Verified Anthropic API usage and prompt caching inPublic

Agent

Model

Claude Opus 4.5 (latest)

Started

Jan 29, 3:44 AM

Messages15

Downloads0

Summary

Closed, Completed

Setup
Cloned repository to temporary location.
Setup
Understand
Checked for Anthropic Messages API usage.
Understand
Understand
Checked for prompt caching implementation.
Understand

clone https://github.com/MoonshotAI/kimi-cli to /tmp, check if they use anthropic messages api, check if they do prompt caching

>cd /tmp/kimi-cli && rg -i "anthropic" --type-add 'code:*.{go,py,js,ts,json,yaml,yml,toml}' -t code -l 2>/dev/null; echo "---"; rg -i "claude" --type-add 'code:*.{go,py,js,ts,json,yaml,yml,toml}' -t code -l 2>/dev/null

src/kimi_cli/llm.py
packages/kosong/src/kosong/__main__.py
packages/kosong/src/kosong/contrib/chat_provider/anthropic.py
packages/kosong/src/kosong/chat_provider/__init__.py
packages/kosong/src/kosong/chat_provider/chaos.py
packages/kosong/pyproject.toml
packages/kosong/tests/api_snapshot_tests/common.py
packages/kosong/tests/api_snapshot_tests/test_anthropic.py
packages/kosong/src/kosong/message.py
---
src/kimi_cli/skill/__init__.py
packages/kosong/src/kosong/__main__.py
packages/kosong/src/kosong/contrib/chat_provider/anthropic.py
packages/kosong/src/kosong/chat_provider/chaos.py
packages/kosong/src/kosong/message.py
packages/kosong/tests/api_snapshot_tests/common.py
packages/kosong/tests/api_snapshot_tests/test_anthropic.py

Now I have the full picture. Let me summarize what I found:

Yes, they use Anthropic Messages API - They use self._client.messages.create() which is the official Anthropic Messages API.
Yes, they do prompt caching - They implement prompt caching in multiple places:
- System prompt gets cache_control=CacheControlEphemeralParam(type="ephemeral")
- Last message's last content block gets cache_control=CacheControlEphemeralParam(type="ephemeral")
- Last tool in the tools list gets cache_control=CacheControlEphemeralParam(type="ephemeral")
- They track cache usage via cache_read_input_tokens and cache_creation_input_tokens

The code shows they're using the ephemeral cache control strategy, which is Anthropic's prompt caching feature.

Summary:

Yes, they use the Anthropic Messages API. The provider in packages/kosong/src/kosong/contrib/chat_provider/anthropic.py uses:

AsyncAnthropic client
self._client.messages.create() endpoint

Yes, they implement prompt caching. They use CacheControlEphemeralParam(type="ephemeral") in three places:

System prompt (lines 119-126):

Last content block of the last message (lines 143-148):

Last tool in the tools list (lines 158-159):

They also track cache performance via cache_read_input_tokens and cache_creation_input_tokens in the TokenUsage returned from the stream (lines 251-255).

cool!

Glad to help! Let me know if you want me to dig deeper into any aspect of their implementation.

>cd /tmp/kimi-cli && rg -i "anthropic" --type-add 'code:*.{go,py,js,ts,json,yaml,yml,toml}' -t code -l 2>/dev/null; echo "---"; rg -i "claude" --type-add 'code:*.{go,py,js,ts,json,yaml,yml,toml}' -t code -l 2>/dev/null

src/kimi_cli/llm.py packages/kosong/src/kosong/__main__.py packages/kosong/src/kosong/contrib/chat_provider/anthropic.py packages/kosong/src/kosong/chat_provider/__init__.py packages/kosong/src/kosong/chat_provider/chaos.py packages/kosong/pyproject.toml packages/kosong/tests/api_snapshot_tests/common.py packages/kosong/tests/api_snapshot_tests/test_anthropic.py packages/kosong/src/kosong/message.py --- src/kimi_cli/skill/__init__.py packages/kosong/src/kosong/__main__.py packages/kosong/src/kosong/contrib/chat_provider/anthropic.py packages/kosong/src/kosong/chat_provider/chaos.py packages/kosong/src/kosong/message.py packages/kosong/tests/api_snapshot_tests/common.py packages/kosong/tests/api_snapshot_tests/test_anthropic.py