fix(reasoning/qwen3): route bare-text 'thinking process:' preamble to reasoning_content (#570)

[raullenchai]

Summary

Closes #570.

Qwen3ReasoningParser now recognises bare-text "thinking process" preambles at the start of an assistant turn and routes them to reasoning_content instead of letting them leak into the user-facing content field.

Root cause

Qwen3 chat templates inject <think>\n after the assistant generation marker when enable_thinking=True, putting the model in implicit-think mode — the model is supposed to emit its chain-of-thought followed by </think> and then the user-facing answer.

In practice the model occasionally restates the channel boundary inline with a bare-text preamble like:

Here's a thinking process:

1.  **Analyze User Input:**
   - **Route:** Seattle to San Diego
2.  **Evaluate Each Option:**
   - Portland, OR: World-renowned food scene.

When the request also runs out of max_tokens before the model reaches </think>, neither tag appears anywhere in the model output. The previous Qwen3ReasoningParser.extract_reasoning "no end token, no start token" branch then returned (None, model_output) — routing the whole chain-of-thought into the user-facing content field while reasoning_content stayed empty. Any OpenAI-compatible client consuming /v1/chat/completions saw raw reasoning leak into the answer.

The bug reproduces both on multi-turn and single-turn prompts; the trigger is "model emits bare-text thinking preamble AND output is truncated before </think>", which becomes more likely as conversation length and max_tokens interact.

Fix

Extend Qwen3ReasoningParser to recognise bare-text "thinking process" preambles and route them to reasoning_content with cleaned_text cleared. Applies to both non-streaming extract_reasoning and the finalize_streaming correction path — matching the existing streaming heuristic in BaseThinkingReasoningParser that already routes pre-tag text to reasoning while </think> has not yet been seen.

The regex is scoped narrowly to explicit scratchpad labels — preambles that end with : and use known scratchpad nouns. Conversational openers (Let me think..., I need to analyze..., I'll analyze...) are deliberately NOT matched: they are common direct-answer phrasings when enable_thinking=False, and matching them would blank out valid responses.

Recognised preambles (matched at ^\s*, case-insensitive, all require terminating :):

• Here's a thinking process:, Here is my reasoning:, Here's the chain-of-thought:, Here is the scratchpad:, Here's the thought process:, Here is the reasoning process:
• Thinking step by step:, Thinking out loud:, Thinking through this:, Thinking carefully:, Thinking aloud:
• Step by step: / Step-by-step:
• My thought process: / My reasoning process:

Why server-side is the right layer

Two alternative paths were considered:

1. Patch the chat template to keep <think> open across continuation turns. The chat template is owned by the upstream HF tokenizer config (Qwen3.6-A3B), so we'd have to maintain a local override; any HF refresh would silently revert the fix. Captured for future upstream contribution; not blocked on it.
2. Inject <think> server-side at the prompt boundary. Would mask the symptom but bare-text preambles also occur in single-turn prompts (the trigger is "truncated before </think>", not "continuation turn"). Server-side parser remains the right place to be defensive.

Regression coverage

• tests/test_reasoning_parsers.py::TestQwen3 — bare-text-fallback positive cases (12 explicit-label variants), the explicit-tag passthrough, the streaming finalizer behavior, and a dedicated test_ambiguous_phrases_not_misclassified negative test that pins the conservative scope (9 conversational openers that must stay as content).
• tests/test_chat_route_tool_tag_leak.py::TestBareThinkingProcessLeakRegression — drives the real _finalize_content_and_reasoning orchestration used by /v1/chat/completions, including the exact text captured from a failing request against qwen3.6-35b-4bit.

Scope

• No version bump (parser fix; release SOP 11-gate walk happens at the next bundled cut).
• Diff limited to one parser file + two test files (+335 / -2).
• No alias changes, no template changes, no upstream pin changes.

Test plan

• python3.12 -m pytest tests/test_reasoning_parsers.py tests/test_chat_route_tool_tag_leak.py — 142 passed
• Broader sweep: python3.12 -m pytest tests/ -k "reason or qwen or think or finalize or tool_tag" --ignore=tests/integrations --ignore=tests/test_streaming_simulator.py — 1003 passed
• ruff check + ruff format --check on touched files — clean
• Direct verification on the captured failing payload via _finalize_content_and_reasoning — reasoning_content populates with the chain-of-thought, content clears

[github-actions]

PR #573 validation scorecard

Title: fix(reasoning/qwen3): route bare-text 'thinking process:' preamble to reasoning_content (#570)
Author: raullenchai
Diff: 3 file(s), +425/-2 LOC, blast radius: medium

Verdict: MERGE-SAFE

step	status	summary	time
fetch	PASS	3 files, +425/-2 LOC, blast=medium	0.9s
test_plan_check	PASS	all 4 test-plan item(s) checked	0.0s
cl_description_quality	PASS	title OK + body has rationale (4856 chars)	0.0s
codex_review	skip	codex CLI not found on PATH (install: npm i -g @openai/codex)	0.0s
supply_chain	PASS	no hooks touched, no suspicious patterns, deps clean	0.0s
lint	PASS	clean (3 file(s))	0.0s
targeted_tests	PASS	exit 2 (in 2 target file(s))	2.0s