Feature/vertex multi publisher

[ggrandez]

Summary

This PR improves Vertex AI support in ai-proxy with multi-publisher routing and protocol conversion robustness, especially for Claude/Mistral scenarios.

Main outcomes:

• Adds automatic publisher-aware routing for Vertex models (Anthropic, Mistral, Codestral, Google/default).
• Improves request conversion/sanitization for Claude-origin payloads before Vertex forwarding.
• Extends tests to cover multi-publisher routing, endpoint selection, and compatibility behavior.
• Adds technical documentation for the Claude protocol fix in Vertex provider.

What changed

1) Vertex multi-publisher routing

• Introduced publisher routing strategy in Vertex provider:
  • routeRawPredict for publishers requiring rawPredict-style endpoints (e.g. Anthropic/Mistral).
  • routeOpenAICompatible for models using Vertex OpenAI-compatible endpoint.
  • routeNativeVertex fallback for Express Mode Google/native flow.
• Added model-prefix based auto-detection and optional explicit override via vertexPublisher.
• Added generic rawPredict path builders for arbitrary publishers (not only Anthropic).

2) Request conversion and protocol handling

• Added Claude-request pre-conversion in Vertex request-body flow when Claude conversion context flag is present.
• Added Claude request sanitization for Vertex compatibility (including schema cleanup and unsupported key handling).
• Improved Claude/OpenAI conversion behavior:
  • Finish reason normalization (strings.ToLower before mapping).
  • Reasoning extraction from tagged text content and clean text reconstruction.
• Ensured OpenAI-compatible path rewrites include proper publisher/model formatting (google/ prefix for unqualified models).

3) Tests

• Added/expanded tests for:
  • Publisher resolution and routing decisions.
  • rawPredict path generation (standard + express mode).
  • OpenAI-compatible endpoint path generation.
  • Multi-publisher request-body behavior (Mistral rawPredict, Llama/DeepSeek/Qwen/Gemini openai-compatible, Claude backward compatibility, publisher override scenarios).
  • Claude→OpenAI tools conversion and schema handling.

4) Documentation

• Added a technical write-up for Vertex Claude protocol fix:
  • plugins/wasm-go/extensions/ai-proxy/VERTEX_CLAUDE_PROTOCOL_FIX.md

Scope (files changed)

• plugins/wasm-go/extensions/ai-proxy/provider/vertex.go
• plugins/wasm-go/extensions/ai-proxy/provider/provider.go
• plugins/wasm-go/extensions/ai-proxy/provider/claude_to_openai.go
• plugins/wasm-go/extensions/ai-proxy/provider/vertex_test.go
• plugins/wasm-go/extensions/ai-proxy/test/vertex.go
• plugins/wasm-go/extensions/ai-proxy/provider/claude_to_openai_test.go
• plugins/wasm-go/extensions/ai-proxy/main.go
• plugins/wasm-go/extensions/ai-proxy/main_test.go

AI Coding Disclosure (required)

AI tools used

• Claude Code
• GitHub Copilot

Prompts / instructions used with AI tools

1. "Implement Vertex AI multi-publisher support for chat completions, with automatic routing by model prefix and support for Anthropic/Mistral/Codestral plus OpenAI-compatible fallback."
2. "Add configurable vertexPublisher override to force publisher routing regardless of model prefix."
3. "Fix Claude /v1/messages compatibility in Vertex by converting/sanitizing request body before provider-specific handling."
4. "Improve Claude/OpenAI conversion around finish_reason normalization and reasoning text extraction."
5. "Add comprehensive tests for routing, rawPredict paths, OpenAI-compatible endpoint behavior, and backward compatibility."

AI Coding Summary

• Key decisions made
  • Route by publisher strategy instead of single-provider assumptions.
  • Keep backward compatibility for Claude flows while enabling broader publisher support.
  • Use OpenAI-compatible endpoint as default in standard mode; use native Vertex route where appropriate in express mode.
• Major changes implemented
  • New routing resolution and generic rawPredict path construction.
  • Request conversion/sanitization improvements for Claude-origin payloads.
  • New/expanded unit and integration-style tests across multi-publisher scenarios.
  • Added detailed technical fix documentation.
• Important considerations / limitations
  • Behavior depends on model naming and mapping conventions; explicit vertexPublisher can override detection.
  • Express Mode and standard mode have different routing constraints.
  • Exact AI prompt text should be included verbatim by author if stricter traceability is required.

Commits included (author: Gad Grandez)

Hash	Date	Message
ad519f3	2026-01-22	Añadir soporte para la conversión de respuestas de Claude y mejorar la validación de herramientas
d858cd8	2026-01-23	Eliminar el manejo de conversión de respuesta de Claude en el proveedor Vertex
ed9b8ad	2026-01-23	Eliminar claves "$schema" no soportadas de los parámetros
a158cf8	2026-01-24	Agregar conversión de solicitudes de Claude a OpenAI en el cuerpo de la solicitud
34ede50	2026-01-24	Agregar corrección para la conversión de protocolo Claude en Vertex AI
930c48c	2026-01-24	Agregar función para sanitizar valores de esquema y eliminar claves no soportadas
d49d72a	2026-01-24	Agregar función para sanitizar solicitudes de Claude antes de enviarlas a Vertex AI
9bc4cd4	2026-01-24	Convert finish reason to lowercase in openAIFinishReasonToClaude function
eae18fd	2026-01-24	Agregar extracción y limpieza de razonamiento de texto en la conversión de respuesta
0b85e95	2026-02-12	feat(vertex): implement multi-publisher support and enhance routing logic
90b8d37	2026-02-13	feat(vertex): add automatic publisher prefix for models in multi-publisher support tests

[CLAassistant]

All committers have signed the CLA.

[ggrandez]

Hi! CLA is now signed by all committers and this PR is ready for review.

Could a maintainer approve the pending workflow runs from this fork so CI can execute?

@johnlanni @rinfx @wydream — any feedback is welcome, thanks!

[ggrandez]

Hi @johnlanni @rinfx @wydream 👋 Just a friendly ping on this PR! The branch is recently updated with main and ready to go. Could one of you please approve the CI workflows so the checks can run? I'd really appreciate your review whenever you have a moment. Thanks!

---

大家好 @johnlanni @rinfx @wydream 👋 友好地提醒一下这个 PR！分支已与 main 保持同步，随时可以合并。能否请您批准 CI 工作流，以便检查可以运行？有时间的话非常感谢您的审阅！谢谢！