Keep code blocks intact when you transform Markdown

Agno now exposes advance_code_fence in agno.utils.markdown, the same fence tracking Agno's page chunker uses. A sync_pages transform sees every line of a page, code included, and advance_code_fence lets it change the prose and leave code examples exactly as written.

Feed advance_code_fence one line at a time with the state it returned for the previous line. This transform decodes HTML entities in prose and skips anything inside a fence. Agno calls a transform with the page's text and the page's path as a keyword argument, so the function accepts both.

import html
 
from agno.utils.markdown import advance_code_fence
 
def normalize_prose(markdown: str, path: str = "") -> str:
    output, fence = [], None
    for line in markdown.splitlines(keepends=True):
        was_code = fence is not None
        fence, delimiter = advance_code_fence(line.rstrip("\r\n"), fence)
        output.append(line if was_code or delimiter else html.unescape(line))
    return "".join(output)
 
report = knowledge.sync_pages(url="https://docs.agno.com/llms.txt", transform=normalize_prose)

Given Use <Agent> here., a Python block containing x = a < b, and Done & dusted., normalize_prose returns Use <Agent> here. and Done & dusted. and leaves the code line as x = a < b.

In advance_code_fence, a closing fence has to use the same character as the opening fence and be at least as long. A four-backtick block can therefore show a three-backtick example safely. advance_code_fence also counts fences indented inside Markdown or MDX.

See the cookbook, and learn more about transforming published Markdown in the documentation.