fix: correct regex replacement logic in _clean_content()

[Dabao21]

Problem

In agent_loop.py, the _clean_content() function used a heuristic '\\n' in p to decide whether to replace with '\n\n' or '':

for p in [r'<file_content>[\s\S]*?</file_content>', r'<tool_(?:use|call)>[\s\S]*?</tool_(?:use|call)>', r'(\r?\n){3,}']:
    text = re.sub(p, '\n\n' if '\\n' in p else '', text)

In Python raw strings, \n is stored as the literal two characters \ and n, so '\\n' in r'(\r?\n){3,}' evaluates to True. This means all three patterns were replaced with '\n\n', including <file_content> and <tool_use> tags which should be replaced with an empty string '' to remove them entirely.

Fix

Replace the heuristic with explicit pattern-replacement pairs:

for p, repl in [(r'<file_content>[\s\S]*?</file_content>', ''), (r'<tool_(?:use|call)>[\s\S]*?</tool_(?:use|call)>', ''), (r'(\r?\n){3,}', '\n\n')]:
    text = re.sub(p, repl, text)

Impact

• <file_content> and <tool_use> tags are now correctly stripped (replaced with empty string) instead of leaving extra newlines
• Reduces unnecessary token consumption in compressed history messages