AI Agent Hub
Back to plugins
🖥️

dsh-ark

Client Updated 2026.08.28

Run the following command in DeepSeek Harness:

dsh plugin install leike0813/dsh-ark

Paste the following prompt into your AI chat to install this plugin:

In DeepSeek Harness, open your terminal and run dsh plugin install leike0813/dsh-ark to add the plugin (project: https://github.com/leike0813/dsh-ark).

About this plugin

The standard, code, and cordis modes in DeepSeek Harness ship with rich personas and large tool schemas, yet deepseek-v4-pro-0813 tends to drift or misbehave when exposed to those expansive schemas. The minimal preset, with only bash and str_replace_editor, is where the model stays most stable. dsh-ark bridges exactly this gap: it freezes the model-visible wire surface of all three Ark presets to the minimal tool set, while wrapping the extended capabilities of standard, code, and cordis into the dsh-ark CLI. The model triggers a bash command when it needs an extended capability, and the CLI proxies the execution on its side. You keep minimal-level stability without giving up the full capability catalog.

The core mechanism revolves around a CLI side channel. At session start, a generated shellrc exports DSH_ARK_MODE and DSH_ARK_WORKSPACE, so the model simply invokes the relevant dsh-ark subcommand inside bash and the CLI loads the matching mode directory to execute instructions, skills, tools, MCP endpoints, and host-bridge operations. Phase 2 adds an authenticated Unix-socket host bridge under $DSH_HOME/ark, letting host-plugin-provided tools appear dynamically as source: proxy entries in the tool list, with long-running delegation tracked via a bridgeJobId polling loop and no daemon process ever spawned.

This plugin is best suited for developers running deepseek-v4-pro-0813 alongside DeepSeek Harness who want the predictability of the minimal preset while still having on-demand access to the full standard, code, or cordis capability catalogs-especially teams that switch between work modes frequently and prefer not to maintain separate model configurations for each.

Use Cases

  • deepseek-v4-pro-0813 drifts under large tool schemas; you need minimal wire stability without losing extended capabilities
  • You frequently switch between standard, code, and cordis modes in the same project and want a single frozen model config
  • Host-plugin tools need to bypass the model wire and be invoked on demand via proxy, keeping the main dialogue clean

Best For

  • Developers pairing DeepSeek Harness with deepseek-v4-pro-0813 and needing predictable model behaviour
  • Teams that want minimal-preset stability while retaining the full standard, code, or cordis capability catalog
  • Engineering leads managing host-plugin tool injection and wanting to isolate model requests from extended execution