diff --git a/scripts/langchain/pr_verifier.py b/scripts/langchain/pr_verifier.py
index 0a127ea7b..93f02ca06 100755
--- a/scripts/langchain/pr_verifier.py
+++ b/scripts/langchain/pr_verifier.py
@@ -675,12 +675,31 @@ def _fallback_evaluation(
     )
 
 
+def _coerce_response_content(content: object) -> str:
+    """Return text from provider response blocks without losing a safe fallback."""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        text_blocks = [
+            block["text"]
+            for block in content
+            if isinstance(block, dict) and isinstance(block.get("text"), str)
+        ]
+        if text_blocks:
+            # Concatenate without a separator: a provider may split one JSON
+            # document across blocks, and an inserted newline inside a string
+            # literal would make the reassembled payload invalid JSON.
+            return "".join(text_blocks)
+    return json.dumps(content, default=str)
+
+
 def _parse_llm_response(
-    content: str, provider: str, *, client: object | None = None
+    content: object, provider: str, *, client: object | None = None
 ) -> EvaluationResult:
+    content_text = _coerce_response_content(content)
     repair = _build_verifier_repair_callback(client) if client is not None else None
     parsed = parse_structured_output(
-        content,
+        content_text,
         EvaluationPayload,
         repair=repair,
         max_repair_attempts=SCHEMA_REPAIR_POLICY.max_attempts,
@@ -704,7 +723,7 @@ def _parse_llm_response(
             summary=None,
             provider_used=provider,
             used_llm=True,
-            raw_content=content,
+            raw_content=content_text,
             error=error,
         )
 
@@ -717,7 +736,7 @@ def _parse_llm_response(
         summary=payload.summary,
         provider_used=provider,
         used_llm=True,
-        raw_content=parsed.raw_content or content,
+        raw_content=parsed.raw_content or content_text,
     )
 
 
@@ -725,11 +744,14 @@ def _build_verifier_repair_callback(client: object) -> Callable[[str, str, str],
     repair = build_repair_callback(client)
 
     def _repair(schema_json: str, validation_errors: str, raw_response: str) -> str | None:
-        return repair(
+        repaired = repair(
             schema_json,
             validation_errors,
             _cap_prompt_text(raw_response, EVAL_SCHEMA_REPAIR_BUDGET_TOKENS),
         )
+        if not repaired:
+            return None
+        return _coerce_response_content(repaired)
 
     return _repair
 
diff --git a/tests/scripts/test_pr_verifier_structured_output.py b/tests/scripts/test_pr_verifier_structured_output.py
index b6013a2df..c708d893d 100644
--- a/tests/scripts/test_pr_verifier_structured_output.py
+++ b/tests/scripts/test_pr_verifier_structured_output.py
@@ -83,6 +83,89 @@ def test_comparison_runner_repairs_malformed_output() -> None:
     assert mock_client.invoke.call_count == 2
 
 
+def test_comparison_runner_handles_structured_response_content() -> None:
+    payload = json.dumps(_valid_payload())
+    response = _response_with(
+        [
+            {"type": "thinking", "thinking": "reviewing", "signature": "opaque"},
+            {"type": "text", "text": payload},
+        ]
+    )
+    mock_client = mock.MagicMock()
+    mock_client.invoke.return_value = response
+
+    runner = pr_verifier.ComparisonRunner(
+        context="context",
+        diff=None,
+        prompt="prompt",
+        clients=[(mock_client, "anthropic", "claude-sonnet")],
+    )
+
+    result = runner.run_single(mock_client, "anthropic", "claude-sonnet")
+
+    assert result.verdict == "PASS"
+    assert result.raw_content == payload
+
+
+def test_comparison_runner_concatenates_split_text_blocks() -> None:
+    payload = json.dumps(_valid_payload())
+    # Split inside the summary string literal, where an inserted newline would
+    # be an unescaped control character and invalidate the JSON.
+    split = payload.index("Looks good.") + 5
+    response = _response_with(
+        [
+            {"type": "text", "text": payload[:split]},
+            {"type": "text", "text": payload[split:]},
+        ]
+    )
+    mock_client = mock.MagicMock()
+    mock_client.invoke.return_value = response
+
+    runner = pr_verifier.ComparisonRunner(
+        context="context",
+        diff=None,
+        prompt="prompt",
+        clients=[(mock_client, "anthropic", "claude-sonnet")],
+    )
+
+    result = runner.run_single(mock_client, "anthropic", "claude-sonnet")
+
+    assert result.verdict == "PASS"
+    assert result.raw_content == payload
+    assert mock_client.invoke.call_count == 1
+
+
+def test_comparison_runner_normalizes_structured_repair_response() -> None:
+    payload = json.dumps(_valid_payload())
+    malformed = _response_with(
+        [
+            {"type": "thinking", "thinking": "reviewing", "signature": "opaque"},
+            {"type": "text", "text": "```json\n" + payload + "\n```"},
+        ]
+    )
+    repaired = _response_with(
+        [
+            {"type": "thinking", "thinking": "repairing", "signature": "opaque"},
+            {"type": "text", "text": payload},
+        ]
+    )
+    mock_client = mock.MagicMock()
+    mock_client.invoke.side_effect = [malformed, repaired]
+
+    runner = pr_verifier.ComparisonRunner(
+        context="context",
+        diff=None,
+        prompt="prompt",
+        clients=[(mock_client, "anthropic", "claude-sonnet")],
+    )
+
+    result = runner.run_single(mock_client, "anthropic", "claude-sonnet")
+
+    assert result.verdict == "PASS"
+    assert result.raw_content == payload
+    assert mock_client.invoke.call_count == 2
+
+
 def test_evaluate_pr_valid_output_no_repair(monkeypatch: pytest.MonkeyPatch) -> None:
     payload = _valid_payload()
     good = json.dumps(payload)