ravi-arnan · Aug 18, 2026 · Jul 28, 2026
diff --git a/ROADMAP.md b/ROADMAP.md
@@ -43,6 +43,33 @@ Status: **Phase 1 (Web) + Phase 2 (Android alarm) selesai 2026-05-29.** Core req
 - [x] **Signed release APK** — keystore di `.archive/`, config di `keystore.properties`
 - [x] **First-launch MIUI onboarding** — auto-open PermissionsSheet sekali per-user via SharedPreferences
 
+### AI
+- [x] **Pindah upstream AI dari GitHub Models ke Groq** — done 2026-07-29
+  - **Kenapa**: GitHub Models mengumumkan pensiun per **2026-07-30** (header `sunset`,
+    ketahuan waktu ngerjain `notulen`). Fitur AI Jadwal bakal mati diam-diam kalau dibiarkan.
+    Kebetulan PAT `models:read`-nya sendiri belum pernah dibuat, jadi tidak ada yang hilang.
+  - Worker: `UPSTREAM` → `https://api.groq.com/openai/v1/chat/completions`,
+    model dipin `openai/gpt-oss-120b`, secret `GITHUB_MODELS_TOKEN` → `GROQ_API_KEY`.
+    Sisa worker (verifikasi Firebase ID token, CORS allowlist, pin model) tidak berubah —
+    Groq OpenAI-compatible.
+  - Pilihan model diuji langsung: `llama-3.3-70b-versatile` dan `qwen/qwen3.6-27b`
+    **menolak** `response_format: json_schema`. Dari dua yang bisa, 120b benar soal tanggal
+    relatif ("Selasa terdekat" → 2026-08-04) sementara 20b menjawab hari Sabtu.
+  - **Bug decoder Groq yang ketemu & diakali**: array integer tanpa constraint kehilangan
+    pemisah antar elemen. "ingatkan 1 hari dan 1 jam sebelum" konsisten balik `[144060]`
+    (1440 dan 60 nyambung) — lalu di-snap `normalizeEvent` jadi 1440 doang, pengingat kedua
+    **hilang diam-diam**. Diuji: salah di tiap run dengan `items: {type: integer}`, benar di
+    tiap run dengan enum string, dan benar juga tanpa `response_format` sama sekali → jadi
+    decoder-nya, bukan modelnya. Schema sekarang pakai enum string; `normalizeEvent` sudah
+    `Number()`-in jadi tidak ada perubahan hilir.
+  - `worker/verify-upstream.sh` — cek yang bisa dijalankan ulang: nembak Groq langsung
+    (tanpa worker/Firebase) dan meng-assert schema diterima, tanggal relatif benar, dan
+    pengingat tidak nyambung. Hijau 2026-07-29.
+  - **BUTUH RAVI**: `npx wrangler secret put GROQ_API_KEY` (key dari console.groq.com,
+    free tier) lalu `npx wrangler deploy`, dan rebuild + deploy web. Sampai itu dijalankan
+    fitur AI Jadwal tetap mati — sama seperti sebelumnya, cuma sekarang penyebabnya bukan
+    layanan yang sudah pensiun.
+
 ## Belum
 
 ### Operasional / distribusi

diff --git a/docs/DEPLOY.md b/docs/DEPLOY.md
@@ -1,5 +1,10 @@
 # myKalender - Deployment Guide
 
+> **Dokumen ini menggambarkan arsitektur yang sudah ditinggalkan** (Next.js di Vercel +
+> Turso + GitHub Models). Yang jalan sekarang: web Vite di Firebase Hosting, data di
+> Firestore, AI lewat Cloudflare Worker ke Groq. Untuk langkah deploy yang berlaku lihat
+> "Catatan teknis" di `ROADMAP.md` dan `worker/README.md`. Disimpan sebagai catatan sejarah.
+
 Deploy myKalender ke Vercel (free tier). Estimasi waktu: 15-30 menit.
 
 ---

diff --git a/docs/PLAN.md b/docs/PLAN.md
@@ -1,5 +1,9 @@
 # myKalender - Implementation Plan
 
+> **Rencana awal, bukan deskripsi sistem sekarang.** Stack berubah di tengah jalan:
+> Next.js/Vercel/Turso/GitHub Models → Vite + Firebase (Hosting, Firestore, Auth) +
+> Android native, dengan AI lewat Cloudflare Worker ke Groq. `ROADMAP.md` yang berlaku.
+
 > Step-by-step implementation plan untuk MVP. Mengacu ke [PRD](./PRD.md) dan [DESIGN](../DESIGN.md). Estimasi total 2-3 minggu.
 
 ---

diff --git a/web/src/components/AiChatPanel.tsx b/web/src/components/AiChatPanel.tsx
@@ -303,7 +303,7 @@ export function AiChatPanel({ onClose }: AiChatPanelProps) {
                 AI Jadwal
               </p>
               <p className="text-[10px] text-muted-soft mt-0.5">
-                gpt-4o-mini via GitHub Models
+                gpt-oss-120b via Groq
               </p>
             </div>
           </div>

diff --git a/web/src/lib/ai-schedule.ts b/web/src/lib/ai-schedule.ts
@@ -1,12 +1,14 @@
 // Client for natural-language schedule parsing.
-// The GitHub Models PAT is NOT in the client bundle: requests go through a
-// Cloudflare Worker proxy (see /worker) that holds the PAT server-side and
+// The upstream API key is NOT in the client bundle: requests go through a
+// Cloudflare Worker proxy (see /worker) that holds the key server-side and
 // verifies the caller's Firebase ID token. VITE_AI_PROXY_URL is the worker URL
 // (public, not a secret).
 
 import { auth } from "./firebase";
 
-const MODEL = "openai/gpt-4o-mini";
+// Groq, since GitHub Models was retired on 2026-07-30. The worker pins the
+// model server-side too; this constant only has to agree with it.
+const MODEL = "openai/gpt-oss-120b";
 const TIMEZONE = "Asia/Makassar";
 
 export type AiRecurrence = "none" | "daily" | "weekdays" | "weekly" | "monthly";
@@ -22,6 +24,11 @@ export interface AiParsedEvent {
   reminderOffsetsMinutes: number[]; // each 0/5/10/20/30/60/1440
 }
 
+/** What the model actually returns: reminder offsets arrive as strings. */
+type AiRawEvent = Omit<AiParsedEvent, "reminderOffsetsMinutes"> & {
+  reminderOffsetsMinutes: string[];
+};
+
 interface ChatMessage {
   role: "system" | "user" | "assistant";
   content: string;
@@ -55,11 +62,24 @@ const EVENT_SCHEMA = {
             type: "string",
             enum: ["none", "daily", "weekdays", "weekly", "monthly"],
           },
+          // Strings, not integers, and that is deliberate. Groq's constrained
+          // decoder drops the separator between elements of an unconstrained
+          // integer array: asking for two reminders ("1 hari dan 1 jam
+          // sebelum") reliably returns [144060] instead of [1440, 60], which
+          // normalizeEvent then snaps to a single 1440 and the second reminder
+          // is silently lost. Verified 2026-07-29: reproducible on every run
+          // with `items: { type: "integer" }`, correct on every run with an
+          // enum of strings, and correct with no response_format at all, so it
+          // is the decoder rather than the model. normalizeEvent already
+          // coerces with Number(), so nothing downstream cares.
           reminderOffsetsMinutes: {
             type: "array",
-            items: { type: "integer" },
+            items: {
+              type: "string",
+              enum: ["0", "5", "10", "20", "30", "60", "1440"],
+            },
             description:
-              "One or more reminders, minutes before start (each: 0, 5, 10, 20, 30, 60, or 1440). Usually one; use multiple only when the user asks for several alerts.",
+              "One or more reminders, minutes before start, one entry per reminder. Usually one; use multiple only when the user asks for several alerts.",
           },
         },
         required: [
@@ -180,7 +200,7 @@ export async function parseSchedulePrompt(
   const content = json.choices?.[0]?.message?.content;
   if (!content) throw new Error("Respons AI kosong");
 
-  let parsed: { events?: AiParsedEvent[] };
+  let parsed: { events?: AiRawEvent[] };
   try {
     parsed = JSON.parse(content);
   } catch {
@@ -195,7 +215,7 @@ const VALID_OFFSETS = [0, 5, 10, 20, 30, 60, 1440];
 const DATE_RE = /^\d{4}-\d{2}-\d{2}$/;
 const TIME_RE = /^\d{2}:\d{2}$/;
 
-function normalizeEvent(e: AiParsedEvent): AiParsedEvent {
+function normalizeEvent(e: AiRawEvent): AiParsedEvent {
   if (!DATE_RE.test(e.startDate)) {
     throw new Error(`AI mengembalikan tanggal tidak valid: ${e.startDate}`);
   }

diff --git a/web/src/routes/index.tsx b/web/src/routes/index.tsx
@@ -382,7 +382,7 @@ function Features() {
       icon: <Sparkles size={20} />,
       title: "AI Schedule Generator",
       body:
-        "Ketik aktivitas mingguan kamu, AI parse jadi event terstruktur. Powered by gpt-4o-mini via GitHub Models.",
+        "Ketik aktivitas mingguan kamu, AI parse jadi event terstruktur. Powered by gpt-oss-120b via Groq.",
     },
     {
       icon: <Users size={20} />,

diff --git a/worker/README.md b/worker/README.md
@@ -1,13 +1,18 @@
 # mykalender-ai — AI proxy (Cloudflare Worker)
 
-Server-side proxy for the AI Schedule Generator. Keeps the GitHub Models PAT
-off the client: the web app sends a Firebase ID token, the worker verifies it
-and forwards the request to GitHub Models with the secret PAT.
+Server-side proxy for the AI Schedule Generator. Keeps the upstream API key off
+the client: the web app sends a Firebase ID token, the worker verifies it and
+forwards the request upstream with the secret key.
+
+Upstream is **Groq** (`openai/gpt-oss-120b`). It was GitHub Models until
+2026-07-29; GitHub announced that service's retirement for 2026-07-30, so it
+moved. Groq is also free-tier and OpenAI-compatible, so only the base URL, the
+model name and the secret name changed.
 
 ## Deploy (one-time)
 
-Needs a (free) Cloudflare account and a GitHub PAT with scope `models:read`
-(create at https://github.com/settings/tokens).
+Needs a (free) Cloudflare account and a Groq API key from
+https://console.groq.com (free tier, no card).
 
 ```bash
 cd worker
@@ -16,8 +21,8 @@ npm install
 # 1. Log in to Cloudflare (opens a browser)
 npx wrangler login
 
-# 2. Store the GitHub PAT as a secret (paste when prompted)
-npx wrangler secret put GITHUB_MODELS_TOKEN
+# 2. Store the Groq key as a secret (paste when prompted)
+npx wrangler secret put GROQ_API_KEY
 
 # 3. Deploy — prints the worker URL, e.g.
 #    https://mykalender-ai.<your-subdomain>.workers.dev
@@ -37,11 +42,37 @@ firebase deploy --only hosting`).
 
 - `wrangler.toml [vars]` — `FIREBASE_PROJECT_ID` and `ALLOWED_ORIGINS`
   (comma-separated CORS allowlist). Not secret.
-- `GITHUB_MODELS_TOKEN` — the PAT, stored as a Wrangler secret.
+- `GROQ_API_KEY` — the Groq key, stored as a Wrangler secret.
 
 ## Local dev
 
 ```bash
-echo "GITHUB_MODELS_TOKEN=ghp_xxx" > .dev.vars   # gitignored
+echo "GROQ_API_KEY=gsk_xxx" > .dev.vars   # gitignored
 npx wrangler dev
 ```
+
+## Checking the upstream
+
+`./verify-upstream.sh` sends the real schema and a real Indonesian prompt
+straight to Groq, bypassing the worker and Firebase. It is the fastest way to
+tell "the model or the API changed" apart from "our auth broke".
+
+```bash
+GROQ_API_KEY=gsk_xxx ./verify-upstream.sh
+```
+
+## Two things that look wrong until you know why
+
+**Reminder offsets are strings in the JSON schema, not integers.** Groq's
+constrained decoder drops the separator between elements of an unconstrained
+integer array. Asking for two reminders returns `[144060]` instead of
+`[1440, 60]`, and the client then snaps that to a single 1440 — the second
+reminder disappears silently. With an enum of strings it is correct on every
+run, and with no `response_format` at all it is also correct, so the fault is
+in the decoder rather than the model. `normalizeEvent` coerces with `Number()`,
+so nothing downstream noticed the change. Do not "tidy" these back to integers.
+
+**Only the gpt-oss models are candidates.** `llama-3.3-70b-versatile` and
+`qwen/qwen3.6-27b` reject `response_format: json_schema` outright. Between the
+two that work, 120b got relative dates ("Selasa terdekat") right where 20b
+returned a Saturday.
diff --git a/worker/src/index.js b/worker/src/index.js
@@ -1,12 +1,20 @@
-// Cloudflare Worker: server-side proxy for the GitHub Models AI feature.
+// Cloudflare Worker: server-side proxy for the AI Schedule Generator.
 //
-// Keeps the GitHub PAT (GITHUB_MODELS_TOKEN secret) off the client. The web
+// Keeps the upstream API key (GROQ_API_KEY secret) off the client. The web
 // app sends a Firebase ID token instead; this worker verifies it against
 // Google's public keys so only signed-in users of *this* Firebase project can
 // use the proxy (the worker URL itself is public, in the client bundle).
-
-const UPSTREAM = "https://models.github.ai/inference/chat/completions";
-const MODEL = "openai/gpt-4o-mini";
+//
+// Upstream was GitHub Models until 2026-07-29. GitHub Models announced its
+// retirement for 2026-07-30 (a `sunset` response header), so this moved to
+// Groq's OpenAI-compatible endpoint, which is also free-tier and already holds
+// the key used elsewhere. Do not go back.
+
+const UPSTREAM = "https://api.groq.com/openai/v1/chat/completions";
+// Of the models Groq serves, the gpt-oss pair are the ones that accept
+// `response_format: json_schema`. 120b beat 20b on relative-date arithmetic
+// ("Selasa terdekat") in side-by-side runs, so it is the one pinned here.
+const MODEL = "openai/gpt-oss-120b";
 const JWK_URL =
   "https://www.googleapis.com/service_accounts/v1/jwk/securetoken@system.gserviceaccount.com";
 
@@ -21,7 +29,7 @@ export default {
     if (request.method !== "POST") {
       return json({ error: "Method not allowed" }, 405, cors);
     }
-    if (!env.GITHUB_MODELS_TOKEN) {
+    if (!env.GROQ_API_KEY) {
       return json({ error: "Proxy belum dikonfigurasi (secret kosong)" }, 500, cors);
     }
 
@@ -49,7 +57,7 @@ export default {
     const upstream = await fetch(UPSTREAM, {
       method: "POST",
       headers: {
-        Authorization: `Bearer ${env.GITHUB_MODELS_TOKEN}`,
+        Authorization: `Bearer ${env.GROQ_API_KEY}`,
         "Content-Type": "application/json",
         Accept: "application/json",
       },

diff --git a/worker/verify-upstream.sh b/worker/verify-upstream.sh
@@ -0,0 +1,118 @@
+#!/usr/bin/env bash
+# Send the real schema and a real Indonesian prompt straight to Groq, skipping
+# the worker and Firebase, then assert the parts that have actually broken
+# before: the model must accept `response_format: json_schema`, and multiple
+# reminders must come back as separate entries rather than one concatenated
+# number (see README, "Two things that look wrong until you know why").
+#
+# Usage: GROQ_API_KEY=gsk_xxx ./verify-upstream.sh
+set -euo pipefail
+
+: "${GROQ_API_KEY:?set GROQ_API_KEY (console.groq.com) before running}"
+MODEL="${MODEL:-openai/gpt-oss-120b}"
+
+# A Wednesday, so "Selasa terdekat" is unambiguous and checkable.
+TODAY="2026-07-29"
+
+request=$(MODEL="$MODEL" TODAY="$TODAY" python3 - <<'PY'
+import json, os
+
+schema = {
+    "type": "object",
+    "additionalProperties": False,
+    "properties": {
+        "events": {
+            "type": "array",
+            "items": {
+                "type": "object",
+                "additionalProperties": False,
+                "properties": {
+                    "title": {"type": "string"},
+                    "description": {"type": ["string", "null"]},
+                    "startDate": {"type": "string"},
+                    "startTime": {"type": "string"},
+                    "endTime": {"type": "string"},
+                    "allDay": {"type": "boolean"},
+                    "recurrence": {
+                        "type": "string",
+                        "enum": ["none", "daily", "weekdays", "weekly", "monthly"],
+                    },
+                    "reminderOffsetsMinutes": {
+                        "type": "array",
+                        "items": {
+                            "type": "string",
+                            "enum": ["0", "5", "10", "20", "30", "60", "1440"],
+                        },
+                    },
+                },
+                "required": [
+                    "title", "description", "startDate", "startTime",
+                    "endTime", "allDay", "recurrence", "reminderOffsetsMinutes",
+                ],
+            },
+        }
+    },
+    "required": ["events"],
+}
+
+system = (
+    "Kamu adalah AI scheduling assistant untuk myKalender, aplikasi kalender "
+    "personal user Indonesia.\n"
+    f"Tanggal hari ini: {os.environ['TODAY']} (Rabu), zona waktu Asia/Makassar "
+    "(WITA, UTC+8).\n"
+    "Mode: FRESH. Parse deskripsi natural language dari user jadi daftar event "
+    "terstruktur.\n\n"
+    "Aturan:\n"
+    "- Output WAJIB valid JSON sesuai schema. Field tambahan dilarang.\n"
+    "- startDate: ISO YYYY-MM-DD di zona waktu user.\n"
+    "- startTime/endTime: HH:mm 24-jam. Untuk allDay, set startTime=00:00 endTime=23:59.\n"
+    "- Jika user nyebut hari tanpa tanggal (misal 'Senin'), pakai Senin terdekat ke depan.\n"
+    "- reminderOffsetsMinutes: menit sebelum mulai, satu entri per pengingat. Default ['20'].\n"
+)
+user = (
+    "kuliah STKI tiap Selasa jam 10 pagi sampai 12 siang di ruang 301, "
+    "ingatkan 1 hari dan 1 jam sebelum."
+)
+
+print(json.dumps({
+    "model": os.environ["MODEL"],
+    "messages": [
+        {"role": "system", "content": system},
+        {"role": "user", "content": user},
+    ],
+    "response_format": {
+        "type": "json_schema",
+        "json_schema": {"name": "schedule_events", "strict": True, "schema": schema},
+    },
+    "temperature": 0.2,
+}))
+PY
+)
+
+response=$(curl -sS https://api.groq.com/openai/v1/chat/completions \
+  -H "Authorization: Bearer $GROQ_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d "$request" --max-time 90)
+
+echo "$response" | python3 -c '
+import json, sys
+
+body = json.load(sys.stdin)
+if "choices" not in body:
+    sys.exit("upstream rejected the request: " + json.dumps(body)[:500])
+
+events = json.loads(body["choices"][0]["message"]["content"])["events"]
+print(json.dumps(events, indent=2, ensure_ascii=False))
+
+assert events, "no events parsed"
+first = events[0]
+assert first["startDate"] == "2026-08-04", (
+    "relative date wrong: nearest Selasa after 2026-07-29 is 2026-08-04, got "
+    + first["startDate"]
+)
+offsets = first["reminderOffsetsMinutes"]
+assert sorted(offsets, key=int) == ["60", "1440"], (
+    "two reminders should stay two entries, got " + repr(offsets)
+)
+print("\nOK: schema accepted, relative date correct, reminders not concatenated.")
+'
diff --git a/worker/wrangler.toml b/worker/wrangler.toml
@@ -2,8 +2,8 @@ name = "mykalender-ai"
 main = "src/index.js"
 compatibility_date = "2025-04-01"
 
-# Non-secret config. The GitHub PAT is stored as a secret instead:
-#   npx wrangler secret put GITHUB_MODELS_TOKEN
+# Non-secret config. The Groq API key is stored as a secret instead:
+#   npx wrangler secret put GROQ_API_KEY
 [vars]
 FIREBASE_PROJECT_ID = "mykalender-cad8f"
 ALLOWED_ORIGINS = "https://mykalender-cad8f.web.app,https://mykalender-cad8f.firebaseapp.com,http://localhost:5173"