Address PR review feedback on model routing and validation logic

Author: matej

claudecode/github_action_audit.py

@@ -314,7 +314,7 @@ def run_prompt(self, repo_dir: Path, prompt: str, model: Optional[str] = None) -
                         error_details += f"Stderr: {result.stderr}\n"
                         error_details += f"Stdout: {result.stdout[:500]}..."
                         return False, error_details, {}
-                    time.sleep(5 * attempt)
+                    time.sleep(5 * (attempt + 1))
                     continue
 
                 success, parsed_result = parse_json_with_fallbacks(result.stdout, "Claude Code output")
@@ -348,9 +348,9 @@ def run_prompt(self, repo_dir: Path, prompt: str, model: Optional[str] = None) -
         except Exception as e:
             return False, f"Claude Code execution error: {str(e)}", {}
 
-    def run_code_review(self, repo_dir: Path, prompt: str) -> Tuple[bool, str, Dict[str, Any]]:
+    def run_code_review(self, repo_dir: Path, prompt: str, model: Optional[str] = None) -> Tuple[bool, str, Dict[str, Any]]:
         """Run code review prompt and normalize to findings payload."""
-        success, error_msg, parsed = self.run_prompt(repo_dir, prompt, model=DEFAULT_CLAUDE_MODEL)
+        success, error_msg, parsed = self.run_prompt(repo_dir, prompt, model=model or DEFAULT_CLAUDE_MODEL)
         if not success:
             return False, error_msg, {}
         if isinstance(parsed, dict) and 'findings' in parsed:

claudecode/review_orchestrator.py

@@ -67,7 +67,11 @@ def _run_phase(self, repo_dir: Path, prompt: str, model: str, phase_name: str) -
         if hasattr(self.claude_runner, "run_prompt"):
             raw_result = self.claude_runner.run_prompt(repo_dir, prompt, model=model)
         if not (isinstance(raw_result, tuple) and len(raw_result) == 3) and hasattr(self.claude_runner, "run_code_review"):
-            raw_result = self.claude_runner.run_code_review(repo_dir, prompt)
+            try:
+                raw_result = self.claude_runner.run_code_review(repo_dir, prompt, model=model)
+            except TypeError:
+                # Backward compatibility for legacy mocks/runners that don't accept model parameter.
+                raw_result = self.claude_runner.run_code_review(repo_dir, prompt)
         if not (isinstance(raw_result, tuple) and len(raw_result) == 3):
             return False, {}, f"Invalid runner response for {phase_name}"
 
@@ -217,7 +221,10 @@ def run(
             return False, {}, f"Validation phase failed: {err_v}"
 
         validated: List[Dict[str, Any]] = []
-        decisions = validation_result.get("validated_findings", [])
+        has_validation_output = isinstance(validation_result, dict) and "validated_findings" in validation_result
+        decisions = validation_result.get("validated_findings", []) if isinstance(validation_result, dict) else []
+        if not isinstance(decisions, list):
+            decisions = []
         for decision in decisions:
             if not isinstance(decision, dict):
                 continue
@@ -235,10 +242,9 @@ def run(
                 finding["confidence"] = float(confidence)
                 validated.append(finding)
 
-        # Fallback: keep merged findings if validation returned no decisions
-        if decisions and not validated:
-            validated = []
-        elif not decisions:
+        # If validator did not return decisions at all, preserve candidates.
+        # If it explicitly returned validated_findings (including empty list), trust validator output.
+        if not has_validation_output:
             validated = all_candidates
 
         # Apply existing filtering pipeline