Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.
Python ai-agentsai-toolsbrowser-agentbrowser-automationbrowser-usecdpcloud-browsercomputer-useheadless-browserllmmcpplaywrightpythonremote-browserstealth-browserweb-automationweb-scraping
improved csv file generation
Saurav Panda8270346ccfed
diff --git a/browser_use/filesystem/file_system.py b/browser_use/filesystem/file_system.pyindex 1e46a47d9..4ca6b5049 100644--- a/browser_use/filesystem/file_system.py+++ b/browser_use/filesystem/file_system.py@@ -1,5 +1,7 @@import asyncioimport base64+import csv+import ioimport osimport reimport shutil@@ -164,12 +166,68 @@ class JsonFile(BaseFile):class CsvFile(BaseFile):- """CSV file implementation"""+ """CSV file implementation with automatic RFC 4180 normalization.++ LLMs frequently produce malformed CSV (missing quotes around fields with commas,+ inconsistent empty fields, unescaped internal quotes). This class parses the raw+ content through Python's csv module on every write to guarantee well-formed output.+ """@propertydef extension(self) -> str:return 'csv'+ @staticmethod+ def _normalize_csv(raw: str) -> str:+ """Parse and re-serialize CSV content to fix quoting, empty fields, and escaping.++ Handles common LLM mistakes: unquoted fields containing commas,+ unescaped quotes inside fields, inconsistent empty fields,+ trailing/leading blank lines, and double-escaped JSON output+ (literal backslash-n and backslash-quote instead of real newlines/quotes).+ """+ stripped = raw.strip()+ if not stripped:+ return raw++ # Detect double-escaped LLM tool call output: if the content has no real+ # newlines but contains literal \n sequences, the entire string is likely+ # double-escaped JSON. Unescape \" → " first, then \n → newline.+ if '\n' not in stripped and '\\n' in stripped:+ stripped = stripped.replace('\\"', '"')+ stripped = stripped.replace('\\n', '\n')++ reader = csv.reader(io.StringIO(stripped))+ rows: list[list[str]] = []+ for row in reader:+ # Skip completely empty rows (artifacts of blank lines)+ if row:+ rows.append(row)++ if not rows:+ return raw++ out = io.StringIO()+ writer = csv.writer(out, lineterminator='\n')+ writer.writerows(rows)+ # Strip trailing newline so callers (write_file action) control line endings+ return out.getvalue().rstrip('\n')++ def write_file_content(self, content: str) -> None:+ """Normalize CSV content before storing."""+ self.update_content(self._normalize_csv(content))++ def append_file_content(self, content: str) -> None:+ """Normalize the appended CSV rows and merge with existing content."""+ normalized_new = self._normalize_csv(content)+ if not normalized_new.strip():+ return+ existing = self.content+ if existing and not existing.endswith('\n'):+ existing += '\n'+ combined = existing + normalized_new+ self.update_content(self._normalize_csv(combined))+class JsonlFile(BaseFile):"""JSONL (JSON Lines) file implementation"""diff --git a/examples/features/csv_file_generation.py b/examples/features/csv_file_generation.pynew file mode 100644index 000000000..efecc5480--- /dev/null+++ b/examples/features/csv_file_generation.py@@ -0,0 +1,51 @@+"""+Generate CSV files with automatic normalization.++The agent's file system automatically normalizes CSV output using Python's csv module,+so fields containing commas, quotes, or empty values are properly handled per RFC 4180.+This means the agent doesn't need to worry about manual quoting — it's fixed at the+infrastructure level.++Common LLM mistakes that are auto-corrected:+- Unquoted fields containing commas (e.g. "San Francisco, CA" without quotes)+- Unescaped double quotes inside fields+- Inconsistent empty field handling+- Stray blank lines+"""++import asyncio+import os+import sys++sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))++from dotenv import load_dotenv++load_dotenv()++from browser_use import Agent, ChatBrowserUse+++async def main():+ agent = Agent(+ task=(+ 'Go to https://en.wikipedia.org/wiki/List_of_largest_cities and extract the top 10 cities. '+ 'Create a CSV file called "top_cities.csv" with columns: rank, city name, country, population. '+ 'Make sure to include all cities even if some data is missing — leave those cells empty.'+ ),+ llm=ChatBrowserUse(model='bu-2-0'),+ )++ history = await agent.run()++ # Check the generated CSV file+ if agent.file_system:+ csv_file = agent.file_system.get_file('top_cities.csv')+ if csv_file:+ print('\nGenerated CSV content:')+ print(csv_file.content)+ print(f'\nFile saved to: {agent.file_system.get_dir() / csv_file.full_name}')+++if __name__ == '__main__':+ asyncio.run(main())diff --git a/tests/ci/infrastructure/test_filesystem.py b/tests/ci/infrastructure/test_filesystem.pyindex 55ea7e685..4472fb779 100644--- a/tests/ci/infrastructure/test_filesystem.py+++ b/tests/ci/infrastructure/test_filesystem.py@@ -1286,3 +1286,131 @@ class TestFileSystemIntegration:assert file_obj.content == f'Content for file {i}'fs.nuke()+++class TestCsvNormalization:+ """Test CSV normalization that fixes common LLM output mistakes."""++ def test_normalize_quotes_fields_with_commas(self):+ """LLMs often forget to quote fields that contain commas."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('name,city\n"Smith, John","San Francisco, CA"')+ assert csv_file.content == 'name,city\n"Smith, John","San Francisco, CA"'++ def test_normalize_escapes_internal_quotes(self):+ """Fields with double quotes inside must be escaped per RFC 4180."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('name,quote\nJohn,"He said ""hello"""')+ assert csv_file.content == 'name,quote\nJohn,"He said ""hello"""'++ def test_normalize_handles_empty_fields(self):+ """Empty fields between commas should be preserved."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('a,b,c\n1,,3\n,,\n4,5,')+ assert csv_file.content == 'a,b,c\n1,,3\n,,\n4,5,'++ def test_normalize_strips_blank_lines(self):+ """Leading/trailing blank lines should be stripped."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('\n\na,b\n1,2\n\n')+ assert csv_file.content == 'a,b\n1,2'++ def test_normalize_preserves_valid_csv(self):+ """Already-valid CSV should pass through unchanged."""+ valid = 'name,age,city\nJohn,30,New York\nJane,25,London'+ csv_file = CsvFile(name='test')+ csv_file.write_file_content(valid)+ assert csv_file.content == valid++ def test_normalize_empty_content(self):+ """Empty or whitespace-only content should pass through."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('')+ assert csv_file.content == ''+ csv_file.write_file_content(' \n ')+ assert csv_file.content == ' \n '++ def test_normalize_on_append(self):+ """Appending rows should produce normalized combined output."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('name,city\nJohn,Boston')+ csv_file.append_file_content('\n"Jane","New York, NY"')+ assert csv_file.content == 'name,city\nJohn,Boston\nJane,"New York, NY"'++ def test_normalize_append_with_leading_newlines(self):+ """LLMs often prefix appended content with newlines."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('a,b\n1,2')+ csv_file.append_file_content('\n\n3,4')+ assert csv_file.content == 'a,b\n1,2\n3,4'++ async def test_normalize_through_filesystem_write(self):+ """CSV normalization works through the FileSystem.write_file path."""+ with tempfile.TemporaryDirectory() as tmp_dir:+ fs = FileSystem(base_dir=tmp_dir, create_default_files=False)++ await fs.write_file('data.csv', 'name,address\nJohn,"123 Main St, Apt 4"')+ file_obj = fs.get_file('data.csv')+ assert isinstance(file_obj, CsvFile)+ assert file_obj.content == 'name,address\nJohn,"123 Main St, Apt 4"'++ disk_content = (fs.data_dir / 'data.csv').read_text()+ assert disk_content == 'name,address\nJohn,"123 Main St, Apt 4"'++ fs.nuke()++ async def test_normalize_through_filesystem_append(self):+ """CSV normalization works through the FileSystem.append_file path."""+ with tempfile.TemporaryDirectory() as tmp_dir:+ fs = FileSystem(base_dir=tmp_dir, create_default_files=False)++ await fs.write_file('data.csv', 'name,score\nAlice,95')+ await fs.append_file('data.csv', '\n"Bob, Jr.",88')++ file_obj = fs.get_file('data.csv')+ assert file_obj is not None+ assert file_obj.content == 'name,score\nAlice,95\n"Bob, Jr.",88'++ fs.nuke()++ def test_normalize_single_column(self):+ """Single-column CSV should work correctly."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('names\nAlice\nBob\nCharlie')+ assert csv_file.content == 'names\nAlice\nBob\nCharlie'++ def test_normalize_quoted_newlines_in_fields(self):+ """Fields with embedded newlines should be properly quoted."""+ csv_file = CsvFile(name='test')+ csv_file.write_file_content('name,bio\nJohn,"Line 1\nLine 2"')+ assert 'Line 1\nLine 2' in csv_file.content+ assert csv_file.content == 'name,bio\nJohn,"Line 1\nLine 2"'++ def test_normalize_double_escaped_newlines(self):+ """LLM tool calls often produce literal \\n instead of real newlines."""+ csv_file = CsvFile(name='test')+ # Simulate double-escaped content: literal \n and \" (as they arrive from LLM)+ csv_file.write_file_content('name,city\\n1,Jakarta\\n2,Dhaka')+ assert csv_file.content == 'name,city\n1,Jakarta\n2,Dhaka'++ def test_normalize_double_escaped_quotes_and_newlines(self):+ """The exact failure mode from the bug: literal \\n and \\" in population values."""+ csv_file = CsvFile(name='test')+ # This is what the LLM actually sends: literal \n for row breaks,+ # literal \" around fields with commas+ content = 'rank,city,country,population\\n1,Jakarta,Indonesia,\\"41,913,860\\"\\n2,Dhaka,Bangladesh,\\"36,585,479\\"'+ csv_file.write_file_content(content)+ # Should unescape and produce proper CSV+ lines = csv_file.content.split('\n')+ assert len(lines) == 3+ assert lines[0] == 'rank,city,country,population'+ assert lines[1] == '1,Jakarta,Indonesia,"41,913,860"'+ assert lines[2] == '2,Dhaka,Bangladesh,"36,585,479"'++ def test_normalize_does_not_unescape_when_real_newlines_exist(self):+ """If content has real newlines, don't touch literal \\n inside field values."""+ csv_file = CsvFile(name='test')+ # Content with real newlines AND a field that legitimately contains \n chars+ csv_file.write_file_content('path,desc\n/tmp/a\\nb,test file')+ # Real newlines present → no unescaping, literal \n stays in the field+ assert csv_file.content == 'path,desc\n/tmp/a\\nb,test file'