[hallerite]

Fix XML boolean arguments such as True being marked INVALID_JSON. Coercion follows vLLM v0.32.0 for aliases, nullable/union/enum schemas, numeric conversion, and non-finite JSON. Arguments without schemas remain strings.

Qwen uses shared XML extraction for both renderer and pluggable parsing, preserving whitespace, incomplete-call diagnostics, and token spans. Prime Qwen3 keeps its template separator normalization. The branch includes current main.

Closes #47.

Validation:

• Full suite: 13,497 passed, 126 skipped.
• 35,433 cases compared with vLLM's helper and XML argument-converter code; 960 structural cases compared with the implementation before cleanup.
• Ruff, pre-commit, and type checks pass. Commits are signed.

References: vLLM 0.32 coercion, Qwen grammar.

---

Note

Medium Risk
Changes tool-call parsing and typed argument coercion used across Qwen/GLM-style renderers; behavior shifts for booleans, unions, and schema-less paths, though coverage is extensive and aligned with vLLM.

Overview
Aligns schema-driven XML tool-argument coercion with vLLM 0.32, so values like True for boolean params parse as OK instead of INVALID_JSON. _coerce_arg_value now honors type aliases, unions/enums, null/int/number/boolean rules, and rejects non-finite JSON; without a schema, arguments stay plain strings (including the pluggable qwen3.5 parser).

Qwen 3.5 XML extraction is centralized in renderers/parsing.py: Qwen35ToolParser delegates to _parse_xml_tool_calls, which uses _parse_xml_function_blocks for more tolerant parameter boundaries, bare <function=…> when <tool_call> is missing, vLLM-style chatcmpl-tool-* ids, preserved content whitespace, and a safe fallback that returns full text if extraction throws. Prime Qwen3 still strips parsed assistant content for template parity.

Adds tests/test_arg_coercion.py and updates GLM unknown-tool expectations so unschema’d args remain strings.

^{Reviewed by Cursor Bugbot for commit e2cd1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Delegate Qwen35ToolParser.extract XML parsing to shared _parse_xml_tool_calls helper

Aligns XML argument coercion with vLLM 0.32 by replacing the parser's inline XML extraction in parsers.py with the shared _parse_xml_tool_calls helper. The call passes the token offset and an empty parameter schema index, so this path does not supply tool schemas and argument values stay strings without a schema. Malformed and unclosed blocks are handled by the shared parser, and tool-call metadata is assigned through that path.

• Risk: for tool calls that contain parameters, the shared coercion code references an undefined types variable, so extraction raises NameError instead of returning parsed calls. Check the coercion branch in _parse_xml_tool_calls.

^{Macroscope summarized e2cd1f8.}

[macroscopeapp]

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — This changes shared production tool-call parsing behavior across multiple XML formats, including argument coercion, malformed-call recovery, content handling, and recognition of bare function calls. The extensive tests are useful, but the runtime scope is broader than a narrowly isolated coercion fix.

Notes:

• Macroscope's correctness review did not run, so approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[cursor]

Stale Bugbot comment from a previous run.

[HBX814]

I ran this branch against the original repro from #47 — posting the numbers in case they help move it along.

main still reproduces the issue (at 1a5bd3b):

_coerce_arg_value("True",  {"type": "boolean"}) -> ('True',  True)   # used_json_fallback
_coerce_arg_value("False", {"type": "boolean"}) -> ('False', True)

End-to-end through Qwen35Renderer.parse_response, using @rohitrastogi's exact schema and completion:

status   : ToolCallParseStatus.INVALID_JSON
arguments: {'path': '/', 'max_depth': 2, 'include_files': 'True', 'show_size': 'True'}

This branch fixes it. Same script on fix/qwen35-arg-coercion-vllm-parity:

_coerce_arg_value("True", {"type": "boolean"}) -> (True, False)

status   : ToolCallParseStatus.OK
arguments: {'path': '/', 'max_depth': 2, 'include_files': True, 'show_size': True}

which lines up with what Qwen3CoderDetector._convert_param_value produces on the SGLang side.

Parsing tests on the branch — test_parsers.py, test_parse_response.py, test_parse_response_robustness.py, test_tool_arg_type_preservation.py, test_kimi_k25_tool_schema.py — 359 passed, 18 skipped.

One thing worth flagging before this can land: the branch is 55 commits behind origin/main (last merge from main was 2026-05-26), so newer test files such as tests/test_glm_tool_name_validation.py don't exist on it yet. Worth merging main in so CI runs against the current suite.

Environment: transformers 5.15.0, Qwen/Qwen3.5-0.8B tokenizer, no GPU.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a5a182f. Configure here.}