Skip to content

Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.

Detect LLM output truncation instead of failing with misleading parse errors

Saurav Panda5f690c207e08
diff --git a/browser_use/llm/anthropic/chat.py b/browser_use/llm/anthropic/chat.py
index b34459954..d466dced3 100644
--- a/browser_use/llm/anthropic/chat.py
+++ b/browser_use/llm/anthropic/chat.py
@@ -387,6 +387,18 @@ class ChatAnthropic(BaseChatModel):
usage = self._get_usage(response)
+ # A tool_use block cut off at max_tokens produces incomplete JSON that fails
+ # validation with a misleading parse error — surface the real cause instead.
+ if response.stop_reason == 'max_tokens':
+ raise ModelProviderError(
+ message=(
+ f'Model output was truncated at max_tokens={self.max_tokens}; the structured'
+ ' output is incomplete. Increase max_tokens or request shorter output.'
+ ),
+ status_code=400,
+ model=self.name,
+ )
+
# Extract the tool use block
for content_block in response.content:
if hasattr(content_block, 'type') and content_block.type == 'tool_use':
@@ -438,5 +450,9 @@ class ChatAnthropic(BaseChatModel):
raise ModelRateLimitError(message=e.message, model=self.name) from e
except APIStatusError as e:
raise ModelProviderError(message=e.message, status_code=e.status_code, model=self.name) from e
+ except ModelProviderError:
+ # Already carries an accurate message and status code (e.g. max_tokens
+ # truncation) — don't re-wrap it with the generic 502.
+ raise
except Exception as e:
raise ModelProviderError(message=str(e), model=self.name) from e
diff --git a/browser_use/llm/google/chat.py b/browser_use/llm/google/chat.py
index 9c79e20f7..01d937125 100644
--- a/browser_use/llm/google/chat.py
+++ b/browser_use/llm/google/chat.py
@@ -193,6 +193,24 @@ class ChatGoogle(BaseChatModel):
return str(response.candidates[0].finish_reason) if hasattr(response.candidates[0], 'finish_reason') else None
return None
+ def _raise_if_output_truncated(self, response: types.GenerateContentResponse) -> None:
+ """Structured output cut off at max_output_tokens produces incomplete JSON that
+ fails parsing with a misleading error — surface the real cause instead.
+
+ Uses status 400 (not in retryable_status_codes): retrying the same request
+ would truncate identically."""
+ stop_reason = self._get_stop_reason(response)
+ if stop_reason and 'MAX_TOKENS' in stop_reason:
+ raise ModelProviderError(
+ message=(
+ f'Model output was truncated at max_output_tokens={self.max_output_tokens};'
+ ' the structured output is incomplete. Increase max_output_tokens or request'
+ ' shorter output.'
+ ),
+ status_code=400,
+ model=self.name,
+ )
+
def _get_usage(self, response: types.GenerateContentResponse) -> ChatInvokeUsage | None:
usage: ChatInvokeUsage | None = None
@@ -374,6 +392,7 @@ class ChatGoogle(BaseChatModel):
self.logger.debug(f'✅ Got structured response in {elapsed:.2f}s')
usage = self._get_usage(response)
+ self._raise_if_output_truncated(response)
# Handle case where response.parsed might be None
if response.parsed is None:
@@ -458,6 +477,7 @@ class ChatGoogle(BaseChatModel):
self.logger.debug(f'✅ Got fallback response in {elapsed:.2f}s')
usage = self._get_usage(response)
+ self._raise_if_output_truncated(response)
# Try to extract JSON from the text response
if response.text:
diff --git a/browser_use/llm/openai/chat.py b/browser_use/llm/openai/chat.py
index 390df689f..5f455f7a7 100644
--- a/browser_use/llm/openai/chat.py
+++ b/browser_use/llm/openai/chat.py
@@ -281,6 +281,19 @@ class ChatOpenAI(BaseChatModel):
usage = self._get_usage(response)
+ # Output cut off at the completion cap produces incomplete JSON that fails
+ # validation with a misleading parse error — surface the real cause instead.
+ if choice.finish_reason == 'length':
+ raise ModelProviderError(
+ message=(
+ f'Model output was truncated at max_completion_tokens={self.max_completion_tokens};'
+ ' the structured output is incomplete. Increase max_completion_tokens or request'
+ ' shorter output.'
+ ),
+ status_code=400,
+ model=self.name,
+ )
+
parsed = output_format.model_validate_json(choice.message.content)
return ChatInvokeCompletion(
diff --git a/tests/ci/test_llm_output_truncation.py b/tests/ci/test_llm_output_truncation.py
new file mode 100644
index 000000000..023882bd8
--- /dev/null
+++ b/tests/ci/test_llm_output_truncation.py
@@ -0,0 +1,65 @@
+"""Regression test: structured output cut off at the completion-token cap must raise a
+clear truncation error, not a misleading JSON parse error ('Unterminated string...')."""
+
+import pytest
+from pydantic import BaseModel
+
+from browser_use.llm.exceptions import ModelProviderError
+from browser_use.llm.messages import UserMessage
+from browser_use.llm.openai.chat import ChatOpenAI
+
+
+class AnswerFormat(BaseModel):
+ answer: str
+
+
+async def test_openai_truncated_structured_output_raises_clear_error(httpserver):
+ """finish_reason='length' with JSON cut mid-string must surface the token cap, not a parse error."""
+ httpserver.expect_request('/v1/chat/completions', method='POST').respond_with_json(
+ {
+ 'id': 'chatcmpl-test',
+ 'object': 'chat.completion',
+ 'created': 0,
+ 'model': 'gpt-4o',
+ 'choices': [
+ {
+ 'index': 0,
+ 'message': {'role': 'assistant', 'content': '{"answer": "this output was cut off mid-sent'},
+ 'finish_reason': 'length',
+ }
+ ],
+ 'usage': {'prompt_tokens': 10, 'completion_tokens': 4096, 'total_tokens': 4106},
+ }
+ )
+
+ llm = ChatOpenAI(model='gpt-4o', api_key='test-key', base_url=httpserver.url_for('/v1'))
+
+ with pytest.raises(ModelProviderError) as exc_info:
+ await llm.ainvoke([UserMessage(content='answer at length')], output_format=AnswerFormat)
+
+ assert 'truncated' in str(exc_info.value), f'expected a truncation error, got: {exc_info.value}'
+ assert 'max_completion_tokens' in str(exc_info.value)
+
+
+async def test_openai_normal_structured_output_still_parses(httpserver):
+ httpserver.expect_request('/v1/chat/completions', method='POST').respond_with_json(
+ {
+ 'id': 'chatcmpl-test',
+ 'object': 'chat.completion',
+ 'created': 0,
+ 'model': 'gpt-4o',
+ 'choices': [
+ {
+ 'index': 0,
+ 'message': {'role': 'assistant', 'content': '{"answer": "complete answer"}'},
+ 'finish_reason': 'stop',
+ }
+ ],
+ 'usage': {'prompt_tokens': 10, 'completion_tokens': 8, 'total_tokens': 18},
+ }
+ )
+
+ llm = ChatOpenAI(model='gpt-4o', api_key='test-key', base_url=httpserver.url_for('/v1'))
+
+ result = await llm.ainvoke([UserMessage(content='answer briefly')], output_format=AnswerFormat)
+ assert result.completion.answer == 'complete answer'