[Fix] Resolve tool argument types through top-level anyOf/oneOf/allOf

[JustinTong0323]

Summary

Tool schemas that compose subschemas with a top-level anyOf/oneOf/allOf (legal JSON Schema / OpenAI spec, e.g. discriminated-union arguments) break most SGLang tool-call parsers: nested object/array/scalar arguments are returned to the client as JSON-encoded strings instead of typed values, or (glm45/glm47 streaming) as invalid JSON.

"parameters": {
  "type": "object",
  "oneOf": [
    { "type": "object",
      "properties": { "kind": {"const": "acme"},
                      "payload": {"type": "object", "properties": {"value": {"type": "string"}}} },
      "required": ["kind", "payload"] },
    { "type": "object", "properties": { "kind": {"const": "other"} }, "required": ["kind"] }
  ]
}

With this schema, calling acme(kind="acme", payload={"value":"hello"}) yields payload == "{\"value\": \"hello\"}" (a string) on the qwen3_coder parser in both streaming and non-streaming modes, and on the glm47 parser when streaming.

Root cause

Every affected detector navigated parameters["properties"][arg] only at the top level. With a top-level combinator there is no properties key, so argument-type inference failed and each parser fell back to treating values as strings:

• qwen3_coder: _get_arguments_config returns the whole schema (with oneOf/type keys) as the param map → every value returned raw, both modes. Scalars are also affected ("count": "7", "verbose": "True").
• glm45/glm47: streaming decides the value type when <arg_value> opens; with no schema info it defaults to string and JSON-quotes the value. Non-streaming survives only because of a json.loads fallback.

A second, pre-existing bug sits on the same path: glm45/glm47 streaming decides whether to emit the outer closing } via _last_arguments.endswith("}"). When the last argument is object/array-typed, the trailing } belongs to the nested value, so the outer object is never closed and streamed arguments is invalid JSON (masked live only when the model emits trailing whitespace inside </arg_value>).

Fix

• New shared get_schema_properties() in function_call/utils.py: returns top-level properties, descending into anyOf/oneOf/allOf branches when absent (first branch wins on duplicate keys, matching oneOf preference order).
• Adopted in every detector with a parse-path property lookup: glm45, glm47, qwen3_coder, step3, minicpm5, hunyuan, spark25, minimax_m2, minimax_m3, dots, poolside_v1, mimo, kimik2 (1–3 lines each).
• glm45/glm47 streaming finalize: close the outer arguments object whenever a parameter was started, instead of the endswith("}") heuristic.

Two deliberate behavior notes:

• minicpm5: for top-level-combinator tools the allowed-props filter was previously bypassed (empty property set); it now activates, so undeclared arguments are dropped as they already are for flat schemas.
• qwen3_coder: a declared-but-empty properties: {} keeps its old semantics (no spurious "param not defined" warnings).

kimik3_structural_tag is intentionally untouched: it builds grammars and already degrades gracefully to AnyTextFormat with top-level combinators.

Test plan

• New registered tests: TestGetSchemaProperties (4) + TestTopLevelCompositeToolSchema (6: glm45/glm47 streaming with oneOf schema, flat-schema streaming brace cases, qwen3_coder stream/non-stream) in test_function_call_parser.py, and TestMinimaxM3TopLevelOneOf (1: scalar coercion under oneOf) in test_minimax_m3_detector.py. Every detector test was verified red on unpatched main; cases that also pass on unfixed code (glm non-streaming, structure-driven minimax streaming) were deliberately dropped.
• Full test_function_call_parser.py + test_minimax_m3_detector.py (270 tests): no regressions.
• GPU A/B on 2×H200, 6 schema shapes (oneOf nested object, anyOf, allOf, array/scalar coercion, flat control, multi-tool mixed) × stream/non-stream, asserting decoded argument types and finish_reason=tool_calls:

model / parser	baseline	this PR
Qwen/Qwen3-Coder-30B-A3B-Instruct + qwen3_coder	2/12	12/12
zai-org/GLM-4.5-Air-FP8 + glm45	7/12	12/12

---

CI States

Latest PR Test (Base): ✅ Run #33100877731
Latest PR Test (Extra): ❌ Run #33100877307
Latest PR Test (AMD ROCm 7.2): ❌ Run #33100877334

[JustinTong0323]

/tag-and-rerun-ci

[JustinTong0323]

/rerun-failed-ci

[JustinTong0323]

/rerun-failed-ci

[JustinTong0323]

/rerun-failed-ci