Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

fix(llm): guard central think-strip against non-string content (follow-up to #4343) (#4611)

LearningCircuit54ef628cbfd4
diff --git a/changelog.d/+harden-central-think-strip.bugfix.md b/changelog.d/+harden-central-think-strip.bugfix.md
new file mode 100644
index 0000000000..7f3d344451
--- /dev/null
+++ b/changelog.d/+harden-central-think-strip.bugfix.md
@@ -0,0 +1 @@
+Harden the central `ProcessingLLMWrapper` think-tag stripping against non-string responses. `remove_think_tags` is text-only, so it now runs only when `response.content` is a `str`; non-string content (e.g. provider content-block lists such as Anthropic's, or `None`) is passed through unchanged instead of raising `TypeError`. This removes the wrapper-level crash that the per-site `str(...)` guards used to absorb before they were dropped. Note: a few direct `.content` consumers still assume string content, so robust list-content handling at those call sites is tracked separately.
diff --git a/changelog.d/+remove-redundant-think-stripping.bugfix.md b/changelog.d/+remove-redundant-think-stripping.bugfix.md
index 938c421524..d4c3564552 100644
--- a/changelog.d/+remove-redundant-think-stripping.bugfix.md
+++ b/changelog.d/+remove-redundant-think-stripping.bugfix.md
@@ -1 +1 @@
-Remove redundant per-site `<think>`-tag stripping now that `get_llm` LLMs are normalized centrally in `ProcessingLLMWrapper`. Direct `self.model.invoke(...).content` is already think-stripped, so ~25 scattered `remove_think_tags(...)` calls across the search strategies, constraint/evidence/knowledge components, and synthesis path were dropped (relying on the single central strip point), keeping explicit handling only where it is load-bearing (agent/`bind_tools` paths, injected LLMs). Empty-answer fallbacks now log a warning so a silent degrade is visible.
+Remove redundant per-site `<think>`-tag stripping now that `get_llm` LLMs are normalized centrally in `ProcessingLLMWrapper`. Direct `self.model.invoke(...).content` is already think-stripped, so 11 scattered stripping calls (6 `remove_think_tags(...)` and 5 `get_llm_response_text(...)`) across the constraint/evidence/knowledge components, the synthesis path, and the document-analysis API were dropped (relying on the single central strip point), keeping explicit handling only where it is load-bearing (agent/`bind_tools` paths, injected LLMs). Empty-answer fallbacks now log a warning so a silent degrade is visible.
diff --git a/src/local_deep_research/config/llm_config.py b/src/local_deep_research/config/llm_config.py
index 953fd73692..1ea4a9e902 100644
--- a/src/local_deep_research/config/llm_config.py
+++ b/src/local_deep_research/config/llm_config.py
@@ -411,9 +411,17 @@ def wrap_llm_without_think_tags(
A bare-string return (some providers/wrappers) is wrapped into an
``AIMessage`` so callers can always rely on ``.content``. Anything
else is passed through unchanged.
+
+ Only *string* ``.content`` is stripped: ``remove_think_tags`` is
+ text-only, and the ``<think>...</think>`` artifact is only ever
+ emitted as plain text by some local models. Non-string content
+ (e.g. provider content-block lists like Anthropic's, or ``None``)
+ is passed through untouched — running the regex on it would raise
+ ``TypeError`` or corrupt the structured content.
"""
if hasattr(response, "content"):
- response.content = remove_think_tags(response.content)
+ if isinstance(response.content, str):
+ response.content = remove_think_tags(response.content)
elif isinstance(response, str):
response = AIMessage(content=remove_think_tags(response))
return response
diff --git a/tests/config/test_llm_config.py b/tests/config/test_llm_config.py
index 9449a490a8..744bd5c220 100644
--- a/tests/config/test_llm_config.py
+++ b/tests/config/test_llm_config.py
@@ -117,6 +117,38 @@ class TestWrapLlmWithoutThinkTags:
wrapper.invoke("test")
mock_remove.assert_called_with("<think>internal</think>visible")
+ def test_wrapper_preserves_nonstring_content(self):
+ """Non-string content (e.g. provider content-block lists) must pass
+ through unchanged instead of raising TypeError in remove_think_tags."""
+ mock_llm = MagicMock()
+ mock_response = MagicMock()
+ blocks = [{"type": "text", "text": "visible"}]
+ mock_response.content = blocks
+ mock_llm.invoke.return_value = mock_response
+
+ with patch(
+ "local_deep_research.config.llm_config.get_setting_from_snapshot",
+ return_value=False,
+ ):
+ wrapper = wrap_llm_without_think_tags(mock_llm)
+ result = wrapper.invoke("test")
+ assert result.content == blocks
+
+ def test_wrapper_preserves_none_content(self):
+ """None content must pass through unchanged rather than raising."""
+ mock_llm = MagicMock()
+ mock_response = MagicMock()
+ mock_response.content = None
+ mock_llm.invoke.return_value = mock_response
+
+ with patch(
+ "local_deep_research.config.llm_config.get_setting_from_snapshot",
+ return_value=False,
+ ):
+ wrapper = wrap_llm_without_think_tags(mock_llm)
+ result = wrapper.invoke("test")
+ assert result.content is None
+
def test_wrapper_delegates_attributes(self):
"""Should delegate attribute access to base LLM."""
mock_llm = MagicMock()