dsh-context-compass

Running long conversations in DeepSeek Harness usually hits one wall well before model quality becomes the issue: you cannot tell whether the context window is about to overflow and a new session would serve you better. dsh-context-compass turns that gut feeling into a color-coded signal next to the session log button, where green, blue, yellow, or red instantly tells you the current health tier and whether to continue or switch. Every figure is drawn from precise harness measurements, including the token meter, model window metadata, and compression-aware projections, rather than educated guesses.

Three layers make up the core experience. Hovering the header badge reveals the window usage bar, per-round token cost, cache hit rate, billing estimate in CNY or USD with peak and off-peak pricing, and a sparkline tracing the last 40 request pressure samples. Typing /compass produces a full text report with optional probes for git status, a user-named handoff document, or a process list; at yellow or red severity it appends a real-state checklist driven by actual filesystem and git queries instead of boilerplate text. On the model side, the context_compass tool lets the agent self-assess structured signals mid-task. The Compass Overview panel, opened from the sidebar, ranks every session by running-first and severity, so you can scan the health of all conversations at a glance and click any row to open and diagnose it instantly.

It is built for developers and power users who juggle multiple DeepSeek sessions spanning code reviews, long-document collaboration, and iterative debugging, and who need a fast, data-driven answer to the question of whether a round can survive one more turn after compaction. The plugin is fully decoupled from any specific knowledge base, every signal is configurable, and it ships under the MIT license so it is ready to use the moment you install it.