bug(cli/mcp): a node name containing </script> breaks out of the inlined JSON and corrupts the offline visualizer page #1514

utils/visualize_graph.py:214 (graph JSON), :135 and :201 (title)

The graph payload is inlined into a <script> block via json.dumps. json.dumps escapes quotes and backslashes but not /, so any node name or file path containing the literal </script> closes the script element early and everything after it is parsed as markup.

Reproduction

$ .venv/bin/python -c "
from codegraphcontext.utils.visualize_graph import build_graph_data, open_in_browser
import tempfile, pathlib
nodes=[{'id':1,'label':'Function','name':'x</script><script>alert(1)</script>','file_path':'/a.py'}]
p = open_in_browser(build_graph_data(nodes, []), title='t',
                    output_path=str(pathlib.Path(tempfile.mkdtemp())/'g.html'))
html = pathlib.Path(p).read_text()
print('script-breakout present:', '</script><script>alert(1)' in html)
"
script-breakout present: True

The rendered document contains the closing tag verbatim inside what was meant to be a JSON string literal:

      "name": "x</script><script>alert(1)</script>",

title is interpolated unescaped at :135 and :201 as well.

Impact

Mostly a robustness problem: a node whose name contains </script> breaks the offline visualization page entirely (blank or malformed graph) with no indication why. Minified or bundled JS, HTML templates and test fixtures containing that string are all realistic sources — CGC already indexes .js and .html files, and this repo itself ships website/public/wasm/*.js.

It is also a genuine (if low-reach) injection: indexing an untrusted repository and opening the generated page executes markup from that repository's identifiers in the local browser.

Suggested fix

Escape </>/& in the serialized JSON before inlining — the standard treatment is replacing < with < (and / in </ with <\/) after json.dumps. HTML-escape title at both interpolation sites.

_{Reproduced with the snippet above.}