fix: universal parsing + OpenRouter + state bug — purpose_agent/purpose_function.py

purpose_agent/purpose_function.py

@@ -225,21 +225,26 @@ class PurposeFunction:
         ]
 
         # Get structured evaluation from LLM
+        from purpose_agent.robust_parser import parse_critic_response
+
         try:
             raw_score = self.llm.generate_structured(
                 messages, schema=PURPOSE_SCORE_SCHEMA, temperature=0.2
             )
-        except Exception as e:
-            logger.error(f"Purpose Function structured output failed: {e}")
-            # Fall back to text-based evaluation
-            raw_score = self._fallback_evaluate(messages)
-
-        # Extract and validate scores
-        phi_before = float(raw_score.get("phi_before", 0.0))
-        phi_after = float(raw_score.get("phi_after", 0.0))
+        except Exception:
+            # Structured output not available — use universal text parser
+            raw = self.llm.generate(messages, temperature=0.2, max_tokens=2000)
+            raw_score = parse_critic_response(raw)
+
+        # Extract and validate scores (safe — parse_critic_response always returns valid keys)
+        def _safe_float(v, d=0.0):
+            try: return float(str(v).rstrip('.'))
+            except (ValueError, TypeError): return d
+        phi_before = _safe_float(raw_score.get("phi_before", 0.0))
+        phi_after = _safe_float(raw_score.get("phi_after", 0.0))
         reasoning = str(raw_score.get("reasoning", ""))
         evidence = str(raw_score.get("evidence", ""))
-        confidence = float(raw_score.get("confidence", 0.5))
+        confidence = _safe_float(raw_score.get("confidence", 0.5))
 
         # Clamp to valid range
         phi_before = max(0.0, min(10.0, phi_before))