From aec7a12f33b783931b407876bc382e432dc44861 Mon Sep 17 00:00:00 2001
From: xiaoyuyu6420 <93528429+xiaoyuyu6420@users.noreply.github.com>
Date: Mon, 31 Aug 2026 11:07:48 +0800
Subject: [PATCH 1/2] fix(provider): exclude cached tokens from Gemini input
 usage accounting

Gemini's prompt_token_count already includes tokens served from cache.
_extract_usage was passing it straight to TokenUsage.input_other while also
setting input_cached to cached_content_token_count, so usage.input double
counted cached tokens and inflated context-occupancy stats. Subtract cached
tokens from input_other, matching the OpenAI provider's accounting.
---
 .../core/provider/sources/gemini_source.py    | 13 +++++--
 tests/test_gemini_source.py                   | 36 +++++++++++++++++++
 2 files changed, 46 insertions(+), 3 deletions(-)

diff --git a/astrbot/core/provider/sources/gemini_source.py b/astrbot/core/provider/sources/gemini_source.py
index f4cba79805..93a775636e 100644
--- a/astrbot/core/provider/sources/gemini_source.py
+++ b/astrbot/core/provider/sources/gemini_source.py
@@ -451,10 +451,17 @@ def _extract_reasoning_content(self, candidate: types.Candidate) -> str:
     def _extract_usage(
         self, usage_metadata: types.GenerateContentResponseUsageMetadata
     ) -> TokenUsage:
-        """Extract usage from candidate"""
+        """Extract usage from candidate.
+
+        `prompt_token_count` includes tokens served from cache, so subtract
+        `cached_content_token_count` to avoid double-counting cached input
+        (matching the OpenAI provider's TokenUsage accounting).
+        """
+        prompt_tokens = usage_metadata.prompt_token_count or 0
+        cached = usage_metadata.cached_content_token_count or 0
         return TokenUsage(
-            input_other=usage_metadata.prompt_token_count or 0,
-            input_cached=usage_metadata.cached_content_token_count or 0,
+            input_other=prompt_tokens - cached,
+            input_cached=cached,
             output=usage_metadata.candidates_token_count or 0,
         )
 
diff --git a/tests/test_gemini_source.py b/tests/test_gemini_source.py
index 4c33de6bca..d33dafd449 100644
--- a/tests/test_gemini_source.py
+++ b/tests/test_gemini_source.py
@@ -131,6 +131,42 @@ def test_gemini_reasoning_only_output_is_allowed():
     )
 
 
+def test_gemini_extract_usage_excludes_cached_tokens_from_input_other():
+    provider = ProviderGoogleGenAI.__new__(ProviderGoogleGenAI)
+
+    usage_metadata = SimpleNamespace(
+        prompt_token_count=100,
+        cached_content_token_count=30,
+        candidates_token_count=50,
+    )
+
+    usage = provider._extract_usage(usage_metadata)
+
+    # prompt_token_count already includes cached tokens; input_other must
+    # exclude them so input (input_other + input_cached) is not inflated.
+    assert usage.input_other == 70
+    assert usage.input_cached == 30
+    assert usage.input == 100
+    assert usage.output == 50
+
+
+def test_gemini_extract_usage_without_cache_keeps_full_prompt_tokens():
+    provider = ProviderGoogleGenAI.__new__(ProviderGoogleGenAI)
+
+    usage_metadata = SimpleNamespace(
+        prompt_token_count=100,
+        cached_content_token_count=0,
+        candidates_token_count=20,
+    )
+
+    usage = provider._extract_usage(usage_metadata)
+
+    assert usage.input_other == 100
+    assert usage.input_cached == 0
+    assert usage.input == 100
+    assert usage.output == 20
+
+
 @pytest.mark.asyncio
 async def test_gemini_get_models_retries_transient_request_error(monkeypatch):
     monkeypatch.setattr(request_retry, "REQUEST_RETRY_WAIT_MIN_S", 0)

From 5a6be73353901e310e97399fb70fdbda993c7de1 Mon Sep 17 00:00:00 2001
From: xiaoyuyu6420 <93528429+xiaoyuyu6420@users.noreply.github.com>
Date: Tue, 1 Sep 2026 08:52:43 +0800
Subject: [PATCH 2/2] docs(provider): clarify _extract_usage reads response
 usage metadata

Address Sourcery review: the method receives GenerateContentResponseUsageMetadata,
not a candidate, so correct the docstring's opening line.
---
 astrbot/core/provider/sources/gemini_source.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/astrbot/core/provider/sources/gemini_source.py b/astrbot/core/provider/sources/gemini_source.py
index 93a775636e..d8ec6dbffc 100644
--- a/astrbot/core/provider/sources/gemini_source.py
+++ b/astrbot/core/provider/sources/gemini_source.py
@@ -451,7 +451,7 @@ def _extract_reasoning_content(self, candidate: types.Candidate) -> str:
     def _extract_usage(
         self, usage_metadata: types.GenerateContentResponseUsageMetadata
     ) -> TokenUsage:
-        """Extract usage from candidate.
+        """Extract usage from response metadata.
 
         `prompt_token_count` includes tokens served from cache, so subtract
         `cached_content_token_count` to avoid double-counting cached input