duhai-vision

When Codex or DSH handles image tasks, the vision entry is often fixed: quota boundaries are unclear, token accounting is opaque, and there is no traceable fallback after a failure. Duhai Vision decouples image recognition from the built-in channel into a swappable vision adapter layer, giving batch OCR, long documents, UI screenshots, and historical publications clear routing rules, call budgets, and graceful degradation.

Every visual task defaults to PaddleOCR-VL 1.6, covering people, buildings, objects, scenes, UI, charts, documents, and complex layouts. It returns structured JSON (text, tables, formulas, seals, uncertain items) back to the Agent for further reasoning and verification. Qwen3-VL-Plus is used only when the user explicitly requests it or Paddle is unavailable as a semantic fallback; Codex Native is the last resort and must be announced with a reason. Each task allows at most two external calls, and the full trail—provider, model, page count, latency, uncertainties—is recorded so unknown values are never written as zero. The same repository serves both the Codex Skill and the DSH Plugin, sharing identical routing rules, executors, and return structures without separate maintenance.

Who it is for: developers who regularly need AI to read documents, tables, scans, UI screenshots, or batch images; teams that care about the free quota (currently 3000 pages per user per model per day on AI Studio) and cost control; and engineers who need an auditable, swappable vision pipeline. Duhai Vision does not replace the Agent reasoning or tool orchestration layer—it makes vision extraction an independent, observable, degradable piece of infrastructure.