Issue 1 feat plugin add native anthropic streaming and prompt caching support

[An-Dongsun]

Closes #1

• feat(plugin): add native Anthropic streaming and prompt caching support #1

Summary

Kong ai-proxy를 Full Fork하여 Native Anthropic Streaming과 Prompt Caching을 지원하는 ai-anthropic-native 플러그인 추가

• Native Anthropic SSE 스트리밍 지원 (OpenAI 변환 스킵)
• Prompt Caching 지원 (cache_control 자동 주입)
• 캐시 메트릭 응답 헤더 노출 (X-Cache-Creation-Input-Tokens, X-Cache-Read-Input-Tokens)
• Anthropic 직접 호출 및 AWS Bedrock 지원

Changes

영역	설명
kong/llm/	LLM 프레임워크 (드라이버, 어댑터, 필터)
kong/plugins/ai-anthropic-native/	새 플러그인 (handler, schema)
kong/llm/plugin/shared-filters/inject-cache-control.lua	Prompt caching cache_control 주입
kong/llm/plugin/shared-filters/expose-cache-metrics.lua	캐시 메트릭 헤더 노출

Configuration Example

plugin: ai-anthropic-native
config:
  model:
    provider: "bedrock"
    name: "anthropic.claude-3-5-sonnet-20241022-v2:0"
  native_streaming: true
  prompt_caching:
    enabled: true
    cache_system_prompt: true
    cache_tools: true
    ttl: "5m"
  expose_cache_metrics: true

[Copilot]

Pull request overview

This PR adds a custom Kong plugin ai-anthropic-native that enables native Anthropic SDK compatibility for AWS Bedrock Claude models. It implements native streaming in Anthropic SSE format (bypassing OpenAI conversion), prompt caching with automatic cache_control injection, and cache metrics exposure via response headers.

Changes:

• New plugin with native Anthropic streaming support (preserves Anthropic SSE format instead of converting to OpenAI)
• Prompt caching support with automatic cache_control injection into system prompts and tools
• Cache metrics exposure through custom response headers
• Extended Thinking (Claude 4) support via thinking parameter conversion
• Full LLM framework fork from Kong AI Proxy including drivers, adapters, and shared filters

Reviewed changes

Copilot reviewed 38 out of 38 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
kong/plugins/ai-anthropic-native/handler.lua	Plugin handler registering custom filters including inject-cache-control and expose-cache-metrics
kong/plugins/ai-anthropic-native/schema.lua	Plugin schema with native_streaming, prompt_caching, and expose_cache_metrics configuration
kong/llm/plugin/shared-filters/inject-cache-control.lua	Filter to inject cache_control into system prompts and tools for Anthropic prompt caching
kong/llm/plugin/shared-filters/expose-cache-metrics.lua	Filter to expose cache metrics as response headers
kong/llm/drivers/anthropic.lua	Anthropic driver with native streaming preserve mode support
kong/llm/drivers/bedrock.lua	Bedrock driver with Bedrock EventStream to Anthropic SSE conversion and Extended Thinking support
kong/llm/adapters/anthropic.lua	Anthropic format adapter for native format passthrough
kong/llm/adapters/bedrock.lua	Bedrock format adapter
kong/llm/adapters/gemini.lua	Gemini format adapter
spec/ai-anthropic-native/*.lua	Schema and unit tests for the plugin
kong.yml.template	Configuration template with examples
kong-plugin-ai-anthropic-native-0.1.0-1.rockspec	LuaRocks package specification

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.