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

[slivanovich]

Overview:

Improves DeepSeek V4 reasoning parsing when a model starts emitting DSML tool-call markup before closing an open <think> block. The parser now treats the DeepSeek DSML tool section marker as a reasoning exit point, preventing tool-call markup from being counted or streamed as reasoning content.

Details:

Rust - parsers/v1/src/reasoning/mod.rs

• Added DEEPSEEK_TOOL_SECTION_BEGIN for the DeepSeek DSML tool-call marker
• Configured the DeepSeek V4 reasoning parser to force-exit reasoning when <｜DSML｜ starts
• Added regression coverage for:
  • Batch parsing where DSML starts before </think>
  • Dangling partial DSML prefixes at the end of reasoning text
  • Streaming chunks where the DSML marker is split across deltas
  • Non-DSML lone < text that should remain reasoning content

Rust - parsers/v1/src/reasoning/base_parser.rs

• Updated truncated reasoning handling to withhold suffix bytes that may still become a configured tool-start marker
• Adjusted streaming overlap handling so even a one-character tool marker prefix is buffered until the next chunk disambiguates it
• Preserves existing behavior for parsers without tool-start markers

Rust - parsers/v1/src/tool_calling/dsml/parser.rs

• Changed DSML end-position detection to return None until the closing DSML block marker arrives
• Updated tests so incomplete DSML blocks stay buffered instead of reporting chunk.len() as a completed boundary

Rust - parsers/v1/src/tool_calling/parsers.rs

• Propagated the DSML parser's Option<usize> end-position result through the generic tool parser dispatch

Where should the reviewer start?

• parsers/v1/src/reasoning/mod.rs:23 - DeepSeek DSML marker constant.
• parsers/v1/src/reasoning/mod.rs:238-242 - DeepSeek V4 reasoning parser registration with the DSML tool-start marker.
• parsers/v1/src/reasoning/base_parser.rs:231-240 and parsers/v1/src/reasoning/base_parser.rs:368-378 - Partial tool-start buffering in batch and streaming reasoning paths.
• parsers/v1/src/tool_calling/dsml/parser.rs:51-63 - DSML end-position detection returning None for incomplete blocks.
• parsers/v1/src/reasoning/mod.rs:378-438 - DeepSeek V4 regression tests covering DSML reasoning exits and split-marker streaming.

Related PRs:

• [Bug] Qwen3 reasoning detector silently swallows tool_call when </think> is missing sgl-project/sglang#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)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[rmccorm4]

Hi @slivanovich, thanks for raising this!

Two quick questions:

1. Do you have some e2e curl requests that reproduce the leak? (@KrishnanPrash @indrajit96 we need to improve coverage here)
2. If you try setting DYN_ENABLE_EXPERIMENTAL_PARSERS_V2=1 in your deployment, do you still see the leak in your setup?

[slivanovich]

Hi @slivanovich, thanks for raising this!

Two quick questions:

1. Do you have some e2e curl requests that reproduce the leak? (@KrishnanPrash @indrajit96 we need to improve coverage here)
2. If you try setting DYN_ENABLE_EXPERIMENTAL_PARSERS_V2=1 in your deployment, do you still see the leak in your setup?

Thanks for the quick response!

1. This problem seems to be probabilistic and appear while processing long context requests. So there is no exact curl that can show the bug. In out installation of DeepSeek-V4-Flash i've reproduced this on the very long context load test (N sessions up to 1M and more context).
2. No, i've not tried to use v2 parsers yet. However, as i can see (probably missing smth) there is only v2 tool parser for DeepSeek models (no reasoning parser). This fix is about "reasoning do not end while receiving tool start in it". Also the tool start token for DeepSeek models (dsml token) is not "atomic" in terms of generation process ("<", "|DSML|", ...), so part of this pr handles this particular case (cause empirically considered that .with_tool_start_token(DEEPSEEK_TOOL_SECTION_BEGIN) is not enough).

[rmccorm4]

/ok to test 6c720ca

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[github-actions]

📊 Conformance matrix rendered — view in CI summary

[slivanovich]

Have resolved @dynamo-review-agent[bot] comments

[devin-ai-integration]

Devin Review found 1 new potential issue.

[indrajit96]

CHanges look good to me, just one small request for the additional prompt-injected batch scenario test described above.
Thanks for finding and fixing this!
We also recently merged foundational Kimi K3 changes that touch the shared parser logic.
Could you merge the latest main into this branch and run the parser tests again to verify everything works together?

[devin-ai-integration]

Devin Review found 1 new potential issue.

[slivanovich]

CHanges look good to me, just one small request for the additional prompt-injected batch scenario test described above. Thanks for finding and fixing this! We also recently merged foundational Kimi K3 changes that touch the shared parser logic. Could you merge the latest main into this branch and run the parser tests again to verify everything works together?

Hi, thanks for the response!
Done: test for the tool call without opener, also add some into jail.rs part (V3.2 uses quite the same DSML notation, but anyway)

yeah, base parser fix has merged faster)

[indrajit96]

/ok to test d3f8d64