[original4422]

Summary

Preserve literal null, NULL and mixed-case null text when Qwen3-Coder tool arguments explicitly have a string schema. The converter currently handles null before checking that schema, so both nonstream and streaming calls produce JSON null instead of the requested string.

 explicit string argument: <parameter=value>null</parameter>
- {"value": null}
+ {"value": "null"}

Keep existing null conversion for nullable, nonstring and unspecified schemas. Add public parser regressions for nonstream and aggregated streaming at several chunk boundaries, alongside ordinary-value controls.

Evidence

• Before: the new regression file reports 12 failures and 68 passes on current main.
  After: all 80 cases pass; four existing Qwen3-Coder parser/CLI tests also pass.
• Repository-pinned Black 21.12b0 and Flake8 6.1.0 checks pass directly for both changed files with the configured arguments. The configured pre-commit bootstrap timed out twice while installing its Black environment, before running hooks; the complete pre-commit pipeline is not claimed as passed.
• Tested with the original package on macOS ARM64 / Python 3.12.14 using isolated CPU dependencies (torch 2.13.0, pydantic 2.14.3). This is not the full inference requirements environment. Two unrelated original API/SSE tests require missing uvloop; no GPU, model, provider or full server suite was run.

AI-assisted with OpenAI Codex; reproduction, implementation and review were performed by the authorized coding agent. Existing parser attribution is unchanged.

Merge Danger

Door: two-way

Blast Radius: parser

Only explicitly string-typed Qwen3-Coder null-like arguments change; nullable and unspecified null handling is covered by controls.

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-14T17:06:07.320140Z	1ef7361	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.