UNPKG

@openai/agents-core

Version:

The OpenAI Agents SDK is a lightweight yet powerful framework for building multi-agent workflows.

764 lines 31.1 kB
import { ModelBehaviorError, ToolCallError, UserError } from "./errors.js"; import { getTransferMessage } from "./handoff.js"; import { RunHandoffCallItem, RunHandoffOutputItem, RunMessageOutputItem, RunReasoningItem, RunToolApprovalItem, RunToolCallItem, RunToolCallOutputItem, } from "./items.js"; import logger from "./logger.js"; import { getLastTextFromOutputMessage } from "./utils/messages.js"; import { withFunctionSpan, withHandoffSpan } from "./tracing/createSpans.js"; import { getSchemaAndParserFromInputType } from "./utils/tools.js"; import { safeExecute } from "./utils/safeExecute.js"; import { addErrorToCurrentSpan } from "./tracing/context.js"; import { RunItemStreamEvent } from "./events.js"; import { z } from '@openai/zod/v3'; import { toSmartString } from "./utils/smartString.js"; import { isZodObject } from "./utils/index.js"; /** * @internal */ export function processModelResponse(modelResponse, agent, tools, handoffs) { const items = []; const runHandoffs = []; const runFunctions = []; const runComputerActions = []; const runMCPApprovalRequests = []; const toolsUsed = []; const handoffMap = new Map(handoffs.map((h) => [h.toolName, h])); const functionMap = new Map(tools.filter((t) => t.type === 'function').map((t) => [t.name, t])); const computerTool = tools.find((t) => t.type === 'computer'); const mcpToolMap = new Map(tools .filter((t) => t.type === 'hosted_tool' && t.providerData?.type === 'mcp') .map((t) => t) .map((t) => [t.providerData.server_label, t])); for (const output of modelResponse.output) { if (output.type === 'message') { if (output.role === 'assistant') { items.push(new RunMessageOutputItem(output, agent)); } } else if (output.type === 'hosted_tool_call') { items.push(new RunToolCallItem(output, agent)); const toolName = output.name; toolsUsed.push(toolName); if (output.providerData?.type === 'mcp_approval_request' || output.name === 'mcp_approval_request') { // Hosted remote MCP server's approval process const providerData = output.providerData; const mcpServerLabel = providerData.server_label; const mcpServerTool = mcpToolMap.get(mcpServerLabel); if (typeof mcpServerTool === 'undefined') { const message = `MCP server (${mcpServerLabel}) not found in Agent (${agent.name})`; addErrorToCurrentSpan({ message, data: { mcp_server_label: mcpServerLabel }, }); throw new ModelBehaviorError(message); } // Do this approval later: // We support both onApproval callback (like the Python SDK does) and HITL patterns. const approvalItem = new RunToolApprovalItem({ type: 'hosted_tool_call', // We must use this name to align with the name sent from the servers name: providerData.name, id: providerData.id, status: 'in_progress', providerData, }, agent); runMCPApprovalRequests.push({ requestItem: approvalItem, mcpTool: mcpServerTool, }); if (!mcpServerTool.providerData.on_approval) { // When onApproval function exists, it confirms the approval right after this. // Thus, this approval item must be appended only for the next turn interruption patterns. items.push(approvalItem); } } } else if (output.type === 'reasoning') { items.push(new RunReasoningItem(output, agent)); } else if (output.type === 'computer_call') { items.push(new RunToolCallItem(output, agent)); toolsUsed.push('computer_use'); if (!computerTool) { addErrorToCurrentSpan({ message: 'Model produced computer action without a computer tool.', data: { agent_name: agent.name, }, }); throw new ModelBehaviorError('Model produced computer action without a computer tool.'); } runComputerActions.push({ toolCall: output, computer: computerTool, }); } if (output.type !== 'function_call') { continue; } toolsUsed.push(output.name); const handoff = handoffMap.get(output.name); if (handoff) { items.push(new RunHandoffCallItem(output, agent)); runHandoffs.push({ toolCall: output, handoff: handoff, }); } else { const functionTool = functionMap.get(output.name); if (!functionTool) { addErrorToCurrentSpan({ message: `Tool ${output.name} not found in agent ${agent.name}.`, data: { tool_name: output.name, agent_name: agent.name, }, }); throw new ModelBehaviorError(`Tool ${output.name} not found in agent ${agent.name}.`); } items.push(new RunToolCallItem(output, agent)); runFunctions.push({ toolCall: output, tool: functionTool, }); } } return { newItems: items, handoffs: runHandoffs, functions: runFunctions, computerActions: runComputerActions, mcpApprovalRequests: runMCPApprovalRequests, toolsUsed: toolsUsed, hasToolsOrApprovalsToRun() { return (runHandoffs.length > 0 || runFunctions.length > 0 || runMCPApprovalRequests.length > 0 || runComputerActions.length > 0); }, }; } export const nextStepSchema = z.discriminatedUnion('type', [ z.object({ type: z.literal('next_step_handoff'), newAgent: z.any(), }), z.object({ type: z.literal('next_step_final_output'), output: z.string(), }), z.object({ type: z.literal('next_step_run_again'), }), z.object({ type: z.literal('next_step_interruption'), data: z.record(z.string(), z.any()), }), ]); class SingleStepResult { originalInput; modelResponse; preStepItems; newStepItems; nextStep; constructor( /** * The input items i.e. the items before run() was called. May be muted by handoff input filters */ originalInput, /** * The model response for the current step */ modelResponse, /** * The items before the current step was executed */ preStepItems, /** * The items after the current step was executed */ newStepItems, /** * The next step to execute */ nextStep) { this.originalInput = originalInput; this.modelResponse = modelResponse; this.preStepItems = preStepItems; this.newStepItems = newStepItems; this.nextStep = nextStep; } /** * The items generated during the agent run (i.e. everything generated after originalInput) */ get generatedItems() { return this.preStepItems.concat(this.newStepItems); } } /** * @internal */ export function maybeResetToolChoice(agent, toolUseTracker, modelSettings) { if (agent.resetToolChoice && toolUseTracker.hasUsedTools(agent)) { return { ...modelSettings, toolChoice: undefined }; } return modelSettings; } /** * @internal */ export async function executeInterruptedToolsAndSideEffects(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) { // call_ids for function tools const functionCallIds = originalPreStepItems .filter((item) => item instanceof RunToolApprovalItem && 'callId' in item.rawItem && item.rawItem.type === 'function_call') .map((item) => item.rawItem.callId); // Run function tools that require approval after they get their approval results const functionToolRuns = processedResponse.functions.filter((run) => { return functionCallIds.includes(run.toolCall.callId); }); const functionResults = await executeFunctionToolCalls(agent, functionToolRuns, runner, state); // Create the initial set of the output items const newItems = functionResults.map((r) => r.runItem); // Run MCP tools that require approval after they get their approval results const mcpApprovalRuns = processedResponse.mcpApprovalRequests.filter((run) => { return (run.requestItem.type === 'tool_approval_item' && run.requestItem.rawItem.type === 'hosted_tool_call' && run.requestItem.rawItem.providerData?.type === 'mcp_approval_request'); }); for (const run of mcpApprovalRuns) { // the approval_request_id "mcpr_123..." const approvalRequestId = run.requestItem.rawItem.id; const approved = state._context.isToolApproved({ // Since this item name must be the same with the one sent from Responses API server toolName: run.requestItem.rawItem.name, callId: approvalRequestId, }); if (typeof approved !== 'undefined') { const providerData = { approve: approved, approval_request_id: approvalRequestId, reason: undefined, }; // Tell Responses API server the approval result in the next turn newItems.push(new RunToolCallItem({ type: 'hosted_tool_call', name: 'mcp_approval_response', providerData, }, agent)); } } const checkToolOutput = await checkForFinalOutputFromTools(agent, functionResults, state); // Exclude the tool approval items, which should not be sent to Responses API, // from the SingleStepResult's preStepItems const preStepItems = originalPreStepItems.filter((item) => { return !(item instanceof RunToolApprovalItem); }); if (checkToolOutput.isFinalOutput) { runner.emit('agent_end', state._context, agent, checkToolOutput.finalOutput); agent.emit('agent_end', state._context, checkToolOutput.finalOutput); return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_final_output', output: checkToolOutput.finalOutput, }); } else if (checkToolOutput.isInterrupted) { return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_interruption', data: { interruptions: checkToolOutput.interruptions, }, }); } // we only ran new tools and side effects. We need to run the rest of the agent return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' }); } /** * @internal */ export async function executeToolsAndSideEffects(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) { const preStepItems = originalPreStepItems; let newItems = processedResponse.newItems; const [functionResults, computerResults] = await Promise.all([ executeFunctionToolCalls(agent, processedResponse.functions, runner, state), executeComputerActions(agent, processedResponse.computerActions, runner, state._context), ]); newItems = newItems.concat(functionResults.map((r) => r.runItem)); newItems = newItems.concat(computerResults); // run hosted MCP approval requests if (processedResponse.mcpApprovalRequests.length > 0) { for (const approvalRequest of processedResponse.mcpApprovalRequests) { const toolData = approvalRequest.mcpTool .providerData; const requestData = approvalRequest.requestItem.rawItem .providerData; if (toolData.on_approval) { // synchronously handle the approval process here const approvalResult = await toolData.on_approval(state._context, approvalRequest.requestItem); const approvalResponseData = { approve: approvalResult.approve, approval_request_id: requestData.id, reason: approvalResult.reason, }; newItems.push(new RunToolCallItem({ type: 'hosted_tool_call', name: 'mcp_approval_response', providerData: approvalResponseData, }, agent)); } else { // receive a user's approval on the next turn newItems.push(approvalRequest.requestItem); const approvalItem = { type: 'hosted_mcp_tool_approval', tool: approvalRequest.mcpTool, runItem: new RunToolApprovalItem({ type: 'hosted_tool_call', name: requestData.name, id: requestData.id, arguments: requestData.arguments, status: 'in_progress', providerData: requestData, }, agent), }; functionResults.push(approvalItem); // newItems.push(approvalItem.runItem); } } } // process handoffs if (processedResponse.handoffs.length > 0) { return await executeHandoffCalls(agent, originalInput, preStepItems, newItems, newResponse, processedResponse.handoffs, runner, state._context); } const checkToolOutput = await checkForFinalOutputFromTools(agent, functionResults, state); if (checkToolOutput.isFinalOutput) { runner.emit('agent_end', state._context, agent, checkToolOutput.finalOutput); agent.emit('agent_end', state._context, checkToolOutput.finalOutput); return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_final_output', output: checkToolOutput.finalOutput, }); } else if (checkToolOutput.isInterrupted) { return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_interruption', data: { interruptions: checkToolOutput.interruptions, }, }); } // check if the agent produced any messages const messageItems = newItems.filter((item) => item instanceof RunMessageOutputItem); // we will use the last content output as the final output const potentialFinalOutput = messageItems.length > 0 ? getLastTextFromOutputMessage(messageItems[messageItems.length - 1].rawItem) : undefined; // if there is no output we just run again if (!potentialFinalOutput) { return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' }); } if (agent.outputType === 'text' && !processedResponse.hasToolsOrApprovalsToRun()) { return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_final_output', output: potentialFinalOutput, }); } else if (agent.outputType !== 'text' && potentialFinalOutput) { // Structured output schema => always leads to a final output if we have text const { parser } = getSchemaAndParserFromInputType(agent.outputType, 'final_output'); const [error] = await safeExecute(() => parser(potentialFinalOutput)); if (error) { addErrorToCurrentSpan({ message: 'Invalid output type', data: { error: String(error), }, }); throw new ModelBehaviorError('Invalid output type'); } return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_final_output', output: potentialFinalOutput }); } return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' }); } /** * @internal */ export function getToolCallOutputItem(toolCall, output) { return { type: 'function_call_result', name: toolCall.name, callId: toolCall.callId, status: 'completed', output: { type: 'text', text: toSmartString(output), }, }; } /** * @internal */ export async function executeFunctionToolCalls(agent, toolRuns, runner, state) { async function runSingleTool(toolRun) { let parsedArgs = toolRun.toolCall.arguments; if (toolRun.tool.parameters) { if (isZodObject(toolRun.tool.parameters)) { parsedArgs = toolRun.tool.parameters.parse(parsedArgs); } else { parsedArgs = JSON.parse(parsedArgs); } } const needsApproval = await toolRun.tool.needsApproval(state._context, parsedArgs, toolRun.toolCall.callId); if (needsApproval) { const approval = state._context.isToolApproved({ toolName: toolRun.tool.name, callId: toolRun.toolCall.callId, }); if (approval === false) { // rejected return withFunctionSpan(async (span) => { const response = 'Tool execution was not approved.'; span.setError({ message: response, data: { tool_name: toolRun.tool.name, error: `Tool execution for ${toolRun.toolCall.callId} was manually rejected by user.`, }, }); span.spanData.output = response; return { type: 'function_output', tool: toolRun.tool, output: response, runItem: new RunToolCallOutputItem(getToolCallOutputItem(toolRun.toolCall, response), agent, response), }; }, { data: { name: toolRun.tool.name, }, }); } if (approval !== true) { // this approval process needs to be done in the next turn return { type: 'function_approval', tool: toolRun.tool, runItem: new RunToolApprovalItem(toolRun.toolCall, agent), }; } } return withFunctionSpan(async (span) => { if (runner.config.traceIncludeSensitiveData) { span.spanData.input = toolRun.toolCall.arguments; } try { runner.emit('agent_tool_start', state._context, agent, toolRun.tool, { toolCall: toolRun.toolCall, }); agent.emit('agent_tool_start', state._context, toolRun.tool, { toolCall: toolRun.toolCall, }); const result = await toolRun.tool.invoke(state._context, toolRun.toolCall.arguments); // Use string data for tracing and event emitter const stringResult = toSmartString(result); runner.emit('agent_tool_end', state._context, agent, toolRun.tool, stringResult, { toolCall: toolRun.toolCall }); agent.emit('agent_tool_end', state._context, toolRun.tool, stringResult, { toolCall: toolRun.toolCall }); if (runner.config.traceIncludeSensitiveData) { span.spanData.output = stringResult; } return { type: 'function_output', tool: toolRun.tool, output: result, runItem: new RunToolCallOutputItem(getToolCallOutputItem(toolRun.toolCall, result), agent, result), }; } catch (error) { span.setError({ message: 'Error running tool', data: { tool_name: toolRun.tool.name, error: String(error), }, }); throw error; } }, { data: { name: toolRun.tool.name, }, }); } try { const results = await Promise.all(toolRuns.map(runSingleTool)); return results; } catch (e) { throw new ToolCallError(`Failed to run function tools: ${e}`, e, state); } } /** * @internal */ // Internal helper: dispatch a computer action and return a screenshot (sync/async) async function _runComputerActionAndScreenshot(computer, toolCall) { const action = toolCall.action; let screenshot; // Dispatch based on action type string (assume action.type exists) switch (action.type) { case 'click': await computer.click(action.x, action.y, action.button); break; case 'double_click': await computer.doubleClick(action.x, action.y); break; case 'drag': await computer.drag(action.path.map((p) => [p.x, p.y])); break; case 'keypress': await computer.keypress(action.keys); break; case 'move': await computer.move(action.x, action.y); break; case 'screenshot': screenshot = await computer.screenshot(); break; case 'scroll': await computer.scroll(action.x, action.y, action.scroll_x, action.scroll_y); break; case 'type': await computer.type(action.text); break; case 'wait': await computer.wait(); break; default: action; // ensures that we handle every action we know of // Unknown action, just take screenshot break; } if (typeof screenshot !== 'undefined') { return screenshot; } // Always return screenshot as base64 string if (typeof computer.screenshot === 'function') { screenshot = await computer.screenshot(); if (typeof screenshot !== 'undefined') { return screenshot; } } throw new Error('Computer does not implement screenshot()'); } /** * @internal */ export async function executeComputerActions(agent, actions, runner, runContext, customLogger = undefined) { const _logger = customLogger ?? logger; const results = []; for (const action of actions) { const computer = action.computer.computer; const toolCall = action.toolCall; // Hooks: on_tool_start (global + agent) runner.emit('agent_tool_start', runContext, agent, action.computer, { toolCall, }); if (typeof agent.emit === 'function') { agent.emit('agent_tool_start', runContext, action.computer, { toolCall }); } // Run the action and get screenshot let output; try { output = await _runComputerActionAndScreenshot(computer, toolCall); } catch (err) { _logger.error('Failed to execute computer action:', err); output = ''; } // Hooks: on_tool_end (global + agent) runner.emit('agent_tool_end', runContext, agent, action.computer, output, { toolCall, }); if (typeof agent.emit === 'function') { agent.emit('agent_tool_end', runContext, action.computer, output, { toolCall, }); } // Always return a screenshot as a base64 data URL const imageUrl = output ? `data:image/png;base64,${output}` : ''; const rawItem = { type: 'computer_call_result', callId: toolCall.callId, output: { type: 'computer_screenshot', data: imageUrl }, }; results.push(new RunToolCallOutputItem(rawItem, agent, imageUrl)); } return results; } /** * @internal */ export async function executeHandoffCalls(agent, originalInput, preStepItems, newStepItems, newResponse, runHandoffs, runner, runContext) { newStepItems = [...newStepItems]; if (runHandoffs.length === 0) { logger.warn('Incorrectly called executeHandoffCalls with no handoffs. This should not happen. Moving on.'); return new SingleStepResult(originalInput, newResponse, preStepItems, newStepItems, { type: 'next_step_run_again' }); } if (runHandoffs.length > 1) { // multiple handoffs. Ignoring all but the first one by adding reject responses for those const outputMessage = 'Multiple handoffs detected, ignoring this one.'; for (let i = 1; i < runHandoffs.length; i++) { newStepItems.push(new RunToolCallOutputItem(getToolCallOutputItem(runHandoffs[i].toolCall, outputMessage), agent, outputMessage)); } } const actualHandoff = runHandoffs[0]; return withHandoffSpan(async (handoffSpan) => { const handoff = actualHandoff.handoff; const newAgent = await handoff.onInvokeHandoff(runContext, actualHandoff.toolCall.arguments); handoffSpan.spanData.to_agent = newAgent.name; if (runHandoffs.length > 1) { const requestedAgents = runHandoffs.map((h) => h.handoff.agentName); handoffSpan.setError({ message: 'Multiple handoffs requested', data: { requested_agents: requestedAgents, }, }); } newStepItems.push(new RunHandoffOutputItem(getToolCallOutputItem(actualHandoff.toolCall, getTransferMessage(newAgent)), agent, newAgent)); runner.emit('agent_handoff', runContext, agent, newAgent); agent.emit('agent_handoff', runContext, newAgent); const inputFilter = handoff.inputFilter ?? runner.config.handoffInputFilter; if (inputFilter) { logger.debug('Filtering inputs for handoff'); if (typeof inputFilter !== 'function') { handoffSpan.setError({ message: 'Invalid input filter', data: { details: 'not callable', }, }); } const handoffInputData = { inputHistory: Array.isArray(originalInput) ? [...originalInput] : originalInput, preHandoffItems: [...preStepItems], newItems: [...newStepItems], }; const filtered = inputFilter(handoffInputData); originalInput = filtered.inputHistory; preStepItems = filtered.preHandoffItems; newStepItems = filtered.newItems; } return new SingleStepResult(originalInput, newResponse, preStepItems, newStepItems, { type: 'next_step_handoff', newAgent }); }, { data: { from_agent: agent.name, }, }); } const NOT_FINAL_OUTPUT = { isFinalOutput: false, isInterrupted: undefined, }; /** * @internal */ export async function checkForFinalOutputFromTools(agent, toolResults, state) { if (toolResults.length === 0) { return NOT_FINAL_OUTPUT; } const interruptions = toolResults .filter((r) => r.runItem instanceof RunToolApprovalItem) .map((r) => r.runItem); if (interruptions.length > 0) { return { isFinalOutput: false, isInterrupted: true, interruptions, }; } if (agent.toolUseBehavior === 'run_llm_again') { return NOT_FINAL_OUTPUT; } const firstToolResult = toolResults[0]; if (agent.toolUseBehavior === 'stop_on_first_tool') { if (firstToolResult?.type === 'function_output') { const stringOutput = toSmartString(firstToolResult.output); return { isFinalOutput: true, isInterrupted: undefined, finalOutput: stringOutput, }; } return NOT_FINAL_OUTPUT; } const toolUseBehavior = agent.toolUseBehavior; if (typeof toolUseBehavior === 'object') { const stoppingTool = toolResults.find((r) => toolUseBehavior.stopAtToolNames.includes(r.tool.name)); if (stoppingTool?.type === 'function_output') { const stringOutput = toSmartString(stoppingTool.output); return { isFinalOutput: true, isInterrupted: undefined, finalOutput: stringOutput, }; } return NOT_FINAL_OUTPUT; } if (typeof toolUseBehavior === 'function') { return toolUseBehavior(state._context, toolResults); } throw new UserError(`Invalid toolUseBehavior: ${toolUseBehavior}`, state); } export function addStepToRunResult(result, step) { for (const item of step.newStepItems) { let itemName; if (item instanceof RunMessageOutputItem) { itemName = 'message_output_created'; } else if (item instanceof RunHandoffCallItem) { itemName = 'handoff_requested'; } else if (item instanceof RunHandoffOutputItem) { itemName = 'handoff_occurred'; } else if (item instanceof RunToolCallItem) { itemName = 'tool_called'; } else if (item instanceof RunToolCallOutputItem) { itemName = 'tool_output'; } else if (item instanceof RunReasoningItem) { itemName = 'reasoning_item_created'; } else if (item instanceof RunToolApprovalItem) { itemName = 'tool_approval_requested'; } else { logger.warn('Unknown item type: ', item); continue; } result._addItem(new RunItemStreamEvent(itemName, item)); } } export class AgentToolUseTracker { #agentToTools = new Map(); addToolUse(agent, toolNames) { this.#agentToTools.set(agent, toolNames); } hasUsedTools(agent) { return this.#agentToTools.has(agent); } toJSON() { return Object.fromEntries(Array.from(this.#agentToTools.entries()).map(([agent, toolNames]) => { return [agent.name, toolNames]; })); } } //# sourceMappingURL=runImplementation.js.map