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

[hallerite]

Summary

Closes #47. Replaces renderers' XML-style tool-call arg coercion with a byte-for-byte port of vLLM's shared extract_types_from_schema + coerce_to_schema_type helpers from vllm/tool_parsers/utils.py. The previous json.loads-centric coercion 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.

The fix

Renderers now mirrors vLLM's current shape (refactored upstream in vLLM #43025, 2026-05-19):

1. _extract_types_from_schema recursively flattens the JSON-Schema fragment to a set of types — handles top-level type (string or list), infers types from enum values, and recurses through anyOf / oneOf / allOf. Defaults to ["string"] when no type info is available.
2. _coerce_arg_value walks a priority-ordered ladder, returning the first successful coercion:

Priority	Type	Coercion
1	null	case-insensitive "null" ⇒ None
2	integer (+ aliases int, uint, long, …)	int(text), else continue
3	number (+ aliases float, double, …)	float(text), demote to int when whole, else continue
4	boolean (+ alias bool)	text.lower().strip() in {"true","1"} ⇒ True; {"false","0"} ⇒ False; else continue
5	object (+ aliases dict)	json.loads(text), else continue
6	array (+ aliases arr, list, sequence)	json.loads(text), else continue
7	string (+ aliases str, text, enum, …)	return verbatim — always-succeeding terminal

When the schema doesn't include string and no typed branch accepts the value, a terminal json.loads is tried before falling back to raw text + INVALID_JSON.

Renderers-specific deviation

_coerce_arg_value returns (value, used_fallback). The flag is True iff every typed branch declined and the terminal json.loads also failed — i.e. the value couldn't be expressed as anything the schema permits. The parser propagates that to ToolCallParseStatus.INVALID_JSON for the verifier / RL-loss signal. vLLM has no such signal (its callers don't need one).

Other structural changes (kept from the original branch)

• parse_qwen35 refactored to use shared _parse_xml_function_blocks, mirroring vLLM's three regex patterns (closed + unclosed alternations, <parameter> lookahead recovery).
• Back-off path: completions without <tool_call> markers but with <function=…> blocks are still surfaced (vLLM qwen3coder_tool_parser.py:269-271).
• Content text no longer stripped (vLLM :316-319).
• Tool-call ids in vLLM's chatcmpl-tool-<16hex> format.
• Exception catch-all: any unhandled extraction error surfaces full text as content with no tool calls (vLLM :327-331).

Behavior changes vs renderers' prior coercion

• Union schemas (anyOf / oneOf / allOf) recursively flatten — Union[str, X] with bare-string wire values now stays OK, subsuming the fix from fix(parsing): accept bare-string args for union schemas with string #63 as a side effect.
• No top-level null short-circuit. "null" only coerces to None when "null" is in the schema's type set (Optional[X] ⇒ anyOf [X, null], or type: ["X", "null"]). Pure string-typed "null" stays the string "null". This drops the deliberate XML-ambiguity carve-out that earlier commits on this branch had carried.
• Bool degeneracy: <parameter=x>yes</parameter> for a bool-typed param now returns the raw text "yes" + INVALID_JSON, not False + INVALID_JSON. Matches current vLLM (the bool branch declines and the priority loop continues to the terminal).
• Number demotion: 1e3 ⇒ int 1000, 1.0 ⇒ int 1, regardless of source-string shape. Pure vLLM rule (val.is_integer()), not the prior "." in text heuristic.
• No ast.literal_eval anywhere — vLLM doesn't use it. Python-literal dicts like {'k': 1} for an object param now flag INVALID_JSON.
• "binary" dropped as a boolean alias (not in vLLM's _TYPE_ALIASES).

Reference parser sources

• vLLM extract_types_from_schema / coerce_to_schema_type: vllm/tool_parsers/utils.py:453,526
• vLLM Qwen3CoderToolParser._convert_param_value: vllm/tool_parsers/qwen3coder_tool_parser.py:123 (routes through the shared helpers)
• SGLang Qwen3CoderDetector: python/sglang/srt/function_call/qwen3_coder_detector.py (functionally close; vLLM is the strict reference)

Test plan

• pytest tests/test_arg_coercion.py — 51 unit tests covering every branch of the new ladder, including the True / False regression through full parse_qwen35 round-trip and explicit cases for the behavior changes above.
• pytest tests/test_tool_arg_type_preservation.py — 48/48 cases pass across all five XML renderers + two JSON-format controls. String-typed "null" now round-trips as "null" (was None on the prior branch); the Optional[str] shape was never exercised here.
• Full suite: 1874 passed, 53 skipped, 1 xfailed.

🤖 Generated with Claude Code

---

Note

Medium Risk
Parsing behavior changes across Qwen3.5 and shared _coerce_arg_value (used by GLM/MiniMax/Laguna too): null coercion, union schemas, and untyped params now differ from the old json.loads-first logic—intentional but can shift verifier/RL signals for edge-case completions.

Overview
Replaces XML tool-call argument coercion with vLLM’s schema-driven type ladder (null → integer → number → boolean → object → array → string), including recursive anyOf/oneOf/allOf flattening. That fixes cases like <parameter=…>True</parameter> on boolean fields (no longer INVALID_JSON) and keeps string-typed "null" as a string unless null is in the schema.

Qwen3.5 parsing is refactored to shared _parse_xml_function_blocks using vLLM’s tolerant regexes (missing </parameter>, unclosed <function>, backoff when <tool_call> is absent). Content before tools is no longer stripped; calls get chatcmpl-tool-<16hex> ids; extraction errors return full text with no partial tool calls. Qwen35ToolParser delegates to the same helper (schema-less path = verbatim strings).

Adds tests/test_arg_coercion.py and updates type-preservation test docs for the new null/string rules.

^{Reviewed by Cursor Bugbot for commit af8d6e1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix XML tool-call argument coercion to match vLLM's type-ladder semantics

• Rewrites _coerce_arg_value in parsing.py to mirror vLLM's priority-ordered type ladder: null > integer > number > boolean > object > array > string, with case-insensitive boolean parsing and null coercion only when declared in the schema.
• Adds _parse_xml_function_blocks, a tolerant parser that accepts missing </parameter> and unclosed </function> tags, strips a single leading/trailing newline from parameter bodies, and assigns vLLM-style chatcmpl-tool-<16 hex> ids via the new _make_tool_call_id helper.
• Updates parse_qwen35 to preserve content whitespace, recover tool calls from bare <function=...> blocks when <tool_call> wrappers are absent, and fall back to returning raw content on extraction errors.
• Refactors _parse_xml_tool_calls and Qwen35ToolParser.extract to delegate to _parse_xml_function_blocks, producing explicit UNCLOSED_BLOCK or MALFORMED_STRUCTURE diagnostics instead of silently dropping malformed calls.
• Behavioral Change: untyped parameters are now returned as raw strings rather than being JSON-coerced; whole floats declared as integer are demoted to int.

^{Macroscope summarized af8d6e1.}

[macroscopeapp]

Approvability

Verdict: Needs human review

Significant refactor of tool-call argument parsing with new type coercion logic, schema handling, and changed parse status determination. The runtime behavior changes and unresolved review comment about status precedence warrant human review.

No code changes detected at af8d6e1. Prior analysis still applies.

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

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

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

^{Reviewed by Cursor Bugbot for commit 8bd06f6. Configure here.}