fix(parsing): vLLM-parity type ladder for XML tool-call arg coercion

[hallerite]

Summary

Closes #47. Renderers' XML-style parsers (Qwen3.5, GLM, MiniMax, Laguna) flagged <parameter=x>True</parameter> as INVALID_JSON for boolean params because json.loads("True") fails — even though both SGLang's Qwen3CoderDetector and vLLM's Qwen3CoderToolParser accept Pythonic literals via case-folded comparison. Models freely emit True/False at inference because the reference parsers normalize them, but the same outputs came back malformed through renderers.

This swaps _coerce_arg_value from a string-or-json.loads dispatch to the full _convert_param_value ladder both reference parsers share:

Declared type	Coercion
(none) — no schema passed	json.loads, fallback to raw text + INVALID_JSON (historical behavior preserved)
string / str / text / varchar / char / enum	return verbatim
int / uint / long / short / unsigned	int(text), fall back to string + INVALID_JSON
num / float	float(text) with SGLang's source-string int-demotion heuristic, fall back to string + INVALID_JSON
boolean / bool / binary	text.lower() == "true"; non-true/false → False + INVALID_JSON
object / array / dict* / list* / anyOf	json.loads, then ast.literal_eval fallback (Python literals like {'k': 1}), then string + INVALID_JSON
Unknown	ast.literal_eval, fall back to string + INVALID_JSON
null literal (any case)	None — except for declared string-family types (see deviation below)

INVALID_JSON is preserved as the verifier / RL-loss signal whenever the value had to degenerate (e.g. yes for a bool, abc for an int).

Deliberate deviation from vLLM/SGLang

Both reference parsers null-coerce "null" before checking type, so a string-typed arg of "null" returns Python None. Renderers keeps "null" verbatim for type: "string" so the existing tests/test_tool_arg_type_preservation.py::test_string_arg_preserves_type[*-string-null] contract holds across all five XML renderers. The XML wire format already can't distinguish the string "null" from JSON null, but when the schema explicitly says type: "string" we honour it.

Reference parser sources

• SGLang: python/sglang/srt/function_call/qwen3_coder_detector.py
• vLLM: vllm/tool_parsers/qwen3coder_tool_parser.py::_convert_param_value

The two are functionally near-identical for this code path; vLLM is slightly more permissive (extra anyOf → object branch). Renderers now matches vLLM's superset, with the one string-null deviation noted above.

Test plan

• pytest tests/test_arg_coercion.py — 43 new unit tests cover every branch of the ladder, including the issue's True/False regression through full parse_qwen35 round-trip.
• pytest tests/test_tool_arg_type_preservation.py — 30/30 string-preservation cases pass across all five XML renderers (the string-null case still round-trips verbatim).
• Full suite: 1371 passed, 53 skipped, 1 xfailed.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes tool-call extraction and argument coercion for XML-based renderers, which may alter parsed values/statuses (e.g., "null"→None, booleans/int/float casting) and affect downstream tool execution or training masks if assumptions differ.

Overview
Brings Qwen3.5-style XML tool-call parsing into closer parity with vLLM/SGLang by replacing the prior json.loads-centric coercion with a schema-driven type ladder (case-insensitive null, bool/int/float handling, and ast.literal_eval fallback for Python literals).

Refactors Qwen3.5 extraction to use shared regex-based <function=…> block parsing (_parse_xml_function_blocks), improving tolerance for malformed/unclosed tags and adding vLLM-style behaviors like unstripped content prefixes, backoff parsing when <tool_call> wrappers are missing, and per-call ids in chatcmpl-tool-… format. Adds comprehensive unit/integration coverage for the new coercion ladder and edge cases, and updates the existing string-type preservation test to account for XML "null" ambiguity.

^{Reviewed by Cursor Bugbot for commit b58fa63. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix XML tool-call argument coercion to match vLLM type ladder in Qwen3.5 parser

• Replaces json.loads-based arg coercion in parsing.py with a schema-aware type ladder: null short-circuit, string passthrough, int/float parsing, boolean via lower() == 'true', and object/array via json.loads with ast.literal_eval fallback.
• Adds _parse_xml_function_blocks to parse <function=> blocks without requiring a <tool_call> wrapper, with tolerance for missing closing tags and per-call id/status assignment.
• Refactors Qwen35ToolParser.extract() in parsers.py to delegate to _parse_xml_function_blocks, enabling multiple ParsedToolCall entries per <tool_call> and more precise status codes (UNCLOSED_BLOCK, MALFORMED_STRUCTURE).
• Tool-call ids now follow the vLLM chatcmpl-tool-<16 hex> format via _make_tool_call_id.
• Behavioral Change: content text is no longer stripped (whitespace preserved); null literal is coerced to None even when the declared type is string; on any parse exception, the full post-think text is returned as content with no tool calls.

^{Macroscope summarized b58fa63.}

[cursor]

Unused compiled regex _TOOL_CALL_BLOCK_RE is dead code

Low Severity

_TOOL_CALL_BLOCK_RE is defined at module level as a compiled regex but is never referenced anywhere in the codebase. A grep confirms the only occurrence is its definition. The sibling patterns _FUNCTION_BLOCK_RE and _PARAMETER_BLOCK_RE are both actively used, but this one appears to have been carried over from the vLLM source listing without an actual consumer. It adds unnecessary clutter and a small startup cost for the unused re.compile.

 

^{Reviewed by Cursor Bugbot for commit a338f9d. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR modifies core parsing logic with significant runtime behavior changes to argument type coercion. An unresolved medium-severity comment identifies a potential regression where Nemotron3 argument parsing may break due to the no-schema path now returning verbatim strings instead of attempting JSON parsing.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

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

There are 2 total unresolved issues (including 1 from previous review).

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

^{Reviewed by Cursor Bugbot for commit 96a1c98. Configure here.}

[cursor]

No-schema coercion change breaks Nemotron3 argument parsing

Medium Severity

The no-schema path in _coerce_arg_value changed from attempting json.loads (with text fallback) to returning verbatim strings. Nemotron3Renderer.parse_response deliberately does not forward tools to parse_qwen35 (marked # noqa: ARG002), so every parameter hits the no-schema branch. Previously "42" parsed to 42 and "true" to True via json.loads; now they stay as raw strings "42" and "true". This silently changes the types in the arguments dict for all Nemotron3 tool calls.

 

^{Reviewed by Cursor Bugbot for commit 96a1c98. Configure here.}