[Bugfix] Support root JSON Schema combinators in tool parsers by xeophon · Pull Request #36483 · sgl-project/sglang

xeophon · 2026-08-26T10:40:28Z

Motivation

Fixes #32286 and generalizes the MiniMax-specific approach in #32299.

Several XML-style tool-call detectors infer argument types only from tool.function.parameters["properties"]. Valid JSON Schemas that declare properties below a root-level allOf, anyOf, or oneOf therefore hide those type hints. In the reported MiniMax-M3 case, numbers become strings and arrays can leak raw parser tags instead of producing typed arguments.

This change exposes those properties at the shared FunctionCallParser layer without changing the original tool schema or adding parser-specific JSON Schema logic.

Modifications

Add get_tool_parser_property_hints, an explicitly lossy detector-facing property projection:
- resolve local JSON Pointer references along the projection path;
- keep direct root properties authoritative;
- project properties below root-level allOf, anyOf, and oneOf;
- retain branch-only or identical declarations;
- treat absence from another branch as non-conflicting;
- collapse differing declarations for the same property name to {} so detectors keep their conservative string fallback.
Construct detector_tools once in FunctionCallParser using Pydantic copies only when the projected properties differ.
Pass detector_tools to non-streaming parsing, streaming increments, and stream finalization.
Keep self.tools and the original schemas unchanged for structural tags, constrained decoding, and validation.
Leave the individual tool-format detectors unchanged.
Add registered CPU regressions for the real disjoint number / stringList / numberList schema shape.

The projection is intentionally a static type-hint view. It is not complete JSON Schema evaluation and does not select or validate branches at runtime. Exact branch-dependent interpretation would require buffering a complete tool call before emitting arguments, which is incompatible with incremental streaming.

Accuracy Tests

Not applicable to model inference. This changes deterministic post-processing of tool-call arguments and does not change model execution or generated tokens.

The regression verifies:

disjoint properties under allOf, anyOf, and oneOf;
direct-property precedence, identical declarations, explicit conflicts, and local references;
preservation of the original schema;
conservative string fallback for conflicting declarations;
number and array parsing without XML-tag leakage;
parser reuse without per-call schema state;
streaming and non-streaming behavior across 11 detector formats.

Speed Tests and Profiling

Not applicable to model execution. The projection is computed once when a FunctionCallParser is constructed; the streaming parse path performs no branch validation or schema mutation.

Validation

Focused root-combinator regression: 9 tests passed, 69 subtests passed.
Ruff formatting and scoped correctness checks passed on the changed files.
git diff --check passed.
Final diff against main contains two production files and one test file.

Checklist

Format your code according to the Format code with pre-commit.
Add unit tests according to the Run and add unit tests.
Update documentation according to Write documentations. (N/A: internal parser behavior only.)
Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. (N/A: deterministic parser-only fix.)
Follow the SGLang code style guidance.

Review and Merge Process

Ping Merge Oncalls to start the process. See the PR Merge Process.
Get approvals from CODEOWNERS and other reviewers.
Trigger CI tests with comments or contact authorized users to do so.
- Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

CI States

Latest PR Test (Base): ❌ Run #32994620403
Latest PR Test (Extra): ❌ Run #32994619524
Latest PR Test (AMD ROCm 7.2): ❌ Run #32994620722

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ebe3a651e0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c14cd5e650

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1ec852ea9a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 00ca197bd8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7224bcaf12

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 884beac524

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-08-26T13:45:00Z

+                errors = validator.iter_errors(candidate_arguments)
+                if all(error.validator == "required" for error in errors):
+                    candidates.append((candidate_properties, candidate_arguments))


Reject incomplete branches once arguments are complete

When alternatives are distinguished by required properties rather than a discriminator, this accepts an invalid branch even after the complete arguments have parsed. For example, with one branch requiring an object-valued payload and another requiring query, an input containing only payload validates the first candidate after coercion, while the second candidate is also retained because its sole error is required; since two candidates remain, payload is returned as a string. Ignore required-only errors only while complete is false, and require full validation at completion.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-08-26T13:45:00Z

+                        candidate_schema = base_schema | {
+                            "allOf": [*base_schema["allOf"], branch]
+                        }


Keep unevaluatedProperties adjacent to the selected branch

For a Draft 2020-12 schema with root oneOf/anyOf and unevaluatedProperties: false, moving the root remainder and selected branch into separate allOf subschemas changes validation semantics: the unevaluatedProperties keyword in the remainder cannot see properties evaluated by the sibling branch, so a valid branch instance is rejected as containing unevaluated fields. No candidate is then selected, leaving branch-specific object or array arguments in their projected string form. Construct the candidate so the selected branch remains adjacent to root-level unevaluatedProperties.

Useful? React with 👍 / 👎.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3018725ef9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-08-26T17:33:44Z

+                elif current_schema != candidate_schema:
+                    properties[name] = {}


Preserve common types across non-identical branch schemas

When every anyOf/oneOf branch declares the same argument as an object or array but applies different nested constraints, the schemas are unequal and this collapses the property hint to {}. Detectors such as Qwen then infer string, so a valid nested argument remains JSON-encoded text even though every alternative agrees on its container type—a common case for discriminated unions with branch-specific payload fields. Combine the candidate schemas under the relevant combinator, or at least preserve their common inferred type, instead of requiring exact schema equality.

Useful? React with 👍 / 👎.

nvpohanh · 2026-08-27T06:25:45Z

@JustinTong0323 Could you help to review this? Thanks!

xeophon · 2026-08-28T11:35:18Z

fixed with #36626

fix: support root JSON Schema combinators in tool parsers

ebe3a65

xeophon requested a review from JustinTong0323 as a code owner August 26, 2026 10:40

chatgpt-codex-connector Bot reviewed Aug 26, 2026

View reviewed changes

Comment thread python/sglang/srt/function_call/utils.py Outdated

fix: resolve local refs in root schema combinators

c14cd5e

chatgpt-codex-connector Bot reviewed Aug 26, 2026

View reviewed changes

Comment thread python/sglang/srt/function_call/utils.py Outdated

fix: resolve refs in projected property schemas

1ec852e

chatgpt-codex-connector Bot reviewed Aug 26, 2026

View reviewed changes

Comment thread python/sglang/srt/function_call/utils.py

Comment thread python/sglang/srt/function_call/utils.py Outdated

fix: preserve unconstrained tool argument strings

00ca197

chatgpt-codex-connector Bot reviewed Aug 26, 2026

View reviewed changes

Comment thread python/sglang/srt/function_call/function_call_parser.py Outdated

Comment thread python/sglang/srt/function_call/function_call_parser.py Outdated

fix: preserve root schema branch streaming

7224bca

chatgpt-codex-connector Bot reviewed Aug 26, 2026

View reviewed changes

Comment thread python/sglang/srt/function_call/function_call_parser.py Outdated

fix: find alternatives beneath allOf

884beac

chatgpt-codex-connector Bot reviewed Aug 26, 2026

View reviewed changes

fix: project root combinator properties for tool parsers

3018725

chatgpt-codex-connector Bot reviewed Aug 26, 2026

View reviewed changes

nvpohanh mentioned this pull request Aug 27, 2026

[Bug] MiniMax-M3 tool-call parser: top-level oneOf parameters schema not resolved (numbers become strings, arrays leak raw tags) #32286

Open

2 tasks done

xeophon closed this Aug 28, 2026

		elif current_schema != candidate_schema:
		properties[name] = {}

Conversation

xeophon commented Aug 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Modifications

Accuracy Tests

Speed Tests and Profiling

Validation

Checklist

Review and Merge Process

CI States

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

nvpohanh commented Aug 27, 2026

Uh oh!

xeophon commented Aug 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

xeophon commented Aug 26, 2026 •

edited

Loading