AI Agent Hub
Back to plugins
🖥️

dsh-git

Client Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install ZicanC/dsh-git

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

Run dsh plugin install ZicanC/dsh-git in your terminal, browse the source at https://github.com/ZicanC/dsh-git, then refresh the DeepSeek Harness Web UI to start using the plugin.

About this plugin

Agent conversation history in DeepSeek Harness is a one-way stream: once a turn goes off track, keeping the good reasoning before and after it usually means manual copy-paste and re-prompting. dsh-git treats each completed Prompt+Answer pair as an atomic commit, letting you cherry-pick turns from any branch, drag them into a custom order, and continue asking questions as a new branch. To the agent, the reconstructed history looks exactly like one it experienced firsthand rather than a context block appended to the prompt.

Core capabilities include a per-Session and per-Workspace conversation graph that visualizes forks, merge edges, and the active HEAD node; a cross-branch Context Tray where selected turns are explicitly ordered by drag; and a one-click merge-branch action that composes the selected turns into a well-formed seed on the Host side and opens a new Session. Graph state persists with the DSH profile, merge provenance is written into the Session log for durability, and the entire workflow operates without modifying the dsh executable or requiring a build step.

Best suited for researchers iterating on prompts, developers comparing multiple reasoning paths, and any DeepSeek Harness user who wants to manage agent conversations with the same flexibility as a version-controlled codebase.

Use Cases

  • Cherry-pick completed turns from different conversation branches
  • Inspect Session and Workspace conversation graphs with forks and merges
  • Merge selected turns into a new branch and continue asking

Best For

  • DeepSeek Harness users
  • Researchers iterating prompts frequently
  • Developers comparing multiple reasoning paths