feat(ai-proxy-multi): add anthropic-native provider for native Anthropic Messages API

[iakuf]

Summary

Add a new anthropic-native driver that speaks the native Anthropic Messages API protocol (/v1/messages) directly.

The existing anthropic provider uses the OpenAI-compatible driver, which works for basic use cases but cannot handle Anthropic-specific protocol details. The new anthropic-native provider addresses this.

Key differences from openai-base driver

• SSE event types: handles message_start, content_block_delta, message_delta, message_stop (no [DONE] sentinel)
• Token usage fields: reads input_tokens/output_tokens instead of prompt_tokens/completion_tokens
• Request format: removes stream_options (not supported by Anthropic), injects anthropic-version header automatically
• Response format: extracts text from content[].text array instead of choices[].message.content
• System prompt: accepts top-level system field and flexible content types (string or array of content blocks)

Changes

• apisix/plugins/ai-drivers/anthropic-native.lua: new driver implementation
• apisix/plugins/ai-drivers/schema.lua: register anthropic-native provider with its own request schema
• t/assets/anthropic-native-response.json: mock response fixture for tests
• t/plugin/ai-proxy-multi.anthropic-native.t: 26 test cases covering non-streaming, streaming, token usage, TTFT, error handling, and schema validation
• docs/en/latest/plugins/ai-proxy-multi.md: document the new provider with a usage example

Testing

All 26 tests pass:

[Baoyuantop]

Following the pattern of anthropic.lua or vertex-ai.lua, make the module entry point return openai-base.new(...) or an equivalent instance object.

[Baoyuantop]

Requires the inclusion of the Apache Software Foundation standard license header.

[Copilot]

Pull request overview

Adds a new anthropic-native AI driver to ai-proxy-multi so APISIX can proxy directly to Anthropic’s native Messages API (/v1/messages) rather than using the OpenAI-compatible protocol, including native SSE event handling and token usage extraction.

Changes:

• Introduces a new anthropic-native driver implementing native request/response and SSE semantics.
• Registers anthropic-native in the AI drivers schema with a dedicated (minimal) request schema.
• Adds tests, fixtures, and documentation for the new provider.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
apisix/plugins/ai-drivers/anthropic-native.lua	New native Anthropic Messages API driver (request shaping, SSE parsing, usage/TTFT extraction).
apisix/plugins/ai-drivers/schema.lua	Adds anthropic-native to provider registry and assigns it a request schema.
t/plugin/ai-proxy-multi.anthropic-native.t	Adds an extensive test suite covering non-streaming, streaming SSE, usage, and error handling.
t/assets/anthropic-native-response.json	Adds a mock native Anthropic response fixture used by tests.
docs/en/latest/plugins/ai-proxy-multi.md	Documents the new provider and adds a usage example.

---

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

[Copilot]

Documentation clarity: this section lists providers using brand names, but "Anthropic-native" is the config key and is shown elsewhere as anthropic-native. Consider using consistent formatting (e.g., backticks and lowercase anthropic-native) to avoid confusion about the exact provider value.

[Copilot]

anthropic-native is required to work with the ai-proxy-multi health-check / endpoint resolution logic. ai-proxy-multi.resolve_endpoint() falls back to reading ai_driver.host/ai_driver.port when override.endpoint is not set, but this driver module doesn't define those fields (unlike the other providers). This will make endpoint resolution/health checks fail when users rely on the default Anthropic endpoint. Define host, port (and ideally path) on the exported driver object (or return an instantiated driver like the other providers) so the default endpoint can be resolved consistently.

[Copilot]

Spelling: "health chekcs" should be "health checks" (also appears in the page description metadata).

Suggested change

	description: The ai-proxy-multi Plugin extends the capabilities of ai-proxy with load balancing, retries, fallbacks, and health chekcs, simplifying the integration with OpenAI, DeepSeek, Azure, AIMLAPI, Anthropic, Anthropic-native, OpenRouter, Gemini, Vertex AI, and other OpenAI-compatible APIs.
	description: The ai-proxy-multi Plugin extends the capabilities of ai-proxy with load balancing, retries, fallbacks, and health checks, simplifying the integration with OpenAI, DeepSeek, Azure, AIMLAPI, Anthropic, Anthropic-native, OpenRouter, Gemini, Vertex AI, and other OpenAI-compatible APIs.

[moonming]

Hi @iakuf, thank you for the substantial work on the anthropic-native provider! This is a significant contribution (1233 lines).

Context: There's an ongoing architectural discussion about how to support Anthropic's API:

• This PR (#13050): Adds a native Anthropic driver that speaks the Anthropic Messages API directly
• #13048: Takes a protocol translation approach (Claude ↔ OpenAI format conversion)

Both approaches have merit:

• Native driver: More efficient, preserves Anthropic-specific features (tool use, vision, etc.)
• Protocol translation: Simpler to maintain, single code path for request/response handling

Before we can proceed, we need to decide on the technical direction. I'll make a decision on this and update both PRs. In the meantime, could you:

1. Address any existing review comments
2. Ensure tests cover the key Anthropic API features (streaming, tool use, vision)
3. Rebase if needed

Thank you for your patience while we sort out the architecture! Your implementation quality is appreciated.