llmist 16.2.3 → 16.2.5

package/dist/index.js

@@ -813,7 +813,7 @@ var init_constants = __esm({
     GADGET_ARG_PREFIX = "!!!ARG:";
     DEFAULT_GADGET_OUTPUT_LIMIT = true;
     DEFAULT_GADGET_OUTPUT_LIMIT_PERCENT = 15;
-    CHARS_PER_TOKEN = 4;
+    CHARS_PER_TOKEN = 2;
     FALLBACK_CONTEXT_WINDOW = 128e3;
   }
 });
@@ -2834,6 +2834,7 @@ var CompactionManager;
 var init_manager = __esm({
   "src/agent/compaction/manager.ts"() {
     "use strict";
+    init_logger();
     init_config();
     init_strategies();
     CompactionManager = class {
@@ -2841,15 +2842,19 @@ var init_manager = __esm({
       model;
       config;
       strategy;
+      logger;
       modelLimits;
+      hasWarnedModelNotFound = false;
+      hasWarnedNoTokenCounting = false;
       // Statistics
       totalCompactions = 0;
       totalTokensSaved = 0;
       lastTokenCount = 0;
-      constructor(client, model, config = {}) {
+      constructor(client, model, config = {}, logger2) {
         this.client = client;
         this.model = model;
         this.config = resolveCompactionConfig(config);
+        this.logger = logger2 ?? createLogger({ name: "llmist:compaction" });
         if (typeof config.strategy === "object" && "compact" in config.strategy) {
           this.strategy = config.strategy;
         } else {
@@ -2867,13 +2872,16 @@ var init_manager = __esm({
         if (!this.config.enabled) {
           return null;
         }
-        if (!this.modelLimits) {
-          this.modelLimits = this.client.modelRegistry.getModelLimits(this.model);
-          if (!this.modelLimits) {
-            return null;
-          }
+        if (!this.resolveModelLimits()) {
+          return null;
         }
         if (!this.client.countTokens) {
+          if (!this.hasWarnedNoTokenCounting) {
+            this.hasWarnedNoTokenCounting = true;
+            this.logger.warn("Compaction skipped: client does not support token counting", {
+              model: this.model
+            });
+          }
           return null;
         }
         const messages = conversation.getMessages();
@@ -2904,11 +2912,8 @@ var init_manager = __esm({
        * @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;
-          }
+        if (!this.resolveModelLimits()) {
+          return null;
         }
         const historyMessages = precomputed?.historyMessages ?? conversation.getHistoryMessages();
         const baseMessages = precomputed?.baseMessages ?? conversation.getBaseMessages();
@@ -2950,6 +2955,42 @@ var init_manager = __esm({
         }
         return event;
       }
+      /**
+       * Feed API-reported input token count for reactive threshold checking.
+       * Call this after each LLM response with the actual inputTokens from usage.
+       */
+      updateUsage(inputTokens) {
+        this.lastTokenCount = inputTokens;
+      }
+      /**
+       * Check if compaction should trigger based on API-reported usage.
+       * Unlike checkAndCompact() which uses estimated token counts,
+       * this uses the ground-truth token count from the last LLM response.
+       */
+      shouldCompactFromUsage() {
+        if (!this.config.enabled) return false;
+        if (!this.resolveModelLimits()) return false;
+        const usagePercent = this.lastTokenCount / this.modelLimits.contextWindow * 100;
+        return usagePercent >= this.config.triggerThresholdPercent;
+      }
+      /**
+       * Resolve and cache model limits from registry. Warns once if not found.
+       * @returns true if limits are available, false otherwise
+       */
+      resolveModelLimits() {
+        if (this.modelLimits) return true;
+        this.modelLimits = this.client.modelRegistry.getModelLimits(this.model);
+        if (!this.modelLimits) {
+          if (!this.hasWarnedModelNotFound) {
+            this.hasWarnedModelNotFound = true;
+            this.logger.warn("Compaction skipped: model not found in registry", {
+              model: this.model
+            });
+          }
+          return false;
+        }
+        return true;
+      }
       /**
        * Get compaction statistics.
        */
@@ -5298,6 +5339,45 @@ var init_activation = __esm({
   }
 });
 
+// src/skills/load-skill-gadget.ts
+import { z as z4 } from "zod";
+function createLoadSkillGadget(registry) {
+  const summaries = registry.getMetadataSummaries();
+  const skillNames = registry.getModelInvocable().map((s) => s.name);
+  const description = [
+    "Load a skill's specialized instructions into context for a task.",
+    "Available skills:",
+    summaries
+  ].join("\n");
+  return createGadget({
+    name: LOAD_SKILL_GADGET_NAME,
+    description,
+    schema: z4.object({
+      skill: z4.enum(skillNames).describe("Name of the skill to load"),
+      arguments: z4.string().optional().describe("Arguments for the skill (e.g., a filename, issue number, or search query)")
+    }),
+    execute: async ({ skill: skillName, arguments: args }) => {
+      const skill = registry.get(skillName);
+      if (!skill) {
+        return `Unknown skill: "${skillName}". Available skills: ${skillNames.join(", ")}`;
+      }
+      const activation = await skill.activate({
+        arguments: args,
+        cwd: process.cwd()
+      });
+      return activation.resolvedInstructions;
+    }
+  });
+}
+var LOAD_SKILL_GADGET_NAME;
+var init_load_skill_gadget = __esm({
+  "src/skills/load-skill-gadget.ts"() {
+    "use strict";
+    init_create_gadget();
+    LOAD_SKILL_GADGET_NAME = "LoadSkill";
+  }
+});
+
 // src/skills/parser.ts
 import fs from "fs";
 import path from "path";
@@ -5791,45 +5871,6 @@ var init_loader = __esm({
   }
 });
 
-// src/skills/use-skill-gadget.ts
-import { z as z4 } from "zod";
-function createUseSkillGadget(registry) {
-  const summaries = registry.getMetadataSummaries();
-  const skillNames = registry.getModelInvocable().map((s) => s.name);
-  const description = [
-    "Activate a skill to get specialized instructions for a task.",
-    "Available skills:",
-    summaries
-  ].join("\n");
-  return createGadget({
-    name: USE_SKILL_GADGET_NAME,
-    description,
-    schema: z4.object({
-      skill: z4.enum(skillNames).describe("Name of the skill to activate"),
-      arguments: z4.string().optional().describe("Arguments for the skill (e.g., a filename, issue number, or search query)")
-    }),
-    execute: async ({ skill: skillName, arguments: args }) => {
-      const skill = registry.get(skillName);
-      if (!skill) {
-        return `Unknown skill: "${skillName}". Available skills: ${skillNames.join(", ")}`;
-      }
-      const activation = await skill.activate({
-        arguments: args,
-        cwd: process.cwd()
-      });
-      return activation.resolvedInstructions;
-    }
-  });
-}
-var USE_SKILL_GADGET_NAME;
-var init_use_skill_gadget = __esm({
-  "src/skills/use-skill-gadget.ts"() {
-    "use strict";
-    init_create_gadget();
-    USE_SKILL_GADGET_NAME = "UseSkill";
-  }
-});
-
 // src/agent/builder-utils.ts
 function formatGadgetCall(gadgetName, invocationId, parameters, prefixes) {
   const startPrefix = prefixes?.start ?? GADGET_START_PREFIX;
@@ -7350,7 +7391,7 @@ var init_constants2 = __esm({
   "src/providers/constants.ts"() {
     "use strict";
     ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS = 4096;
-    FALLBACK_CHARS_PER_TOKEN = 4;
+    FALLBACK_CHARS_PER_TOKEN = 2;
     OPENAI_MESSAGE_OVERHEAD_TOKENS = 4;
     OPENAI_REPLY_PRIMING_TOKENS = 2;
     OPENAI_NAME_FIELD_OVERHEAD_TOKENS = 1;
@@ -9705,6 +9746,7 @@ var init_huggingface_models = __esm({
 
 // src/providers/openai-compatible-provider.ts
 import OpenAI from "openai";
+import { get_encoding } from "tiktoken";
 var ROLE_MAP, OpenAICompatibleProvider;
 var init_openai_compatible_provider = __esm({
   "src/providers/openai-compatible-provider.ts"() {
@@ -9909,11 +9951,38 @@ var init_openai_compatible_provider = __esm({
         }
       }
       /**
-       * Count tokens using character-based fallback estimation.
-       * Most meta-providers don't have a native token counting API.
+       * Count tokens using tiktoken o200k_base encoding.
+       *
+       * While o200k_base isn't model-exact for non-OpenAI models routed through
+       * meta-providers like OpenRouter, BPE tokenizers with 200K vocab produce
+       * counts within 10-20% of true values — far better than the character-based
+       * fallback which can be off by 250% for JSON/code-heavy content.
+       *
+       * Falls back to character-based estimation if tiktoken fails.
        */
       async countTokens(messages, descriptor, _spec) {
+        if (!messages || messages.length === 0) return 0;
         try {
+          const encoding = get_encoding("o200k_base");
+          try {
+            let tokenCount = 0;
+            for (const msg of messages) {
+              const parts = normalizeMessageContent(msg.content);
+              for (const part of parts) {
+                if (part.type === "text") {
+                  tokenCount += encoding.encode(part.text).length;
+                }
+              }
+            }
+            return tokenCount;
+          } finally {
+            encoding.free();
+          }
+        } catch (error) {
+          console.warn(
+            `Token counting with tiktoken failed for ${descriptor.name}, using fallback estimation:`,
+            error
+          );
           let totalChars = 0;
           for (const msg of messages) {
             const parts = normalizeMessageContent(msg.content);
@@ -9924,9 +9993,6 @@ var init_openai_compatible_provider = __esm({
             }
           }
           return Math.ceil(totalChars / FALLBACK_CHARS_PER_TOKEN);
-        } catch (error) {
-          console.warn(`Token counting failed for ${descriptor.name}, using zero estimate:`, error);
-          return 0;
         }
       }
     };
@@ -12588,6 +12654,7 @@ var init_client = __esm({
     "use strict";
     init_builder();
     init_discovery();
+    init_constants();
     init_model_registry();
     init_image();
     init_speech();
@@ -12706,8 +12773,43 @@ var init_client = __esm({
         if (adapter.countTokens) {
           return adapter.countTokens(messages, descriptor, spec);
         }
-        const totalChars = messages.reduce((sum, msg) => sum + (msg.content?.length ?? 0), 0);
-        return Math.ceil(totalChars / 4);
+        try {
+          const { get_encoding: get_encoding2 } = await import("tiktoken");
+          const encoding = get_encoding2("o200k_base");
+          try {
+            let tokenCount = 0;
+            for (const msg of messages) {
+              const content = msg.content;
+              if (typeof content === "string") {
+                tokenCount += encoding.encode(content).length;
+              } else if (Array.isArray(content)) {
+                for (const part of content) {
+                  if (part.type === "text") {
+                    tokenCount += encoding.encode(part.text).length;
+                  }
+                }
+              }
+            }
+            return tokenCount;
+          } finally {
+            encoding.free();
+          }
+        } catch {
+          let totalChars = 0;
+          for (const msg of messages) {
+            const content = msg.content;
+            if (typeof content === "string") {
+              totalChars += content.length;
+            } else if (Array.isArray(content)) {
+              for (const part of content) {
+                if (part.type === "text") {
+                  totalChars += part.text.length;
+                }
+              }
+            }
+          }
+          return Math.ceil(totalChars / CHARS_PER_TOKEN);
+        }
       }
       resolveAdapter(descriptor) {
         const adapter = this.adapters.find((item) => item.supports(descriptor));
@@ -12846,9 +12948,9 @@ var init_builder = __esm({
     init_model_shortcuts();
     init_registry();
     init_activation();
+    init_load_skill_gadget();
     init_loader();
     init_parser();
-    init_use_skill_gadget();
     init_agent();
     init_agent_internal_key();
     init_builder_utils();
@@ -13183,7 +13285,7 @@ ${resolved}`);
         const skillRegistry = this.resolveSkillRegistry();
         if (skillRegistry && skillRegistry.size > 0) {
           if (skillRegistry.getModelInvocable().length > 0) {
-            registry.registerByClass(createUseSkillGadget(skillRegistry));
+            registry.registerByClass(createLoadSkillGadget(skillRegistry));
           }
           const preActivatedBlock = this.resolvePreActivatedInstructions(skillRegistry);
           if (preActivatedBlock) {
@@ -16372,7 +16474,8 @@ var init_agent = __esm({
           this.compactionManager = new CompactionManager(
             this.client,
             this.model,
-            options.compactionConfig
+            options.compactionConfig,
+            this.logger
           );
         }
         this.signal = options.signal;
@@ -16718,6 +16821,22 @@ var init_agent = __esm({
                 this.logger.info("Loop terminated by gadget or processor");
                 break;
               }
+              if (this.compactionManager && result.usage?.inputTokens) {
+                this.compactionManager.updateUsage(result.usage.inputTokens);
+                if (this.compactionManager.shouldCompactFromUsage()) {
+                  this.logger.info("Reactive compaction triggered from API-reported usage", {
+                    inputTokens: result.usage.inputTokens,
+                    iteration: currentIteration
+                  });
+                  const reactiveCompaction = await this.compactionManager.compact(
+                    this.conversation,
+                    currentIteration
+                  );
+                  if (reactiveCompaction) {
+                    yield await this.emitCompactionEvent(reactiveCompaction, currentIteration);
+                  }
+                }
+              }
               if (this.budget !== void 0) {
                 const totalCost = this.tree.getTotalCost();
                 if (totalCost >= this.budget) {
@@ -17587,11 +17706,11 @@ var SimpleSessionManager = class extends BaseSessionManager {
 
 // src/skills/index.ts
 init_activation();
+init_load_skill_gadget();
 init_loader();
 init_parser();
 init_registry2();
 init_skill();
-init_use_skill_gadget();
 
 // src/utils/format.ts
 function truncate(text3, maxLength, suffix = "...") {
@@ -17779,6 +17898,7 @@ export {
   HybridStrategy,
   LLMMessageBuilder,
   LLMist,
+  LOAD_SKILL_GADGET_NAME,
   MODEL_ALIASES,
   MediaStore,
   ModelIdentifierParser,
@@ -17795,7 +17915,6 @@ export {
   SummarizationStrategy,
   TaskCompletionSignal,
   TimeoutException,
-  USE_SKILL_GADGET_NAME,
   audioFromBase64,
   audioFromBuffer,
   collectEvents,
@@ -17808,12 +17927,12 @@ export {
   createGeminiProviderFromEnv,
   createHints,
   createHuggingFaceProviderFromEnv,
+  createLoadSkillGadget,
   createLogger,
   createMediaOutput,
   createOpenAIProviderFromEnv,
   createOpenRouterProviderFromEnv,
   createSubagent,
-  createUseSkillGadget,
   defaultLogger,
   detectAudioMimeType,
   detectImageMimeType,