fix: preserve XML-like tags in write_file/execute_code/patch content (#72797)

[webtecnica]

Description

Fixes #72797: The strip_think_blocks function in agent/agent_runtime_helpers.py and the _strip_reasoning_tags function in cli.py were both using re.IGNORECASE when stripping tool-call XML blocks (<tool_call>, <tool_calls>, <tool_result>, <function_call>, <function_calls>) from assistant text content.

This caused uppercase variants like <TOOL_CALL> and </TOOL_CALL> — which appear as valid literal data in JS/HTML/Python source code being written to files — to be silently stripped, truncating the tool content arguments before they reached the tool.

Root cause

The comment in agent_runtime_helpers.py already documented that case-insensitive matching was intentionally avoided to prevent corrupting literal data (e.g., JS/HTML containing <TOOL_CALL> strings), but the code was incorrect — it used re.IGNORECASE despite the comment.

The cli.py version also silently used re.IGNORECASE without the clarifying comment.

Changes

agent/agent_runtime_helpers.py

• Removed re.IGNORECASE from the tool-call XML block regex patterns
• Added clarifying comment referencing [Bug]: write_file/execute_code/patch silently strip XML-like tags from content arguments — agent must use chr(60)/chr(62) workarounds #72797
• Also fixed the stray close-tag regex which had the same issue

cli.py

• Removed re.IGNORECASE from the same tool-call XML block patterns
• Added clarifying comment referencing [Bug]: write_file/execute_code/patch silently strip XML-like tags from content arguments — agent must use chr(60)/chr(62) workarounds #72797
• Also fixed the stray close-tag regex which had the same issue

Think/reasoning tag patterns (<think>, <thinking>, <reasoning>, <thought>, <REASONING_SCRATCHPAD>) still use re.IGNORECASE — those are expected in any case from reasoning models and should be stripped regardless of case.

Testing

• All existing tests pass (38/38: 9 in test_strip_reasoning_tags_cli.py + 29 in test_run_agent.py related tests)
• Manually verified:
  • <TOOL_CALL> literal data in JS/Python source is preserved
  • Lowercase <tool_call> blocks are still correctly stripped
  • Reasoning blocks (<think>, etc.) continue to be stripped case-insensitively as expected