fix(tool_parser): treat qwen_xml tool-call arg values literally

[key4ng]

Description

Problem

qwen_xml's value decoder ran html.unescape on every tool-call argument value, so any argument containing a literal HTML entity (&, <, ', …) was silently mutated on parse (& → &).

The Qwen3-Coder XML tool format is not HTML-escaped on render — the chat template emits argument values via | tojson | safe / | string, so the model emits literal characters. vLLM's Qwen3CoderToolParser, Qwen-Agent, and Qwen's chat template all pass argument values through verbatim. SMG was the only implementation that unescaped, and it applies no matching escape on its own render side, so its parse was not the inverse of its render.

Impact: a coding agent writing an HTML/XML/Markdown snippet had its argument silently corrupted before it reached the tool:

• Model emits (literal): <parameter=content>\n<a>Tom &amp; Jerry</a>\n</parameter>
• vLLM delivers to the tool: <a>Tom &amp; Jerry</a> ✅
• SMG delivered to the tool: <a>Tom & Jerry</a> ❌

Because the assistant tool call is re-serialized into the next turn's prompt, the mutated value also propagated into subsequent turns.

Fixes #1888.

Solution

Treat argument values literally — drop the html_unescape call from both safe_val (schema-less inference path) and coerce_value (schema-typed path), matching vLLM's qwen3coder parser. JSON / schema-type coercion is unchanged.

Changes

• Remove html_unescape from safe_val and coerce_value in crates/tool_parser/src/parsers/qwen_xml.rs (keep trim + JSON / schema-type coercion).
• Delete the now-dead html_unescape helper (~70 lines).
• Invert the unit + integration tests that asserted entity decoding so they now assert literal preservation.
• Add golden conformance test test_arg_values_with_entities_roundtrip_unchanged covering both the inference and string-schema paths — this class of divergence now fails in CI, not as a benchmark score.

Test Plan

Before: value <a>Tom &amp; Jerry</a> &lt;x&gt; it&#39;s parsed to <a>Tom & Jerry</a> <x> it's (entities decoded).
After: the same value round-trips unchanged.

cargo test -p tool-parser                                            # all pass, incl. new golden test
cargo +nightly fmt --all -- --check                                  # clean
cargo clippy -p tool-parser --all-targets --all-features -- -D warnings  # clean

Checklist

• cargo +nightly fmt passes
• cargo clippy --all-targets --all-features -- -D warnings passes
• (Optional) Documentation updated
• (Optional) Please join us on Slack #sig-smg to discuss, review, and merge PRs

Summary by CodeRabbit

• Bug Fixes
  • Qwen XML tool argument values now preserve HTML-entity-like text exactly as provided, rather than decoding it into rendered characters.
  • Encoded inputs now round-trip consistently in both schema-less parsing and schema-typed string parsing.
  • Mixed XML text and JSON values now keep entity-like substrings unchanged.
• Tests
  • Updated and added coverage to assert literal preservation for named, numeric, and hex entities, including round-trip regression coverage.

[coderabbitai]

📝 Walkthrough

Walkthrough

Removes HTML-entity decoding from Qwen XML tool argument parsing while retaining JSON and schema-based coercion. Unit and integration tests now verify that named, numeric, and hexadecimal entity-like substrings remain unchanged.

Changes

Literal value parsing

Layer / File(s)	Summary
Preserve literal argument values crates/tool_parser/src/parsers/qwen_xml.rs	safe_val and coerce_value no longer decode HTML entities; values are trimmed, JSON-parsed, checked for Python literals, and otherwise returned literally with schema coercion applied when available.
Parser regression coverage crates/tool_parser/src/parsers/qwen_xml.rs	Removes entity-decoding tests and adds coverage for unchanged entities in safe_val, schema-less parsing, and schema-typed string parsing.
XML integration regression coverage crates/tool_parser/tests/tool_parser_qwen_xml.rs	Updates named-entity, numeric/hex-entity, and mixed HTML-and-JSON assertions to expect encoded entity text in parsed values.

Estimated code review effort: 2 (Simple) | ~12 minutes

Possibly related PRs

• lightseekorg/smg#1841: Modified the same Qwen XML argument coercion paths in qwen_xml.rs.

Suggested reviewers: slin1237, CatherineSue, claude

Poem

A rabbit hopped through XML bright,
And kept each entity just right.
“No decoding here,” it said with cheer,
“The literal text belongs right here!”
Tests now guard the encoded trail. 🐇

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The parser now preserves entity text, keeps JSON/schema coercion, and adds coverage for unchanged round-tripping.
Out of Scope Changes check	✅ Passed	The changes stay focused on qwen_xml parsing behavior and related tests, with no unrelated scope added.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: qwen_xml tool-call argument values are now preserved literally instead of being HTML-unescaped.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/qwen-xml-html-unescape

---

_{Comment @coderabbitai help to get the list of available commands.}

[claude]

Clean removal of incorrect HTML-entity decoding in the Qwen XML parser. The change aligns with vLLM's Qwen3CoderToolParser and the Qwen chat template, both of which pass argument values through verbatim. Tests comprehensively cover the schema-less inference path, the schema-typed string path, and the mixed HTML-entity-in-JSON edge case. No issues found.

[gemini-code-assist]

Code Review

This pull request removes HTML-entity decoding from the Qwen XML parser to ensure argument values are treated literally, aligning with vLLM's Qwen3CoderToolParser and Qwen-Agent. Tests have been updated and added to verify that HTML entities are preserved verbatim. The review feedback correctly points out a mismatch in the newly added test test_arg_values_with_entities_roundtrip_unchanged where the function name in the XML template (write_file) does not match the function name defined in tool_with_props (f), which prevents the schema-typed path from being properly exercised.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

In this test, the XML input uses <function=write_file>, but the tools list is created using tool_with_props, which defines a function named "f".

Because of this mismatch, helpers::param_types_for_function(tools, "write_file") returns an empty map, meaning the schema-typed path (coerce_value) does not find the "content" parameter's declared type ("string") and silently falls back to the schema-less safe_val path. As a result, the schema-typed string path is not actually being tested here.

Changing the function name in the XML template to "f" ensures that the schema-typed path is correctly matched and exercised.

        let text = format!(
            "<tool_call>\n<function=f>\n\\\n             <parameter=content>\n{literal}\n</parameter>\n\\\n             </function>\n</tool_call>"
        );

[key4ng]

triggered bfcl eval: https://github.com/lightseekorg/smg/actions/runs/29046920019