Flush dropped reasoning at stream end when stream_reasoning=False

[JustinTong0323]

Motivation

Follow-up to #31787, which fixed dropped reasoning at stream end for InklingDetector. As noted in the review of that PR (#31787 (comment)), the base class BaseReasoningFormatDetector has the same bug.

With stream_reasoning=False, parse_streaming_increment accumulates the whole thinking block in self._buffer and only emits it when the end token (</think>) arrives. If generation is cut short before the end token (e.g. max_tokens mid-thinking), the stream ends with the trace still stuck in _buffer. finish() only flushed it when force_nonempty_content=True; otherwise the buffered reasoning was dropped entirely.

This affects every detector that uses the base streaming path — DeepSeekR1Detector, Qwen3Detector (and its subclasses), Glm45Detector, KimiK2Detector, MistralDetector, Nemotron3Detector, MiniMaxM3Detector, etc. — for streaming requests with stream_reasoning=False (the value the scheduler and responses API use).

Modifications

• BaseReasoningFormatDetector.finish(): flush the buffered trace when the stream ends mid-reasoning. force_nonempty_content still reclassifies it as normal text (unchanged); otherwise it is now emitted as reasoning_text instead of being dropped, matching the non-streaming detect_and_parse path. The leading think-start token is stripped in both cases (extracted into a small _strip_leading_think_start helper).
• CohereCommand4Detector.finish() (new override): this detector keeps _in_reasoning pinned True for its whole run and tracks phase via _reasoning_done, so the base heuristic (keyed on _in_reasoning) would misfile a truncated answer tail as reasoning. It now flushes the buffered tail to the bucket its phase implies — reasoning while still thinking, normal text once the answer block has started — which also recovers the answer tail when a stream stops without <|END_TEXT|>.

I audited the other detectors that override the streaming state machine:

• InklingDetector already overrides finish() (Fix dropped Inkling reasoning at stream end #31787) — unaffected.
• GptOssDetector / MiniMaxAppendThinkDetector never write self._buffer, so the base flush is a no-op for them.
• Apertus2509Detector maintains _in_reasoning honestly, so the base flush correctly recovers its buffered reasoning.

Accuracy Tests

N/A — pure Python parser change, no model forward / kernel code touched.

Speed Tests and Profiling

N/A.

Tests

Added regression cases to test/registered/unit/parser/test_reasoning_parser.py:

• base detector flushes truncated reasoning at finish() with stream_reasoning=False (and strips the opening <think>);
• stream_reasoning=True only drops a lingering partial end-tag fragment, never re-emits streamed reasoning;
• DeepSeekR1Detector forced reasoning (no <think> start token) flushes on truncation;
• CohereCommand4Detector flushes a truncated thinking block as reasoning, and a truncated answer tail as normal text.

Verified each new case fails on the pre-fix code and passes after the fix; the pre-existing force_nonempty_content cases are unchanged.

Checklist

• Format your code according to pre-commit.
• Add unit tests.

---

CI States

Latest PR Test (Base): ✅ Run #30155713666
Latest PR Test (Extra): ❌ Run #30155713585

[gemini-code-assist]

Caution

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

[JustinTong0323]

/tag-and-rerun-ci

[JustinTong0323]

/rerun-failed-ci

[JustinTong0323]

/rerun-failed-ci

[JustinTong0323]

/rerun-failed-ci