different strategy
+       * .withCompaction({
+       *   strategy: 'sliding-window',
+       * })
+       *
+       * // With callback
+       * .withCompaction({
+       *   onCompaction: (event) => {
+       *     console.log(`Saved ${event.tokensBefore - event.tokensAfter} tokens`);
+       *   }
+       * })
+       * ```
+       */
+      withCompaction(config) {
+        this.compactionConfig = { ...config, enabled: config.enabled ?? true };
+        return this;
+      }
+      /**
+       * Disable context compaction.
+       *
+       * By default, compaction is enabled. Use this method to explicitly disable it.
+       *
+       * @returns This builder for chaining
+       *
+       * @example
+       * ```typescript
+       * .withoutCompaction() // Disable automatic compaction
+       * ```
+       */
+      withoutCompaction() {
+        this.compactionConfig = { enabled: false };
+        return this;
+      }
       /**
        * Add a synthetic gadget call to the conversation history.
        *
@@ -5791,7 +6378,8 @@ ${endPrefix}`
           shouldContinueAfterError: this.shouldContinueAfterError,
           defaultGadgetTimeoutMs: this.defaultGadgetTimeoutMs,
           gadgetOutputLimit: this.gadgetOutputLimit,
-          gadgetOutputLimitPercent: this.gadgetOutputLimitPercent
+          gadgetOutputLimitPercent: this.gadgetOutputLimitPercent,
+          compactionConfig: this.compactionConfig
         };
         return new Agent(AGENT_INTERNAL_KEY, options);
       }
@@ -5893,7 +6481,8 @@ ${endPrefix}`
           shouldContinueAfterError: this.shouldContinueAfterError,
           defaultGadgetTimeoutMs: this.defaultGadgetTimeoutMs,
           gadgetOutputLimit: this.gadgetOutputLimit,
-          gadgetOutputLimitPercent: this.gadgetOutputLimitPercent
+          gadgetOutputLimitPercent: this.gadgetOutputLimitPercent,
+          compactionConfig: this.compactionConfig
         };
         return new Agent(AGENT_INTERNAL_KEY, options);
       }
@@ -5908,8 +6497,12 @@ __export(index_exports, {
   AnthropicMessagesProvider: () => AnthropicMessagesProvider,
   BaseGadget: () => BaseGadget,
   BreakLoopException: () => BreakLoopException,
+  CompactionManager: () => CompactionManager,
   ConversationManager: () => ConversationManager,
+  DEFAULT_COMPACTION_CONFIG: () => DEFAULT_COMPACTION_CONFIG,
+  DEFAULT_HINTS: () => DEFAULT_HINTS,
   DEFAULT_PROMPTS: () => DEFAULT_PROMPTS,
+  DEFAULT_SUMMARIZATION_PROMPT: () => DEFAULT_SUMMARIZATION_PROMPT,
   Gadget: () => Gadget,
   GadgetExecutor: () => GadgetExecutor,
   GadgetOutputStore: () => GadgetOutputStore,
@@ -5917,6 +6510,7 @@ __export(index_exports, {
   GeminiGenerativeProvider: () => GeminiGenerativeProvider,
   HookPresets: () => HookPresets,
   HumanInputException: () => HumanInputException,
+  HybridStrategy: () => HybridStrategy,
   LLMMessageBuilder: () => LLMMessageBuilder,
   LLMist: () => LLMist,
   MODEL_ALIASES: () => MODEL_ALIASES,
@@ -5926,8 +6520,10 @@ __export(index_exports, {
   ModelIdentifierParser: () => ModelIdentifierParser,
   ModelRegistry: () => ModelRegistry,
   OpenAIChatProvider: () => OpenAIChatProvider,
+  SlidingWindowStrategy: () => SlidingWindowStrategy,
   StreamParser: () => StreamParser,
   StreamProcessor: () => StreamProcessor,
+  SummarizationStrategy: () => SummarizationStrategy,
   collectEvents: () => collectEvents,
   collectText: () => collectText,
   complete: () => complete,
@@ -5935,6 +6531,7 @@ __export(index_exports, {
   createGadget: () => createGadget,
   createGadgetOutputViewer: () => createGadgetOutputViewer,
   createGeminiProviderFromEnv: () => createGeminiProviderFromEnv,
+  createHints: () => createHints,
   createLogger: () => createLogger,
   createMockAdapter: () => createMockAdapter,
   createMockClient: () => createMockClient,
@@ -5947,7 +6544,10 @@ __export(index_exports, {
   getModelId: () => getModelId,
   getProvider: () => getProvider,
   hasProviderPrefix: () => hasProviderPrefix,
+  iterationProgressHint: () => iterationProgressHint,
   mockLLM: () => mockLLM,
+  parallelGadgetHint: () => parallelGadgetHint,
+  resolveHintTemplate: () => resolveHintTemplate,
   resolveModel: () => resolveModel,
   resolvePromptTemplate: () => resolvePromptTemplate,
   resolveRulesTemplate: () => resolveRulesTemplate,
@@ -6462,6 +7062,51 @@ var HookPresets = class _HookPresets {
       }
     };
   }
+  /**
+   * Tracks context compaction events.
+   *
+   * **Output:**
+   * - Compaction events with 🗜️ emoji
+   * - Strategy name, tokens before/after, and savings
+   * - Cumulative statistics
+   *
+   * **Use cases:**
+   * - Monitoring long-running conversations
+   * - Understanding when and how compaction occurs
+   * - Debugging context management issues
+   *
+   * **Performance:** Minimal overhead. Simple console output.
+   *
+   * @returns Hook configuration that can be passed to .withHooks()
+   *
+   * @example
+   * ```typescript
+   * await LLMist.createAgent()
+   *   .withHooks(HookPresets.compactionTracking())
+   *   .ask("Your prompt");
+   * ```
+   */
+  static compactionTracking() {
+    return {
+      observers: {
+        onCompaction: async (ctx) => {
+          const saved = ctx.event.tokensBefore - ctx.event.tokensAfter;
+          const percent = (saved / ctx.event.tokensBefore * 100).toFixed(1);
+          console.log(
+            `\u{1F5DC}\uFE0F  Compaction (${ctx.event.strategy}): ${ctx.event.tokensBefore} \u2192 ${ctx.event.tokensAfter} tokens (saved ${saved}, ${percent}%)`
+          );
+          console.log(
+            `   Messages: ${ctx.event.messagesBefore} \u2192 ${ctx.event.messagesAfter}`
+          );
+          if (ctx.stats.totalCompactions > 1) {
+            console.log(
+              `   Cumulative: ${ctx.stats.totalCompactions} compactions, ${ctx.stats.totalTokensSaved} tokens saved`
+            );
+          }
+        }
+      }
+    };
+  }
   /**
    * Returns empty hook configuration for clean output without any logging.
    *
@@ -6692,6 +7337,113 @@ init_conversation_manager();
 init_stream_processor();
 init_gadget_output_store();
 
+// src/agent/compaction/index.ts
+init_config();
+init_strategy();
+init_strategies();
+init_manager();
+
+// src/agent/hints.ts
+init_prompt_config();
+function iterationProgressHint(options) {
+  const { timing = "always", showUrgency = true, template } = options ?? {};
+  return {
+    controllers: {
+      beforeLLMCall: async (ctx) => {
+        const iteration = ctx.iteration + 1;
+        const maxIterations = ctx.maxIterations;
+        const progress = iteration / maxIterations;
+        if (timing === "late" && progress < 0.5) {
+          return { action: "proceed" };
+        }
+        if (timing === "urgent" && progress < 0.8) {
+          return { action: "proceed" };
+        }
+        const remaining = maxIterations - iteration;
+        const hintContext = {
+          iteration,
+          maxIterations,
+          remaining
+        };
+        let hint = resolveHintTemplate(
+          template,
+          DEFAULT_HINTS.iterationProgressHint,
+          hintContext
+        );
+        if (showUrgency && progress >= 0.8) {
+          hint += " \u26A0\uFE0F Running low on iterations - focus on completing the task.";
+        }
+        const messages = [...ctx.options.messages];
+        let lastUserIndex = -1;
+        for (let i = messages.length - 1; i >= 0; i--) {
+          if (messages[i].role === "user") {
+            lastUserIndex = i;
+            break;
+          }
+        }
+        if (lastUserIndex >= 0) {
+          messages.splice(lastUserIndex + 1, 0, {
+            role: "user",
+            content: `[System Hint] ${hint}`
+          });
+        } else {
+          messages.push({
+            role: "user",
+            content: `[System Hint] ${hint}`
+          });
+        }
+        return {
+          action: "proceed",
+          modifiedOptions: { messages }
+        };
+      }
+    }
+  };
+}
+function parallelGadgetHint(options) {
+  const {
+    minGadgetsForEfficiency = 2,
+    message = DEFAULT_HINTS.parallelGadgetsHint,
+    enabled = true
+  } = options ?? {};
+  return {
+    controllers: {
+      afterLLMCall: async (ctx) => {
+        if (!enabled) {
+          return { action: "continue" };
+        }
+        if (ctx.gadgetCallCount > 0 && ctx.gadgetCallCount < minGadgetsForEfficiency) {
+          return {
+            action: "append_messages",
+            messages: [
+              {
+                role: "user",
+                content: `[System Hint] ${message}`
+              }
+            ]
+          };
+        }
+        return { action: "continue" };
+      }
+    }
+  };
+}
+function createHints(config) {
+  const hooksToMerge = [];
+  if (config.iterationProgress) {
+    const options = typeof config.iterationProgress === "boolean" ? {} : config.iterationProgress;
+    hooksToMerge.push(iterationProgressHint(options));
+  }
+  if (config.parallelGadgets) {
+    const options = typeof config.parallelGadgets === "boolean" ? {} : config.parallelGadgets;
+    hooksToMerge.push(parallelGadgetHint(options));
+  }
+  if (config.custom) {
+    hooksToMerge.push(...config.custom);
+  }
+  return HookPresets.merge(...hooksToMerge);
+}
+
 // src/index.ts
 init_client();
 init_messages();
@@ -7466,14 +8218,21 @@ function createMockClient(options) {
 
 // src/testing/mock-gadget.ts
 init_gadget();
+
+// src/testing/cli-helpers.ts
+var import_node_stream = require("stream");
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   AgentBuilder,
   AnthropicMessagesProvider,
   BaseGadget,
   BreakLoopException,
+  CompactionManager,
   ConversationManager,
+  DEFAULT_COMPACTION_CONFIG,
+  DEFAULT_HINTS,
   DEFAULT_PROMPTS,
+  DEFAULT_SUMMARIZATION_PROMPT,
   Gadget,
   GadgetExecutor,
   GadgetOutputStore,
@@ -7481,6 +8240,7 @@ init_gadget();
   GeminiGenerativeProvider,
   HookPresets,
   HumanInputException,
+  HybridStrategy,
   LLMMessageBuilder,
   LLMist,
   MODEL_ALIASES,
@@ -7490,8 +8250,10 @@ init_gadget();
   ModelIdentifierParser,
   ModelRegistry,
   OpenAIChatProvider,
+  SlidingWindowStrategy,
   StreamParser,
   StreamProcessor,
+  SummarizationStrategy,
   collectEvents,
   collectText,
   complete,
@@ -7499,6 +8261,7 @@ init_gadget();
   createGadget,
   createGadgetOutputViewer,
   createGeminiProviderFromEnv,
+  createHints,
   createLogger,
   createMockAdapter,
   createMockClient,
@@ -7511,7 +8274,10 @@ init_gadget();
   getModelId,
   getProvider,
   hasProviderPrefix,
+  iterationProgressHint,
   mockLLM,
+  parallelGadgetHint,
+  resolveHintTemplate,
   resolveModel,
   resolvePromptTemplate,
   resolveRulesTemplate,