Give DeepSeek (text-only) models **vision** in Claude Code and Codex by routing image files to any OpenAI-compatible vision endpoint (OpenRouter, SiliconFlow, DashScope, Ollama, llama.cpp, vLLM, LM Studio, …). Zero runtime dependencies, MIT licensed.
Jina AI tools for DeepSeek Harness: 12 model tools (web / arXiv / SSRN search, read, screenshot, embeddings, rerank, classify, PDF, primer) plus a settings-page API key card with live balance check.
Give your text-only LLM eyes: zero-dependency MCP server that reads images via external OpenAI-compatible vision models with automatic provider fallback, plus a one-command patch for DSH image-attachment degradation.
Long tasks shouldn't cost like they need a frontier model. Cheap-model long-task engine: reliability in structure (disk state machine + mechanical gates + T3 zero-prose dispatch + multi-brain audit), not in model strength. Pure Python stdlib. DSH skill.
Design-fidelity QA for DeepSeek Harness: lend any text-only model an eye, then judge whether the implementation matches the mock. Ships the benchmark behind that judgement — four fixtures, 23 injected defects, and every raw model transcript. Retires itself when DeepSeek ships vision.
