@h-ai/ai 0.1.0-alpha.39 → 0.1.0-alpha.40

package/dist/index.js

@@ -1,5 +1,5 @@
-import { aiM, AIConfigSchema, MemoryConfigSchema, RetrievalConfigSchema, KnowledgeConfigSchema, TokenConfigSchema, SummaryConfigSchema, CompressConfigSchema, AudioConfigSchema, resolveModelEntry, resolveAudioModel, resolveModelApi } from './chunk-OZYROB52.js';
-export { A2AConfigSchema, A2ASkillConfigSchema, AIConfigSchema, AUDIO_WS_PATH, ApiTypeSchema, AudioConfigSchema, AudioFormatSchema, AudioModelEntrySchema, AudioProviderSchema, AudioWsClientMessageSchema, AudioWsDoneMessageSchema, AudioWsStartMessageSchema, AudioWsTextMessageSchema, CompressConfigSchema, CompressionStrategySchema, EmbeddingConfigSchema, EntityTypeSchema, FileConfigSchema, KnowledgeConfigSchema, LLMConfigSchema, MCPConfigSchema, MCPServerCapabilitiesSchema, MCPServerConfigSchema, MemoryConfigSchema, MemoryTypeSchema, ModelEntrySchema, ModelScenarioSchema, RetrievalConfigSchema, RetrievalSourceSchema, SummaryConfigSchema, TokenConfigSchema, resolveAudioModel, resolveModelApi, resolveModelEntry } from './chunk-OZYROB52.js';
+import { aiM, AIConfigSchema, MemoryConfigSchema, RetrievalConfigSchema, KnowledgeConfigSchema, TokenConfigSchema, SummaryConfigSchema, CompressConfigSchema, AudioConfigSchema, resolveModelEntry, resolveAudioModel, resolveModelApi } from './chunk-DRSZFSCT.js';
+export { A2AConfigSchema, A2ASkillConfigSchema, AIConfigSchema, AUDIO_WS_PATH, ApiTypeSchema, AudioConfigSchema, AudioFormatSchema, AudioModelEntrySchema, AudioProviderSchema, AudioWsClientMessageSchema, AudioWsDoneMessageSchema, AudioWsStartMessageSchema, AudioWsTextMessageSchema, CompressConfigSchema, CompressionStrategySchema, EmbeddingConfigSchema, EntityTypeSchema, FileConfigSchema, KnowledgeConfigSchema, LLMConfigSchema, MCPConfigSchema, MCPServerCapabilitiesSchema, MCPServerConfigSchema, MemoryConfigSchema, MemoryTypeSchema, ModelEntrySchema, ModelScenarioSchema, RetrievalConfigSchema, RetrievalSourceSchema, SummaryConfigSchema, TokenConfigSchema, resolveAudioModel, resolveModelApi, resolveModelEntry } from './chunk-DRSZFSCT.js';
 import { HaiAIError } from './chunk-P3GND76X.js';
 export { HaiAIError } from './chunk-P3GND76X.js';
 import { core, err, ok } from '@h-ai/core';
@@ -1553,6 +1553,86 @@ function createProvider(name) {
   }
 }
 var logger8 = core.logger.child({ module: "ai", scope: "compress" });
+var CONVERSATION_SUMMARY_PREFIX = "[Conversation Summary]";
+function completeCompression(result, maxTokens) {
+  if (result.compressedTokens > maxTokens) {
+    return err(
+      HaiAIError.CONTEXT_BUDGET_EXCEEDED,
+      aiM("ai_contextBudgetExceeded", {
+        params: {
+          tokens: result.compressedTokens,
+          budget: maxTokens
+        }
+      })
+    );
+  }
+  return ok(result);
+}
+function isGeneratedConversationSummary(message) {
+  return message.role === "system" && typeof message.content === "string" && message.content.startsWith(CONVERSATION_SUMMARY_PREFIX);
+}
+function partitionSummaryMessages(messages, preserveSystem) {
+  const isPreservedSystem = (message) => {
+    return preserveSystem && message.role === "system" && !isGeneratedConversationSummary(message);
+  };
+  return {
+    systemMessages: messages.filter(isPreservedSystem),
+    conversationMessages: messages.filter((message) => !isPreservedSystem(message))
+  };
+}
+function unitizeNonSystemMessages(messages) {
+  const units = [];
+  for (let index = 0; index < messages.length; index += 1) {
+    const message = messages[index];
+    if (message?.role !== "assistant" || !message.tool_calls?.length) {
+      if (message) {
+        units.push([message]);
+      }
+      continue;
+    }
+    const toolCallIds = new Set(message.tool_calls.map((toolCall) => toolCall.id));
+    const unit = [message];
+    let nextIndex = index + 1;
+    while (nextIndex < messages.length) {
+      const nextMessage = messages[nextIndex];
+      if (nextMessage?.role !== "tool" || !toolCallIds.has(nextMessage.tool_call_id)) {
+        break;
+      }
+      unit.push(nextMessage);
+      nextIndex += 1;
+    }
+    units.push(unit);
+    index = nextIndex - 1;
+  }
+  return units;
+}
+function flattenUnits(units) {
+  return units.flatMap((unit) => unit);
+}
+function splitProtectedSuffix(messages, preserveLastN) {
+  const units = unitizeNonSystemMessages(messages);
+  let protectedStart = units.length;
+  let protectedMessageCount = 0;
+  const requiredMessageCount = Math.max(0, preserveLastN);
+  while (protectedStart > 0 && protectedMessageCount < requiredMessageCount) {
+    protectedStart -= 1;
+    protectedMessageCount += units[protectedStart]?.length ?? 0;
+  }
+  let latestUserUnitIndex = -1;
+  for (let index = units.length - 1; index >= 0; index -= 1) {
+    if (units[index]?.some((message) => message.role === "user")) {
+      latestUserUnitIndex = index;
+      break;
+    }
+  }
+  if (latestUserUnitIndex >= 0) {
+    protectedStart = Math.min(protectedStart, latestUserUnitIndex);
+  }
+  return {
+    removableUnits: units.slice(0, protectedStart),
+    protectedUnits: units.slice(protectedStart)
+  };
+}
 function createCompressOperations(config, token, summary, modelMaxTokens) {
   function resolveMaxTokens(optionMaxTokens) {
     const fromOption = optionMaxTokens ?? config.defaultMaxTokens;
@@ -1570,29 +1650,38 @@ function createCompressOperations(config, token, summary, modelMaxTokens) {
         nonSystemMessages.push(msg);
       }
     }
-    const preserved = nonSystemMessages.slice(-preserveLastN);
-    const result = [...systemMessages, ...preserved];
-    const currentTokens = token.estimateMessages(result);
-    if (currentTokens <= maxTokens) {
+    const { removableUnits, protectedUnits } = splitProtectedSuffix(nonSystemMessages, preserveLastN);
+    const emptyMessageTokens = token.estimateMessages([]);
+    const estimateUnitTokens = (unit) => {
+      return Math.max(0, token.estimateMessages(unit) - emptyMessageTokens);
+    };
+    let retainedTokens = token.estimateMessages(systemMessages) + protectedUnits.reduce((total, unit) => total + estimateUnitTokens(unit), 0);
+    let retainedStart = removableUnits.length;
+    let retainedUnits = [...protectedUnits];
+    let finalMessages = [...systemMessages, ...flattenUnits(retainedUnits)];
+    if (retainedTokens > maxTokens) {
       return {
-        messages: result,
-        removedCount: nonSystemMessages.length - preserved.length
+        messages: finalMessages,
+        removedCount: flattenUnits(removableUnits).length
       };
     }
-    let finalMessages = [...systemMessages];
-    let tokens = token.estimateMessages(finalMessages);
-    const addable = [];
-    for (let i = nonSystemMessages.length - 1; i >= 0; i--) {
-      const msgTokens = token.estimateMessages([nonSystemMessages[i]]);
-      if (tokens + msgTokens > maxTokens)
+    for (let index = removableUnits.length - 1; index >= 0; index -= 1) {
+      const unit = removableUnits[index];
+      if (!unit) {
+        continue;
+      }
+      const unitTokens = estimateUnitTokens(unit);
+      if (retainedTokens + unitTokens > maxTokens) {
         break;
-      addable.unshift(nonSystemMessages[i]);
-      tokens += msgTokens;
+      }
+      retainedTokens += unitTokens;
+      retainedStart = index;
     }
-    finalMessages = [...systemMessages, ...addable];
+    retainedUnits = [...removableUnits.slice(retainedStart), ...protectedUnits];
+    finalMessages = [...systemMessages, ...flattenUnits(retainedUnits)];
     return {
       messages: finalMessages,
-      removedCount: nonSystemMessages.length - addable.length
+      removedCount: nonSystemMessages.length - flattenUnits(retainedUnits).length
     };
   }
   async function tryCompress(messages, options) {
@@ -1620,25 +1709,25 @@ function createCompressOperations(config, token, summary, modelMaxTokens) {
         );
         const compressedTokens = token.estimateMessages(compressed);
         logger8.trace("Sliding window compression completed", { originalTokens, compressedTokens, removedCount });
-        return ok({
+        return completeCompression({
           messages: compressed,
           originalTokens,
           compressedTokens,
           removedCount
-        });
+        }, maxTokens);
       }
       if (strategy === "summary") {
-        const systemMessages2 = preserveSystem ? messages.filter((m) => m.role === "system") : [];
-        const nonSystem2 = messages.filter((m) => m.role !== "system" || !preserveSystem);
-        const preserved = nonSystem2.slice(-preserveLastN);
-        const toSummarize2 = nonSystem2.slice(0, nonSystem2.length - preserveLastN);
+        const { systemMessages: systemMessages2, conversationMessages: conversationMessages2 } = partitionSummaryMessages(messages, preserveSystem);
+        const { removableUnits: removableUnits2, protectedUnits: protectedUnits2 } = splitProtectedSuffix(conversationMessages2, preserveLastN);
+        const preserved = flattenUnits(protectedUnits2);
+        const toSummarize2 = flattenUnits(removableUnits2);
         if (toSummarize2.length === 0) {
-          return ok({
+          return completeCompression({
             messages: [...messages],
             originalTokens,
             compressedTokens: originalTokens,
             removedCount: 0
-          });
+          }, maxTokens);
         }
         const summaryResult = await summary.generate(toSummarize2, { model: options?.summaryModel });
         if (!summaryResult.success)
@@ -1646,19 +1735,19 @@ function createCompressOperations(config, token, summary, modelMaxTokens) {
         const summaryText = summaryResult.data;
         const summaryMessage = {
           role: "system",
-          content: `[Conversation Summary]
+          content: `${CONVERSATION_SUMMARY_PREFIX}
 ${summaryText}`
         };
         const compressed = [...systemMessages2, summaryMessage, ...preserved];
         const compressedTokens = token.estimateMessages(compressed);
         logger8.trace("Summary compression completed", { originalTokens, compressedTokens, removedCount: toSummarize2.length });
-        return ok({
+        return completeCompression({
           messages: compressed,
           originalTokens,
           compressedTokens,
           removedCount: toSummarize2.length,
           summary: summaryText
-        });
+        }, maxTokens);
       }
       const { messages: windowResult, removedCount: windowRemoved } = slidingWindow(
         messages,
@@ -1668,17 +1757,17 @@ ${summaryText}`
       );
       const windowTokens = token.estimateMessages(windowResult);
       if (windowTokens <= maxTokens) {
-        return ok({
+        return completeCompression({
           messages: windowResult,
           originalTokens,
           compressedTokens: windowTokens,
           removedCount: windowRemoved
-        });
+        }, maxTokens);
       }
-      const systemMessages = preserveSystem ? messages.filter((m) => m.role === "system") : [];
-      const nonSystem = messages.filter((m) => m.role !== "system" || !preserveSystem);
-      const preservedMessages = nonSystem.slice(-preserveLastN);
-      const toSummarize = nonSystem.slice(0, nonSystem.length - preserveLastN);
+      const { systemMessages, conversationMessages } = partitionSummaryMessages(messages, preserveSystem);
+      const { removableUnits, protectedUnits } = splitProtectedSuffix(conversationMessages, preserveLastN);
+      const preservedMessages = flattenUnits(protectedUnits);
+      const toSummarize = flattenUnits(removableUnits);
       if (toSummarize.length > 0) {
         const summaryResult = await summary.generate(toSummarize, { model: options?.summaryModel });
         if (!summaryResult.success)
@@ -1686,26 +1775,26 @@ ${summaryText}`
         const summaryText = summaryResult.data;
         const summaryMessage = {
           role: "system",
-          content: `[Conversation Summary]
+          content: `${CONVERSATION_SUMMARY_PREFIX}
 ${summaryText}`
         };
         const compressed = [...systemMessages, summaryMessage, ...preservedMessages];
         const compressedTokens = token.estimateMessages(compressed);
         logger8.trace("Hybrid compression completed", { originalTokens, compressedTokens, removedCount: toSummarize.length });
-        return ok({
+        return completeCompression({
           messages: compressed,
           originalTokens,
           compressedTokens,
           removedCount: toSummarize.length,
           summary: summaryText
-        });
+        }, maxTokens);
       }
-      return ok({
+      return completeCompression({
         messages: windowResult,
         originalTokens,
         compressedTokens: windowTokens,
         removedCount: windowRemoved
-      });
+      }, maxTokens);
     } catch (error) {
       logger8.error("Context compression failed", { error });
       return err(HaiAIError.CONTEXT_COMPRESS_FAILED, aiM("ai_contextCompressFailed", { params: { error: String(error) } }), error);
@@ -1857,7 +1946,7 @@ function createContextOperations(compressConfig, tokenOps, compressOps, contextS
         });
         if (!compressResult.success) {
           logger9.warn("Auto-compression failed, keeping original messages", { error: compressResult.error });
-          return ok(void 0);
+          return compressResult;
         }
         if (compressResult.data.summary) {
           state.summaries.push({
@@ -2142,7 +2231,9 @@ function createContextOperations(compressConfig, tokenOps, compressOps, contextS
         const signal = combineSignal2(chatOpts?.signal, internalSignal);
         let streamTurn = null;
         try {
-          await manager.addMessage({ role: "user", content: message });
+          const addResult = await manager.addMessage({ role: "user", content: message });
+          if (!addResult.success)
+            throw addResult.error;
           const messagesResult = manager.getMessages();
           if (!messagesResult.success) {
             throw new Error(`Failed to get messages: ${String(messagesResult.error)}`);
@@ -6915,37 +7006,11 @@ function estimateMessagesTokens(messages, tokenRatio = 0.25) {
   let total = 0;
   for (const msg of messages) {
     total += 4;
-    const content = getMessageContent(msg);
-    total += estimateTextTokens(content, tokenRatio);
+    total += estimateTextTokens(JSON.stringify(msg), tokenRatio);
   }
   total += 2;
   return total;
 }
-function getMessageContent(msg) {
-  if (msg.role === "assistant") {
-    const c = msg.content;
-    if (typeof c === "string")
-      return c;
-    if (Array.isArray(c))
-      return c.filter((p) => p.type === "text").map((p) => p.text ?? "").join(" ");
-    return "";
-  }
-  if (msg.role === "tool") {
-    const c = msg.content;
-    if (typeof c === "string")
-      return c;
-    if (Array.isArray(c))
-      return c.filter((p) => p.type === "text").map((p) => p.text ?? "").join(" ");
-    return "";
-  }
-  const content = msg.content;
-  if (typeof content === "string")
-    return content;
-  if (Array.isArray(content)) {
-    return content.filter((c) => c.type === "text").map((c) => c.text).join(" ");
-  }
-  return "";
-}
 function createTokenOperations(config) {
   const { tokenRatio } = config;
   return {