Support namespaced function tools in the Responses API

[PierreLeGuen]

Motivation

Responses clients such as Codex group client-executed functions under type="namespace". SGLang accepts these declarations but omits their functions from the chat template, and replay can discard the namespace. This prevents a stateless tool-call loop.

Related gateway protocol fixes: smg-project/smg#2451

Modifications

• Flatten namespace members into qualified model-facing names while preserving each function's schema, description, and strictness.
• Restore separate name and namespace fields in non-streaming calls and streaming item events, and preserve them on stateless replay.
• Support named namespace tool selection, reject ambiguous qualified names, and validate namespace membership shape. Custom tools and hosted tools retain their existing behavior.
• Add regression coverage for duplicate member names in different namespaces, dotted top-level names, collisions, forced selection, streaming, and string/content-array tool results.

Accuracy Tests

No kernel or model-forward changes. Responses unit tests on the patched serving runtime: 50 passed, 4 subtests passed.

cd test/registered/unit/entrypoints/openai
python3 -m pytest -q test_response_namespace_tools.py test_serving_responses.py test_serving_responses_stream.py

Live GLM-5.3-Flash checks passed in both streaming and non-streaming modes: force weather.lookup, verify namespace="weather" and the JSON arguments, then replay the returned call with a content-array tool result and store=false (no previous_response_id). The next response correctly uses the supplied result. Codex CLI 0.153.0 also completed a namespaced MCP call, replayed its result, inspected and edited a Python file through the shell, passed all three fixture tests, and returned the exact synthetic marker from the MCP result.

Speed Tests and Profiling

No inference kernel changes; no throughput claim. The change adds function declaration/name translation at the API boundary.

Checklist

• Ruff lint/format, isort, codespell, and git diff --check pass for changed files.
• Added CPU unit tests registered with CI.
• Followed existing Responses serving/test conventions.
• Documentation and model accuracy/speed benchmarks: not applicable to the protocol translation tested here.

---

CI States

Latest PR Test (Base): ❌ Run #34147251332
Latest PR Test (Extra): ❌ Run #34147251186
Latest PR Test (AMD ROCm 7.2): ❌ Run #34147251277

[chatgpt-codex-connector]

Codex Review Summary

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-07T17:25:46.573560Z	02d84e6	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.

[PierreLeGuen]

@JustinTong0323 Could you review this and enable/rerun the required CI? The Base, Extra, and AMD workflows currently stop at the missing run-ci label before tests execute; lint passes. The PR includes regression coverage and validation details, and the companion gateway change is smg-project/smg#2451. Thanks!