Merge pull request #350 from AayushGTM6693/feat/llm

Feat(llm): add LLM gating

Author: mesudip

agent-manager/prisma/schema.prisma

@@ -26,6 +26,7 @@ model Agent {
   triggers           Trigger[] // One-to-many relationship with Trigger
   template           Template?            @relation(fields: [template_id], references: [id], onDelete: SetNull, onUpdate: Cascade)
   secret_key         Bytes                @db.ByteA
+  config             Json?
 }
 
 model Trigger {

agent-manager/src/repository/agent_manager_repository.ts

@@ -24,6 +24,7 @@ export async function fetchAgentConfiguration(agentId: string): Promise<
           configurations: any[]
           agentIndex: number
           agentName: string
+          agentConfig: any
       }
     | undefined
 > {
@@ -46,6 +47,8 @@ export async function fetchAgentConfiguration(agentId: string): Promise<
             const instanceCount = Number(agentInstance.instance)
             const agentIndex = Number(agentInstance.index)
             const agentName = agentInstance.name
+            const agentConfig = (agentInstance as any).config || null
+
             const configurationsData = agentConfigurations.map(
                 (config: { id: string; type: string; data: JsonValue; action: JsonValue }) => ({
                     id: config.id,
@@ -55,7 +58,7 @@ export async function fetchAgentConfiguration(agentId: string): Promise<
                 })
             )
 
-            return { instanceCount, configurations: configurationsData, agentIndex, agentName }
+            return { instanceCount, configurations: configurationsData, agentIndex, agentName, agentConfig }
         }
     } catch (error: any) {
         console.log(`Error fetching agent configuration: ${error}`)

agent-node/.env.example

@@ -1,2 +1,3 @@
 TOKEN=
-WS_URL=ws://localhost:3001
+WS_URL=ws://localhost:3001
+GEMINI_API_KEY=

agent-node/package.json

@@ -18,6 +18,7 @@
     "description": "",
     "dependencies": {
         "@emurgo/cardano-serialization-lib-asmjs": "^11.5.0",
+        "@google/genai": "^1.13.0",
         "@types/ws": "^8.5.10",
         "axios": "^1.6.8",
         "bech32": "^2.0.0",
@@ -29,8 +30,8 @@
         "ws": "^8.18.0"
     },
     "devDependencies": {
-        "@types/luxon": "^3.4.2",
         "@eslint/js": "^9.4.0",
+        "@types/luxon": "^3.4.2",
         "@types/node-cron": "^3.0.11",
         "@types/websocket": "^1.0.10",
         "eslint": "8",

agent-node/src/constants/global.ts

@@ -3,9 +3,16 @@ import { IEventBasedAction } from '../types/eventTriger'
 export const globalState: {
     eventTriggerTypeDetails: IEventBasedAction[]
     agentName: string
+    systemPrompt:string
+    functionLLMSettings: Record<
+        string,
+        { enabled: boolean; userPrefText: string; prefs?: any }
+    >
 } = {
     eventTriggerTypeDetails: [],
     agentName: '',
+    systemPrompt:'',
+    functionLLMSettings: {}
 }
 
 export const globalRootKeyBuffer: { value: Buffer | null } = {

agent-node/src/executor/AgentRunner.ts

@@ -18,7 +18,7 @@ export class AgentRunner {
         this.executor = new Executor(null, managerInterface, txListener)
     }
 
-    invokeFunction(triggerType: TriggerType, instanceIndex: number, method: string, ...args: any) {
+    async invokeFunction(triggerType: TriggerType, instanceIndex: number, method: string, ...args: any) {
         this.executor.invokeFunction(method, ...args).then((result) => {
             saveTxLog(result, this.managerInterface, triggerType, instanceIndex)
         })

agent-node/src/executor/LLMGatedRunner.ts

@@ -0,0 +1,89 @@
+import { TriggerType } from '../service/triggerService'
+import { saveTxLog } from '../utils/agent'
+import { globalState } from '../constants/global'
+import { LLMService } from '../service/LLMService'
+import { EventContext } from './BaseFunction'
+import { AgentRunner } from './AgentRunner'
+
+export class LLMGatedRunner {
+    constructor(private readonly core: AgentRunner) {}
+
+    async invokeFunction(triggerType: TriggerType, instanceIndex: number, method: string, ...args: any) {
+        const extractedArgs = this.extractArgumentValues(args)
+        const shouldGate = this.shouldUseLLMForFunction(method) && this.isCron(triggerType)
+        if (shouldGate) {
+            try {
+                const llm = new LLMService()
+                const decision = await llm.shouldExecuteFunction(
+                    method,
+                    extractedArgs,
+                    {},
+                    this.getUserPreferenceText(method),
+                    this.getSystemPrompt()
+                )
+                if (!decision.should_execute) {
+                    const blocked = [
+                        {
+                            function: method,
+                            arguments: args,
+                            return: {
+                                operation: method,
+                                executed: false,
+                                blocked_by_llm: true,
+                                llm_reasoning: decision.reasoning,
+                                llm_confidence: decision.confidence,
+                                message: `LLM blocked: ${decision.reasoning}`,
+                                timestamp: new Date().toISOString(),
+                            },
+                        },
+                    ]
+                    saveTxLog(blocked, (this.core as any).managerInterface, triggerType, instanceIndex)
+                    return
+                }
+            } catch (e) {
+                console.error(`LLM gating failed, continuing: ${e}`)
+            }
+        }
+        return this.core.invokeFunction(triggerType, instanceIndex, method, ...args)
+    }
+
+    async invokeFunctionWithEventContext(
+        eventFilterContext: any,
+        context: EventContext,
+        triggerType: TriggerType,
+        instanceIndex: number,
+        method: string,
+        parameters: any[]
+    ) {
+        return this.core.invokeFunctionWithEventContext(
+            eventFilterContext,
+            context,
+            triggerType,
+            instanceIndex,
+            method,
+            parameters
+        )
+    }
+
+    async remakeContext(index: number) {
+        return this.core.remakeContext(index)
+    }
+
+    // helpers
+    private isCron(triggerType: TriggerType): boolean {
+        return String(triggerType) === 'CRON'
+    }
+    private shouldUseLLMForFunction(method: string): boolean {
+        const fnCfg = globalState.functionLLMSettings?.[method]
+        return !!(fnCfg && fnCfg.enabled)
+    }
+    private getUserPreferenceText(method: string): string {
+        return globalState.functionLLMSettings?.[method]?.userPrefText || ''
+    }
+    private getSystemPrompt(): string {
+        return (globalState.systemPrompt ?? '').toString()
+    }
+    private extractArgumentValues(args: any[]) {
+        return args.map((a) => (a && typeof a === 'object' && 'value' in a ? a.value : a))
+    }
+}

agent-node/src/index.ts

@@ -14,6 +14,7 @@ import { AgentRunner } from './executor/AgentRunner'
 import { decodeBase64string } from './utils/base64converter'
 import { validateToken } from './utils/validator'
 import { getHandlers } from './executor/AgentFunctions'
+import { LLMGatedRunner } from './executor/LlmGatedRunner'
 
 configDotenv()
 let wsUrl: string = process.env.WS_URL as string
@@ -62,7 +63,7 @@ function connectToManagerWebSocket() {
     const managerInterface = new ManagerInterface(rpcChannel)
     const txListener = new TxListener()
 
-    const agentRunners: Array<AgentRunner> = []
+    const agentRunners: Array<any> = []
 
     rpcChannel.on('methodCall', (method, args) => {
         agentRunners.forEach((runner, index) => {
@@ -71,14 +72,58 @@ function connectToManagerWebSocket() {
     })
 
     const topicHandler = new RpcTopicHandler(managerInterface, txListener)
+
+    // LLM settings extractor from configurations
+    function applyFnSettingsFromConfigurations(message: any) {
+        if (!message?.configurations) return
+        globalState.functionLLMSettings = {}
+        message.configurations.forEach((cfg: any) => {
+            const act = cfg?.action || {}
+            if (act.function_name) {
+                globalState.functionLLMSettings[act.function_name] = {
+                    enabled: !!act.llm_enabled,
+                    userPrefText: act.llm_user_preferences_text || '',
+                    prefs: act.llm_preferences || undefined,
+                }
+            }
+        })
+        console.log('[INIT] LLM settings for:', Object.keys(globalState.functionLLMSettings))
+    }
+
+    // loads the system prompt
+    function applySystemPromptFromMessage(message: any, logCtx: string) {
+        const prompt = message?.agentConfig?.system_prompt ?? message?.config?.system_prompt
+
+        if (typeof prompt === 'string' && prompt.length && prompt !== globalState.systemPrompt) {
+            globalState.systemPrompt = prompt
+        }
+    }
+
     rpcChannel.on('event', (topic, message) => {
+        // initial payload containing configs
+        if (topic === 'initial_config') {
+            applySystemPromptFromMessage(message, 'initial_config')
+            applyFnSettingsFromConfigurations(message)
+            return
+        }
+
+        // config updates from manager
+        if (topic === 'config_updated') {
+            applyFnSettingsFromConfigurations(message)
+            applySystemPromptFromMessage(message, 'initial_config')
+        }
+
         if (topic == 'instance_count') {
+            applySystemPromptFromMessage(message, 'initial_config')
+            applyFnSettingsFromConfigurations(message)
+
             globalRootKeyBuffer.value = message.rootKeyBuffer
             globalState.agentName = message.agentName
             Array(message.instanceCount)
                 .fill('')
                 .forEach(async (item, index) => {
-                    const runner = new AgentRunner(managerInterface, txListener)
+                    const coreRunner = new AgentRunner(managerInterface, txListener)
+                    const runner = new LLMGatedRunner(coreRunner)
                     await runner.remakeContext(index)
                     agentRunners.push(runner)
                 })

agent-node/src/service/LLMService.ts

@@ -0,0 +1,127 @@
+import 'dotenv/config'
+import { GoogleGenAI } from '@google/genai'
+
+const ai = new GoogleGenAI({})
+
+export class LLMService {
+  private apiKey: string
+
+  constructor() {
+    this.apiKey = process.env.GEMINI_API_KEY || ''
+    if (!this.apiKey) {
+      console.warn('No Gemini API key')
+    }
+  }
+
+  async shouldExecuteFunction(
+    functionName: string,
+    functionArgs: any[],
+    structuredPreferences: any,
+    userPreferenceText: any,
+    systemPrompt: string
+  ): Promise<{
+    should_execute: boolean
+    confidence: number
+    reasoning: string
+  }> {
+    if (!this.apiKey) {
+      console.log('LLM not configured')
+      return {
+        should_execute: true,
+        confidence: 0.5,
+        reasoning: 'LLM not configured, default allow',
+      }
+    }
+
+    try {
+      const prompt = this.buildPrompt(
+        functionName,
+        functionArgs,
+        structuredPreferences,
+        userPreferenceText,
+        systemPrompt
+      )
+
+      console.log('Asking LLM...')
+
+      const response: any = await ai.models.generateContent({
+        model: 'gemini-2.5-flash',
+        contents: prompt,
+      })
+
+      console.log('Response from Gemini:', response.text)
+
+      const decision = this.extractJson(response.text)
+      console.log('After parsing:', decision)
+
+      return {
+        should_execute: decision.should_execute,
+        confidence: decision.confidence,
+        reasoning: decision.reasoning,
+      }
+    } catch (error: any) {
+      console.error('LLM failed, error:', error)
+      return {
+        should_execute: true,
+        confidence: 0.6,
+        reasoning: `LLM service failed: ${error.message}`,
+      }
+    }
+  }
+
+  private buildPrompt(
+    functionName: string,
+    functionArgs: any[],
+    structuredPreferences: any,
+    userPreferenceText: any,
+    systemPrompt: string
+  ): string {
+    const baseSystemP = systemPrompt || 'You are a Cardano autonomous agent'
+    console.log('System prompt check:', systemPrompt)
+
+    const context =
+      structuredPreferences && Object.keys(structuredPreferences).length
+        ? `\nContext:\n${JSON.stringify(structuredPreferences, null, 2)} `
+        : ''
+
+    const userPolicy = userPreferenceText
+      ? `\nUser Policy:\n${userPreferenceText}`
+      : ''
+
+    console.log('User policy:', userPolicy)
+    console.log('Context:', context)
+
+    return `
+${baseSystemP}
+
+FUNCTION TO EXECUTE: ${functionName}
+Args: ${JSON.stringify(functionArgs)}${context}${userPolicy}
+
+Analyze this call strictly against "User Policy" and System prompt.
+Return ONLY JSON:
+{"should_execute": true/false, "confidence": 0.0-1.0, "reasoning": "brief"}
+`
+  }
+
+  extractJson(text: string): any {
+    const cleaned = text.replace(/```json|```/g, '').trim()
+
+    const start = cleaned.indexOf('{')
+    const end = cleaned.lastIndexOf('}')
+
+    if (start !== -1 && end !== -1) {
+      const jsonString = cleaned.substring(start, end + 1)
+      try {
+        return JSON.parse(jsonString)
+      } catch (e) {
+        console.error('JSON parse error:', e, jsonString)
+      }
+    }
+
+    return {
+      should_execute: false,
+      confidence: 0.0,
+      reasoning: 'Failed to parse LLM response',
+    }
+  }
+}