feat: drop the bespoke Anthropic client for ai-kit's Groq -> OpenRouter chain

__tests__/ai-chat-rules.test.ts

@@ -1,6 +1,6 @@
 /**
  * Unit tests for the rule-based AI response path in lib/domain/ai-chat.ts.
- * This is the fallback used when ANTHROPIC_API_KEY is not set.
+ * This is the fallback used when no chain provider is configured.
  * ruleBasedResponse is a pure function: no DB, no network.
  */
 import { describe, it, expect, vi } from "vitest";

__tests__/ai-forms.test.ts

@@ -26,7 +26,7 @@ const field = (name: string) => {
 
 const writable = CHECK_IN_FORM.fields.filter((f) => !f.aiExcluded).map((f) => f.name);
 
-/** Stands in for the route's `complete` on a deployment with no ANTHROPIC_API_KEY. */
+/** Stands in for the route's `complete` on a deployment with no chain provider configured. */
 const keylessComplete = (instruction: string, intent: "fill" | "refine") => async () =>
   JSON.stringify({ values: keywordFallback(instruction, intent), message: "" });
 

__tests__/check-in-parse.test.ts

@@ -1,7 +1,7 @@
 /**
  * Unit tests for the keyword-based check-in parser (lib/domain/check-in-parse.ts).
  * keywordParse is pure: no DB, no network, no mocks needed.
- * This is the fallback path used when ANTHROPIC_API_KEY is not set.
+ * This is the fallback path used when no chain provider is configured.
  */
 import { describe, it, expect } from "vitest";
 import { keywordParse } from "@/lib/domain/check-in-parse";

__tests__/health.test.ts

@@ -17,11 +17,15 @@ beforeEach(() => {
 });
 
 describe("GET /api/health", () => {
-  it("returns 200 with { success: true } when the database answers", async () => {
+  it("returns 200 with { success: true } and the LLM chain's health when the database answers", async () => {
     execute.mockResolvedValueOnce([{ "?column?": 1 }]);
     const res = await GET();
     expect(res.status).toBe(200);
-    expect(await res.json()).toEqual({ success: true });
+    const body = await res.json();
+    expect(body.success).toBe(true);
+    // Informational only — asserting shape, not a specific status, since
+    // this must never gate the 200/503 the deploy pipeline checks.
+    expect(body.llm).toMatchObject({ status: expect.any(String) });
   });
 
   it("returns 503 with { success: false } when the database is unreachable", async () => {

app/api/admin/clients/[id]/session-prep/route.ts

@@ -1,7 +1,7 @@
 /**
  * Pre-session AI prep: generates a clinical summary for a practitioner
  * before a session with a specific client.
- * Uses Anthropic API directly (no SDK). Gracefully degrades when key is absent.
+ * Uses the fleet's ai-kit provider chain (Groq -> OpenRouter). Gracefully degrades when no chain provider is configured.
  */
 import { formatEnumValue, roundOne } from "@/lib/utils";
 import { db } from "@/lib/db";
@@ -18,7 +18,7 @@ import {
 } from "@/lib/db/schema";
 import { eq, desc, and, gte, isNull } from "drizzle-orm";
 import type { ProgramPhase } from "@/lib/domain/program";
-import { callClaude } from "@/lib/domain/anthropic";
+import { callLLM } from "@/lib/domain/llm";
 import { localDateString, addDaysISO } from "@/lib/utils";
 import { computeDailyAdherenceTrend } from "@/lib/domain/techniques";
 import { computeCurrentProgramWeek } from "@/lib/domain/check-in";
@@ -140,7 +140,7 @@ export async function GET(_req: Request, { params }: { params: Promise<{ id: str
     activeEnrollment ?? null,
   );
 
-  const aiSummary = await callClaude({
+  const aiSummary = await callLLM({
     messages: [
       {
         role: "user",

app/api/ai/form-assist/route.ts

@@ -11,9 +11,10 @@
  *
  * Why this is hand-written instead of `createFormAssistHandler`: the package's
  * handler owns the request, and this app needs the raw instruction inside
- * `complete` so it can answer from the keyword parser when there is no
- * ANTHROPIC_API_KEY. Owning the request here is what keeps the check-in form
- * working on a deployment with no model at all — the property the previous
+ * `complete` so it can answer from the keyword parser when no chain provider
+ * (GROQ_API_KEY / OPENROUTER_API_KEY) is configured. Owning the request here
+ * is what keeps the check-in form working on a deployment with no model at
+ * all — the property the previous
  * /api/check-in/parse route had and which a straight port would have deleted.
  */
 
@@ -23,7 +24,7 @@ import { runFormAssist, type CompleteFn } from "@fleet/ai-forms";
 import { requireAuth } from "@/lib/api";
 import { AI_FORMS } from "@/lib/config/ai-forms";
 import { API_ERR_INVALID_INPUT, FIELD_MAX_MEDIUM } from "@/lib/constants";
-import { callClaude } from "@/lib/domain/anthropic";
+import { callLLM } from "@/lib/domain/llm";
 import { keywordFallback } from "@/lib/domain/check-in-parse";
 
 const requestSchema = z.object({
@@ -57,7 +58,7 @@ export async function POST(req: Request) {
   }
 
   const complete: CompleteFn = async ({ system, prompt, maxTokens }) => {
-    const text = await callClaude({
+    const text = await callLLM({
       messages: [{ role: "user", content: prompt }],
       system,
       maxTokens,

app/api/cron/ai-digest/route.ts

@@ -3,7 +3,7 @@
  * Runs every Sunday at 19:00 CET (after weekly-report at 17:00).
  * For each client with ≥3 check-ins in the past 7 days, generates an AI
  * narrative summary and stores it in the most recent check-in's aiInsight field.
- * Gracefully degrades when ANTHROPIC_API_KEY is absent.
+ * Gracefully degrades when no chain provider (GROQ_API_KEY / OPENROUTER_API_KEY) is configured.
  */
 import { NextResponse } from "next/server";
 import { db } from "@/lib/db";

app/api/health/route.ts

@@ -3,11 +3,16 @@ export const dynamic = "force-dynamic";
 import { NextResponse } from "next/server";
 import { sql } from "drizzle-orm";
 import { db } from "@/lib/db";
+import { getLLMHealth } from "@/lib/domain/llm";
 
 // Public health check (fleet convention): 200 = the app is up AND its database
 // answers. Deploy monitoring curls this after every push — a status code is the
 // contract, the body is for humans. The body never exposes internals; the
 // detail goes to the logs, where the person fixing it is already looking.
+//
+// `llm` is informational only — it never flips the status code. A dead
+// provider key can't be fixed by a restart, so it must never fail the check
+// that triggers one. @see ai-kit's createHealthTracker, adopted fleet-wide.
 export async function GET() {
   try {
     await db.execute(sql`select 1`);
@@ -16,5 +21,5 @@ export async function GET() {
     return NextResponse.json({ success: false, error: "database unreachable" }, { status: 503 });
   }
 
-  return NextResponse.json({ success: true });
+  return NextResponse.json({ success: true, llm: getLLMHealth() });
 }

lib/constants.ts

@@ -105,8 +105,7 @@ export const COMPANY_ADDRESS = "Surf Your Life · Zollikerstrasse 183, 8008 Zür
 // "today" matches what the user sees on their wall clock, not UTC.
 export const CLINIC_TZ = "Europe/Zurich";
 
-// AI model identifiers — update here when upgrading models
-export const AI_MODEL_FAST = "claude-haiku-4-5-20251001";
+// Embedding model — chat models come from ai-kit's chain, not from here.
 export const EMBEDDING_MODEL = "text-embedding-3-small";
 export const OPENAI_EMBEDDINGS_URL = "https://api.openai.com/v1/embeddings";
 

lib/domain/ai-chat.ts

@@ -1,10 +1,10 @@
 /**
  * AI chat logic for the client portal.
  *
- * When ANTHROPIC_API_KEY is set: calls Claude with check-in context.
- * When not set: returns a rule-based response grounded in the client's actual data.
+ * When GROQ_API_KEY or OPENROUTER_API_KEY is set: calls the chain with
+ * check-in context. When neither is set: returns a rule-based response
+ * grounded in the client's actual data.
  *
- * To enable AI: set ANTHROPIC_API_KEY in your environment.
  * The generateAiReply function handles both paths — no other changes needed.
  */
 
@@ -29,7 +29,7 @@ import {
 } from "@/lib/constants";
 import { summariseCheckIns, computeCurrentProgramWeek } from "@/lib/domain/check-in";
 export { summariseCheckIns, type CheckInSummaryRow } from "@/lib/domain/check-in";
-import { callClaude } from "@/lib/domain/anthropic";
+import { callLLM } from "@/lib/domain/llm";
 import { semanticCheckInSearch } from "@/lib/domain/embeddings";
 import { localDateString, addDaysISO } from "@/lib/utils";
 import { computeDailyAdherenceTrend } from "@/lib/domain/techniques";
@@ -472,9 +472,9 @@ export function ruleBasedResponse(
   return `Here's a quick snapshot: over your last ${stats.count} check-ins, your average energy is ${stats.avgEnergy.toFixed(1)}/10${sleepSummary}, and you've had ${stats.pemCount} PEM episode${stats.pemCount === 1 ? "" : "s"}. Ask me about any of these in more detail. ${dataNote}`;
 }
 
-// ─── AI call (to be enabled when ANTHROPIC_API_KEY is set) ────────────────────
+// ─── AI call (no-op when no chain provider is configured) ─────────────────────
 
-async function callAnthropicApi(
+async function callChainChat(
   userMessage: string,
   context: BuildContextResult,
   history: { role: "user" | "assistant"; content: string }[],
@@ -548,7 +548,7 @@ ${techniqueLines}
 Current program:
 ${programLine}`;
 
-  return callClaude({
+  return callLLM({
     messages: [...history.slice(-AI_CHAT_CONTEXT_WINDOW), { role: "user", content: userMessage }],
     system: systemPrompt,
     maxTokens: 500,
@@ -572,8 +572,8 @@ export async function generateAiReply(
     activeEnrollment,
   } = context;
 
-  // Try AI first (no-op until ANTHROPIC_API_KEY is set)
-  const aiReply = await callAnthropicApi(userMessage, context, history);
+  // Try AI first (no-op if no chain provider is configured)
+  const aiReply = await callChainChat(userMessage, context, history);
   if (aiReply) return aiReply;
 
   // Fall back to rule-based response grounded in actual data

lib/domain/check-in-parse.ts

@@ -2,8 +2,8 @@
  * Parse a free-text description of a day into structured check-in fields,
  * using keyword heuristics and no network call at all.
  *
- * This is what /api/ai/form-assist answers with when ANTHROPIC_API_KEY is not
- * set, so the check-in form still fills itself in on a deployment with no
+ * This is what /api/ai/form-assist answers with when no chain provider is
+ * configured, so the check-in form still fills itself in on a deployment with no
  * model. The AI path lives in lib/config/ai-forms.ts — the field list there is
  * the only description of these fields the model ever sees, which is why there
  * is no hand-written extraction prompt duplicating the list below.

lib/domain/digest.ts

@@ -1,4 +1,4 @@
-import { callClaude } from "@/lib/domain/anthropic";
+import { callLLM } from "@/lib/domain/llm";
 
 type CheckInRow = {
   createdAt: Date;
@@ -63,5 +63,5 @@ Write a concise clinical narrative (3-5 sentences) summarising:
 
 Be factual, empathetic, and clinically precise. No bullet points — flowing prose only.`;
 
-  return callClaude({ messages: [{ role: "user", content: prompt }], maxTokens: 300 });
+  return callLLM({ messages: [{ role: "user", content: prompt }], maxTokens: 300 });
 }

lib/domain/llm.ts

@@ -0,0 +1,93 @@
+/**
+ * Chat completion via the fleet's shared AI provider chain (Groq -> OpenRouter).
+ *
+ * This used to be a bespoke Anthropic-only client, the one hand-rolled
+ * provider in a fleet where everything else installs `ai-kit`. It needed its
+ * own `ANTHROPIC_API_KEY` — a credential this fleet doesn't otherwise carry —
+ * so every caller degraded (silently, by design) on every deployment that
+ * only ever had GROQ_API_KEY / OPENROUTER_API_KEY set, which is all of them.
+ *
+ * Same graceful-degrade contract as before: returns null on any failure
+ * (no key configured, every vendor refused) rather than throwing, because
+ * every caller here already treats a null response as "fall back to the
+ * non-AI path," not as an error to surface.
+ */
+import { freeChain, usableChain, chainFrom, tryChain, createHealthTracker } from "ai-kit";
+
+const health = createHealthTracker({ downAfter: 3 });
+
+export function getLLMHealth() {
+  return health.getHealth();
+}
+
+type Message = { role: "user" | "assistant"; content: string };
+
+type CallOptions = {
+  messages: Message[];
+  system?: string;
+  maxTokens?: number;
+  /** Vendor-specific model id. Ignored — model choice comes from the chain. */
+  model?: string;
+};
+
+/**
+ * Call the chain and return the text of the first choice, or null on
+ * failure (no provider configured, or every provider in the chain refused).
+ */
+export async function callLLM({
+  messages,
+  system,
+  maxTokens = 500,
+}: CallOptions): Promise<string | null> {
+  const links = usableChain(freeChain("SURF"), process.env);
+  if (links.length === 0) return null;
+
+  // One link per vendor — a second model at the same vendor draws on the
+  // same daily budget, so it is not a real fallback. @see ai-kit chain.ts
+  const seen = new Set<string>();
+  const chain = links.filter((link) => {
+    if (seen.has(link.provider.id)) return false;
+    seen.add(link.provider.id);
+    return true;
+  });
+
+  try {
+    return await tryChain(chain, {
+      health,
+      attempt: async (link) => {
+        const [resolved] = chainFrom(undefined, [link]);
+        const res = await fetch(`${link.provider.baseUrl}/chat/completions`, {
+          method: "POST",
+          headers: {
+            "Content-Type": "application/json",
+            Authorization: `Bearer ${process.env[link.provider.keyEnv] ?? ""}`,
+          },
+          body: JSON.stringify({
+            model: resolved?.model ?? link.model,
+            max_tokens: maxTokens,
+            messages: system ? [{ role: "system", content: system }, ...messages] : messages,
+          }),
+        });
+
+        if (!res.ok) {
+          // Body goes on the thrown error for the log only — never returned.
+          const detail = await res.text().catch(() => "");
+          throw new Error(
+            `${link.provider.id} chat failed (${res.status}): ${detail.slice(0, 200)}`,
+          );
+        }
+
+        const body = (await res.json()) as {
+          choices?: Array<{ message?: { content?: string } }>;
+        };
+        const text = body.choices?.[0]?.message?.content;
+        if (!text) throw new Error(`${link.provider.id} returned no content`);
+        return text;
+      },
+    });
+  } catch {
+    // ChainExhaustedError or an empty chain — every caller treats null as
+    // "fall back," matching the contract this replaces.
+    return null;
+  }
+}

package.json

@@ -26,6 +26,7 @@
     "@auth/drizzle-adapter": "^1.11.1",
     "@fleet/ai-forms": "github:bitbaum/ai-forms#v0.1.0",
     "@sentry/nextjs": "^10.66.0",
+    "ai-kit": "github:bitbaum/ai-kit#v0.6.2",
     "bcryptjs": "^3.0.3",
     "class-variance-authority": "^0.7.1",
     "clsx": "^2.1.1",