I've found that at least for Claude Haiku 4.5, the model will call the write_file tool many times in a row without providing the content parameter (3-8 times incorrectly before finally providing content). This issue might be alleviated by refining the tool description, currently in which content is never explicitly described as the content of the file:

WRITE_FILE_TOOL_DESCRIPTION = """Writes to a new file in the filesystem.

Usage:
- The file_path parameter must be an absolute path, not a relative path
- The content parameter must be a string
- The write_file tool will create the a new file.
- Prefer to edit existing files over creating new ones when possible."""

While the purpose of the content parameter might seem obvious, it could be interpreted by a naive LLM as "content type" parameter and thus not seen as a truly necessary parameter.

Also, the LLM might be reluctant to write out such a long parameter value to a tool; in my case, the content is >60k tokens. Anecdotally, the reluctancy to provide the content parameter seems to increase as the content size increases (e.g., 8-10 failed tool calls if the content token count is >100k tokens).