fix(provider): count Anthropic cache_creation tokens in input usage

[xiaoyuyu6420]

Problem

Anthropic's Usage.input_tokens only counts tokens after the last cache breakpoint — it excludes both cache reads and cache writes. ProviderAnthropic._extract_usage maps input_tokens into TokenUsage.input_other and cache_read_input_tokens into input_cached, which is correct for reads — but cache_creation_input_tokens (tokens written into a freshly created cache entry) are dropped entirely.

Per Anthropic's official docs (linked in the code comment):

total_input_tokens = cache_read_input_tokens + cache_creation_input_tokens + input_tokens

So on any request that creates a new cache entry (e.g. first turn after a cache ephemeral_1h breakpoint on a long system prompt), usage.input and usage.total undercount the real input — which feeds context-occupancy stats (current_context_tokens in the tool-loop runner).

The OpenAI provider handles this category correctly: input_other = prompt_tokens - cached, so input_other includes cache-write tokens. Anthropic was the only provider dropping them.

Changes

• astrbot/core/provider/sources/anthropic_source.py: fold cache_creation_input_tokens into input_other in both _extract_usage (non-streaming / message_start) and _update_usage (streaming message_delta), mirroring OpenAI's accounting where input_other holds all non-cache-read input including cache writes.
• tests/test_anthropic_source.py (new): 5 unit tests covering extract/update with and without cache fields, and omitted-field preservation in streaming deltas.

Verification

$ uv run pytest tests/test_anthropic_source.py tests/test_anthropic_kimi_code_provider.py tests/unit/test_provider_stats.py -q
38 passed

Summary by Sourcery

Ensure Anthropic cache writes are included in input token accounting so context and usage statistics remain accurate.

Bug Fixes:

• Correct Anthropic input and total token usage by including cache-creation tokens in non-cached input accounting.

Tests:

• Add unit coverage for Anthropic usage extraction and streaming updates with cache fields and omitted usage fields.

[sourcery-ai]

Hey - I've reviewed your changes and they look great!

Sourcery assessment

Approved.

---

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.}