fix: Add parity in behavior and tests with vllm nemotron_v3 for tools parser

[indrajit96]

Overview:

Port vLLM Nemotron v3 reasoning parser coverage into Dynamo and align Dynamo’s parser naming/disabled-thinking behavior with the vLLM contract.

Why:

vLLM added Nemotron v3 reasoning parser with tests.
Dynamo has a separate Rust implementation of reasoning parsing, so this PR adds matching coverage to make the expected Nemotron v3 behavior explicit and trackable.

vLLM source test file:
https://github.com/vllm-project/vllm/blob/main/tests/reasoning/test_nemotron_v3_reasoning_parser.py

Missing relevant vLLM test cases:

• test_nemotron_v3_reasoning
  • without_start_token
  • without_start_token_streaming
  • with_start_token
  • with_start_token_streaming
• test_nemotron_v3_without_thinking_returns_content
• test_nemotron_v3_force_nonempty_content_returns_content

Details:

• Added nemotron_v3 as a reasoning parser alias for the existing force-reasoning ... parser.

• Added direct vLLM parity tests for Nemotron v3 reasoning extraction:

  • with and without opening
  • non-streaming and streaming paths

• Added postprocessor coverage for vLLM request-flag behavior:

  • enable_thinking=false returns normal content
  • force_nonempty_content=true strips a leading and returns normal content

• Documented nemotron_v3 as the vLLM-compatible alias.

  Newly added vLLM-to-Dynamo test mapping:

  vLLM test case	Newly added Dynamo coverage
  without_start_token	test_nemotron_v3_detect_and_parse_vllm_cases
  with_start_token	test_nemotron_v3_detect_and_parse_vllm_cases
  without_start_token_streaming	test_nemotron_v3_streaming_vllm_cases
  with_start_token_streaming	test_nemotron_v3_streaming_vllm_cases
  test_nemotron_v3_without_thinking_returns_content	postprocessor_parsing_stream_nemotron_v3_enable_thinking_false_returns_content
  test_nemotron_v3_force_nonempty_content_returns_content	postprocessor_parsing_stream_nemotron_v3_force_nonempty_strips_start_token

Additional Dynamo-only regression coverage:

Scenario	Newly added Dynamo coverage
Keep disabled-reasoning <think> strip state independent per streamed choice	postprocessor_parsing_stream_nemotron_v3_force_nonempty_tracks_prefix_per_choice

Where should the reviewer start?

• lib/parsers/src/reasoning/mod.rs
  • parser alias registration
  • vLLM parity tests
• lib/llm/src/preprocessor.rs
  • request-flag handling for nemotron_v3
  • leading <think> stripping when reasoning is disabled
• lib/llm/tests/postprocessor_parsing_stream.rs
  • postprocessor tests for enable_thinking=false and force_nonempty_content=true
• docs/agents/reasoning.md
  • parser-name documentation update

---

Summary by CodeRabbit

• New Features

  • Added support for nemotron_v3 as an alternative name for Nemotron reasoning parsers
  • Implemented per-request reasoning extraction control with improved handling of reasoning markers

• Documentation

  • Updated supported reasoning parsers reference to reflect upstream vLLM mapping for Nemotron

[coderabbitai]

Walkthrough

The PR adds support for disabling reasoning extraction in vLLM Nemotron v3 responses when specific chat template arguments are set. It introduces a new parser alias (nemotron_v3), implements reasoning-stripping logic for streams, and adds corresponding test coverage.

Changes

Nemotron v3 Reasoning Disabling

Layer / File(s)	Summary
Parser Registry lib/parsers/src/reasoning/mod.rs	New nemotron_v3 alias is registered, mapping to ReasoningParserType::DeepseekR1. Test expectations are updated to include the new alias, and old deepseek_v4 streaming test is removed. Two new nemotron_v3 streaming and detection tests are added.
Reasoning Disable Logic lib/llm/src/preprocessor.rs	New reasoning_disabled_by_request flag is computed from parser name and chat_template_args. New helper should_strip_reasoning_start_when_disabled() identifies Nemotron v3-family parsers requiring stripping. New stream transformer strip_leading_reasoning_start_from_stream() robustly buffers and removes leading <think> markers across chunk boundaries.
Preprocessing Integration lib/llm/src/preprocessor.rs	postprocessor_parsing_stream is refactored to gate reasoning extraction by the new disable flag. When disabled for v3 parsers, the stream is routed through the stripping transformer instead of direct reasoning parsing. Test setup adds force_nonempty_content_true argument map and extends test cases with v3 aliases and disabled-reasoning expectations.
Stream Processing Tests lib/llm/tests/postprocessor_parsing_stream.rs	Two new Tokio tests verify Nemotron v3 behavior: one with enable_thinking=false producing empty reasoning, another with force_nonempty_content=true stripping a leading <think> token split across chunk boundaries.
Documentation docs/agents/reasoning.md	Table entry for nemotron3 is updated to specify vLLM upstream name (nemotron_v3) and clarify that the parser accepts both aliases.

---

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately describes the main change: adding vLLM nemotron_v3 parity to Dynamo's reasoning parser, which is the primary focus of all file changes.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The pull request description is comprehensive and well-structured, covering all required sections of the template with detailed context about objectives, changes, test mappings, and specific files for reviewer focus.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

Tip

💬 Introducing Slack Agent: Turn conversations into code.

Slack Agent is built on CodeRabbit's deep understanding of your code, so your team can collaborate across the entire SDLC without losing context.

• Generate code and open pull requests
• Plan features and break down work
• Investigate incidents and troubleshoot customer tickets together
• Automate recurring tasks and respond to alerts with triggers
• Summarize progress and report instantly

Built for teams:

• Shared memory across your entire org—no repeating context
• Per-thread sandboxes to safely plan and execute work
• Governance built-in—scoped access, auditability, and budget controls

One agent for your entire SDLC. Right inside Slack.

👉 Get your free trial and get 200 agent minutes per Slack user (a $50 value).

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@lib/llm/src/preprocessor.rs`:
- Around line 1494-1524: The stream-end path currently discards any bytes left
in state.buffer (so a partial think prefix like "<thi" is lost); update the
branch that returns None at stream completion to check state.buffer and, if
non-empty, emit it as normal content instead of dropping it — e.g. construct an
output string from state.buffer (taking into account stripping think_start_token
if appropriate, using state.think_start_token and the same logic used earlier)
and set choice.delta.content = Some(ChatCompletionMessageContent::Text(...))
before returning the processed_response/state; ensure state.buffer is cleared
afterwards and that existing fields (state.decided, choice.delta.content) are
handled consistently with the other branch.
- Around line 1457-1516: The current strip_leading_reasoning_start_from_stream
uses a single StripReasoningStartState with shared buffer and decided fields,
which incorrectly mixes state across multiple choices in a single streamed
response; update StripReasoningStartState to track per-choice state keyed by
choice.index (e.g., a HashMap<usize, (buffer: String, decided: bool)>), and then
in the processing closure use the choice.index to lookup/create the per-choice
buffer/decided pair, operate on that pair when accumulating/checking
think_start_token, write back the possibly-trimmed output into
choice.delta.content, and remove the per-choice entry when you know the choice
is finished to avoid unbounded growth. Ensure you reference
strip_leading_reasoning_start_from_stream, StripReasoningStartState,
choice.index, state.buffer and state.decided in your changes.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 0a851996-4dea-498b-9735-dd507f91f757

📥 Commits

Reviewing files that changed from the base of the PR and between 297119e and d49d5de.

📒 Files selected for processing (4)

• docs/agents/reasoning.md
• lib/llm/src/preprocessor.rs
• lib/llm/tests/postprocessor_parsing_stream.rs
• lib/parsers/src/reasoning/mod.rs

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-7d4ab284-03bd-4b7f-b8e7-f8dc72cf80d4.docs.buildwithfern.com/dynamo/dev

[keivenchang]

thanks Indrajit — helper extraction, new PRE.6 doc, and the PARSER_CASES rename all landed cleanly. ship it.