From 8f95ca9d989187fb43476d588bf0f54cac9d1259 Mon Sep 17 00:00:00 2001
From: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Date: Fri, 3 Apr 2026 16:35:25 +0800
Subject: [PATCH] fix: filter Gemini thinking parts from user-facing message
 chain (#7196)

Gemini 3 models return thinking parts (part.thought=True) alongside the
actual response text.  _process_content_parts was including these thinking
parts in the message chain sent to the user, effectively leaking internal
reasoning into the output.  On platforms that split long messages (e.g.
aiocqhttp with realtime segmenting), this caused duplicate or triple
replies since the thinking text often mirrors the actual response.

The streaming path already handled this correctly via chunk.text which
skips thinking parts, but the non-streaming path and the final-chunk
processing in streaming both went through _process_content_parts.

Also switch the Gemini 3 model name matching from an exhaustive list to
prefix matching (gemini-3- / gemini-3.) so new variants like gemini-3.1
get proper thinkingLevel config without code changes.

Fixes #7183
---
 .../core/provider/sources/gemini_source.py    | 19 +++++++++----------
 1 file changed, 9 insertions(+), 10 deletions(-)

diff --git a/astrbot/core/provider/sources/gemini_source.py b/astrbot/core/provider/sources/gemini_source.py
index 388ab5c91..489ca358b 100644
--- a/astrbot/core/provider/sources/gemini_source.py
+++ b/astrbot/core/provider/sources/gemini_source.py
@@ -241,15 +241,10 @@ class ProviderGoogleGenAI(Provider):
                 thinking_config = types.ThinkingConfig(
                     thinking_budget=thinking_budget,
                 )
-        elif model_name in [
-            "gemini-3-pro",
-            "gemini-3-pro-preview",
-            "gemini-3-flash",
-            "gemini-3-flash-preview",
-            "gemini-3-flash-lite",
-            "gemini-3-flash-lite-preview",
-        ]:
-            # The thinkingLevel parameter, recommended for Gemini 3 models and onwards
+        elif any(model_name.startswith(p) for p in ("gemini-3-", "gemini-3.")):
+            # The thinkingLevel parameter, recommended for Gemini 3 models and onwards.
+            # Use prefix match so new variants (3.1, 3-flash-lite-preview, etc.) are
+            # covered without needing to keep an exhaustive list up to date.
             # Gemini 2.5 series models don't support thinkingLevel; use thinkingBudget instead.
             thinking_level = self.provider_config.get("gm_thinking_config", {}).get(
                 "level", "HIGH"
@@ -517,7 +512,11 @@ class ProviderGoogleGenAI(Provider):
         ):
             chain.append(Comp.Plain("这是图片"))
         for part in result_parts:
-            if part.text:
+            # Skip thinking parts — their text is already captured via
+            # _extract_reasoning_content above.  Including them here would
+            # leak the model's internal reasoning into the user-facing message,
+            # which also causes duplicate/triple replies on some platforms.
+            if part.text and not part.thought:
                 chain.append(Comp.Plain(part.text))
 
             if (