fix(parsers): tools are leaking in the reasoning content (deepseek)

[slivanovich]

Motivation

DeepSeek V4 can start emitting DSML tool-call markup before closing an open <think> block. When that happens, the reasoning parser should stop treating subsequent text as reasoning; otherwise DSML tool-call markup may be counted or streamed as reasoning content.

This PR makes DeepSeek V4 DSML tool section markers explicit reasoning exit points.

Modifications

Python - python/sglang/srt/parser/reasoning_parser.py

• Added DeepSeekBaseDetector for DeepSeek reasoning parsing behavior with optional tool-start markers.
• Moved streaming DSML tool-start handling out of DeepSeekV4Detector and into the DeepSeek base detector.
• Configured DeepSeek V4 to exit reasoning on <｜DSML｜tool_calls> or <｜DSML｜invoke name=.
• Preserved DeepSeek V4 strict-thinking excluded tokens for constrained reasoning.
• Added streaming buffering for partial DSML tool-start prefixes split across deltas.

Python - test/registered/unit/parser/test_reasoning_parser.py

• Added regression coverage for DeepSeek V4:
  • DSML tool section markers split across streaming deltas.
  • DSML tool section markers emitted in the same delta as <think>.
  • Split DSML invoke markers.
  • Non-tool DSML-like prefixes that should remain reasoning content.
  • </think> taking priority over DSML tool-start text already inside reasoning.

Where should the reviewer start?

• python/sglang/srt/parser/reasoning_parser.py:901 - DeepSeekBaseDetector with reasoning parsing and optional tool-start handling.
• python/sglang/srt/parser/reasoning_parser.py:997-1009 - Partial tool-start buffering for streaming deltas.
• python/sglang/srt/parser/reasoning_parser.py:1025-1052 - DeepSeek V4 detector configuration with DSML tool-start markers.
• test/registered/unit/parser/test_reasoning_parser.py:185-244 - DeepSeek V4 DSML reasoning-exit regression tests.

Related PRs

• [Bug] Qwen3 reasoning detector silently swallows tool_call when </think> is missing #22837 (the same problem with Qwen3)
• [Bugfix] Handle case when kimi ends reasoning with a tool call vllm-project/vllm#33646 (the same problem with Kimi)
• fix(parsers): tools are leaking in the reasoning content (deepseek) ai-dynamo/frontend-crates#144 (the same fix in the Dynamo parsers)

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.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #31493822238
Latest PR Test (Extra): ❌ Run #31493821942

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[slivanovich]

friendly ping @JustinTong0323

[hnyls2002]

Covered by #34458 (5899674). DeepSeekV4Detector now passes tool_start_token, so a DSML block that follows reasoning without a </think> is routed to normal_text and reaches the tool-call detector instead of staying in reasoning_content.

The merged version does this by passing the argument rather than introducing a shared base class, but the behaviour this PR reported is fixed.

You are credited as a co-author on that commit.

If I misread what this PR does and part of it is still missing, please rebase on main and reopen — happy to take it.