dsh-plugins

Everyday use of DeepSeek Harness tends to surface the same set of frustrations: knowledge evaporates between sessions, teams have no way to share and revisit conversation context, long-running Goals die on a timeout or crash and must be re-launched by hand, and model reasoning parameters lack a unified visual control surface. dsh-plugins packages five tools into a single installable bundle that addresses all four gaps in one go.

The core capabilities span memory, collaboration, structure, resilience, and model control. memarc maintains a per-workspace persistent JSON memory store with session archiving, full-text search, batch restore, and permanent delete; session-git exports conversations as Git-committable JSON files so any teammate can one-click import them as live, resumable sessions. session-branches renders a fork and sub-agent relationship tree in a floating sidebar panel, making complex multi-agent workflows navigable at a glance. On the reliability side, goal-auto-resume automatically re-arms a Goal after error, interruption, max-tokens, or crash, with exponential backoff and safety valves. dsh-model-reasoning offers a three-tier settings page (global policy / route overview / per-model detail) for editing reasoning level mapping, context and output limits, and protocol compatibility toggles, plus a session-level intensity slider right next to the model picker in the input box.

This bundle is best suited for individual developers and small teams who rely on DeepSeek Harness for daily work or complex AI-assisted workflows. If you need persistent context across sessions, want session history under version control for team sharing, run long Goals and hate losing progress to a single timeout, or want fine-grained control over each model's reasoning depth, dsh-plugins plugs directly into your stack. It has zero runtime dependencies, installs offline, and requires only DSH 0.1.1-rc.2 or later.