LGTM, I installed this, and it works with Gemma 4. I don't believe there is more testing required.
Thank you, these models are dam good. Pretty much at GPT-5.4-mini level and smarter than Gemini-3-flash.
hoping this gets merged soon

Sometimes the model leaks thought tags or internal reasoning into the final output.

Good fix! I hit this same issue with Gemma 4 31B via the Gemini API and applied the same patch locally.

One thing I noticed: _extract_reasoning() (around line 1721) also has a hardcoded list of inline patterns used to capture reasoning content when no structured reasoning fields are present:

inline_patterns = (
    r"<think>(.*?)</think>",
    r"<thinking>(.*?)</thinking>",
    r"<reasoning>(.*?)</reasoning>",
    r"<REASONING_SCRATCHPAD>(.*?)</REASONING_SCRATCHPAD>",
)

<thought> is missing here too. Without it, Gemma 4 reasoning gets silently discarded instead of being routed to the reasoning display (when show_reasoning: true). The fix in _strip_think_blocks prevents the leak, but the reasoning data is lost instead of properly captured.

Suggested addition:

inline_patterns = (
    r"<think>(.*?)</think>",
    r"<thinking>(.*?)</thinking>",
    r"<thought>(.*?)</thought>",      # Gemma 4
    r"<reasoning>(.*?)</reasoning>",
    r"<REASONING_SCRATCHPAD>(.*?)</REASONING_SCRATCHPAD>",
)

Probably has to do with what @lhear reported.

Merged via PR #8562. Your commits were cherry-picked onto current main with authorship preserved. Added the missing auxiliary_client.py fix and tests on top. Thanks @Unayung!