[Bugfix] Default Qwen3 reasoning parser to prompt-has-open-think

[rivetphilbot]

Summary

The standard Qwen3 chat template injects <think>\n into the assistant turn opener whenever enable_thinking is not explicitly False. That means completion tokens only contain </think> followed by the answer — never an opening <think>.

The current parser default (prompt_has_open_think=False) forces every client to pass chat_template_kwargs={"enable_thinking": True} on each request. When they don't (the common case for OpenAI-compatible clients), the parser sees no opening <think>, falls through to the both-tags-required branch, and dumps the entire completion — reasoning prose, the close tag, and the answer — into content with reasoning left null.

Flip the default to True so the parser matches the chat template's actual behavior out of the box. An explicit enable_thinking=False still works correctly: in that case the template emits a closed <think></think> pair into the prompt, so neither tag appears in the completion and the existing no-prompt-open-think branch is the right code path.

Reproduction

Live on V100 TP=2 with a Qwen3-derived model (deckard-40b-w4a16-hermes), serving config:

--reasoning-parser qwen3 --tool-call-parser qwen3_coder --enable-auto-tool-choice

Before (no chat_template_kwargs):

reasoning: null
content: "\n\n # Train Meeting Problem\n## Step 1: ...\n## Final Answer\nThey meet ..."

The CoT framing leaks into content because the parser couldn't find an opening <think>.

With explicit chat_template_kwargs={"enable_thinking": true} (workaround):

reasoning: "This is a classic relative speed problem. Let me work through it..."
content: "\n\n # Train Meeting Problem\n..."

After this patch (no chat_template_kwargs):

reasoning: "This is a classic relative speed problem. Let me work through it..."  (1377 chars)
content: "\n\n # Train Meeting Problem\n..."  (clean)

Test plan

• Live A/B on V100 TP=2, deckard-40b-w4a16-hermes, before/with-kwargs/after — see reproduction above
• enable_thinking=False path: still hits the no-prompt-open-think branch (template emits closed <think></think>, neither tag in completion)
• Streaming path verification (uses same flag)

Co-Authored-By: RivetOS Claude noreply@rivetos.dev