Fix: < > operators falsely rejected as placeholders; Groq TTS models filtered; reason prints

Agent/agentic_workflow.py

@@ -1030,7 +1030,10 @@ def _groq_chat_models() -> list:
             )
             r.raise_for_status()
             ids = [m.get("id") for m in r.json().get("data", []) if m.get("id")]
-            exclude_tokens = ("whisper", "tts", "playai", "guard")
+            # Non-chat families: ASR, TTS (incl. Orpheus/CanopyLabs voices,
+            # whose ids carry no "tts" token), guard models, embeddings.
+            exclude_tokens = ("whisper", "tts", "playai", "guard", "orpheus",
+                              "canopylabs", "asr", "embed")
             text_models = [
                 i for i in ids
                 if not any(tok in i.lower() for tok in exclude_tokens)
@@ -1055,8 +1058,10 @@ def _groq_text_completion(prompt: str, max_tokens: int = 1024, timeout: float =
     if GROQ_TEXT_MODEL:
         candidates.append(GROQ_TEXT_MODEL)
     candidates.extend(_groq_chat_models())
-    if not candidates:
-        candidates.extend(_GROQ_CANDIDATE_FALLBACKS)
+    # ALWAYS merge the static known-good ids: the live /models listing can
+    # return non-chat families (filtered) or be temporarily empty — the static
+    # tail guarantees resumable candidates. Duplicates are removed below.
+    candidates.extend(_GROQ_CANDIDATE_FALLBACKS)
     # Prefer Qwen-family ids (closest match to the dormant HF primary,
     # Qwen/Qwen2.5-Coder) while keeping the explicit override untouched.
     ordered = candidates[:1]  # explicit override, if set, stays first
@@ -2845,19 +2850,25 @@ def is_sql_valid(sql: str, question: str = None) -> bool:
 
     sql_lower = sql.lower().strip()
     
-    # Quick placeholder check
-    PLACEHOLDER_TOKENS = ["<", ">", "...", "___", "placeholder", "example", "your_"]
+    # Quick placeholder check.
+    # NOTE: "<" and ">" are LEGITIMATE comparison operators ("price > average",
+    # "count < 5", "<>"). They were once listed here and SILENTLY rejected every
+    # greater/less-than question on every database. Real placeholders are caught
+    # by the regexes below (<TABLE_NAME> style) — keep operators legal.
+    PLACEHOLDER_TOKENS = ["...", "___", "placeholder", "example", "your_"]
     PLACEHOLDER_REGEXES = [
         r"<[A-Z_]+>",  # <TABLE_NAME>, <COLUMN>
         r"\b(table_name|column_name|your_\w+)\b"
     ]
-    
+
     for token in PLACEHOLDER_TOKENS:
         if token in sql_lower:
+            print(f"[VALIDATION] Rejected: placeholder token '{token}'")
             return False
-    
+
     for regex in PLACEHOLDER_REGEXES:
         if re.search(regex, sql, re.IGNORECASE):
+            print(f"[VALIDATION] Rejected: placeholder pattern /{regex}/")
             return False
     
     # Reject hardcoded department codes (D001-D009) - indicates incomplete query