Aug 19 - Update: AI model since the old one no longer exist

CLAUDE.md

@@ -163,7 +163,7 @@ part of the tree — see §7. Device registration on the API side lives in
 | `DIGEST_SECRET` | Authenticates all cron endpoints |
 | `REDIS_URL` | Optional. When set, Flask-Limiter uses Redis for shared rate-limit counters across Gunicorn workers. |
 | `GROQ_API_KEY` | Optional. When set, enables the AI chatbot at `/support/chat`. Absent → chat input disabled; customers see a "Submit to Support" fallback only. |
-| `GROQ_MODEL` | Optional. Groq model ID. Defaults to `llama-3.3-70b-versatile`. |
+| `GROQ_MODEL` | Optional. Groq model ID. Defaults to `_DEFAULT_GROQ_MODEL` in `routes/support.py` (`openai/gpt-oss-120b`, verified Aug 2026). **Groq retires models without notice** — when the configured one disappears the API 404s and EVERY question returns the generic "problem reaching the AI assistant" reply, with nothing else broken, so it stays invisible until a customer complains. That is how `llama-3.3-70b-versatile` took the chat down. The error handler in `chat_message()` logs the model name and an explicit "set GROQ_MODEL" hint for exactly this case; fixing it needs no deploy, just the env var. |
 | `ENROLLMENT_NOTIFY_EMAILS` | Optional. Comma-separated extra addresses alerted on a new enrollment, **in addition to** every active `admin` account. For people who should be told but hold no JQC login. |
 | `ENROLLMENT_DIR` | Optional. Directory for enrollment-form JSON submissions. Defaults to `<instance_path>/enrollments` (git-ignored). Created at boot. |
 | `DEFAULT_UI_THEME` | Optional, default `modern` (phase50). The design shown when a user has no stored preference — i.e. new accounts and unauthenticated pages. A stored `users.ui_theme` always wins. Set `classic` to revert the default **without** touching anyone's saved choice. |

app/routes/support.py

@@ -161,6 +161,10 @@ INSPECTOR_FAQS = [
 ]
 
 
+#: Groq model used when GROQ_MODEL is unset. Verified available Aug 2026.
+#: Groq retires models periodically — see the error handler in chat_message().
+_DEFAULT_GROQ_MODEL = 'openai/gpt-oss-120b'
+
 # Soft cap on injected knowledge to keep prompt size (and token cost) reasonable.
 _KB_MAX_CHARS = 6000
 
@@ -397,7 +401,14 @@ def chat_message():
                     messages.append({'role': m['role'], 'content': _redact_pii(m['content'])})
             messages.append({'role': 'user', 'content': _redact_pii(user_message)})
 
-            model = os.environ.get('GROQ_MODEL', 'llama-3.3-70b-versatile')
+            # Default model. Groq RETIRES models without notice, and when the
+            # configured one disappears every question fails with the generic
+            # "problem reaching the AI assistant" reply — invisible until a
+            # customer complains. That is exactly how llama-3.3-70b-versatile
+            # took the chat down (404 model_not_found, Aug 2026). If this
+            # happens again, check the log line below and set GROQ_MODEL to a
+            # current model — no deploy needed.
+            model = os.environ.get('GROQ_MODEL', _DEFAULT_GROQ_MODEL)
             completion = client.chat.completions.create(
                 model=model,
                 messages=messages,
@@ -406,7 +417,17 @@ def chat_message():
             )
             reply = completion.choices[0].message.content.strip()
         except Exception as exc:
-            logger.error('SUPPORT | Groq error: %s', exc)
+            # Always name the model — a bare "Groq error" gives whoever reads
+            # the log nothing to act on, and a retired model is the most likely
+            # cause of a total outage here.
+            if 'model_not_found' in str(exc) or 'does not exist' in str(exc):
+                logger.error(
+                    'SUPPORT | Groq model %r is not available on this account — '
+                    'the assistant is DOWN for every user. Set GROQ_MODEL to a '
+                    'current model (see https://console.groq.com/docs/models). '
+                    'Underlying error: %s', model, exc)
+            else:
+                logger.error('SUPPORT | Groq error (model=%r): %s', model, exc)
             reply = ("I ran into a problem reaching the AI assistant. "
                      "Please try again, or use **Submit to Support** to contact our team.")