feat: add Qwen (Alibaba Cloud DashScope) provider

[leslieeilsel]

Summary

Add comprehensive support for Alibaba Cloud's Qwen models via the DashScope native API (/api/v1), providing full feature parity with other major providers.

Capabilities

Feature	Status	Details
Text Generation	✅	Multi-step tool calling, generation parameters
Multi-Modal (VL)	✅	Automatic endpoint routing for vision-language models
Streaming	✅	DashScope SSE protocol, reasoning/thinking tokens
Structured Output	✅	JSON Object mode + JSON Schema mode (strict)
Embeddings	✅	Configurable dimensions
Image Generation	✅	qwen-image-max, qwen-image-plus
Image Editing	✅	Single-image editing & multi-image fusion
Audio (TTS/STT)	❌	Not supported — DashScope uses WebSocket/async protocols

Architecture

• Uses DashScope's native API (not the OpenAI-compatible mode) for the most complete and up-to-date feature coverage
• Dynamic endpoint routing: Handlers automatically switch between text-generation and multimodal-generation endpoints based on message content (images present → multimodal endpoint)
• DashScope SSE streaming: Custom SSE parsing for DashScope's distinct event format (X-DashScope-SSE header, incremental_output mode)
• Structured output: Supports both response_format: {"type": "json_object"} (broad model support) and response_format: {"type": "json_schema", ...} (strict schema enforcement for newer models)
• Region-aware configuration: Supports International (Singapore), China Mainland (Beijing), Global/US (Virginia) deployment modes with clear documentation on API key and URL requirements

Files

• Provider source: src/Providers/Qwen/ — 14 files (handlers, maps, concerns)
• Configuration: config/prism.php, src/PrismManager.php, src/Enums/Provider.php
• Tests: tests/Providers/Qwen/ — 9 test files, 52 tests with 176 assertions
• Fixtures: tests/Fixtures/qwen/ — 14 fixture files recorded from real Qwen API responses
• Documentation: docs/providers/qwen.md + updates to structured-output.md, text-generation.md, image-generation.md, introduction.md, ProviderSupport.vue

Design Decisions

1. Native API over OpenAI-compatible mode: The DashScope native API is more actively maintained by Alibaba Cloud and provides better feature coverage (multimodal, image editing, structured output modes). The OpenAI-compatible mode lacks some features and has different error response formats.

2. Dynamic multimodal endpoint routing: Rather than requiring users to manually select endpoints, the handlers inspect message content and automatically route to the correct DashScope endpoint. This keeps the user-facing API clean and consistent with other Prism providers.

3. No audio support: DashScope's TTS (CosyVoice) uses WebSocket and STT (Paraformer) uses async REST with task polling — neither is compatible with Prism's synchronous HTTP interface. This is clearly documented with alternatives.

Test Plan

• All 52 Qwen tests pass (176 assertions)
• Text generation: basic prompt, system prompt, multi-step tool calling, generation parameters
• Multi-modal: VL model with images, automatic endpoint routing
• Streaming: basic text, tools, system prompt, max tokens, reasoning/thinking tokens, step events
• Structured output: JSON Object mode, JSON Schema mode, auto mode fallback
• Embeddings: basic input, request format, dimensions option
• Images: generation, provider options, editing (single/multi/base64), error handling
• Error handling: Arrearage, DataInspectionFailed, rate limits (429), overloaded (503)
• Message mapping: user, assistant, tool results, images (URL/path/base64), system prompts
• Tool mapping: standard tools, strict mode, tool choice validation
• Code formatted with Pint
• Fixtures recorded from real Qwen API responses

Made with Cursor