dsh-omc-tui

Running DeepSeek Harness in a terminal usually means the TUI seizes the alternate screen buffer, making scrollback and text selection painfully clunky; and when a text-only coding model hits a screenshot or a UI error image, the user must manually swap the global model and invoke a vision skill, breaking the flow of thought. dsh-omc-tui removes both frictions: it never enters the alternate screen, so every message, tool call, Thinking block, and Diff appends to the regular terminal buffer, preserving native scrollback, selection, and copy behavior. At the same time, a zero-touch autonomous vision pipeline lets the main agent spin up an isolated, ephemeral vision Subagent on its own, parse the image, and feed the result back as a standard tool output-the user never switches models or types a command.\n\nBeyond vision, the plugin covers the key pieces of a terminal workflow: inline Diff approval and permission toggling, Plan and Jobs management, an adaptive status bar (current model, Context watermark, Git state, recent tools and latency, in three density modes), file references, session resume, and context compaction. On the safety front, a built-in Danger Guard watchdog performs structured syntax review before every tool execution, blocking high-destructiveness patterns such as recursive deletes, disk formatting, fork bombs, and deeply nested shell obfuscation, with support for project-level custom rules.\n\nWho it is for: developers who already use DeepSeek Harness in a terminal, value native scrollback and clipboard workflows, and need multimodal image understanding plus fine-grained permission control. The plugin handles only rendering and keyboard interaction; model inference, session persistence, sandboxing, and tool execution all remain with the official Harness services, keeping the boundary clean and the integration cost low.