agency-lang 0.0.66 → 0.0.67

package/dist/lib/templates/backends/graphGenerator/goToNode.d.ts

@@ -1,4 +1,4 @@
-export declare const template = "goToNode(\"{{{nodeName:string}}}\",\n  {\n    messages: __self.messages,\n    __metadata: {\n      graph: __graph,\n      statelogClient,\n      callbacks: __callbacks,\n    },\n    {{#hasData}}\n    data: {{{data:string}}}\n    {{/hasData}}\n    {{^hasData}}\n    data: null\n    {{/hasData}}\n  }\n);";
+export declare const template = "goToNode(\"{{{nodeName:string}}}\",\n  {\n    messages: __self.messages,\n    __metadata: {\n      graph: __graph,\n      // we need to pass in the state log client here because\n      // if we rely on the local state log client\n      // each client in each file has a different trace id.\n      // So we pass in the client to make sure they all use the same trace id\n\n      statelogClient,\n      callbacks: __callbacks,\n    },\n    {{#hasData}}\n    data: {{{data:string}}}\n    {{/hasData}}\n    {{^hasData}}\n    data: null\n    {{/hasData}}\n  }\n);";
 export type TemplateType = {
     nodeName: string;
     hasData: boolean;

package/dist/lib/templates/backends/graphGenerator/goToNode.js

@@ -7,6 +7,11 @@ export const template = `goToNode("{{{nodeName:string}}}",
     messages: __self.messages,
     __metadata: {
       graph: __graph,
+      // we need to pass in the state log client here because
+      // if we rely on the local state log client
+      // each client in each file has a different trace id.
+      // So we pass in the client to make sure they all use the same trace id
+
       statelogClient,
       callbacks: __callbacks,
     },

package/dist/lib/templates/backends/graphGenerator/imports.d.ts

@@ -1,4 +1,4 @@
-export declare const template = "import { fileURLToPath } from \"url\";\nimport process from \"process\";\nimport { z } from \"zod\";\nimport * as readline from \"readline\";\nimport fs from \"fs\";\nimport { StatelogClient, SimpleMachine, goToNode, nanoid } from \"agency-lang\";\nimport * as smoltalk from \"agency-lang\";\n\n/* Code to log to statelog */\nconst statelogHost = \"{{{logHost:string}}}\";\nconst __traceId = nanoid();\nconst statelogConfig = {\n  host: statelogHost,\n  traceId: __traceId,\n  {{#hasApiKey}}\n  apiKey: \"{{{logApiKey?:string}}}\",\n  {{/hasApiKey}}\n  {{^hasApiKey}}\n  apiKey: process.env.STATELOG_API_KEY || \"\",\n  {{/hasApiKey}}\n  projectId: \"{{{logProjectId:string}}}\",\n  debugMode: {{{logDebugMode:boolean}}},\n};\nconst __statelogClient = new StatelogClient(statelogConfig);\n\n/* Code for Smoltalk client */\nconst __model = \"{{{clientDefaultModel:string}}}\";\n\nconst __getClientWithConfig = (config = {}) => {\n  const defaultConfig = {\n    {{#hasOpenAiApiKey}}\n    openAiApiKey: \"{{{clientOpenAiApiKey?:string}}}\",\n    {{/hasOpenAiApiKey}}\n    {{^hasOpenAiApiKey}}\n    openAiApiKey: process.env.OPENAI_API_KEY || \"\",\n    {{/hasOpenAiApiKey}}\n    {{#hasGoogleApiKey}}\n    googleApiKey: \"{{{clientGoogleApiKey?:string}}}\",\n    {{/hasGoogleApiKey}}\n    {{^hasGoogleApiKey}}\n    googleApiKey: process.env.GEMINI_API_KEY || \"\",\n    {{/hasGoogleApiKey}}\n    model: __model,\n    logLevel: \"{{{clientLogLevel:string}}}\",\n  };\n\n  return smoltalk.getClient({ ...defaultConfig, ...config });\n};\n\nlet __client = __getClientWithConfig();\n\n/* Code for SimpleMachine graph */\n\n// enable debug logging\nconst graphConfig = {\n  debug: {\n    log: true,\n    logData: false,\n  },\n  statelog: statelogConfig,\n};\n\nconst graph = new SimpleMachine(graphConfig);\n\n/******** builtins ********/\n\nconst not = (val) => !val;\nconst eq = (a, b) => a === b;\nconst neq = (a, b) => a !== b;\nconst lt = (a, b) => a < b;\nconst lte = (a, b) => a <= b;\nconst gt = (a, b) => a > b;\nconst gte = (a, b) => a >= b;\nconst and = (a, b) => a && b;\nconst or = (a, b) => a || b;\nconst head = (arr) => arr[0];\nconst tail = (arr) => arr.slice(1);\nconst empty = (arr) => arr.length === 0;\n\nasync function _builtinFetch(url, args = {}) {\n  const result = await fetch(url, args);\n  try {\n    const text = await result.text();\n    return text;\n  } catch (e) {\n    throw new Error(`Failed to get text from ${url}: ${e}`);\n  }\n}\n\nasync function _builtinFetchJSON(url, args = {}) {\n  const result = await fetch(url, args);\n  try {\n    const json = await result.json();\n    return json;\n  } catch (e) {\n    throw new Error(`Failed to parse JSON from ${url}: ${e}`);\n  }\n}\n\nfunction _builtinInput(prompt) {\n  const rl = readline.createInterface({\n    input: process.stdin,\n    output: process.stdout,\n  });\n\n  return new Promise((resolve) => {\n    rl.question(prompt, (answer) => {\n      rl.close();\n      resolve(answer);\n    });\n  });\n}\n\nfunction _builtinRead(filename) {\n  const data = fs.readFileSync(filename);\n  const contents = data.toString(\"utf8\");\n  return contents;\n}\n\n/*\n * @param filePath The absolute or relative path to the image file.\n * @returns The Base64 string, or null if an error occurs.\n */\nfunction _builtinReadImage(filePath) {\n  const data = fs.readFileSync(filePath); // Synchronous file reading\n  const base64String = data.toString(\"base64\");\n  return base64String;\n}\n\nfunction _builtinSleep(seconds) {\n  return new Promise((resolve) => {\n    setTimeout(resolve, seconds * 1000);\n  });\n}\n\nfunction printJSON(obj) {\n  console.log(JSON.stringify(obj, null, 2));\n}\n\nexport const __readSkillTool = {\n  name: \"readSkill\",\n  description: `Skills provide specialized knowledge and instructions for particular scenarios.\nUse this tool when you need enhanced guidance for a specific type of task.\n\nArgs:\n    filepath: The name of the skill to read.\n\nReturns:\n    The skill content with specialized instructions, or an error message\n    if the skill is not found.\n`,\n  schema: z.object({\"filepath\": z.string(), })\n};\n\nexport function readSkill({filepath}) {\n  return _builtinRead(filepath);\n}\n\n/******** for internal agency use only ********/\n\nfunction __createReturnObject(result) {\n  // Note: we're *not* using structuredClone here because structuredClone\n  // doesn't call `toJSON`, so it's not cloning our message objects correctly.\n  return JSON.parse(JSON.stringify({\n    messages: result.messages,\n    data: result.data,\n    tokens: __stateStack.globals.__tokenStats\n  }));\n}\n\n\n\n/******** interrupts ********/\n\nexport function interrupt(data) {\n  return {\n    type: \"interrupt\",\n    data,\n  };\n}\n\nexport function isInterrupt(obj) {\n  return obj && obj.type === \"interrupt\";\n}\n\nexport async function respondToInterrupt(\n  _interrupt,\n  _interruptResponse,\n  metadata = {},\n) {\n  const interrupt = structuredClone(_interrupt);\n  const interruptResponse = structuredClone(_interruptResponse);\n\n  __stateStack = StateStack.fromJSON(interrupt.__state || {});\n  __stateStack.deserializeMode();\n\n  const messages = (__stateStack.interruptData.messages || []).map(\n    (json) => {\n      // create message objects from JSON\n      return smoltalk.messageFromJSON(json);\n    },\n  );\n  __stateStack.interruptData.messages = messages;\n  __stateStack.interruptData.interruptResponse = interruptResponse;\n\n  if (interruptResponse.type === \"approve\" && interruptResponse.newArguments) {\n    __stateStack.interruptData.toolCall = {\n      ...__stateStack.interruptData.toolCall,\n      arguments: {\n        ...__stateStack.interruptData.toolCall.arguments,\n        ...interruptResponse.newArguments,\n      },\n    };\n    // Error:\n    // TypeError: Cannot set property arguments of #<ToolCall> which has only a getter\n    //         toolCall.arguments = { ...toolCall.arguments, ...interruptResponse.newArguments };\n    //\n    // const lastMessage = __stateStack.interruptData.messages[__stateStack.interruptData.messages.length - 1];\n    // if (lastMessage && lastMessage.role === \"assistant\") {\n    //   const toolCall = lastMessage.toolCalls?.[lastMessage.toolCalls.length - 1];\n    //   if (toolCall) {\n    //     toolCall.arguments = { ...toolCall.arguments, ...interruptResponse.newArguments };\n    //   }\n    // }\n  }\n\n  // start at the last node we visited\n  const nodesTraversed = __stateStack.interruptData.nodesTraversed || [];\n  const nodeName = nodesTraversed[nodesTraversed.length - 1];\n  const __result = await graph.run(nodeName, {\n    messages: messages,\n    __metadata: {\n      graph: graph,\n      statelogClient: __statelogClient,\n      __stateStack: __stateStack,\n      __callbacks: metadata.callbacks,\n    },\n\n    // restore args from the state stack\n    data: \"<from-stack>\",\n  });\n  return __createReturnObject(__result);\n}\n\nexport async function approveInterrupt(\n  interrupt,\n  metadata = {},\n) {\n  return await respondToInterrupt(interrupt, { type: \"approve\" }, metadata);\n}\n\nexport async function modifyInterrupt(\n  interrupt,\n  newArguments,\n  metadata = {},\n) {\n  return await respondToInterrupt(\n    interrupt,\n    { type: \"approve\", newArguments },\n    metadata,\n  );\n}\n\nexport async function rejectInterrupt(\n  interrupt,\n  metadata = {},\n) {\n  return await respondToInterrupt(interrupt, { type: \"reject\" }, metadata);\n}\n\n/****** StateStack and related functions for serializing/deserializing execution state during interrupts ********/\n\n// See docs for notes on how this works.\nclass StateStack {\n  stack = [];\n  mode = \"serialize\";\n  globals = {};\n  other = {};\n  interruptData = {};\n\n  deserializeStackLength = 0;\n\n  constructor(\n    stack = [],\n    mode = \"serialize\",\n  ) {\n    this.stack = stack;\n    this.mode = mode;\n  }\n\n  getNewState() {\n    if (this.mode === \"deserialize\" && this.deserializeStackLength <= 0) {\n      console.log(\"Forcing mode to serialize, nothing left to deserialize\");\n      this.mode = \"serialize\";\n    }\n    if (this.mode === \"serialize\") {\n      const newState = {\n        args: {},\n        locals: {},\n        step: 0,\n      };\n      this.stack.push(newState);\n      return newState;\n    } else if (this.mode === \"deserialize\") {\n      this.deserializeStackLength -= 1;\n      const item = this.stack.shift();\n      this.stack.push(item);\n      return item;\n    }\n    return null;\n  }\n\n  deserializeMode() {\n    this.mode = \"deserialize\";\n    this.deserializeStackLength = this.stack.length;\n  }\n\n  pop() {\n    return this.stack.pop();\n  }\n\n  toJSON() {\n    return structuredClone({\n      stack: this.stack,\n      globals: this.globals,\n      other: this.other,\n      interruptData: this.interruptData,\n      mode: this.mode,\n      deserializeStackLength: this.deserializeStackLength,\n    });\n  }\n\n  static fromJSON(json) {\n    const stateStack = new StateStack([], \"serialize\");\n    stateStack.stack = json.stack || [];\n    stateStack.globals = json.globals || {};\n    stateStack.other = json.other || {};\n    stateStack.interruptData = json.interruptData || {};\n    stateStack.mode = json.mode || \"serialize\";\n    stateStack.deserializeStackLength = json.deserializeStackLength || 0;\n    return stateStack;\n  }\n}\n\nlet __stateStack = new StateStack();\n\n__stateStack.globals.__tokenStats = {\n  usage: {\n    inputTokens: 0,\n    outputTokens: 0,\n    cachedInputTokens: 0,\n    totalTokens: 0,\n  },\n  cost: {\n    inputCost: 0,\n    outputCost: 0,\n    totalCost: 0,\n    currency: \"USD\",\n  },\n};\n\nfunction __updateTokenStats(\n  usage,\n  cost,\n) {\n  if (!usage || !cost) return;\n  const tokenStats = __stateStack.globals.__tokenStats;\n  tokenStats.usage.inputTokens += usage.inputTokens || 0;\n  tokenStats.usage.outputTokens += usage.outputTokens || 0;\n  tokenStats.usage.cachedInputTokens += usage.cachedInputTokens || 0;\n  tokenStats.usage.totalTokens += usage.totalTokens || 0;\n\n  tokenStats.cost.inputCost += cost.inputCost || 0;\n  tokenStats.cost.outputCost += cost.outputCost || 0;\n  tokenStats.cost.totalCost += cost.totalCost || 0;\n}\n\n/**** Streaming callback and lock ****/\nfunction isGenerator(variable) {\n  const toString = Object.prototype.toString.call(variable);\n  return (\n    toString === \"[object Generator]\" || toString === \"[object AsyncGenerator]\"\n  );\n}\n\nlet __callbacks = {};\n\nasync function __callHook(name, data) {\n  if (__callbacks[name]) {\n    await __callbacks[name](data);\n  }\n}\n\nlet onStreamLock = false;\n\nfunction __cloneArray(arr) {\n  if (arr == undefined) return [];\n  return [...arr];\n}\n\nconst handleStreamingResponse = async (__completion) => {\n  if (isGenerator(__completion)) {\n    if (!__callbacks.onStream) {\n      console.log(\n        \"No onStream callback provided for streaming response, returning response synchronously\",\n      );\n      statelogClient.debug(\n        \"Got streaming response but no onStream callback provided, returning response synchronously\",\n        {\n          prompt: __prompt,\n          callbacks: Object.keys(__callbacks),\n        },\n      );\n      let syncResult = \"\";\n      for await (const chunk of __completion) {\n        switch (chunk.type) {\n          case \"tool_call\":\n            __toolCalls.push(chunk.toolCall);\n            break;\n          case \"done\":\n            syncResult = chunk.result;\n            break;\n          case \"error\":\n            console.error(`Error in LLM response stream: ${chunk.error}`);\n            break;\n          default:\n            break;\n        }\n      }\n      __completion = { success: true, value: syncResult };\n    } else {\n      // try to acquire lock\n      let count = 0;\n      // wait 60 seconds to acquire lock\n      while (onStreamLock && count < 10 * 60) {\n        await _builtinSleep(0.1);\n        count++;\n      }\n      if (onStreamLock) {\n        console.log(`Couldn't acquire lock, ${count}`);\n      }\n      onStreamLock = true;\n\n      for await (const chunk of __completion) {\n        switch (chunk.type) {\n          case \"text\":\n            __callbacks.onStream({ type: \"text\", text: chunk.text });\n            break;\n          case \"tool_call\":\n            __toolCalls.push(chunk.toolCall);\n            __callbacks.onStream({\n              type: \"tool_call\",\n              toolCall: chunk.toolCall,\n            });\n            break;\n          case \"done\":\n            __callbacks.onStream({ type: \"done\", result: chunk.result });\n            __completion = { success: true, value: chunk.result };\n            break;\n          case \"error\":\n            __callbacks.onStream({ type: \"error\", error: chunk.error });\n            break;\n        }\n      }\n\n      onStreamLock = false;\n    }\n  }\n};\n\n\n/**** Message thread handling ****/\n\nclass MessageThread {\n  messages = [];\n  children = [];\n\n  constructor(messages = []) {\n    this.messages = messages;\n    this.children = [];\n  }\n\n  addMessage(message) {\n    this.messages.push(message);\n  }\n\n  cloneMessages() {\n    return this.messages.map(m => m.toJSON()).map(m => smoltalk.messageFromJSON(m));\n  }\n\n  getMessages() {\n    return this.messages;\n  }\n\n  setMessages(messages) {\n    this.messages = messages;\n  }\n\n  newChild() {\n    const child = new MessageThread();\n    this.children.push(child);\n    return child;\n  }\n\n  newSubthreadChild() {\n    const child = new MessageThread(this.cloneMessages());\n    this.children.push(child);\n    return child;\n  }\n\n  toJSON() {\n    return {\n      messages: this.messages.map(m => m.toJSON()),\n      children: this.children.map((child) => child.toJSON()),\n    };\n  }\n}";
+export declare const template = "import { fileURLToPath } from \"url\";\nimport process from \"process\";\nimport { z } from \"zod\";\nimport * as readline from \"readline\";\nimport fs from \"fs\";\nimport { StatelogClient, SimpleMachine, goToNode, nanoid } from \"agency-lang\";\nimport * as smoltalk from \"agency-lang\";\n\n/* Code to log to statelog */\nconst statelogHost = \"{{{logHost:string}}}\";\nconst __traceId = nanoid();\nconst statelogConfig = {\n  host: statelogHost,\n  traceId: __traceId,\n  {{#hasApiKey}}\n  apiKey: \"{{{logApiKey?:string}}}\",\n  {{/hasApiKey}}\n  {{^hasApiKey}}\n  apiKey: process.env.STATELOG_API_KEY || \"\",\n  {{/hasApiKey}}\n  projectId: \"{{{logProjectId:string}}}\",\n  debugMode: {{{logDebugMode:boolean}}},\n};\nconst __statelogClient = new StatelogClient(statelogConfig);\n\n/* Code for Smoltalk client */\nconst __model = \"{{{clientDefaultModel:string}}}\";\n\nconst __getClientWithConfig = (config = {}) => {\n  const defaultConfig = {\n    {{#hasOpenAiApiKey}}\n    openAiApiKey: \"{{{clientOpenAiApiKey?:string}}}\",\n    {{/hasOpenAiApiKey}}\n    {{^hasOpenAiApiKey}}\n    openAiApiKey: process.env.OPENAI_API_KEY || \"\",\n    {{/hasOpenAiApiKey}}\n    {{#hasGoogleApiKey}}\n    googleApiKey: \"{{{clientGoogleApiKey?:string}}}\",\n    {{/hasGoogleApiKey}}\n    {{^hasGoogleApiKey}}\n    googleApiKey: process.env.GEMINI_API_KEY || \"\",\n    {{/hasGoogleApiKey}}\n    model: __model,\n    logLevel: \"{{{clientLogLevel:string}}}\",\n  };\n\n  return smoltalk.getClient({ ...defaultConfig, ...config });\n};\n\nlet __client = __getClientWithConfig();\n\n/* Code for SimpleMachine graph */\n\n// enable debug logging\nconst graphConfig = {\n  debug: {\n    log: true,\n    logData: false,\n  },\n  statelog: statelogConfig,\n};\n\nconst graph = new SimpleMachine(graphConfig);\n\n/******** builtins ********/\n\nconst not = (val) => !val;\nconst eq = (a, b) => a === b;\nconst neq = (a, b) => a !== b;\nconst lt = (a, b) => a < b;\nconst lte = (a, b) => a <= b;\nconst gt = (a, b) => a > b;\nconst gte = (a, b) => a >= b;\nconst and = (a, b) => a && b;\nconst or = (a, b) => a || b;\nconst head = (arr) => arr[0];\nconst tail = (arr) => arr.slice(1);\nconst empty = (arr) => arr.length === 0;\n\nasync function _builtinFetch(url, args = {}) {\n  const result = await fetch(url, args);\n  try {\n    const text = await result.text();\n    return text;\n  } catch (e) {\n    throw new Error(`Failed to get text from ${url}: ${e}`);\n  }\n}\n\nasync function _builtinFetchJSON(url, args = {}) {\n  const result = await fetch(url, args);\n  try {\n    const json = await result.json();\n    return json;\n  } catch (e) {\n    throw new Error(`Failed to parse JSON from ${url}: ${e}`);\n  }\n}\n\nfunction _builtinInput(prompt) {\n  const rl = readline.createInterface({\n    input: process.stdin,\n    output: process.stdout,\n  });\n\n  return new Promise((resolve) => {\n    rl.question(prompt, (answer) => {\n      rl.close();\n      resolve(answer);\n    });\n  });\n}\n\nfunction _builtinRead(filename) {\n  const data = fs.readFileSync(filename);\n  const contents = data.toString(\"utf8\");\n  return contents;\n}\n\n/*\n * @param filePath The absolute or relative path to the image file.\n * @returns The Base64 string, or null if an error occurs.\n */\nfunction _builtinReadImage(filePath) {\n  const data = fs.readFileSync(filePath); // Synchronous file reading\n  const base64String = data.toString(\"base64\");\n  return base64String;\n}\n\nfunction _builtinSleep(seconds) {\n  return new Promise((resolve) => {\n    setTimeout(resolve, seconds * 1000);\n  });\n}\n\nfunction printJSON(obj) {\n  console.log(JSON.stringify(obj, null, 2));\n}\n\nexport const __readSkillTool = {\n  name: \"readSkill\",\n  description: `Skills provide specialized knowledge and instructions for particular scenarios.\nUse this tool when you need enhanced guidance for a specific type of task.\n\nArgs:\n    filepath: The name of the skill to read.\n\nReturns:\n    The skill content with specialized instructions, or an error message\n    if the skill is not found.\n`,\n  schema: z.object({\"filepath\": z.string(), })\n};\n\nexport function readSkill({filepath}) {\n  return _builtinRead(filepath);\n}\n\n/******** for internal agency use only ********/\n\nfunction __createReturnObject(result) {\n  // Note: we're *not* using structuredClone here because structuredClone\n  // doesn't call `toJSON`, so it's not cloning our message objects correctly.\n  return JSON.parse(JSON.stringify({\n    messages: result.messages,\n    data: result.data,\n    tokens: __stateStack.globals.__tokenStats\n  }));\n}\n\n\n\n/******** interrupts ********/\n\nexport function interrupt(data) {\n  return {\n    type: \"interrupt\",\n    data,\n  };\n}\n\nexport function isInterrupt(obj) {\n  return obj && obj.type === \"interrupt\";\n}\n\nexport async function respondToInterrupt(\n  _interrupt,\n  _interruptResponse,\n  metadata = {},\n) {\n  const interrupt = structuredClone(_interrupt);\n  const interruptResponse = structuredClone(_interruptResponse);\n\n  __stateStack = StateStack.fromJSON(interrupt.__state || {});\n  __stateStack.deserializeMode();\n\n  const messages = (__stateStack.interruptData.messages || []).map(\n    (json) => {\n      // create message objects from JSON\n      return smoltalk.messageFromJSON(json);\n    },\n  );\n  __stateStack.interruptData.messages = messages;\n  __stateStack.interruptData.interruptResponse = interruptResponse;\n\n  if (interruptResponse.type === \"approve\" && interruptResponse.newArguments) {\n    __stateStack.interruptData.toolCall = {\n      ...__stateStack.interruptData.toolCall,\n      arguments: {\n        ...__stateStack.interruptData.toolCall.arguments,\n        ...interruptResponse.newArguments,\n      },\n    };\n    // Error:\n    // TypeError: Cannot set property arguments of #<ToolCall> which has only a getter\n    //         toolCall.arguments = { ...toolCall.arguments, ...interruptResponse.newArguments };\n    //\n    // const lastMessage = __stateStack.interruptData.messages[__stateStack.interruptData.messages.length - 1];\n    // if (lastMessage && lastMessage.role === \"assistant\") {\n    //   const toolCall = lastMessage.toolCalls?.[lastMessage.toolCalls.length - 1];\n    //   if (toolCall) {\n    //     toolCall.arguments = { ...toolCall.arguments, ...interruptResponse.newArguments };\n    //   }\n    // }\n  }\n\n  // start at the last node we visited\n  const nodesTraversed = __stateStack.interruptData.nodesTraversed || [];\n  const nodeName = nodesTraversed[nodesTraversed.length - 1];\n  const __result = await graph.run(nodeName, {\n    messages: messages,\n    __metadata: {\n      graph: graph,\n      // we need to pass in the state log client here because\n      // if we rely on the local state log client\n      // each client in each file has a different trace id.\n      // So we pass in the client to make sure they all use the same trace id\n      statelogClient: __statelogClient,\n      __stateStack: __stateStack,\n      __callbacks: metadata.callbacks,\n    },\n\n    // restore args from the state stack\n    data: \"<from-stack>\",\n  });\n  return __createReturnObject(__result);\n}\n\nexport async function approveInterrupt(\n  interrupt,\n  metadata = {},\n) {\n  return await respondToInterrupt(interrupt, { type: \"approve\" }, metadata);\n}\n\nexport async function modifyInterrupt(\n  interrupt,\n  newArguments,\n  metadata = {},\n) {\n  return await respondToInterrupt(\n    interrupt,\n    { type: \"approve\", newArguments },\n    metadata,\n  );\n}\n\nexport async function rejectInterrupt(\n  interrupt,\n  metadata = {},\n) {\n  return await respondToInterrupt(interrupt, { type: \"reject\" }, metadata);\n}\n\n/****** StateStack and related functions for serializing/deserializing execution state during interrupts ********/\n\n// See docs for notes on how this works.\nclass StateStack {\n  stack = [];\n  mode = \"serialize\";\n  globals = {};\n  other = {};\n  interruptData = {};\n\n  deserializeStackLength = 0;\n\n  constructor(\n    stack = [],\n    mode = \"serialize\",\n  ) {\n    this.stack = stack;\n    this.mode = mode;\n  }\n\n  getNewState() {\n    if (this.mode === \"deserialize\" && this.deserializeStackLength <= 0) {\n      console.log(\"Forcing mode to serialize, nothing left to deserialize\");\n      this.mode = \"serialize\";\n    }\n    if (this.mode === \"serialize\") {\n      const newState = {\n        args: {},\n        locals: {},\n        step: 0,\n      };\n      this.stack.push(newState);\n      return newState;\n    } else if (this.mode === \"deserialize\") {\n      this.deserializeStackLength -= 1;\n      const item = this.stack.shift();\n      this.stack.push(item);\n      return item;\n    }\n    return null;\n  }\n\n  deserializeMode() {\n    this.mode = \"deserialize\";\n    this.deserializeStackLength = this.stack.length;\n  }\n\n  pop() {\n    return this.stack.pop();\n  }\n\n  toJSON() {\n    return structuredClone({\n      stack: this.stack,\n      globals: this.globals,\n      other: this.other,\n      interruptData: this.interruptData,\n      mode: this.mode,\n      deserializeStackLength: this.deserializeStackLength,\n    });\n  }\n\n  static fromJSON(json) {\n    const stateStack = new StateStack([], \"serialize\");\n    stateStack.stack = json.stack || [];\n    stateStack.globals = json.globals || {};\n    stateStack.other = json.other || {};\n    stateStack.interruptData = json.interruptData || {};\n    stateStack.mode = json.mode || \"serialize\";\n    stateStack.deserializeStackLength = json.deserializeStackLength || 0;\n    return stateStack;\n  }\n}\n\nlet __stateStack = new StateStack();\n\n__stateStack.globals.__tokenStats = {\n  usage: {\n    inputTokens: 0,\n    outputTokens: 0,\n    cachedInputTokens: 0,\n    totalTokens: 0,\n  },\n  cost: {\n    inputCost: 0,\n    outputCost: 0,\n    totalCost: 0,\n    currency: \"USD\",\n  },\n};\n\nfunction __updateTokenStats(\n  usage,\n  cost,\n) {\n  if (!usage || !cost) return;\n  const tokenStats = __stateStack.globals.__tokenStats;\n  tokenStats.usage.inputTokens += usage.inputTokens || 0;\n  tokenStats.usage.outputTokens += usage.outputTokens || 0;\n  tokenStats.usage.cachedInputTokens += usage.cachedInputTokens || 0;\n  tokenStats.usage.totalTokens += usage.totalTokens || 0;\n\n  tokenStats.cost.inputCost += cost.inputCost || 0;\n  tokenStats.cost.outputCost += cost.outputCost || 0;\n  tokenStats.cost.totalCost += cost.totalCost || 0;\n}\n\n/**** Streaming callback and lock ****/\nfunction isGenerator(variable) {\n  const toString = Object.prototype.toString.call(variable);\n  return (\n    toString === \"[object Generator]\" || toString === \"[object AsyncGenerator]\"\n  );\n}\n\nlet __callbacks = {};\n\nasync function __callHook(name, data) {\n  if (__callbacks[name]) {\n    await __callbacks[name](data);\n  }\n}\n\nlet onStreamLock = false;\n\nfunction __cloneArray(arr) {\n  if (arr == undefined) return [];\n  return [...arr];\n}\n\nconst handleStreamingResponse = async (__completion, statelogClient, __prompt, __toolCalls) => {\n  if (isGenerator(__completion)) {\n    if (!__callbacks.onStream) {\n      console.log(\n        \"No onStream callback provided for streaming response, returning response synchronously\",\n      );\n      statelogClient.debug(\n        \"Got streaming response but no onStream callback provided, returning response synchronously\",\n        {\n          prompt: __prompt,\n          callbacks: Object.keys(__callbacks),\n        },\n      );\n      let syncResult = \"\";\n      for await (const chunk of __completion) {\n        switch (chunk.type) {\n          case \"tool_call\":\n            __toolCalls.push(chunk.toolCall);\n            break;\n          case \"done\":\n            syncResult = chunk.result;\n            break;\n          case \"error\":\n            console.error(`Error in LLM response stream: ${chunk.error}`);\n            break;\n          default:\n            break;\n        }\n      }\n      return { success: true, value: syncResult };\n    } else {\n      // try to acquire lock\n      let count = 0;\n      // wait 60 seconds to acquire lock\n      while (onStreamLock && count < 10 * 60) {\n        await _builtinSleep(0.1);\n        count++;\n      }\n      if (onStreamLock) {\n        console.log(`Couldn't acquire lock, ${count}`);\n      }\n      onStreamLock = true;\n\n      for await (const chunk of __completion) {\n        switch (chunk.type) {\n          case \"text\":\n            __callbacks.onStream({ type: \"text\", text: chunk.text });\n            break;\n          case \"tool_call\":\n            __toolCalls.push(chunk.toolCall);\n            __callbacks.onStream({\n              type: \"tool_call\",\n              toolCall: chunk.toolCall,\n            });\n            break;\n          case \"done\":\n            __callbacks.onStream({ type: \"done\", result: chunk.result });\n            return { success: true, value: chunk.result };\n          case \"error\":\n            __callbacks.onStream({ type: \"error\", error: chunk.error });\n            break;\n        }\n      }\n\n      onStreamLock = false;\n    }\n  }\n};\n\n\n/**** Message thread handling ****/\n\nclass MessageThread {\n  messages = [];\n  children = [];\n\n  constructor(messages = []) {\n    this.messages = messages;\n    this.children = [];\n  }\n\n  addMessage(message) {\n    this.messages.push(message);\n  }\n\n  cloneMessages() {\n    return this.messages.map(m => m.toJSON()).map(m => smoltalk.messageFromJSON(m));\n  }\n\n  getMessages() {\n    return this.messages;\n  }\n\n  setMessages(messages) {\n    this.messages = messages;\n  }\n\n  newChild() {\n    const child = new MessageThread();\n    this.children.push(child);\n    return child;\n  }\n\n  newSubthreadChild() {\n    const child = new MessageThread(this.cloneMessages());\n    this.children.push(child);\n    return child;\n  }\n\n  toJSON() {\n    return {\n      messages: this.messages.map(m => m.toJSON()),\n      children: this.children.map((child) => child.toJSON()),\n    };\n  }\n}";
 export type TemplateType = {
     logHost: string;
     hasApiKey: boolean;

package/dist/lib/templates/backends/graphGenerator/imports.js

@@ -235,6 +235,10 @@ export async function respondToInterrupt(
     messages: messages,
     __metadata: {
       graph: graph,
+      // we need to pass in the state log client here because
+      // if we rely on the local state log client
+      // each client in each file has a different trace id.
+      // So we pass in the client to make sure they all use the same trace id
       statelogClient: __statelogClient,
       __stateStack: __stateStack,
       __callbacks: metadata.callbacks,
@@ -402,7 +406,7 @@ function __cloneArray(arr) {
   return [...arr];
 }
 
-const handleStreamingResponse = async (__completion) => {
+const handleStreamingResponse = async (__completion, statelogClient, __prompt, __toolCalls) => {
   if (isGenerator(__completion)) {
     if (!__callbacks.onStream) {
       console.log(
@@ -431,7 +435,7 @@ const handleStreamingResponse = async (__completion) => {
             break;
         }
       }
-      __completion = { success: true, value: syncResult };
+      return { success: true, value: syncResult };
     } else {
       // try to acquire lock
       let count = 0;
@@ -459,8 +463,7 @@ const handleStreamingResponse = async (__completion) => {
             break;
           case "done":
             __callbacks.onStream({ type: "done", result: chunk.result });
-            __completion = { success: true, value: chunk.result };
-            break;
+            return { success: true, value: chunk.result };
           case "error":
             __callbacks.onStream({ type: "error", error: chunk.error });
             break;

package/dist/lib/templates/backends/typescriptGenerator/promptFunction.d.ts

@@ -1,4 +1,4 @@
-export declare const template = "\nasync function _{{{variableName:string}}}({{{argsStr:string}}}) {\n  const __prompt = {{{promptCode:string}}};\n  const startTime = performance.now();\n  let __messages = __metadata?.messages || [];\n\n  // These are to restore state after interrupt.\n  // TODO I think this could be implemented in a cleaner way.\n  let __toolCalls = __stateStack.interruptData?.toolCall ? [__stateStack.interruptData.toolCall] : [];\n  const __interruptResponse = __stateStack.interruptData?.interruptResponse || null;\n  const __tools = {{{tools}}};\n\n  {{#hasResponseFormat}}\n  // Need to make sure this is always an object\n  const __responseFormat = z.object({\n     response: {{{zodSchema:string}}}\n  });\n  {{/hasResponseFormat}}\n  {{^hasResponseFormat}}\n  const __responseFormat = undefined;\n  {{/hasResponseFormat}}\n  \n  const __client = __getClientWithConfig({{{clientConfig:string}}});\n  let responseMessage;\n\n  if (__toolCalls.length === 0) {\n    __messages.push(smoltalk.userMessage(__prompt));\n  \n  \n    await __callHook(\"onLLMCallStart\", { prompt: __prompt, tools: __tools, model: __client.getModel() });\n    let __completion = await __client.text({\n      messages: __messages,\n      tools: __tools,\n      responseFormat: __responseFormat,\n      stream: {{{isStreaming:boolean}}}\n    });\n\n    const endTime = performance.now();\n\n    await handleStreamingResponse(__completion);\n\n    statelogClient.promptCompletion({\n      messages: __messages,\n      completion: __completion,\n      model: __client.getModel(),\n      timeTaken: endTime - startTime,\n      tools: __tools,\n      responseFormat: __responseFormat\n    });\n\n    if (!__completion.success) {\n      throw new Error(\n        `Error getting response from ${__model}: ${__completion.error}`\n      );\n    }\n\n    responseMessage = __completion.value;\n    __toolCalls = responseMessage.toolCalls || [];\n\n    if (__toolCalls.length > 0) {\n      // Add assistant's response with tool calls to message history\n      __messages.push(smoltalk.assistantMessage(responseMessage.output, { toolCalls: __toolCalls }));\n    }\n\n    __updateTokenStats(responseMessage.usage, responseMessage.cost);\n    await __callHook(\"onLLMCallEnd\", { result: responseMessage, usage: responseMessage.usage, cost: responseMessage.cost, timeTaken: endTime - startTime });\n\n  }\n\n  // Handle function calls\n  if (__toolCalls.length > 0) {\n    let toolCallStartTime, toolCallEndTime;\n    let haltExecution = false;\n    let haltToolCall = {}\n    let haltInterrupt = null;\n\n    // Process each tool call\n    for (const toolCall of __toolCalls) {\n      {{{functionCalls:string}}}\n    }\n\n    if (haltExecution) {\n      statelogClient.debug(`Tool call interrupted execution.`, {\n        messages: __messages,\n        model: __client.getModel(),\n      });\n\n      __stateStack.interruptData = {\n        messages: __messages.map((msg) => msg.toJSON()),\n        nodesTraversed: __graph.getNodesTraversed(),\n        toolCall: haltToolCall,\n      };\n      haltInterrupt.__state = __stateStack.toJSON();\n      return haltInterrupt;\n    }\n  \n    const nextStartTime = performance.now();\n    await __callHook(\"onLLMCallStart\", { prompt: __prompt, tools: __tools, model: __client.getModel() });\n    let __completion = await __client.text({\n      messages: __messages,\n      tools: __tools,\n      responseFormat: __responseFormat,\n      stream: {{{isStreaming:boolean}}}\n    });\n\n    const nextEndTime = performance.now();\n\n    await handleStreamingResponse(__completion);\n\n    statelogClient.promptCompletion({\n      messages: __messages,\n      completion: __completion,\n      model: __client.getModel(),\n      timeTaken: nextEndTime - nextStartTime,\n      tools: __tools,\n      responseFormat: __responseFormat,\n    });\n\n    if (!__completion.success) {\n      throw new Error(\n        `Error getting response from ${__model}: ${__completion.error}`\n      );\n    }\n    responseMessage = __completion.value;\n    __updateTokenStats(responseMessage.usage, responseMessage.cost);\n    await __callHook(\"onLLMCallEnd\", { result: responseMessage, usage: responseMessage.usage, cost: responseMessage.cost, timeTaken: nextEndTime - nextStartTime });\n  }\n\n  // Add final assistant response to history\n  // not passing tool calls back this time\n  __messages.push(smoltalk.assistantMessage(responseMessage.output));\n  {{#hasResponseFormat}}\n  try {\n  const result = JSON.parse(responseMessage.output || \"\");\n  return result.response;\n  } catch (e) {\n    return responseMessage.output;\n    // console.error(\"Error parsing response for variable '{{{variableName:string}}}':\", e);\n    // console.error(\"Full completion response:\", JSON.stringify(__completion, null, 2));\n    // throw e;\n  }\n  {{/hasResponseFormat}}\n\n  {{^hasResponseFormat}}\n  return responseMessage.output;\n  {{/hasResponseFormat}}\n}\n\n{{#isAsync}}\n__self.{{{variableName:string}}} = _{{{variableName:string}}}({{{funcCallParams:string}}});\n{{/isAsync}}\n\n{{^isAsync}}\n__self.{{{variableName:string}}} = await _{{{variableName:string}}}({{{funcCallParams:string}}});\n\n// return early from node if this is an interrupt\nif (isInterrupt(__self.{{{variableName:string}}})) {\n  {{#nodeContext}}\n  return { messages: {{{messagesVar:string}}} , data: __self.{{{variableName:string}}} };\n  {{/nodeContext}}\n   {{^nodeContext}}\n   return  __self.{{{variableName:string}}};\n   {{/nodeContext}}\n}\n{{/isAsync}}";
+export declare const template = "\nasync function _{{{variableName:string}}}({{{argsStr:string}}}) {\n  const __prompt = {{{promptCode:string}}};\n  const startTime = performance.now();\n  let __messages = __metadata?.messages || [];\n\n  // These are to restore state after interrupt.\n  // TODO I think this could be implemented in a cleaner way.\n  let __toolCalls = __stateStack.interruptData?.toolCall ? [__stateStack.interruptData.toolCall] : [];\n  const __interruptResponse = __stateStack.interruptData?.interruptResponse || null;\n  const __tools = {{{tools}}};\n\n  {{#hasResponseFormat}}\n  // Need to make sure this is always an object\n  const __responseFormat = z.object({\n     response: {{{zodSchema:string}}}\n  });\n  {{/hasResponseFormat}}\n  {{^hasResponseFormat}}\n  const __responseFormat = undefined;\n  {{/hasResponseFormat}}\n  \n  const __client = __getClientWithConfig({{{clientConfig:string}}});\n  let responseMessage;\n\n  if (__toolCalls.length === 0) {\n    __messages.push(smoltalk.userMessage(__prompt));\n  \n  \n    await __callHook(\"onLLMCallStart\", { prompt: __prompt, tools: __tools, model: __client.getModel() });\n    let __completion = await __client.text({\n      messages: __messages,\n      tools: __tools,\n      responseFormat: __responseFormat,\n      stream: {{{isStreaming:boolean}}}\n    });\n\n    const endTime = performance.now();\n\n    {{#isStreaming}}\n    __completion = await handleStreamingResponse(__completion, statelogClient, __prompt, __toolCalls);\n    {{/isStreaming}}\n\n    statelogClient.promptCompletion({\n      messages: __messages,\n      completion: __completion,\n      model: __client.getModel(),\n      timeTaken: endTime - startTime,\n      tools: __tools,\n      responseFormat: __responseFormat\n    });\n\n    if (!__completion.success) {\n      throw new Error(\n        `Error getting response from ${__model}: ${__completion.error}`\n      );\n    }\n\n    responseMessage = __completion.value;\n    __toolCalls = responseMessage.toolCalls || [];\n\n    if (__toolCalls.length > 0) {\n      // Add assistant's response with tool calls to message history\n      __messages.push(smoltalk.assistantMessage(responseMessage.output, { toolCalls: __toolCalls }));\n    }\n\n    __updateTokenStats(responseMessage.usage, responseMessage.cost);\n    await __callHook(\"onLLMCallEnd\", { result: responseMessage, usage: responseMessage.usage, cost: responseMessage.cost, timeTaken: endTime - startTime });\n\n  }\n\n  // Handle function calls\n  if (__toolCalls.length > 0) {\n    let toolCallStartTime, toolCallEndTime;\n    let haltExecution = false;\n    let haltToolCall = {}\n    let haltInterrupt = null;\n\n    // Process each tool call\n    for (const toolCall of __toolCalls) {\n      {{{functionCalls:string}}}\n    }\n\n    if (haltExecution) {\n      statelogClient.debug(`Tool call interrupted execution.`, {\n        messages: __messages,\n        model: __client.getModel(),\n      });\n\n      __stateStack.interruptData = {\n        messages: __messages.map((msg) => msg.toJSON()),\n        nodesTraversed: __graph.getNodesTraversed(),\n        toolCall: haltToolCall,\n      };\n      haltInterrupt.__state = __stateStack.toJSON();\n      return haltInterrupt;\n    }\n  \n    const nextStartTime = performance.now();\n    await __callHook(\"onLLMCallStart\", { prompt: __prompt, tools: __tools, model: __client.getModel() });\n    let __completion = await __client.text({\n      messages: __messages,\n      tools: __tools,\n      responseFormat: __responseFormat,\n      stream: {{{isStreaming:boolean}}}\n    });\n\n    const nextEndTime = performance.now();\n\n    {{#isStreaming}}\n    __completion = await handleStreamingResponse(__completion, statelogClient, __prompt, __toolCalls);\n    {{/isStreaming}}\n\n    statelogClient.promptCompletion({\n      messages: __messages,\n      completion: __completion,\n      model: __client.getModel(),\n      timeTaken: nextEndTime - nextStartTime,\n      tools: __tools,\n      responseFormat: __responseFormat,\n    });\n\n    if (!__completion.success) {\n      throw new Error(\n        `Error getting response from ${__model}: ${__completion.error}`\n      );\n    }\n    responseMessage = __completion.value;\n    __updateTokenStats(responseMessage.usage, responseMessage.cost);\n    await __callHook(\"onLLMCallEnd\", { result: responseMessage, usage: responseMessage.usage, cost: responseMessage.cost, timeTaken: nextEndTime - nextStartTime });\n  }\n\n  // Add final assistant response to history\n  // not passing tool calls back this time\n  __messages.push(smoltalk.assistantMessage(responseMessage.output));\n  {{#hasResponseFormat}}\n  try {\n  const result = JSON.parse(responseMessage.output || \"\");\n  return result.response;\n  } catch (e) {\n    return responseMessage.output;\n    // console.error(\"Error parsing response for variable '{{{variableName:string}}}':\", e);\n    // console.error(\"Full completion response:\", JSON.stringify(__completion, null, 2));\n    // throw e;\n  }\n  {{/hasResponseFormat}}\n\n  {{^hasResponseFormat}}\n  return responseMessage.output;\n  {{/hasResponseFormat}}\n}\n\n{{#isAsync}}\n__self.{{{variableName:string}}} = _{{{variableName:string}}}({{{funcCallParams:string}}});\n{{/isAsync}}\n\n{{^isAsync}}\n__self.{{{variableName:string}}} = await _{{{variableName:string}}}({{{funcCallParams:string}}});\n\n// return early from node if this is an interrupt\nif (isInterrupt(__self.{{{variableName:string}}})) {\n  {{#nodeContext}}\n  return { messages: {{{messagesVar:string}}} , data: __self.{{{variableName:string}}} };\n  {{/nodeContext}}\n   {{^nodeContext}}\n   return  __self.{{{variableName:string}}};\n   {{/nodeContext}}\n}\n{{/isAsync}}";
 export type TemplateType = {
     variableName: string;
     argsStr: string;

package/dist/lib/templates/backends/typescriptGenerator/promptFunction.js

@@ -41,7 +41,9 @@ async function _{{{variableName:string}}}({{{argsStr:string}}}) {
 
     const endTime = performance.now();
 
-    await handleStreamingResponse(__completion);
+    {{#isStreaming}}
+    __completion = await handleStreamingResponse(__completion, statelogClient, __prompt, __toolCalls);
+    {{/isStreaming}}
 
     statelogClient.promptCompletion({
       messages: __messages,
@@ -109,7 +111,9 @@ async function _{{{variableName:string}}}({{{argsStr:string}}}) {
 
     const nextEndTime = performance.now();
 
-    await handleStreamingResponse(__completion);
+    {{#isStreaming}}
+    __completion = await handleStreamingResponse(__completion, statelogClient, __prompt, __toolCalls);
+    {{/isStreaming}}
 
     statelogClient.promptCompletion({
       messages: __messages,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agency-lang",
-  "version": "0.0.66",
+  "version": "0.0.67",
   "description": "The Agency language",
   "main": "lib/index.js",
   "scripts": {
@@ -70,4 +70,4 @@
     "typescript": "^5.9.3",
     "vitest": "^4.0.16"
   }
-}
+}