mirror of
https://github.com/rowboatlabs/rowboat.git
synced 2026-06-06 19:35:44 +02:00
909 lines
No EOL
32 KiB
TypeScript
909 lines
No EOL
32 KiB
TypeScript
// External dependencies
|
|
import { Agent, AgentInputItem, run, tool, Tool } from "@openai/agents";
|
|
import { RECOMMENDED_PROMPT_PREFIX } from "@openai/agents-core/extensions";
|
|
import { aisdk } from "@openai/agents-extensions";
|
|
import { createOpenAI } from "@ai-sdk/openai";
|
|
import { CoreMessage, embed, generateText } from "ai";
|
|
import { ObjectId } from "mongodb";
|
|
import { z } from "zod";
|
|
|
|
// Internal dependencies
|
|
import { embeddingModel } from '../lib/embedding';
|
|
import { getMcpClient } from "./mcp";
|
|
import { dataSourceDocsCollection, dataSourcesCollection } from "./mongodb";
|
|
import { qdrantClient } from '../lib/qdrant';
|
|
import { EmbeddingRecord } from "./types/datasource_types";
|
|
import { ConnectedEntity, sanitizeTextWithMentions, Workflow, WorkflowAgent, WorkflowPrompt, WorkflowTool } from "./types/workflow_types";
|
|
import { CHILD_TRANSFER_RELATED_INSTRUCTIONS, RAG_INSTRUCTIONS } from "./agent_instructions";
|
|
import { PrefixLogger } from "./utils";
|
|
import { Message, AssistantMessage, AssistantMessageWithToolCalls, ToolMessage } from "./types/types";
|
|
|
|
const PROVIDER_API_KEY = process.env.PROVIDER_API_KEY || process.env.OPENAI_API_KEY || '';
|
|
const PROVIDER_BASE_URL = process.env.PROVIDER_BASE_URL || undefined;
|
|
const MODEL = process.env.PROVIDER_DEFAULT_MODEL || 'gpt-4o';
|
|
|
|
const openai = createOpenAI({
|
|
apiKey: PROVIDER_API_KEY,
|
|
baseURL: PROVIDER_BASE_URL,
|
|
});
|
|
|
|
export const ZUsage = z.object({
|
|
tokens: z.object({
|
|
total: z.number(),
|
|
prompt: z.number(),
|
|
completion: z.number(),
|
|
}),
|
|
});
|
|
|
|
const ZOutMessage = z.union([
|
|
AssistantMessage,
|
|
AssistantMessageWithToolCalls,
|
|
ToolMessage,
|
|
]);
|
|
|
|
// Helper to handle mock tool responses
|
|
async function invokeMockTool(
|
|
logger: PrefixLogger,
|
|
toolName: string,
|
|
args: string,
|
|
description: string,
|
|
mockInstructions: string
|
|
): Promise<string> {
|
|
logger = logger.child(`invokeMockTool`);
|
|
logger.log(`toolName: ${toolName}`);
|
|
logger.log(`args: ${args}`);
|
|
logger.log(`description: ${description}`);
|
|
logger.log(`mockInstructions: ${mockInstructions}`);
|
|
|
|
const messages: CoreMessage[] = [{
|
|
role: "system" as const,
|
|
content: `You are simulating the execution of a tool called '${toolName}'. Here is the description of the tool: ${description}. Here are the instructions for the mock tool: ${mockInstructions}. Generate a realistic response as if the tool was actually executed with the given parameters.`
|
|
}, {
|
|
role: "user" as const,
|
|
content: `Generate a realistic response for the tool '${toolName}' with these parameters: ${args}. The response should be concise and focused on what the tool would actually return.`
|
|
}];
|
|
|
|
const { text } = await generateText({
|
|
model: openai(MODEL),
|
|
messages,
|
|
});
|
|
logger.log(`generated text: ${text}`);
|
|
|
|
return text;
|
|
}
|
|
|
|
// Helper to handle RAG tool calls
|
|
async function invokeRagTool(
|
|
logger: PrefixLogger,
|
|
projectId: string,
|
|
query: string,
|
|
sourceIds: string[],
|
|
returnType: 'chunks' | 'content',
|
|
k: number
|
|
): Promise<{
|
|
title: string;
|
|
name: string;
|
|
content: string;
|
|
docId: string;
|
|
sourceId: string;
|
|
}[]> {
|
|
logger = logger.child(`invokeRagTool`);
|
|
logger.log(`projectId: ${projectId}`);
|
|
logger.log(`query: ${query}`);
|
|
logger.log(`sourceIds: ${sourceIds.join(', ')}`);
|
|
logger.log(`returnType: ${returnType}`);
|
|
logger.log(`k: ${k}`);
|
|
|
|
// Create embedding for question
|
|
const { embedding } = await embed({
|
|
model: embeddingModel,
|
|
value: query,
|
|
});
|
|
|
|
// Fetch all data sources for this project
|
|
const sources = await dataSourcesCollection.find({
|
|
projectId: projectId,
|
|
active: true,
|
|
}).toArray();
|
|
const validSourceIds = sources
|
|
.filter(s => sourceIds.includes(s._id.toString())) // id should be in sourceIds
|
|
.filter(s => s.active) // should be active
|
|
.map(s => s._id.toString());
|
|
logger.log(`valid source ids: ${validSourceIds.join(', ')}`);
|
|
|
|
// if no sources found, return empty response
|
|
if (validSourceIds.length === 0) {
|
|
logger.log(`no valid source ids found, returning empty response`);
|
|
return [];
|
|
}
|
|
|
|
// Perform vector search
|
|
const qdrantResults = await qdrantClient.query("embeddings", {
|
|
query: embedding,
|
|
filter: {
|
|
must: [
|
|
{ key: "projectId", match: { value: projectId } },
|
|
{ key: "sourceId", match: { any: validSourceIds } },
|
|
],
|
|
},
|
|
limit: k,
|
|
with_payload: true,
|
|
});
|
|
logger.log(`found ${qdrantResults.points.length} results`);
|
|
|
|
// if return type is chunks, return the chunks
|
|
let results = qdrantResults.points.map((point) => {
|
|
const { title, name, content, docId, sourceId } = point.payload as z.infer<typeof EmbeddingRecord>['payload'];
|
|
return {
|
|
title,
|
|
name,
|
|
content,
|
|
docId,
|
|
sourceId,
|
|
};
|
|
});
|
|
|
|
if (returnType === 'chunks') {
|
|
logger.log(`returning chunks`);
|
|
return results;
|
|
}
|
|
|
|
// otherwise, fetch the doc contents from mongodb
|
|
const docs = await dataSourceDocsCollection.find({
|
|
_id: { $in: results.map(r => new ObjectId(r.docId)) },
|
|
}).toArray();
|
|
logger.log(`fetched docs: ${docs.length}`);
|
|
|
|
// map the results to the docs
|
|
results = results.map(r => {
|
|
const doc = docs.find(d => d._id.toString() === r.docId);
|
|
return {
|
|
...r,
|
|
content: doc?.content || '',
|
|
};
|
|
});
|
|
|
|
return results;
|
|
}
|
|
|
|
// Helper to handle MCP tool calls
|
|
async function invokeMcpTool(
|
|
logger: PrefixLogger,
|
|
projectId: string,
|
|
name: string,
|
|
input: any,
|
|
mcpServerURL: string,
|
|
mcpServerName: string
|
|
) {
|
|
logger = logger.child(`invokeMcpTool`);
|
|
logger.log(`projectId: ${projectId}`);
|
|
logger.log(`name: ${name}`);
|
|
logger.log(`input: ${JSON.stringify(input)}`);
|
|
logger.log(`mcpServerURL: ${mcpServerURL}`);
|
|
logger.log(`mcpServerName: ${mcpServerName}`);
|
|
|
|
const client = await getMcpClient(mcpServerURL, mcpServerName || '');
|
|
const result = await client.callTool({
|
|
name,
|
|
arguments: input,
|
|
});
|
|
logger.log(`mcp tool result: ${JSON.stringify(result)}`);
|
|
await client.close();
|
|
return result;
|
|
}
|
|
|
|
// Helper to create RAG tool
|
|
function createRagTool(
|
|
logger: PrefixLogger,
|
|
config: z.infer<typeof WorkflowAgent>,
|
|
projectId: string
|
|
): Tool {
|
|
if (!config.ragDataSources?.length) {
|
|
throw new Error(`data sources not found for agent ${config.name}`);
|
|
}
|
|
|
|
return tool({
|
|
name: "rag_search",
|
|
description: config.description,
|
|
parameters: z.object({
|
|
query: z.string().describe("The query to search for")
|
|
}),
|
|
async execute(input: { query: string }) {
|
|
const results = await invokeRagTool(
|
|
logger,
|
|
projectId,
|
|
input.query,
|
|
config.ragDataSources || [],
|
|
config.ragReturnType || 'chunks',
|
|
config.ragK || 3
|
|
);
|
|
return JSON.stringify({
|
|
results,
|
|
});
|
|
}
|
|
});
|
|
}
|
|
|
|
// Helper to create a mock tool
|
|
function createMockTool(
|
|
logger: PrefixLogger,
|
|
config: z.infer<typeof WorkflowTool>,
|
|
): Tool {
|
|
return tool({
|
|
name: config.name,
|
|
description: config.description,
|
|
parameters: z.object({
|
|
query: z.string().describe("The query to search for")
|
|
}),
|
|
async execute(input: { query: string }) {
|
|
try {
|
|
const result = await invokeMockTool(
|
|
logger,
|
|
config.name,
|
|
JSON.stringify(input),
|
|
config.description,
|
|
config.mockInstructions || ''
|
|
);
|
|
return JSON.stringify({
|
|
result,
|
|
});
|
|
} catch (error) {
|
|
logger.log(`Error executing mock tool ${config.name}:`, error);
|
|
return JSON.stringify({
|
|
error: `Mock tool execution failed: ${error}`,
|
|
});
|
|
}
|
|
}
|
|
});
|
|
}
|
|
|
|
// Helper to create an mcp tool
|
|
function createMcpTool(
|
|
logger: PrefixLogger,
|
|
config: z.infer<typeof WorkflowTool>,
|
|
projectId: string
|
|
): Tool {
|
|
const { name, description, parameters, mcpServerName, mcpServerURL } = config;
|
|
|
|
return tool({
|
|
name,
|
|
description,
|
|
strict: false,
|
|
parameters: {
|
|
type: 'object',
|
|
properties: parameters.properties,
|
|
required: parameters.required || [],
|
|
additionalProperties: true,
|
|
},
|
|
async execute(input: any) {
|
|
try {
|
|
const result = await invokeMcpTool(logger, projectId, name, input, mcpServerURL || '', mcpServerName || '');
|
|
return JSON.stringify({
|
|
result,
|
|
});
|
|
} catch (error) {
|
|
logger.log(`Error executing mcp tool ${name}:`, error);
|
|
return JSON.stringify({
|
|
error: `Tool execution failed: ${error}`,
|
|
});
|
|
}
|
|
}
|
|
});
|
|
}
|
|
|
|
// Helper to create an agent
|
|
function createAgent(
|
|
logger: PrefixLogger,
|
|
config: z.infer<typeof WorkflowAgent>,
|
|
tools: Record<string, Tool>,
|
|
projectTools: z.infer<typeof WorkflowTool>[],
|
|
workflow: z.infer<typeof Workflow>,
|
|
promptConfig: Record<string, z.infer<typeof WorkflowPrompt>>,
|
|
): { agent: Agent, entities: z.infer<typeof ConnectedEntity>[] } {
|
|
const agentLogger = logger.child(`createAgent: ${config.name}`);
|
|
|
|
// Combine instructions and examples
|
|
let instructions = `${RECOMMENDED_PROMPT_PREFIX}
|
|
|
|
## Your Name
|
|
${config.name}
|
|
|
|
## Description
|
|
${config.description}
|
|
|
|
## Instructions
|
|
${config.instructions}
|
|
|
|
${config.examples ? ('# Examples\n' + config.examples) : ''}
|
|
|
|
${'-'.repeat(100)}
|
|
|
|
${CHILD_TRANSFER_RELATED_INSTRUCTIONS}
|
|
`;
|
|
|
|
let { sanitized, entities } = sanitizeTextWithMentions(instructions, workflow, projectTools);
|
|
agentLogger.log(`instructions: ${JSON.stringify(sanitized)}`);
|
|
agentLogger.log(`mentions: ${JSON.stringify(entities)}`);
|
|
|
|
// // add prompts to instructions
|
|
// for (const e of entities) {
|
|
// if (e.type === 'prompt') {
|
|
// const prompt = promptConfig[e.name];
|
|
// if (prompt) {
|
|
// compiledInstructions = compiledInstructions + '\n\n# ' + prompt.name + '\n' + prompt.prompt;
|
|
// }
|
|
// }
|
|
// }
|
|
|
|
const agentTools = entities.filter(e => e.type === 'tool').map(e => tools[e.name]).filter(Boolean) as Tool[];
|
|
|
|
// Add RAG tool if needed
|
|
if (config.ragDataSources?.length) {
|
|
const ragTool = createRagTool(logger, config, workflow.projectId);
|
|
agentTools.push(ragTool);
|
|
|
|
// update instructions to include RAG instructions
|
|
sanitized = sanitized + '\n\n' + ('-'.repeat(100)) + '\n\n' + RAG_INSTRUCTIONS(ragTool.name);
|
|
agentLogger.log(`added rag instructions`);
|
|
}
|
|
|
|
// Create the agent
|
|
const agent = new Agent({
|
|
name: config.name,
|
|
instructions: sanitized,
|
|
tools: agentTools,
|
|
model: aisdk(openai(config.model)),
|
|
// model: config.model,
|
|
modelSettings: {
|
|
temperature: 0.0,
|
|
}
|
|
});
|
|
agentLogger.log(`created agent`);
|
|
|
|
return {
|
|
agent,
|
|
entities,
|
|
};
|
|
}
|
|
|
|
// Convert messages to agent input items
|
|
function convertMsgsInput(messages: z.infer<typeof Message>[]): AgentInputItem[] {
|
|
const msgs: AgentInputItem[] = [];
|
|
|
|
for (const msg of messages) {
|
|
if (msg.role === 'assistant' && msg.content) {
|
|
msgs.push({
|
|
role: 'assistant',
|
|
content: [{
|
|
type: 'output_text',
|
|
text: `Sender agent: ${msg.agentName}\nContent: ${msg.content}`,
|
|
}],
|
|
status: 'completed',
|
|
});
|
|
} else if (msg.role === 'user') {
|
|
msgs.push({
|
|
role: 'user',
|
|
content: msg.content,
|
|
});
|
|
} else if (msg.role === 'system') {
|
|
msgs.push({
|
|
role: 'system',
|
|
content: msg.content,
|
|
});
|
|
}
|
|
}
|
|
|
|
return msgs;
|
|
}
|
|
|
|
// Helper to determine the next agent name based on control settings
|
|
function getNextAgentName(
|
|
logger: PrefixLogger,
|
|
stack: string[],
|
|
agentConfig: Record<string, z.infer<typeof WorkflowAgent>>,
|
|
workflow: z.infer<typeof Workflow>,
|
|
): string {
|
|
logger = logger.child(`getNextAgentName`);
|
|
logger.log(`stack: ${stack.join(', ')}`);
|
|
|
|
// get the last agent from the stack
|
|
// if stack is empty, use the start agent
|
|
const lastAgentName = stack.pop() || workflow.startAgent;
|
|
|
|
return lastAgentName;
|
|
|
|
// TODO: control-type logic is being ignored for now
|
|
// if control type is retain, return last agent
|
|
// const lastAgentName = stack.pop() || workflow.startAgent;
|
|
// const lastAgentConfig = agentConfig[lastAgentName];
|
|
// if (!lastAgentConfig) {
|
|
// logger.log(`last agent ${lastAgentName} not found in agent config, returning start agent: ${workflow.startAgent}`);
|
|
// return workflow.startAgent;
|
|
// }
|
|
// switch (lastAgentConfig.controlType) {
|
|
// case 'retain':
|
|
// logger.log(`last agent ${lastAgentName} control type is retain, returning last agent: ${lastAgentName}`);
|
|
// return lastAgentName;
|
|
// case 'relinquish_to_parent':
|
|
// const parentAgentName = stack.pop() || workflow.startAgent;
|
|
// logger.log(`last agent ${lastAgentName} control type is relinquish_to_parent, returning most recent parent: ${parentAgentName}`);
|
|
// return parentAgentName;
|
|
// case 'relinquish_to_start':
|
|
// logger.log(`last agent ${lastAgentName} control type is relinquish_to_start, returning start agent: ${workflow.startAgent}`);
|
|
// return workflow.startAgent;
|
|
// }
|
|
}
|
|
|
|
// Logs an event and then yields it
|
|
async function* emitEvent(
|
|
logger: PrefixLogger,
|
|
event: z.infer<typeof ZOutMessage> | z.infer<typeof ZUsage>,
|
|
): AsyncIterable<z.infer<typeof ZOutMessage> | z.infer<typeof ZUsage>> {
|
|
logger.log(`-> emitting event: ${JSON.stringify(event)}`);
|
|
yield event;
|
|
return;
|
|
}
|
|
|
|
// Emits an agent -> agent transfer event
|
|
function createTransferEvents(
|
|
fromAgent: string,
|
|
toAgent: string,
|
|
): [z.infer<typeof AssistantMessageWithToolCalls>, z.infer<typeof ToolMessage>] {
|
|
const toolCallId = crypto.randomUUID();
|
|
const m1: z.infer<typeof Message> = {
|
|
role: 'assistant',
|
|
content: null,
|
|
toolCalls: [{
|
|
id: toolCallId,
|
|
type: 'function',
|
|
function: {
|
|
name: 'transfer_to_agent',
|
|
arguments: JSON.stringify({ assistant: toAgent }),
|
|
},
|
|
}],
|
|
agentName: fromAgent,
|
|
};
|
|
|
|
const m2: z.infer<typeof Message> = {
|
|
role: 'tool',
|
|
content: JSON.stringify({ assistant: toAgent }),
|
|
toolCallId: toolCallId,
|
|
toolName: 'transfer_to_agent',
|
|
};
|
|
|
|
return [m1, m2];
|
|
}
|
|
|
|
// Tracks agent to agent transfer counts
|
|
class AgentTransferCounter {
|
|
private calls: Record<string, number> = {};
|
|
|
|
increment(fromAgent: string, toAgent: string): void {
|
|
const key = `${fromAgent}:${toAgent}`;
|
|
this.calls[key] = (this.calls[key] || 0) + 1;
|
|
}
|
|
|
|
get(fromAgent: string, toAgent: string): number {
|
|
const key = `${fromAgent}:${toAgent}`;
|
|
return this.calls[key] || 0;
|
|
}
|
|
}
|
|
|
|
class UsageTracker {
|
|
private usage: {
|
|
total: number;
|
|
prompt: number;
|
|
completion: number;
|
|
} = { total: 0, prompt: 0, completion: 0 };
|
|
|
|
increment(total: number, prompt: number, completion: number): void {
|
|
this.usage.total += total;
|
|
this.usage.prompt += prompt;
|
|
this.usage.completion += completion;
|
|
}
|
|
|
|
get(): { total: number, prompt: number, completion: number } {
|
|
return this.usage;
|
|
}
|
|
|
|
asEvent(): z.infer<typeof ZUsage> {
|
|
return {
|
|
tokens: this.usage,
|
|
};
|
|
}
|
|
}
|
|
|
|
function ensureSystemMessage(logger: PrefixLogger, messages: z.infer<typeof Message>[]) {
|
|
logger = logger.child(`ensureSystemMessage`);
|
|
|
|
// ensure that a system message is set
|
|
if (messages.length > 0 && messages[0]?.role !== 'system') {
|
|
messages.unshift({
|
|
role: 'system',
|
|
content: 'You are a helpful assistant.',
|
|
});
|
|
logger.log(`added system message: ${messages[0]?.content}`);
|
|
}
|
|
|
|
// ensure that system message isn't blank
|
|
if (messages.length > 0 && messages[0]?.role === 'system' && !messages[0].content) {
|
|
messages[0].content = 'You are a helpful assistant.';
|
|
logger.log(`updated system message: ${messages[0].content}`);
|
|
}
|
|
}
|
|
|
|
function mapConfig(workflow: z.infer<typeof Workflow>, projectTools: z.infer<typeof WorkflowTool>[]): {
|
|
agentConfig: Record<string, z.infer<typeof WorkflowAgent>>;
|
|
toolConfig: Record<string, z.infer<typeof WorkflowTool>>;
|
|
promptConfig: Record<string, z.infer<typeof WorkflowPrompt>>;
|
|
} {
|
|
const agentConfig: Record<string, z.infer<typeof WorkflowAgent>> = workflow.agents.reduce((acc, agent) => ({
|
|
...acc,
|
|
[agent.name]: agent
|
|
}), {});
|
|
const toolConfig: Record<string, z.infer<typeof WorkflowTool>> = [
|
|
...workflow.tools,
|
|
...projectTools,
|
|
].reduce((acc, tool) => ({
|
|
...acc,
|
|
[tool.name]: tool
|
|
}), {});
|
|
const promptConfig: Record<string, z.infer<typeof WorkflowPrompt>> = workflow.prompts.reduce((acc, prompt) => ({
|
|
...acc,
|
|
[prompt.name]: prompt
|
|
}), {});
|
|
return { agentConfig, toolConfig, promptConfig };
|
|
}
|
|
|
|
async function* emitGreetingTurn(logger: PrefixLogger, workflow: z.infer<typeof Workflow>): AsyncIterable<z.infer<typeof ZOutMessage> | z.infer<typeof ZUsage>> {
|
|
// find the greeting prompt
|
|
const prompt = workflow.prompts.find(p => p.type === 'greeting')?.prompt || 'How can I help you today?';
|
|
logger.log(`greeting turn: ${prompt}`);
|
|
|
|
// emit greeting turn
|
|
yield* emitEvent(logger, {
|
|
role: 'assistant',
|
|
content: prompt,
|
|
agentName: workflow.startAgent,
|
|
responseType: 'external',
|
|
});
|
|
|
|
// emit final usage information
|
|
yield* emitEvent(logger, new UsageTracker().asEvent());
|
|
}
|
|
|
|
function createAgentCallStack(messages: z.infer<typeof Message>[]): string[] {
|
|
const stack: string[] = [];
|
|
for (const msg of messages) {
|
|
if (msg.role === 'assistant' && msg.agentName) {
|
|
// skip duplicate entries
|
|
if (stack.length > 0 && stack[stack.length - 1] === msg.agentName) {
|
|
continue;
|
|
}
|
|
// add to stack
|
|
stack.push(msg.agentName);
|
|
}
|
|
}
|
|
return stack;
|
|
}
|
|
|
|
function createTools(logger: PrefixLogger, workflow: z.infer<typeof Workflow>, toolConfig: Record<string, z.infer<typeof WorkflowTool>>): Record<string, Tool> {
|
|
const tools: Record<string, Tool> = {};
|
|
for (const [toolName, config] of Object.entries(toolConfig)) {
|
|
if (config.isMcp) {
|
|
tools[toolName] = createMcpTool(logger, config, workflow.projectId);
|
|
logger.log(`created mcp tool: ${toolName}`);
|
|
} else if (config.mockTool) {
|
|
tools[toolName] = createMockTool(logger, config);
|
|
logger.log(`created mock tool: ${toolName}`);
|
|
} else {
|
|
logger.log(`unsupported tool type: ${toolName}`);
|
|
}
|
|
}
|
|
return tools;
|
|
}
|
|
|
|
function createAgents(
|
|
logger: PrefixLogger,
|
|
workflow: z.infer<typeof Workflow>,
|
|
agentConfig: Record<string, z.infer<typeof WorkflowAgent>>,
|
|
tools: Record<string, Tool>,
|
|
projectTools: z.infer<typeof WorkflowTool>[],
|
|
promptConfig: Record<string, z.infer<typeof WorkflowPrompt>>,
|
|
): { agents: Record<string, Agent>, mentions: Record<string, z.infer<typeof ConnectedEntity>[]> } {
|
|
const agents: Record<string, Agent> = {};
|
|
const mentions: Record<string, z.infer<typeof ConnectedEntity>[]> = {};
|
|
|
|
// create agents
|
|
for (const [agentName, config] of Object.entries(agentConfig)) {
|
|
const { agent, entities } = createAgent(
|
|
logger,
|
|
config,
|
|
tools,
|
|
projectTools,
|
|
workflow,
|
|
promptConfig,
|
|
);
|
|
agents[agentName] = agent;
|
|
mentions[agentName] = entities;
|
|
logger.log(`created agent: ${agentName}`);
|
|
}
|
|
|
|
// set handoffs
|
|
for (const [agentName, agent] of Object.entries(agents)) {
|
|
const connectedAgentNames = (mentions[agentName] || []).filter(e => e.type === 'agent').map(e => e.name);
|
|
agent.handoffs = connectedAgentNames.map(e => agents[e]).filter(Boolean) as Agent[];
|
|
logger.log(`set handoffs for ${agentName}: ${connectedAgentNames.join(',')}`);
|
|
}
|
|
|
|
return { agents, mentions };
|
|
}
|
|
|
|
// Main function to stream an agentic response
|
|
// using OpenAI Agents SDK
|
|
export async function* streamResponse(
|
|
workflow: z.infer<typeof Workflow>,
|
|
projectTools: z.infer<typeof WorkflowTool>[],
|
|
messages: z.infer<typeof Message>[],
|
|
): AsyncIterable<z.infer<typeof ZOutMessage> | z.infer<typeof ZUsage>> {
|
|
// set up logging
|
|
let logger = new PrefixLogger(`agent-loop`)
|
|
logger.log('projectId', workflow.projectId);
|
|
logger.log('workflow', workflow.name);
|
|
|
|
// ensure valid system message
|
|
ensureSystemMessage(logger, messages);
|
|
|
|
// if there is only a system message, emit greeting turn and return
|
|
if (messages.length === 1 && messages[0]?.role === 'system') {
|
|
yield* emitGreetingTurn(logger, workflow);
|
|
return;
|
|
}
|
|
|
|
// create map of agent, tool and prompt configs
|
|
const { agentConfig, toolConfig, promptConfig } = mapConfig(workflow, projectTools);
|
|
|
|
// create agent call stack from input messages
|
|
const stack = createAgentCallStack(messages);
|
|
|
|
// create tools
|
|
const tools = createTools(logger, workflow, toolConfig);
|
|
|
|
// create agents
|
|
const { agents } = createAgents(logger, workflow, agentConfig, tools, projectTools, promptConfig);
|
|
|
|
// track agent to agent calls
|
|
const transferCounter = new AgentTransferCounter();
|
|
|
|
// track usage
|
|
const usageTracker = new UsageTracker();
|
|
|
|
// get next agent name
|
|
let agentName = getNextAgentName(logger, stack, agentConfig, workflow);
|
|
|
|
// set up initial state for loop
|
|
logger.log('@@ starting agent turn @@');
|
|
let iter = 0;
|
|
const turnMsgs: z.infer<typeof Message>[] = [...messages];
|
|
|
|
// loop indefinitely
|
|
turnLoop: while (true) {
|
|
// increment loop counter
|
|
iter++;
|
|
|
|
// set up logging
|
|
const loopLogger = logger.child(`iter-${iter}`);
|
|
|
|
// log agent info
|
|
loopLogger.log(`agent name: ${agentName}`);
|
|
loopLogger.log(`stack: ${stack.join(', ')}`);
|
|
if (!agents[agentName]) {
|
|
throw new Error(`agent not found in agent config!`);
|
|
}
|
|
const agent: Agent = agents[agentName]!;
|
|
|
|
// convert messages to agents sdk compatible input
|
|
const inputs = convertMsgsInput(turnMsgs);
|
|
|
|
// run the agent
|
|
const result = await run(agent, inputs, {
|
|
stream: true,
|
|
});
|
|
|
|
// handle streaming events
|
|
for await (const event of result) {
|
|
const eventLogger = loopLogger.child(event.type);
|
|
// eventLogger.log(`----------> event: ${JSON.stringify(event)}`);
|
|
|
|
switch (event.type) {
|
|
case 'raw_model_stream_event':
|
|
if (event.data.type === 'response_done') {
|
|
for (const output of event.data.response.output) {
|
|
// handle tool call invocation
|
|
// except for transfer_to_* tool calls
|
|
if (output.type === 'function_call' && !output.name.startsWith('transfer_to')) {
|
|
const m: z.infer<typeof Message> = {
|
|
role: 'assistant',
|
|
content: null,
|
|
toolCalls: [{
|
|
id: output.callId,
|
|
type: 'function',
|
|
function: {
|
|
name: output.name,
|
|
arguments: output.arguments,
|
|
},
|
|
}],
|
|
agentName: agentName,
|
|
};
|
|
|
|
// add message to turn
|
|
turnMsgs.push(m);
|
|
|
|
// emit event
|
|
yield* emitEvent(eventLogger, m);
|
|
}
|
|
}
|
|
|
|
// update usage information
|
|
usageTracker.increment(
|
|
event.data.response.usage.totalTokens,
|
|
event.data.response.usage.inputTokens,
|
|
event.data.response.usage.outputTokens
|
|
);
|
|
eventLogger.log(`updated usage information: ${JSON.stringify(usageTracker.get())}`);
|
|
}
|
|
break;
|
|
case 'run_item_stream_event':
|
|
// handle handoff event
|
|
if (event.name === 'handoff_occurred' && event.item.type === 'handoff_output_item') {
|
|
// skip if its the same agent
|
|
if (agentName === event.item.targetAgent.name) {
|
|
eventLogger.log(`skipping handoff to same agent: ${agentName}`);
|
|
break;
|
|
}
|
|
|
|
// emit transfer tool call invocation
|
|
const [transferStart, transferComplete] = createTransferEvents(agentName, event.item.targetAgent.name);
|
|
|
|
// add messages to turn
|
|
turnMsgs.push(transferStart);
|
|
turnMsgs.push(transferComplete);
|
|
|
|
// emit events
|
|
yield* emitEvent(eventLogger, transferStart);
|
|
yield* emitEvent(eventLogger, transferComplete);
|
|
|
|
// update transfer counter
|
|
transferCounter.increment(agentName, event.item.targetAgent.name);
|
|
|
|
// add current agent to stack
|
|
stack.push(agentName);
|
|
|
|
// set this as the new agent name
|
|
agentName = event.item.targetAgent.name;
|
|
loopLogger.log(`switched to agent: ${agentName}`);
|
|
}
|
|
|
|
// handle tool call result
|
|
if (event.item.type === 'tool_call_output_item' &&
|
|
event.item.rawItem.type === 'function_call_result' &&
|
|
event.item.rawItem.status === 'completed' &&
|
|
event.item.rawItem.output.type === 'text') {
|
|
const m: z.infer<typeof Message> = {
|
|
role: 'tool',
|
|
content: event.item.rawItem.output.text,
|
|
toolCallId: event.item.rawItem.callId,
|
|
toolName: event.item.rawItem.name,
|
|
};
|
|
|
|
// add message to turn
|
|
turnMsgs.push(m);
|
|
|
|
// emit event
|
|
yield* emitEvent(eventLogger, m);
|
|
}
|
|
|
|
// handle model response message output
|
|
if (event.item.type === 'message_output_item' &&
|
|
event.item.rawItem.type === 'message' &&
|
|
event.item.rawItem.status === 'completed') {
|
|
// check response visibility
|
|
const isInternal = agentConfig[agentName]?.outputVisibility === 'internal';
|
|
for (const content of event.item.rawItem.content) {
|
|
if (content.type === 'output_text') {
|
|
// create message
|
|
const msg: z.infer<typeof Message> = {
|
|
role: 'assistant',
|
|
content: content.text,
|
|
agentName: agentName,
|
|
responseType: isInternal ? 'internal' : 'external',
|
|
};
|
|
|
|
// add message to turn
|
|
turnMsgs.push(msg);
|
|
|
|
// emit event
|
|
yield* emitEvent(eventLogger, msg);
|
|
}
|
|
}
|
|
|
|
// if this is an internal agent, switch to previous agent
|
|
if (isInternal) {
|
|
const current = agentName;
|
|
agentName = getNextAgentName(logger, stack, agentConfig, workflow);
|
|
|
|
// emit transfer tool call invocation
|
|
const [transferStart, transferComplete] = createTransferEvents(current, agentName);
|
|
|
|
// add messages to turn
|
|
turnMsgs.push(transferStart);
|
|
turnMsgs.push(transferComplete);
|
|
|
|
// emit events
|
|
yield* emitEvent(eventLogger, transferStart);
|
|
yield* emitEvent(eventLogger, transferComplete);
|
|
|
|
// update transfer counter
|
|
transferCounter.increment(current, agentName);
|
|
|
|
// add current agent to stack
|
|
stack.push(current);
|
|
|
|
// set this as the new agent name
|
|
loopLogger.log(`switched to agent (reason: internal agent put out a message): ${agentName}`);
|
|
|
|
// run the turn from the previous agent
|
|
continue turnLoop;
|
|
}
|
|
break;
|
|
}
|
|
break;
|
|
default:
|
|
break;
|
|
}
|
|
}
|
|
|
|
// if the last message was a text response by a user-facing agent, complete the turn
|
|
// loopLogger.log(`iter end, turnMsgs: ${JSON.stringify(turnMsgs)}, agentName: ${agentName}`);
|
|
const lastMessage = turnMsgs[turnMsgs.length - 1];
|
|
if (agentConfig[agentName]?.outputVisibility === 'user_facing' &&
|
|
lastMessage?.role === 'assistant' &&
|
|
lastMessage?.content !== null &&
|
|
lastMessage?.agentName === agentName
|
|
) {
|
|
loopLogger.log(`last message was by a user_facing agent, breaking out of parent loop`);
|
|
break turnLoop;
|
|
}
|
|
}
|
|
|
|
// emit usage information
|
|
yield* emitEvent(logger, usageTracker.asEvent());
|
|
}
|
|
|
|
// this is a sync version of streamResponse
|
|
export async function getResponse(
|
|
workflow: z.infer<typeof Workflow>,
|
|
projectTools: z.infer<typeof WorkflowTool>[],
|
|
messages: z.infer<typeof Message>[],
|
|
): Promise<{
|
|
messages: z.infer<typeof ZOutMessage>[],
|
|
usage: z.infer<typeof ZUsage>,
|
|
}> {
|
|
const out: z.infer<typeof ZOutMessage>[] = [];
|
|
let usage: z.infer<typeof ZUsage> = {
|
|
tokens: {
|
|
total: 0,
|
|
prompt: 0,
|
|
completion: 0,
|
|
},
|
|
};
|
|
for await (const event of streamResponse(workflow, projectTools, messages)) {
|
|
if ('role' in event && event.role === 'assistant') {
|
|
out.push(event);
|
|
}
|
|
if ('tokens' in event) {
|
|
usage = event;
|
|
}
|
|
}
|
|
return { messages: out, usage };
|
|
} |