fix(intelligence): replace Workers AI model deprecated 2026-05-30

src/api/routes/mcp.js

@@ -10,6 +10,7 @@
 
 import { Hono } from "hono";
 import { dispatchToolCall } from "../../mcp/tool-dispatcher.js";
+import { resolveAiModel, extractAiText } from "../../lib/ai-model.js";
 import { MCP_TOOLS } from "../../mcp/tool-registry.js";
 
 const mcpRoutes = new Hono();
@@ -463,8 +464,7 @@ mcpRoutes.post("/sampling/sample", async (c) => {
       aiMessages.push({ role: msg.role, content: textContent });
     }
 
-    const model =
-      modelPreferences?.hints?.[0]?.name || "@cf/meta/llama-3.1-8b-instruct";
+    const model = modelPreferences?.hints?.[0]?.name || resolveAiModel(c.env);
 
     const aiResult = await c.env.AI.run(model, {
       messages: aiMessages,
@@ -474,7 +474,7 @@ mcpRoutes.post("/sampling/sample", async (c) => {
     return c.json({
       model,
       role: "assistant",
-      content: { type: "text", text: aiResult.response },
+      content: { type: "text", text: extractAiText(aiResult) },
       stopReason: "endTurn",
     });
   } catch (error) {

src/api/routes/prompts.js

@@ -16,6 +16,7 @@
  */
 
 import { Hono } from "hono";
+import { resolveAiModel, extractAiText } from "../../lib/ai-model.js";
 
 export const promptRoutes = new Hono();
 
@@ -356,7 +357,7 @@ promptRoutes.post("/execute", async (c) => {
       // Direct AI proxy via Workers AI
       if (c.env.AI) {
         const aiResult = await c.env.AI.run(
-          c.env.AI_MODEL_PRIMARY || "@cf/meta/llama-4-scout-17b-16e-instruct",
+          resolveAiModel(c.env),
           {
             messages: [
               { role: "system", content: composedPrompt },
@@ -365,7 +366,7 @@ promptRoutes.post("/execute", async (c) => {
             max_tokens: body.maxTokens || 4096,
           }
         );
-        result = aiResult.response;
+        result = extractAiText(aiResult);
         executedBy = "chittyconnect/workers-ai";
       } else {
         throw new Error("No AI binding or dispatch target available");

src/intelligence/cognitive-coordination.js

@@ -8,6 +8,7 @@
  */
 
 import { ContextConsciousness } from "./context-consciousness.js";
+import { resolveAiModel, extractAiText } from "../lib/ai-model.js";
 import { MemoryCloude } from "./memory-cloude.js";
 
 /**
@@ -366,7 +367,7 @@ Respond in JSON format: {
   "risks": ["..."]
 }`;
 
-      const response = await this.env.AI.run("@cf/meta/llama-3.1-8b-instruct", {
+      const response = await this.env.AI.run(resolveAiModel(this.env), {
         messages: [
           {
             role: "system",
@@ -380,7 +381,7 @@ Respond in JSON format: {
         ],
       });
 
-      const analysis = JSON.parse(response.response);
+      const analysis = JSON.parse(extractAiText(response));
       console.log(
         `[Cognitive-Coordination™] Task complexity: ${analysis.complexity}`,
       );
@@ -492,7 +493,7 @@ Respond in JSON format: {
   "recommendations": ["..."]
 }`;
 
-      const response = await this.env.AI.run("@cf/meta/llama-3.1-8b-instruct", {
+      const response = await this.env.AI.run(resolveAiModel(this.env), {
         messages: [
           {
             role: "system",
@@ -506,7 +507,7 @@ Respond in JSON format: {
         ],
       });
 
-      const synthesis = JSON.parse(response.response);
+      const synthesis = JSON.parse(extractAiText(response));
 
       return {
         success: true,

src/intelligence/context-consciousness.js

@@ -7,6 +7,8 @@
  * @module intelligence/context-consciousness
  */
 
+import { resolveAiModel, extractAiText } from "../lib/ai-model.js";
+
 export class ContextConsciousness {
   constructor(env) {
     this.env = env;
@@ -197,7 +199,7 @@ Look for:
 
 Respond in JSON format: {"anomalies": [{"type": "...", "description": "...", "severity": "low|medium|high"}]}`;
 
-      const response = await this.env.AI.run("@cf/meta/llama-3.1-8b-instruct", {
+      const response = await this.env.AI.run(resolveAiModel(this.env), {
         messages: [
           {
             role: "system",
@@ -208,7 +210,7 @@ Respond in JSON format: {"anomalies": [{"type": "...", "description": "...", "se
         ],
       });
 
-      const result = JSON.parse(response.response);
+      const result = JSON.parse(extractAiText(response));
       return result.anomalies || [];
     } catch (error) {
       console.warn(

src/intelligence/intent-predictor.js

@@ -7,6 +7,8 @@
  * @module intelligence/intent-predictor
  */
 
+import { resolveAiModel, extractAiText } from "../lib/ai-model.js";
+
 export class IntentPredictor {
   constructor(env, deps = {}) {
     this.env = env;
@@ -582,7 +584,7 @@ Prediction: ${JSON.stringify(prediction)}
 
 Return strict JSON with keys:
 intent, candidateIntents, suggestedServices, preloadData, nextActions, confidence, signals, historySummary`;
-      const response = await this.ai.run("@cf/meta/llama-3.1-8b-instruct", {
+      const response = await this.ai.run(resolveAiModel(this.env), {
         messages: [
           {
             role: "system",
@@ -596,7 +598,7 @@ intent, candidateIntents, suggestedServices, preloadData, nextActions, confidenc
         ],
       });
 
-      const text = String(response?.response || "").trim();
+      const text = extractAiText(response).trim();
       const json = this.extractJson(text);
       if (!json) return null;
 

src/intelligence/memory-cloude.js

@@ -9,6 +9,8 @@
  * @canon chittycanon://gov/governance#core-types
  */
 
+import { resolveAiModel, extractAiText } from "../lib/ai-model.js";
+
 export class MemoryCloude {
   constructor(env) {
     this.env = env;
@@ -360,7 +362,7 @@ export class MemoryCloude {
 
     try {
       // Use AI to generate summary
-      const response = await this.env.AI.run("@cf/meta/llama-3.1-8b-instruct", {
+      const response = await this.env.AI.run(resolveAiModel(this.env), {
         messages: [
           {
             role: "system",
@@ -374,7 +376,12 @@ export class MemoryCloude {
         ],
       });
 
-      const summary = response.response;
+      const summary = extractAiText(response);
+      if (!summary) {
+        // An empty envelope must not be cached as a valid summary — fall through
+        // to the catch below so the failure stays visible instead of persisting "".
+        throw new Error("Workers AI returned no summary text");
+      }
 
       // Store summary
       await this.kv.put(`session:${sessionId}:summary`, summary, {

src/intelligence/relationship-engine.js

@@ -7,6 +7,8 @@
  * @module intelligence/relationship-engine
  */
 
+import { resolveAiModel, extractAiText } from "../lib/ai-model.js";
+
 export class RelationshipEngine {
   constructor(env) {
     this.env = env;
@@ -448,7 +450,7 @@ export class RelationshipEngine {
     if (!this.ai) return null;
 
     try {
-      const response = await this.ai.run("@cf/meta/llama-3.1-8b-instruct", {
+      const response = await this.ai.run(resolveAiModel(this.env), {
         messages: [
           {
             role: "system",
@@ -469,7 +471,7 @@ export class RelationshipEngine {
         ],
       });
 
-      return response?.response || null;
+      return extractAiText(response) || null;
     } catch (error) {
       console.warn(
         "[RelationshipEngine] Summary generation failed:",

src/lib/ai-model.js

@@ -0,0 +1,38 @@
+/**
+ * Workers AI model resolution for the ChittyConnect intelligence layer.
+ *
+ * `@cf/meta/llama-3.1-8b-instruct` aliases `@cf/meta/infire-llama-3.1-8b-instruct`,
+ * which Cloudflare deprecated on 2026-05-30. Calls to it fail with error 5028, so
+ * every intelligence-layer AI path that hardcoded it has been silently degraded
+ * since that date. Resolve the model through this module instead of inlining an id,
+ * so the next deprecation is a single env var away from being handled.
+ */
+
+export const AI_MODEL_DEFAULT = "@cf/meta/llama-4-scout-17b-16e-instruct";
+
+/**
+ * @param {{AI_MODEL_PRIMARY?: string}} env
+ * @returns {string} model id to pass to env.AI.run()
+ */
+export function resolveAiModel(env) {
+  return env?.AI_MODEL_PRIMARY || AI_MODEL_DEFAULT;
+}
+
+/**
+ * Workers AI returns a chat-completions envelope that carries the generated text
+ * both as a top-level `response` string and under `choices[0].message.content`.
+ * Read both so a model whose envelope omits either shape does not degrade to
+ * `undefined` — an undefined summary is a silent failure, which is worse than the
+ * loud one it would replace.
+ *
+ * @param {unknown} result value returned by env.AI.run()
+ * @returns {string} generated text, or "" when the envelope carries none
+ */
+export function extractAiText(result) {
+  if (typeof result === "string") return result;
+  if (!result || typeof result !== "object") return "";
+  if (typeof result.response === "string") return result.response;
+  const choice = Array.isArray(result.choices) ? result.choices[0] : null;
+  const content = choice?.message?.content;
+  return typeof content === "string" ? content : "";
+}

tests/lib/ai-model.test.js

@@ -0,0 +1,89 @@
+/**
+ * Tests for Workers AI model resolution.
+ *
+ * The envelopes below are not hand-written fixtures — they are the verbatim
+ * payloads returned by production Workers AI on 2026-09-04 via
+ * POST https://connect.chitty.cc/api/thirdparty/cloudflare/ai/run, recorded so
+ * the parser is exercised against the shape the platform actually emits rather
+ * than the shape we assume it emits.
+ */
+
+import { describe, it, expect } from "vitest";
+import {
+  AI_MODEL_DEFAULT,
+  resolveAiModel,
+  extractAiText,
+} from "../../src/lib/ai-model.js";
+
+// Recorded live: model "@cf/meta/llama-4-scout-17b-16e-instruct", prompt "reply with the word ok"
+const LIVE_SCOUT_ENVELOPE = {
+  choices: [
+    {
+      finish_reason: "stop",
+      index: 0,
+      logprobs: null,
+      message: {
+        annotations: null,
+        audio: null,
+        content: "ok",
+        function_call: null,
+        reasoning: null,
+        refusal: null,
+        role: "assistant",
+      },
+      routed_experts: null,
+      stop_reason: null,
+      token_ids: null,
+    },
+  ],
+  created: 1788561543,
+  id: "chatcmpl-a318096f-ca2c-423a-a8bc-4813ca68c36f",
+  model: "@cf/meta/llama-4-scout-17b-16e-instruct",
+  object: "chat.completion",
+  response: "ok",
+  tool_calls: [],
+  usage: { prompt_tokens: 15, completion_tokens: 2, total_tokens: 17 },
+};
+
+describe("resolveAiModel", () => {
+  it("defaults to a model that is live in production, not the 2026-05-30 deprecated one", () => {
+    expect(resolveAiModel({})).toBe(AI_MODEL_DEFAULT);
+    expect(AI_MODEL_DEFAULT).not.toContain("llama-3.1-8b-instruct");
+  });
+
+  it("honours the AI_MODEL_PRIMARY override so the next deprecation needs no code change", () => {
+    expect(
+      resolveAiModel({
+        AI_MODEL_PRIMARY: "@cf/meta/llama-3.3-70b-instruct-fp8-fast",
+      }),
+    ).toBe("@cf/meta/llama-3.3-70b-instruct-fp8-fast");
+  });
+
+  it("tolerates a missing env rather than throwing inside a catch-wrapped AI path", () => {
+    expect(resolveAiModel(undefined)).toBe(AI_MODEL_DEFAULT);
+  });
+});
+
+describe("extractAiText", () => {
+  it("reads the live production envelope", () => {
+    expect(extractAiText(LIVE_SCOUT_ENVELOPE)).toBe("ok");
+  });
+
+  it("still reads the envelope when the compat `response` field is absent", () => {
+    const { response: _dropped, ...choicesOnly } = LIVE_SCOUT_ENVELOPE;
+    expect(extractAiText(choicesOnly)).toBe("ok");
+  });
+
+  it("returns empty string — never undefined — for an envelope carrying no text", () => {
+    // A caller that stores undefined turns a loud failure into a silent one.
+    for (const empty of [
+      {},
+      { choices: [] },
+      { choices: [{ message: {} }] },
+      null,
+      7,
+    ]) {
+      expect(extractAiText(empty)).toBe("");
+    }
+  });
+});