@illuma-ai/agents 1.0.90 → 1.0.93

package/src/utils/title.js

@@ -0,0 +1,119 @@
+import { ChatPromptTemplate } from '@langchain/core/prompts';
+import { RunnableLambda, RunnableSequence } from '@langchain/core/runnables';
+import { ContentTypes } from '@/common';
+const defaultTitlePrompt = `Analyze this conversation and provide:
+1. The detected language of the conversation
+2. A concise title in the detected language (5 words or less, no punctuation or quotation)
+
+{convo}`;
+const titleSchema = {
+    type: 'object',
+    properties: {
+        title: {
+            type: 'string',
+            description: 'A concise title for the conversation in 5 words or less, without punctuation or quotation',
+        },
+    },
+    required: ['title'],
+};
+const combinedSchema = {
+    type: 'object',
+    properties: {
+        language: {
+            type: 'string',
+            description: 'The detected language of the conversation',
+        },
+        title: {
+            type: 'string',
+            description: 'A concise title for the conversation in 5 words or less, without punctuation or quotation',
+        },
+    },
+    required: ['language', 'title'],
+};
+export const createTitleRunnable = async (model, _titlePrompt) => {
+    // Disabled since this works fine
+    // eslint-disable-next-line @typescript-eslint/ban-ts-comment
+    /* @ts-ignore */
+    const titleLLM = model.withStructuredOutput(titleSchema);
+    // eslint-disable-next-line @typescript-eslint/ban-ts-comment
+    /* @ts-ignore */
+    const combinedLLM = model.withStructuredOutput(combinedSchema);
+    const titlePrompt = ChatPromptTemplate.fromTemplate(_titlePrompt ?? defaultTitlePrompt).withConfig({ runName: 'TitlePrompt' });
+    const titleOnlyInnerChain = RunnableSequence.from([titlePrompt, titleLLM]);
+    const combinedInnerChain = RunnableSequence.from([titlePrompt, combinedLLM]);
+    /** Wrap titleOnlyChain in RunnableLambda to create parent span */
+    const titleOnlyChain = new RunnableLambda({
+        func: async (input, config) => {
+            const result = await titleOnlyInnerChain.invoke(input, config);
+            return result;
+        },
+    }).withConfig({ runName: 'TitleOnlyChain' });
+    /** Wrap combinedChain in RunnableLambda to create parent span */
+    const combinedChain = new RunnableLambda({
+        func: async (input, config) => {
+            const result = await combinedInnerChain.invoke(input, config);
+            return result;
+        },
+    }).withConfig({ runName: 'TitleLanguageChain' });
+    /** Runnable to add default values if needed */
+    const addDefaults = new RunnableLambda({
+        func: (result) => ({
+            language: result?.language ?? 'English',
+            title: result?.title ?? '',
+        }),
+    }).withConfig({ runName: 'AddDefaults' });
+    const combinedChainInner = RunnableSequence.from([
+        combinedChain,
+        addDefaults,
+    ]);
+    /** Wrap combinedChainWithDefaults in RunnableLambda to create parent span */
+    const combinedChainWithDefaults = new RunnableLambda({
+        func: async (input, config) => {
+            return await combinedChainInner.invoke(input, config);
+        },
+    }).withConfig({ runName: 'CombinedChainWithDefaults' });
+    return new RunnableLambda({
+        func: async (input, config) => {
+            const invokeInput = { convo: input.convo };
+            if (input.skipLanguage) {
+                return (await titleOnlyChain.invoke(invokeInput, config));
+            }
+            return await combinedChainWithDefaults.invoke(invokeInput, config);
+        },
+    }).withConfig({ runName: 'TitleGenerator' });
+};
+const defaultCompletionPrompt = `Provide a concise, 5-word-or-less title for the conversation, using title case conventions. Only return the title itself.
+
+Conversation:
+{convo}`;
+export const createCompletionTitleRunnable = async (model, titlePrompt) => {
+    const completionPrompt = ChatPromptTemplate.fromTemplate(titlePrompt ?? defaultCompletionPrompt).withConfig({ runName: 'CompletionTitlePrompt' });
+    /** Runnable to extract content from model response */
+    const extractContent = new RunnableLambda({
+        func: (response) => {
+            let content = '';
+            if (typeof response.content === 'string') {
+                content = response.content;
+            }
+            else if (Array.isArray(response.content)) {
+                content = response.content
+                    .filter((part) => part.type === ContentTypes.TEXT)
+                    .map((part) => part.text)
+                    .join('');
+            }
+            return { title: content.trim() };
+        },
+    }).withConfig({ runName: 'ExtractTitle' });
+    const innerChain = RunnableSequence.from([
+        completionPrompt,
+        model,
+        extractContent,
+    ]);
+    /** Wrap in RunnableLambda to create a parent span for LangFuse */
+    return new RunnableLambda({
+        func: async (input, config) => {
+            return await innerChain.invoke(input, config);
+        },
+    }).withConfig({ runName: 'CompletionTitleChain' });
+};
+//# sourceMappingURL=title.js.map

package/src/utils/title.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"title.js","sourceRoot":"","sources":["title.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,kBAAkB,EAAE,MAAM,yBAAyB,CAAC;AAC7D,OAAO,EAAE,cAAc,EAAE,gBAAgB,EAAE,MAAM,2BAA2B,CAAC;AAI7E,OAAO,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAExC,MAAM,kBAAkB,GAAG;;;;QAInB,CAAC;AAET,MAAM,WAAW,GAAG;IAClB,IAAI,EAAE,QAAQ;IACd,UAAU,EAAE;QACV,KAAK,EAAE;YACL,IAAI,EAAE,QAAQ;YACd,WAAW,EACT,2FAA2F;SAC9F;KACF;IACD,QAAQ,EAAE,CAAC,OAAO,CAAC;CACX,CAAC;AAEX,MAAM,cAAc,GAAG;IACrB,IAAI,EAAE,QAAQ;IACd,UAAU,EAAE;QACV,QAAQ,EAAE;YACR,IAAI,EAAE,QAAQ;YACd,WAAW,EAAE,2CAA2C;SACzD;QACD,KAAK,EAAE;YACL,IAAI,EAAE,QAAQ;YACd,WAAW,EACT,2FAA2F;SAC9F;KACF;IACD,QAAQ,EAAE,CAAC,UAAU,EAAE,OAAO,CAAC;CACvB,CAAC;AAEX,MAAM,CAAC,MAAM,mBAAmB,GAAG,KAAK,EACtC,KAA0B,EAC1B,YAAqB,EACF,EAAE;IACrB,iCAAiC;IACjC,6DAA6D;IAC7D,gBAAgB;IAChB,MAAM,QAAQ,GAAG,KAAK,CAAC,oBAAoB,CAAC,WAAW,CAAC,CAAC;IACzD,6DAA6D;IAC7D,gBAAgB;IAChB,MAAM,WAAW,GAAG,KAAK,CAAC,oBAAoB,CAAC,cAAc,CAAC,CAAC;IAE/D,MAAM,WAAW,GAAG,kBAAkB,CAAC,YAAY,CACjD,YAAY,IAAI,kBAAkB,CACnC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,aAAa,EAAE,CAAC,CAAC;IAEzC,MAAM,mBAAmB,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE,QAAQ,CAAC,CAAC,CAAC;IAC3E,MAAM,kBAAkB,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE,WAAW,CAAC,CAAC,CAAC;IAE7E,kEAAkE;IAClE,MAAM,cAAc,GAAG,IAAI,cAAc,CAAC;QACxC,IAAI,EAAE,KAAK,EACT,KAAwB,EACxB,MAAgC,EACJ,EAAE;YAC9B,MAAM,MAAM,GAAG,MAAM,mBAAmB,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;YAC/D,OAAO,MAA2B,CAAC;QACrC,CAAC;KACF,CAAC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,gBAAgB,EAAE,CAAC,CAAC;IAE7C,iEAAiE;IACjE,MAAM,aAAa,GAAG,IAAI,cAAc,CAAC;QACvC,IAAI,EAAE,KAAK,EACT,KAAwB,EACxB,MAAgC,EACc,EAAE;YAChD,MAAM,MAAM,GAAG,MAAM,kBAAkB,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;YAC9D,OAAO,MAA6C,CAAC;QACvD,CAAC;KACF,CAAC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,oBAAoB,EAAE,CAAC,CAAC;IAEjD,+CAA+C;IAC/C,MAAM,WAAW,GAAG,IAAI,cAAc,CAAC;QACrC,IAAI,EAAE,CACJ,MAAuD,EAClB,EAAE,CAAC,CAAC;YACzC,QAAQ,EAAE,MAAM,EAAE,QAAQ,IAAI,SAAS;YACvC,KAAK,EAAE,MAAM,EAAE,KAAK,IAAI,EAAE;SAC3B,CAAC;KACH,CAAC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,aAAa,EAAE,CAAC,CAAC;IAE1C,MAAM,kBAAkB,GAAG,gBAAgB,CAAC,IAAI,CAAC;QAC/C,aAAa;QACb,WAAW;KACZ,CAAC,CAAC;IAEH,6EAA6E;IAC7E,MAAM,yBAAyB,GAAG,IAAI,cAAc,CAAC;QACnD,IAAI,EAAE,KAAK,EACT,KAAwB,EACxB,MAAgC,EACc,EAAE;YAChD,OAAO,MAAM,kBAAkB,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACxD,CAAC;KACF,CAAC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,2BAA2B,EAAE,CAAC,CAAC;IAExD,OAAO,IAAI,cAAc,CAAC;QACxB,IAAI,EAAE,KAAK,EACT,KAIC,EACD,MAAgC,EACkC,EAAE;YACpE,MAAM,WAAW,GAAG,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC;YAE3C,IAAI,KAAK,CAAC,YAAY,EAAE,CAAC;gBACvB,OAAO,CAAC,MAAM,cAAc,CAAC,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,CAEvD,CAAC;YACJ,CAAC;YAED,OAAO,MAAM,yBAAyB,CAAC,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,CAAC;QACrE,CAAC;KACF,CAAC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,gBAAgB,EAAE,CAAC,CAAC;AAC/C,CAAC,CAAC;AAEF,MAAM,uBAAuB,GAAG;;;QAGxB,CAAC;AAET,MAAM,CAAC,MAAM,6BAA6B,GAAG,KAAK,EAChD,KAA0B,EAC1B,WAAoB,EACD,EAAE;IACrB,MAAM,gBAAgB,GAAG,kBAAkB,CAAC,YAAY,CACtD,WAAW,IAAI,uBAAuB,CACvC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,uBAAuB,EAAE,CAAC,CAAC;IAEnD,sDAAsD;IACtD,MAAM,cAAc,GAAG,IAAI,cAAc,CAAC;QACxC,IAAI,EAAE,CAAC,QAAmB,EAAqB,EAAE;YAC/C,IAAI,OAAO,GAAG,EAAE,CAAC;YACjB,IAAI,OAAO,QAAQ,CAAC,OAAO,KAAK,QAAQ,EAAE,CAAC;gBACzC,OAAO,GAAG,QAAQ,CAAC,OAAO,CAAC;YAC7B,CAAC;iBAAM,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE,CAAC;gBAC3C,OAAO,GAAG,QAAQ,CAAC,OAAO;qBACvB,MAAM,CACL,CAAC,IAAI,EAAqD,EAAE,CAC1D,IAAI,CAAC,IAAI,KAAK,YAAY,CAAC,IAAI,CAClC;qBACA,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;qBACxB,IAAI,CAAC,EAAE,CAAC,CAAC;YACd,CAAC;YACD,OAAO,EAAE,KAAK,EAAE,OAAO,CAAC,IAAI,EAAE,EAAE,CAAC;QACnC,CAAC;KACF,CAAC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,cAAc,EAAE,CAAC,CAAC;IAE3C,MAAM,UAAU,GAAG,gBAAgB,CAAC,IAAI,CAAC;QACvC,gBAAgB;QAChB,KAAK;QACL,cAAc;KACf,CAAC,CAAC;IAEH,kEAAkE;IAClE,OAAO,IAAI,cAAc,CAAC;QACxB,IAAI,EAAE,KAAK,EACT,KAAwB,EACxB,MAAgC,EACJ,EAAE;YAC9B,OAAO,MAAM,UAAU,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QAChD,CAAC;KACF,CAAC,CAAC,UAAU,CAAC,EAAE,OAAO,EAAE,sBAAsB,EAAE,CAAC,CAAC;AACrD,CAAC,CAAC"}

package/src/utils/tokens.js

@@ -0,0 +1,92 @@
+import { Tokenizer } from 'ai-tokenizer';
+import { ContentTypes } from '@/common/enum';
+const tokenizers = {};
+async function getTokenizer(encoding = 'o200k_base') {
+    const cached = tokenizers[encoding];
+    if (cached) {
+        return cached;
+    }
+    const data = encoding === 'claude'
+        ? await import('ai-tokenizer/encoding/claude')
+        : await import('ai-tokenizer/encoding/o200k_base');
+    const instance = new Tokenizer(data);
+    tokenizers[encoding] = instance;
+    return instance;
+}
+export function encodingForModel(model) {
+    if (model.toLowerCase().includes('claude')) {
+        return 'claude';
+    }
+    return 'o200k_base';
+}
+export function getTokenCountForMessage(message, getTokenCount) {
+    const tokensPerMessage = 3;
+    const processValue = (value) => {
+        if (Array.isArray(value)) {
+            for (const item of value) {
+                if (!item ||
+                    !item.type ||
+                    item.type === ContentTypes.ERROR ||
+                    item.type === ContentTypes.IMAGE_URL) {
+                    continue;
+                }
+                if (item.type === ContentTypes.TOOL_CALL && item.tool_call != null) {
+                    const toolName = item.tool_call?.name || '';
+                    if (toolName != null && toolName && typeof toolName === 'string') {
+                        numTokens += getTokenCount(toolName);
+                    }
+                    const args = item.tool_call?.args || '';
+                    if (args != null && args && typeof args === 'string') {
+                        numTokens += getTokenCount(args);
+                    }
+                    const output = item.tool_call?.output || '';
+                    if (output != null && output && typeof output === 'string') {
+                        numTokens += getTokenCount(output);
+                    }
+                    continue;
+                }
+                const nestedValue = item[item.type];
+                if (!nestedValue) {
+                    continue;
+                }
+                processValue(nestedValue);
+            }
+        }
+        else if (typeof value === 'string') {
+            numTokens += getTokenCount(value);
+        }
+        else if (typeof value === 'number') {
+            numTokens += getTokenCount(value.toString());
+        }
+        else if (typeof value === 'boolean') {
+            numTokens += getTokenCount(value.toString());
+        }
+    };
+    let numTokens = tokensPerMessage;
+    processValue(message.content);
+    return numTokens;
+}
+/**
+ * Creates a token counter function using the specified encoding.
+ * Lazily loads the encoding data on first use via dynamic import.
+ */
+export const createTokenCounter = async (encoding = 'o200k_base') => {
+    const tok = await getTokenizer(encoding);
+    const countTokens = (text) => tok.count(text);
+    return (message) => getTokenCountForMessage(message, countTokens);
+};
+/** Utility to manage the token encoder lifecycle explicitly. */
+export const TokenEncoderManager = {
+    async initialize() {
+        // No-op: ai-tokenizer is synchronously initialized from bundled data.
+    },
+    reset() {
+        for (const key of Object.keys(tokenizers)) {
+            delete tokenizers[key];
+        }
+    },
+    isInitialized() {
+        return Object.keys(tokenizers).length > 0;
+    },
+};
+//# sourceMappingURL=tokens.js.map

package/src/utils/tokens.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"tokens.js","sourceRoot":"","sources":["tokens.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,SAAS,EAAE,MAAM,cAAc,CAAC;AAEzC,OAAO,EAAE,YAAY,EAAE,MAAM,eAAe,CAAC;AAI7C,MAAM,UAAU,GAA6C,EAAE,CAAC;AAEhE,KAAK,UAAU,YAAY,CACzB,WAAyB,YAAY;IAErC,MAAM,MAAM,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC;IACpC,IAAI,MAAM,EAAE,CAAC;QACX,OAAO,MAAM,CAAC;IAChB,CAAC;IACD,MAAM,IAAI,GACR,QAAQ,KAAK,QAAQ;QACnB,CAAC,CAAC,MAAM,MAAM,CAAC,8BAA8B,CAAC;QAC9C,CAAC,CAAC,MAAM,MAAM,CAAC,kCAAkC,CAAC,CAAC;IACvD,MAAM,QAAQ,GAAG,IAAI,SAAS,CAAC,IAAI,CAAC,CAAC;IACrC,UAAU,CAAC,QAAQ,CAAC,GAAG,QAAQ,CAAC;IAChC,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,MAAM,UAAU,gBAAgB,CAAC,KAAa;IAC5C,IAAI,KAAK,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAE,CAAC;QAC3C,OAAO,QAAQ,CAAC;IAClB,CAAC;IACD,OAAO,YAAY,CAAC;AACtB,CAAC;AAED,MAAM,UAAU,uBAAuB,CACrC,OAAoB,EACpB,aAAuC;IAEvC,MAAM,gBAAgB,GAAG,CAAC,CAAC;IAE3B,MAAM,YAAY,GAAG,CAAC,KAAc,EAAQ,EAAE;QAC5C,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;YACzB,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;gBACzB,IACE,CAAC,IAAI;oBACL,CAAC,IAAI,CAAC,IAAI;oBACV,IAAI,CAAC,IAAI,KAAK,YAAY,CAAC,KAAK;oBAChC,IAAI,CAAC,IAAI,KAAK,YAAY,CAAC,SAAS,EACpC,CAAC;oBACD,SAAS;gBACX,CAAC;gBAED,IAAI,IAAI,CAAC,IAAI,KAAK,YAAY,CAAC,SAAS,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,EAAE,CAAC;oBACnE,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC;oBAC5C,IAAI,QAAQ,IAAI,IAAI,IAAI,QAAQ,IAAI,OAAO,QAAQ,KAAK,QAAQ,EAAE,CAAC;wBACjE,SAAS,IAAI,aAAa,CAAC,QAAQ,CAAC,CAAC;oBACvC,CAAC;oBAED,MAAM,IAAI,GAAG,IAAI,CAAC,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC;oBACxC,IAAI,IAAI,IAAI,IAAI,IAAI,IAAI,IAAI,OAAO,IAAI,KAAK,QAAQ,EAAE,CAAC;wBACrD,SAAS,IAAI,aAAa,CAAC,IAAI,CAAC,CAAC;oBACnC,CAAC;oBAED,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,EAAE,MAAM,IAAI,EAAE,CAAC;oBAC5C,IAAI,MAAM,IAAI,IAAI,IAAI,MAAM,IAAI,OAAO,MAAM,KAAK,QAAQ,EAAE,CAAC;wBAC3D,SAAS,IAAI,aAAa,CAAC,MAAM,CAAC,CAAC;oBACrC,CAAC;oBACD,SAAS;gBACX,CAAC;gBAED,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAEpC,IAAI,CAAC,WAAW,EAAE,CAAC;oBACjB,SAAS;gBACX,CAAC;gBAED,YAAY,CAAC,WAAW,CAAC,CAAC;YAC5B,CAAC;QACH,CAAC;aAAM,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;YACrC,SAAS,IAAI,aAAa,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;aAAM,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;YACrC,SAAS,IAAI,aAAa,CAAC,KAAK,CAAC,QAAQ,EAAE,CAAC,CAAC;QAC/C,CAAC;aAAM,IAAI,OAAO,KAAK,KAAK,SAAS,EAAE,CAAC;YACtC,SAAS,IAAI,aAAa,CAAC,KAAK,CAAC,QAAQ,EAAE,CAAC,CAAC;QAC/C,CAAC;IACH,CAAC,CAAC;IAEF,IAAI,SAAS,GAAG,gBAAgB,CAAC;IACjC,YAAY,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IAC9B,OAAO,SAAS,CAAC;AACnB,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAG,KAAK,EACrC,WAAyB,YAAY,EACM,EAAE;IAC7C,MAAM,GAAG,GAAG,MAAM,YAAY,CAAC,QAAQ,CAAC,CAAC;IACzC,MAAM,WAAW,GAAG,CAAC,IAAY,EAAU,EAAE,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IAC9D,OAAO,CAAC,OAAoB,EAAU,EAAE,CACtC,uBAAuB,CAAC,OAAO,EAAE,WAAW,CAAC,CAAC;AAClD,CAAC,CAAC;AAEF,gEAAgE;AAChE,MAAM,CAAC,MAAM,mBAAmB,GAAG;IACjC,KAAK,CAAC,UAAU;QACd,sEAAsE;IACxE,CAAC;IAED,KAAK;QACH,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,EAAE,CAAC;YAC1C,OAAO,UAAU,CAAC,GAAmB,CAAC,CAAC;QACzC,CAAC;IACH,CAAC;IAED,aAAa;QACX,OAAO,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC;IAC5C,CAAC;CACF,CAAC"}

package/src/utils/tokens.ts

@@ -1,142 +1,118 @@
-import { getEncoding, type Tiktoken } from 'js-tiktoken';
-import type { BaseMessage } from '@langchain/core/messages';
-import { ContentTypes } from '@/common/enum';
-
-export function getTokenCountForMessage(
-  message: BaseMessage,
-  getTokenCount: (text: string) => number
-): number {
-  const tokensPerMessage = 3;
-
-  const processValue = (value: unknown): void => {
-    if (Array.isArray(value)) {
-      for (const item of value) {
-        if (
-          !item ||
-          !item.type ||
-          item.type === ContentTypes.ERROR ||
-          item.type === ContentTypes.IMAGE_URL
-        ) {
-          continue;
-        }
-
-        if (item.type === ContentTypes.TOOL_CALL && item.tool_call != null) {
-          const toolName = item.tool_call?.name || '';
-          if (toolName != null && toolName && typeof toolName === 'string') {
-            numTokens += getTokenCount(toolName);
-          }
-
-          const args = item.tool_call?.args || '';
-          if (args != null && args && typeof args === 'string') {
-            numTokens += getTokenCount(args);
-          }
-
-          const output = item.tool_call?.output || '';
-          if (output != null && output && typeof output === 'string') {
-            numTokens += getTokenCount(output);
-          }
-          continue;
-        }
-
-        const nestedValue = item[item.type];
-
-        if (!nestedValue) {
-          continue;
-        }
-
-        processValue(nestedValue);
-      }
-    } else if (typeof value === 'string') {
-      numTokens += getTokenCount(value);
-    } else if (typeof value === 'number') {
-      numTokens += getTokenCount(value.toString());
-    } else if (typeof value === 'boolean') {
-      numTokens += getTokenCount(value.toString());
-    }
-  };
-
-  let numTokens = tokensPerMessage;
-  processValue(message.content);
-  return numTokens;
-}
-
-let encoderPromise: Promise<Tiktoken | null> | undefined;
-let tokenCounterPromise: Promise<(message: BaseMessage) => number> | undefined;
-
-/**
- * Simple character-based token estimation (~4 chars per token).
- * Used as fallback when tiktoken encoder fails to load.
- */
-const estimateTokens = (text: string): number => Math.ceil(text.length / 4);
-
-async function getSharedEncoder(): Promise<Tiktoken | null> {
-  if (encoderPromise) {
-    return encoderPromise;
-  }
-  // Use bundled tokenizer data - no external network calls
-  encoderPromise = Promise.resolve().then(() => {
-    try {
-      return getEncoding('o200k_base');
-    } catch {
-      // Fallback to null - will use character estimation
-      return null;
-    }
-  });
-  return encoderPromise;
-}
-
-/**
- * Creates a singleton token counter function that reuses the same encoder instance.
- * Falls back to character-based estimation (~4 chars per token) if encoder fails.
- * This ensures token counting never blocks LLM calls.
- */
-export const createTokenCounter = async (): Promise<
-  (message: BaseMessage) => number
-> => {
-  if (tokenCounterPromise) {
-    return tokenCounterPromise;
-  }
-
-  tokenCounterPromise = (async (): Promise<
-    (message: BaseMessage) => number
-  > => {
-    const enc = await getSharedEncoder();
-    const countTokens = enc
-      ? (text: string): number => enc.encode(text).length
-      : estimateTokens;
-    return (message: BaseMessage): number =>
-      getTokenCountForMessage(message, countTokens);
-  })();
-
-  return tokenCounterPromise;
-};
-
-/**
- * Utility to manage the token encoder lifecycle explicitly.
- * Useful for applications that need fine-grained control over resource management.
- */
-export const TokenEncoderManager = {
-  /**
-   * Pre-initializes the encoder. This can be called during app startup
-   * to avoid lazy loading delays later.
-   */
-  async initialize(): Promise<void> {
-    await getSharedEncoder();
-  },
-
-  /**
-   * Clears the cached encoder and token counter.
-   * Useful for testing or when you need to force a fresh reload.
-   */
-  reset(): void {
-    encoderPromise = undefined;
-    tokenCounterPromise = undefined;
-  },
-
-  /**
-   * Checks if the encoder has been initialized.
-   */
-  isInitialized(): boolean {
-    return encoderPromise !== undefined;
-  },
-};
+import { Tokenizer } from 'ai-tokenizer';
+import type { BaseMessage } from '@langchain/core/messages';
+import { ContentTypes } from '@/common/enum';
+
+export type EncodingName = 'o200k_base' | 'claude';
+
+const tokenizers: Partial<Record<EncodingName, Tokenizer>> = {};
+
+async function getTokenizer(
+  encoding: EncodingName = 'o200k_base'
+): Promise<Tokenizer> {
+  const cached = tokenizers[encoding];
+  if (cached) {
+    return cached;
+  }
+  const data =
+    encoding === 'claude'
+      ? await import('ai-tokenizer/encoding/claude')
+      : await import('ai-tokenizer/encoding/o200k_base');
+  const instance = new Tokenizer(data);
+  tokenizers[encoding] = instance;
+  return instance;
+}
+
+export function encodingForModel(model: string): EncodingName {
+  if (model.toLowerCase().includes('claude')) {
+    return 'claude';
+  }
+  return 'o200k_base';
+}
+
+export function getTokenCountForMessage(
+  message: BaseMessage,
+  getTokenCount: (text: string) => number
+): number {
+  const tokensPerMessage = 3;
+
+  const processValue = (value: unknown): void => {
+    if (Array.isArray(value)) {
+      for (const item of value) {
+        if (
+          !item ||
+          !item.type ||
+          item.type === ContentTypes.ERROR ||
+          item.type === ContentTypes.IMAGE_URL
+        ) {
+          continue;
+        }
+
+        if (item.type === ContentTypes.TOOL_CALL && item.tool_call != null) {
+          const toolName = item.tool_call?.name || '';
+          if (toolName != null && toolName && typeof toolName === 'string') {
+            numTokens += getTokenCount(toolName);
+          }
+
+          const args = item.tool_call?.args || '';
+          if (args != null && args && typeof args === 'string') {
+            numTokens += getTokenCount(args);
+          }
+
+          const output = item.tool_call?.output || '';
+          if (output != null && output && typeof output === 'string') {
+            numTokens += getTokenCount(output);
+          }
+          continue;
+        }
+
+        const nestedValue = item[item.type];
+
+        if (!nestedValue) {
+          continue;
+        }
+
+        processValue(nestedValue);
+      }
+    } else if (typeof value === 'string') {
+      numTokens += getTokenCount(value);
+    } else if (typeof value === 'number') {
+      numTokens += getTokenCount(value.toString());
+    } else if (typeof value === 'boolean') {
+      numTokens += getTokenCount(value.toString());
+    }
+  };
+
+  let numTokens = tokensPerMessage;
+  processValue(message.content);
+  return numTokens;
+}
+
+/**
+ * Creates a token counter function using the specified encoding.
+ * Lazily loads the encoding data on first use via dynamic import.
+ */
+export const createTokenCounter = async (
+  encoding: EncodingName = 'o200k_base'
+): Promise<(message: BaseMessage) => number> => {
+  const tok = await getTokenizer(encoding);
+  const countTokens = (text: string): number => tok.count(text);
+  return (message: BaseMessage): number =>
+    getTokenCountForMessage(message, countTokens);
+};
+
+/** Utility to manage the token encoder lifecycle explicitly. */
+export const TokenEncoderManager = {
+  async initialize(): Promise<void> {
+    // No-op: ai-tokenizer is synchronously initialized from bundled data.
+  },
+
+  reset(): void {
+    for (const key of Object.keys(tokenizers)) {
+      delete tokenizers[key as EncodingName];
+    }
+  },
+
+  isInitialized(): boolean {
+    return Object.keys(tokenizers).length > 0;
+  },
+};

package/src/utils/toolCallContinuation.ts

@@ -0,0 +1,55 @@
+/**
+ * toolCallContinuation — Utilities for detecting and classifying LLM finish
+ * reasons across providers. Used to detect max_tokens truncation and normalize
+ * stop reason extraction from provider-specific metadata formats.
+ *
+ * @module
+ */
+
+import type { AIMessageChunk } from '@langchain/core/messages';
+import { FinishReasons } from '@/common';
+
+/**
+ * Extract the finish/stop reason from an AIMessageChunk's response_metadata.
+ * Centralizes provider-specific metadata access patterns.
+ *
+ * Provider formats:
+ * - OpenAI/Azure: `response_metadata.finish_reason`
+ * - Anthropic direct: `response_metadata.stop_reason`
+ * - Bedrock invoke: `response_metadata.stopReason`
+ * - Bedrock streaming: `response_metadata.messageStop.stopReason`
+ * - VertexAI/Google: `response_metadata.finishReason`
+ *
+ * @param chunk - The final accumulated AIMessageChunk
+ * @returns The finish reason string, or undefined if not found
+ */
+export function extractFinishReason(
+  chunk: AIMessageChunk | undefined
+): string | undefined {
+  if (!chunk || !('response_metadata' in chunk)) return undefined;
+
+  const meta = chunk.response_metadata as Record<string, unknown>;
+  if (!meta) return undefined;
+
+  // Bedrock streaming nests stopReason inside messageStop: { stopReason: '...' }
+  const messageStop = meta.messageStop as Record<string, unknown> | undefined;
+
+  return (
+    (meta.finish_reason as string | undefined) ?? // OpenAI/Azure
+    (meta.stop_reason as string | undefined) ?? // Anthropic direct API
+    (meta.stopReason as string | undefined) ?? // Bedrock invoke (non-streaming)
+    (messageStop?.stopReason as string | undefined) ?? // Bedrock streaming
+    (meta.finishReason as string | undefined) // VertexAI/Google
+  );
+}
+
+/**
+ * Check if a finish reason indicates the response was truncated by token limits.
+ *
+ * @param reason - The finish reason from extractFinishReason()
+ * @returns true if the response was cut short by max_tokens
+ */
+export function isMaxTokensFinish(reason: string | undefined): boolean {
+  if (!reason) return false;
+  return reason === FinishReasons.MAX_TOKENS || reason === FinishReasons.LENGTH;
+}