fix(provider): exclude cached tokens from Gemini input usage accounting

[xiaoyuyu6420]

Problem

ProviderGoogleGenAI._extract_usage builds TokenUsage by passing Gemini's prompt_token_count straight into input_other while also setting input_cached to cached_content_token_count.

Per the Gemini API docs, prompt_token_count already includes the tokens served from cache. So TokenUsage.input (input_other + input_cached) double-counts cached tokens, inflating context-occupancy stats (e.g. current_context_tokens in the tool-loop runner).

The OpenAI provider handles this correctly (input_other = prompt_tokens - cached, see openai_source.py), so the two providers are inconsistent.

Changes

• astrbot/core/provider/sources/gemini_source.py: subtract cached_content_token_count from prompt_token_count when building input_other, matching the OpenAI provider's accounting.
• tests/test_gemini_source.py: add two unit tests covering usage with and without cached tokens.

Verification

$ uv run pytest tests/test_gemini_source.py tests/unit/test_provider_stats.py -q
10 passed

Summary by Sourcery

Prevent cached Gemini prompt tokens from inflating reported input usage.

Bug Fixes:

• Correct Gemini input token usage accounting so cached prompt tokens are not double-counted.

Tests:

• Add coverage for Gemini usage accounting with and without cached tokens.

[sourcery-ai]

Hey - I've found 1 issue

Prompt for AI Agents

Please address the comments from this code review:

## Individual Comments

### Comment 1
<location path="astrbot/core/provider/sources/gemini_source.py" line_range="454" />
<code_context>
         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
</code_context>
<issue_to_address>
**nitpick:** The expanded docstring still says the method extracts usage from a candidate, but `_extract_usage` accepts and reads `GenerateContentResponseUsageMetadata`; callers pass `result.usage_metadata` or `chunk.usage_metadata`. The description therefore misstates the method's input and can mislead maintainers about its contract.

**Suggested fix:** Change the first line to `Extract usage from response metadata.`

```suggestion
        """Extract usage from response metadata.
```
</issue_to_address>

Sourcery assessment

Needs a human reviewer. This changes the token counts used for Gemini usage and potentially cost or billing calculations; if the API's cached count is not a subset of the prompt count, it could record incorrect or negative non-cached usage. Reverting restores the old calculation, but any already-recorded usage or charges would need to be recomputed or corrected.

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

_{Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.}