[qimcis]

Description

The v1 GLM-4.7 parser decodes < > & " ' in every tool-call argument value, so when the model writes an entity on purpose, the client gets the raw character instead. Found this to be an issue when we were serving glm-5.3 and seeing agents writing python and html, e.g. cleaning up scraped HTML:

generated: txt = re.sub(r""", '"', txt)
delivered: txt = re.sub(r""", '"', txt)    -> SyntaxError: unterminated triple-quoted string literal

and attributes like title="He said "hi" & waved 'bye'." come through as title="He said "hi" & waved 'bye'.". In this case, the decoded value is still a valid string, so nothing errors on the serving side and it only shows up downstream

the main cause is decode_xml_entities in parse_tool_call_block, where the glm chat template writes string arguments into <arg_value> raw (only non-strings go through tojson), so there is no escaping layer to undo: the text between the markers already is the value.

This removes the decode, passes raw_value straight to coerce_value, and flips the existing test_xml_entity_decoding into test_xml_entities_kept_verbatim. This matches the v2 parser after #249 as well as vLLM's and SGLang's GLM-4.7 parsers, which also keep entities verbatim

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[qimcis]

PR redundant now, as change is now in #290