fix(ai): prioritize Groq's most stable model, stop repeating mining disclaimer (#9 )

User decision after the Groq reliability fix (previous PR): prioritize Groq's smallest/fastest model, llama-3.1-8b-instant, over gpt-oss-120b — it's long out of preview and has historically the best free-tier availability, at the cost of gpt-oss-120b's better Turkish/Kurdish fluency (which is why it was chosen originally). Model order is now llama-3.1-8b-instant -> gpt-oss-120b -> llama-3.3-70b-versatile before the (currently uncredited) Anthropic fallback. Also fixed a real UX complaint: asking about the Mining Simulation made the model repeat "this isn't real mining / it's just an estimate" in nearly every paragraph. Added an explicit instruction to state that once, briefly, near the start of the answer, and not repeat it — verified live, the disclaimer now appears once per response instead of several times.
fix(ai): Groq intermittent failures + sync git with actually-deployed source (#8 )
2026-07-23 21:45:40 +00:00 · 2026-07-21 07:52:59 -07:00 · 2026-07-21 07:33:17 -07:00 · 2026-07-12 19:26:02 -07:00
4 changed files with 117 additions and 49 deletions
@@ -1,6 +1,6 @@
 {
  "name": "pezkuwi-telegram-miniapp",
-  "version": "1.0.243",
+  "version": "1.0.242",
  "type": "module",
  "description": "Pezkuwichain Telegram Mini App - Forum, Announcements, Rewards",
  "author": "Pezkuwichain Team",
@@ -1,5 +1,5 @@
 {
-  "version": "1.0.243",
-  "buildTime": "2026-07-19T16:57:33.951Z",
-  "buildNumber": 1784480253951
+  "version": "1.0.242",
+  "buildTime": "2026-07-21T14:51:23.002Z",
+  "buildNumber": 1784645483003
 }
@@ -150,6 +150,7 @@ Inside the Trust Score card there is a small "Mining Simulation" square with a d
 - If your Trust Score is 0, the counter cannot start — the app shows a warning that your trust score must be greater than 0. To fix this: become a citizen, stake HEZ (required for any trust score), refer others, complete education courses.
 - Colors: red square = inactive (tap to start), gold = actively mining.
 - The small Telegram icon next to it opens the official channel: https://t.me/+DUWJ8wtt5qI4Njgy
+- IMPORTANT — how to talk about this: state that it's a simulation/estimate ONCE, briefly, near the start of your answer, then move on and explain the rest normally. Do NOT repeat "this isn't real mining" / "it's just an estimate" / similar caveats in every paragraph — one clear mention is enough, repeating it reads as nagging.

 "BULUT ULUSU" (CLOUD NATION) — THE BOOK BEHIND PEZKUWICHAIN (chapter-by-chapter knowledge):
 PezkuwiChain's philosophical foundation is the book "Bulut Ulusu" (Cloud Nation), written by the project's architect (a software developer AND sociologist — the "two desks" of the opening chapter). Available in Turkish, English and Kurdish. Core theses:
@@ -228,6 +229,56 @@ async function allowed(ip: string): Promise<boolean> {
  return true;
 }

+// Groq's openai/gpt-oss-120b returns intermittent errors under load (429/5xx,
+// and even a raw 502 from the edge runtime once) unrelated to the question
+// asked — confirmed live 2026-07-21 by hitting this endpoint repeatedly with
+// trivial questions ("merhaba") and seeing a ~50% failure rate with 15s+
+// between calls (ruling out our own per-IP rate limiter above), which 2
+// retries on the same model did not fully fix.
+//
+// 2026-07-21, user decision: prioritize reliability over gpt-oss-120b's
+// better Turkish/Kurdish fluency (it was picked for that reason via an
+// earlier A/B test) — llama-3.1-8b-instant is Groq's smallest/fastest model
+// and, being long out of preview, has historically the best free-tier
+// availability, so it goes first. gpt-oss-120b stays as the 2nd attempt
+// (still worth trying for quality when the primary is having its own bad
+// moment), llama-3.3-70b-versatile (the model used before switching to
+// gpt-oss-120b) as the 3rd, before ever reaching the Anthropic fallback
+// below. Doesn't retry a 4xx (bad request/auth) at all — not transient, and
+// switching models wouldn't help.
+const GROQ_MODELS = ['llama-3.1-8b-instant', 'openai/gpt-oss-120b', 'llama-3.3-70b-versatile'];
+
+async function callGroqWithRetry(messages: unknown[], retriesPerModel = 2): Promise<string | null> {
+  for (const model of GROQ_MODELS) {
+    for (let attempt = 0; attempt <= retriesPerModel; attempt++) {
+      try {
+        const gr = await fetch('https://api.groq.com/openai/v1/chat/completions', {
+          method: 'POST',
+          headers: { Authorization: 'Bearer ' + GROQ_API_KEY, 'Content-Type': 'application/json' },
+          body: JSON.stringify({
+            model,
+            temperature: 0.3,
+            max_tokens: 900,
+            messages: [{ role: 'system', content: SYSTEM }, ...messages],
+          }),
+        });
+        if (gr.ok) {
+          const gd = await gr.json();
+          const reply = gd.choices?.[0]?.message?.content || null;
+          if (reply) return reply;
+        } else {
+          console.error('[AI] Groq error:', model, gr.status, await gr.text());
+          if (gr.status !== 429 && gr.status < 500) break; // non-transient, try next model instead
+        }
+      } catch (e) {
+        console.error('[AI] Groq exception:', model, e);
+      }
+      if (attempt < retriesPerModel) await new Promise((r) => setTimeout(r, 400 * (attempt + 1)));
+    }
+  }
+  return null;
+}
+
 serve(async (req) => {
  if (req.method === 'OPTIONS') return new Response('ok', { headers: CORS });
  if (req.method !== 'POST') return J({ error: 'method' }, 405);
@@ -255,26 +306,7 @@ serve(async (req) => {
    let answer: string | null = null;

    if (GROQ_API_KEY) {
-      try {
-        const gr = await fetch('https://api.groq.com/openai/v1/chat/completions', {
-          method: 'POST',
-          headers: { Authorization: 'Bearer ' + GROQ_API_KEY, 'Content-Type': 'application/json' },
-          body: JSON.stringify({
-            model: 'openai/gpt-oss-120b',
-            temperature: 0.3,
-            max_tokens: 900,
-            messages: [{ role: 'system', content: SYSTEM }, ...messages],
-          }),
-        });
-        if (gr.ok) {
-          const gd = await gr.json();
-          answer = gd.choices?.[0]?.message?.content || null;
-        } else {
-          console.error('[AI] Groq error:', gr.status, await gr.text());
-        }
-      } catch (e) {
-        console.error('[AI] Groq exception:', e);
-      }
+      answer = await callGroqWithRetry(messages);
    }

    if (!answer && ANTHROPIC_API_KEY) {
@@ -230,6 +230,7 @@ Inside the Trust Score card there is a small "Mining Simulation" square with a d
 - If your Trust Score is 0, the counter cannot start — the app shows a warning that your trust score must be greater than 0. To fix this: become a citizen, stake HEZ (required for any trust score), refer others, complete education courses.
 - Colors: red square = inactive (tap to start), gold = actively mining.
 - The small Telegram icon next to it opens the official channel: https://t.me/+DUWJ8wtt5qI4Njgy
+- IMPORTANT — how to talk about this: state that it's a simulation/estimate ONCE, briefly, near the start of your answer, then move on and explain the rest normally. Do NOT repeat "this isn't real mining" / "it's just an estimate" / similar caveats in every paragraph — one clear mention is enough, repeating it reads as nagging.

 "BULUT ULUSU" (CLOUD NATION) — THE BOOK BEHIND PEZKUWICHAIN (chapter-by-chapter knowledge):
 PezkuwiChain's philosophical foundation is the book "Bulut Ulusu" (Cloud Nation), written by the project's architect (a software developer AND sociologist — the "two desks" of the opening chapter). Available in Turkish, English and Kurdish. Core theses:
@@ -269,9 +270,66 @@ Phase 4 (2028+): Cross-chain governance federation, decentralized identity, stab

 License: Apache 2.0, Copyright 2026 Kurdistan Tech Institute. Lead Architect: SatoshiQaziMuhammed.`;

-// ── Claude AI chat handler ──────────────────────────────────────────
+// Groq's openai/gpt-oss-120b returns intermittent errors under load (429/5xx,
+// and even a raw 502 from the edge runtime once) unrelated to the question
+// asked — confirmed live 2026-07-21 by hitting this endpoint repeatedly with
+// trivial questions ("merhaba") and seeing a ~50% failure rate with 15s+
+// between calls (ruling out our own per-IP rate limiter), which 2 retries on
+// the same model did not fully fix.
+//
+// 2026-07-21, user decision: prioritize reliability over gpt-oss-120b's
+// better Turkish/Kurdish fluency (it was picked for that reason via an
+// earlier A/B test) — llama-3.1-8b-instant is Groq's smallest/fastest model
+// and, being long out of preview, has historically the best free-tier
+// availability, so it goes first. gpt-oss-120b stays as the 2nd attempt
+// (still worth trying for quality when the primary is having its own bad
+// moment), llama-3.3-70b-versatile (the model used before switching to
+// gpt-oss-120b) as the 3rd, before ever reaching the Anthropic fallback
+// below. Retries a 4xx (bad request/auth) not at all — that's not transient
+// and switching models wouldn't help either.
+const GROQ_MODELS = ['llama-3.1-8b-instant', 'openai/gpt-oss-120b', 'llama-3.3-70b-versatile'];
+
+async function callGroqWithRetry(
+  systemPrompt: string,
+  userMessage: string,
+  retriesPerModel = 2
+): Promise<string | null> {
+  for (const model of GROQ_MODELS) {
+    for (let attempt = 0; attempt <= retriesPerModel; attempt++) {
+      try {
+        const gr = await fetch('https://api.groq.com/openai/v1/chat/completions', {
+          method: 'POST',
+          headers: { Authorization: 'Bearer ' + GROQ_API_KEY, 'Content-Type': 'application/json' },
+          body: JSON.stringify({
+            model,
+            temperature: 0.3,
+            max_tokens: 900,
+            messages: [
+              { role: 'system', content: systemPrompt },
+              { role: 'user', content: userMessage },
+            ],
+          }),
+        });
+        if (gr.ok) {
+          const gd = await gr.json();
+          const reply = gd.choices?.[0]?.message?.content || null;
+          if (reply) return reply;
+        } else {
+          console.error('[AI] Groq error:', model, gr.status, await gr.text());
+          if (gr.status !== 429 && gr.status < 500) break; // non-transient, try next model instead
+        }
+      } catch (e) {
+        console.error('[AI] Groq exception:', model, e);
+      }
+      if (attempt < retriesPerModel) await new Promise((r) => setTimeout(r, 400 * (attempt + 1)));
+    }
+  }
+  return null;
+}
+
+// ── AI chat handler ──────────────────────────────────────────
 async function handleAIChat(token: string, chatId: number, userMessage: string, userName: string) {
-  if (!ANTHROPIC_API_KEY) {
+  if (!GROQ_API_KEY && !ANTHROPIC_API_KEY) {
    await sendTelegramRequest(token, 'sendMessage', {
      chat_id: chatId,
      text: 'AI assistant is being configured. Please try again later.',
@@ -290,29 +348,7 @@ async function handleAIChat(token: string, chatId: number, userMessage: string,
    let aiReply: string | null = null;

    if (GROQ_API_KEY) {
-      try {
-        const gr = await fetch('https://api.groq.com/openai/v1/chat/completions', {
-          method: 'POST',
-          headers: { Authorization: 'Bearer ' + GROQ_API_KEY, 'Content-Type': 'application/json' },
-          body: JSON.stringify({
-            model: 'openai/gpt-oss-120b',
-            temperature: 0.3,
-            max_tokens: 900,
-            messages: [
-              { role: 'system', content: CLAUDE_SYSTEM_PROMPT },
-              { role: 'user', content: userMessage },
-            ],
-          }),
-        });
-        if (gr.ok) {
-          const gd = await gr.json();
-          aiReply = gd.choices?.[0]?.message?.content || null;
-        } else {
-          console.error('[AI] Groq error:', gr.status, await gr.text());
-        }
-      } catch (e) {
-        console.error('[AI] Groq exception:', e);
-      }
+      aiReply = await callGroqWithRetry(CLAUDE_SYSTEM_PROMPT, userMessage);
    }

    if (!aiReply && ANTHROPIC_API_KEY) {