Route Workers AI helpers through llm-providers

web/src/kernel/router.ts

@@ -1,3 +1,4 @@
+import { createLLMProviderFactory } from '@stackbilt/llm-providers';
 import { getProcedure, findNearMiss, procedureKey, PROCEDURE_MIN_SUCCESSES, PROCEDURE_MIN_SUCCESS_RATE, getConversationHistory } from './memory/index.js';
 import { askGroq, askGroqWithLogprobs } from '../groq.js';
 import type { KernelIntent, ExecutionPlan, Executor } from './types.js';
@@ -73,21 +74,21 @@ async function classifyWithWorkersAI(
   systemPrompt: string,
   userPrompt: string,
 ): Promise<string> {
-  const result = await ai.run('@cf/meta/llama-3.2-3b-instruct', {
+  const result = await createLLMProviderFactory({
+    cloudflare: { ai },
+    fallbackRules: [],
+    enableCircuitBreaker: true,
+    enableRetries: true,
+  }).generateResponse({
+    model: '@cf/meta/llama-3.1-8b-instruct',
+    systemPrompt,
     messages: [
-      { role: 'system', content: systemPrompt },
       { role: 'user', content: userPrompt },
     ],
-    max_tokens: 200,
+    maxTokens: 200,
     temperature: 0.1,
-  }) as { response?: unknown };
-  const raw = result.response;
-  if (typeof raw === 'string') return raw;
-  if (raw == null) return '';
-  // Workers AI sometimes returns structured responses (objects with tool_calls,
-  // arrays of segments, etc.). Coerce to string so downstream .trim()/JSON.parse
-  // callers don't crash on non-string payloads.
-  return typeof raw === 'object' ? JSON.stringify(raw) : String(raw);
+  });
+  return result.message ?? '';
 }
 
 

web/src/kernel/scheduled/conversation-facts.ts

@@ -5,6 +5,7 @@
 // Workers AI llama-3.1-8b (free tier) as fallback.
 
 import { type EdgeEnv } from '../dispatch.js';
+import { buildLLMProviderFactory } from '../provider-factory.js';
 import { recordMemory as recordMemoryAdapter } from '../memory-adapter.js';
 import { askGroq } from '../../groq.js';
 import { pushFactsToMindSpring, type FactEntry } from './mindspring-notebook.js';
@@ -97,13 +98,14 @@ async function askAi(
   }
   // Workers AI fallback — llama-3.1-8b is on the genuine free tier
   if (env.ai) {
-    const result = await env.ai.run(
-      '@cf/meta/llama-3.1-8b-instruct' as Parameters<Ai['run']>[0],
-      { messages: [{ role: 'system', content: system }, { role: 'user', content: user }] },
-    );
-    if (typeof result === 'string') return result;
-    const obj = result as { response?: string; choices?: Array<{ message?: { content?: string } }> };
-    return obj.choices?.[0]?.message?.content ?? obj.response ?? '';
+    const result = await buildLLMProviderFactory(env).generateResponse({
+      model: '@cf/meta/llama-3.1-8b-instruct',
+      systemPrompt: system,
+      messages: [{ role: 'user', content: user }],
+      maxTokens: 1024,
+      temperature: 0.2,
+    });
+    return result.message ?? '';
   }
   throw new Error('[conv-facts] No LLM provider available (groqApiKey and env.ai both missing)');
 }

web/src/kernel/scheduled/dreaming/llm.ts

@@ -1,6 +1,7 @@
 // Shared LLM helper — Groq first (free, 70B quality), Workers AI 70B fallback
 
 import type { EdgeEnv } from '../../dispatch.js';
+import { buildLLMProviderFactory } from '../../provider-factory.js';
 import { askGroq } from '../../../groq.js';
 
 export async function askWorkersAiOrGroq(
@@ -20,11 +21,14 @@ export async function askWorkersAiOrGroq(
   }
   // Workers AI fallback — only fires if Groq is unavailable or throws
   if (env.ai) {
-    const result = await env.ai.run(
-      '@cf/meta/llama-3.3-70b-instruct-fp8-fast' as Parameters<Ai['run']>[0],
-      { messages: [{ role: 'system', content: system }, { role: 'user', content: user }] },
-    ) as { response?: string; choices?: Array<{ message?: { content?: string } }> };
-    return result.choices?.[0]?.message?.content ?? result.response ?? '';
+    const result = await buildLLMProviderFactory(env).generateResponse({
+      model: '@cf/meta/llama-3.3-70b-instruct-fp8-fast',
+      systemPrompt: system,
+      messages: [{ role: 'user', content: user }],
+      maxTokens: 2048,
+      temperature: 0.2,
+    });
+    return result.message ?? '';
   }
   throw new Error('[dreaming] No LLM provider available (groqApiKey and env.ai both missing)');
 }

web/src/kernel/scheduled/social-engage.ts

@@ -2,6 +2,7 @@
 // Runs every 6 hours. Likes replies, follows back real accounts,
 // replies to substantive comments with Workers AI.
 
+import { createLLMProviderFactory } from '@stackbilt/llm-providers';
 import { type EdgeEnv } from '../dispatch.js';
 import {
   getNotifications,
@@ -153,21 +154,24 @@ async function generateReply(
   incomingText: string,
   authorHandle: string,
 ): Promise<string | null> {
-  const result = await ai.run('@cf/meta/llama-3.1-8b-instruct' as Parameters<Ai['run']>[0], {
+  const result = await createLLMProviderFactory({
+    cloudflare: { ai },
+    fallbackRules: [],
+    enableCircuitBreaker: true,
+    enableRetries: true,
+  }).generateResponse({
+    model: '@cf/meta/llama-3.1-8b-instruct',
+    systemPrompt: `You are the operator's social media voice. Direct, builder-energy, anti-corporate. No emoji spam. No "excited to announce." Keep replies under 200 chars. Be genuine and conversational. If you can't add value, return SKIP.`,
     messages: [
-      {
-        role: 'system',
-        content: `You are the operator's social media voice. Direct, builder-energy, anti-corporate. No emoji spam. No "excited to announce." Keep replies under 200 chars. Be genuine and conversational. If you can't add value, return SKIP.`,
-      },
       {
         role: 'user',
         content: `@${authorHandle} replied to our Bluesky post: "${incomingText}"\n\nWrite a brief, genuine reply. Return ONLY the reply text, or SKIP if there's nothing meaningful to add.`,
       },
     ],
-    max_tokens: 100,
-  }) as { response?: string };
+    maxTokens: 100,
+  });
 
-  const reply = result.response?.trim();
+  const reply = result.message?.trim();
   if (!reply || reply === 'SKIP' || reply.length < 5) return null;
 
   // Safety: truncate to 300 chars (Bluesky limit)