v0.3.0 — Persistent previews and faster vision routing

Highlights

• Keep original DSH image thumbnails visible after sending
• Archive processed images locally under the Git-ignored .dsh-vision-router/images/ directory using SHA-256 filenames
• Rewrite images only at the text-only provider adapter boundary
• Format visual context and explicit tool results as Markdown
• Cache Ollama model discovery and identical vision analyses
• Keep local Gemma loaded with Ollama keep_alive
• Prevent DeepSeek from searching for or re-analyzing images after vision context is ready
• Run real Gemma tests serially to avoid local 26B model contention

Verified

• Native paste/upload preview remains visible
• Gemma4:26b analyzes the attachment locally
• DeepSeek-V4-Flash produces a formatted answer in one agent step
• Automated test suite: 3/3 passing
• Package dry-run and secret scan passing

Documentation in English and Chinese, plus the Chinese technical article, is updated for this release.

v0.2.0 — Native DSH image attachments

What's new

• Drag or paste images directly into DSH Web while using a text-only DeepSeek model
• Route authorized attachment bytes through local Ollama/Gemma before the main model call
• Bridge DSH image admission without exposing internal attachment paths
• Keep explicit inspect_image support for paths and data URLs
• Add multi-provider configuration, privacy guidance, bilingual README updates, and a Chinese technical article
• Add real Ollama model discovery, ToolRuntime, and attachment bridge tests

Privacy

Automatic routing remains local-only by default. Remote image fallback stays disabled unless explicitly enabled. Real API keys are not included in this repository or release.

v0.1.0 — Initial release

Highlights

• Zero-config discovery of local vision-capable Ollama models
• Real inspect_image tool integration with the DeepSeek Harness Cordis runtime
• Multiple prioritized Ollama and OpenAI-compatible providers
• Privacy-aware local-first routing; remote fallback is disabled by default
• English and Simplified Chinese documentation

Compatibility

• DeepSeek Harness: 0.1.0-rc.6
• Node.js: 22.19+