Introduction

When DeepSeek Harness (DSH) users compare different models solving the same programming task, they often need to manually switch between multiple sessions, or lack a traceable chain of evidence. The dsh-evidence-arena plugin provides an evidence-first multi-model programming comparison workbench. It allows isolated parallel evaluation without modifying the source code of the official DSH Host, Session, or chat UI.

Plugin Overview

This is an independent community project maintained by shengshifantang. The plugin is installed via the npm registry. Its core value is to provide parallel execution of multiple Builders, independent Reviewers, and reproducible checkpoints. In the end, only human-confirmed candidate code is written back to the original workspace.

Core Features

  • Parallel isolated evaluation: Supports 2 or 3 Builders solving the same task in isolated Git worktrees, with each Builder having an independent DSH SDK runtime.
  • Result inspectability: Provides per-file diffs, independent Reviewer verdicts (logic and safety checks), and reproducible checkpoints. All results are sealed evidence.
  • Evidence-first mechanism: Only explicitly confirmed candidates are written back to the original workspace. Unconfirmed run results are retained as evidence only.
  • Independent runtime environment: The Reviewer runs in a zero-tool sealed evidence environment, and DSH_HOME points to an empty directory, ensuring that host configuration files and credentials are not exposed to the model-facing Shell.
  • Compatibility verification: Verified compatibility snapshots with the DSH release series are available as of 2026-08-19.

Installation and Enablement

  1. Install the plugin:
    npm install dsh-evidence-arena
  1. Enable the workbench:
    After installation, find the A/B action at the bottom of the DSH sidebar and open it. The plugin does not register a new /arena slash command, nor does it install any dependencies.

Typical Usage

  1. Open the plugin workbench and select the corresponding task. Only then will settings and pre-checks be loaded.
  2. Expand Run candidate result, confirm execution, open the one-time loopback preview, and inspect the logs.
  3. Based on the preview and logs, manually make a determination (pass/fail/uncertain).
  4. Run it in the terminal to download a versioned JSON evidence report for review or sharing.

Applicable Scenarios and Notes

  • Applicable scenarios: Model performance comparison, code generation quality evaluation, and strategy verification for specific engineering tasks.
  • Compatibility notes: The plugin has clear compatibility boundaries. You should use a DSH release series that matches its build. After upgrading DSH, it is recommended to rerun installation and browser smoke tests rather than assuming that private interfaces remain permanently stable.
  • Security notes: The plugin integrates with the host only through the plugin manifest, Host RPC, Workspace Registry, sidebar.footer.action, and shell.overlay. The released payload does not copy the full official DSH runtime; it depends only on the SDK/sub-agent bootstrap packages missing from the 5 standard Web closures.

Short Conclusion

This plugin is a structured evaluation tool in the DeepSeek Harness ecosystem and is suitable for developers who need rigorous comparison and evidence archiving. For more information, refer to the GitHub repository: shengshifantang/dsh-evidence-arena.