Foreword¶

DeepSeek Harness (DSH) adopts an “everything is a plugin” architecture, allowing functionalities to be extended as needed. However, its default behavior does not emphasize execution discipline: the system prompt is empty, failed actions are easily repeated, installed Skills may not be invoked, and pending tools are often ignored. When developers actually run tasks, common phenomena include the agent repeatedly retrying the same problem, not actively reflecting, and not consolidating successful experiences.

cleverer-dsh is a workflow plugin suite designed for DSH, maintained by Classicoke. Its goal is to address the above gaps without modifying DSH’s source code, using 11 plugins and 6 built-in Skills. The suite has zero external dependencies. The README notes that all 478 unit tests pass, with 100% statement and line coverage.

What This Is¶

cleverer-dsh is positioned as a DSH execution-discipline plugin suite: through the collaboration of a discipline layer, tool layer, and Skill layer, it provides stable constraints for Harness in areas such as failure interception, task planning, memory deduplication, and experience consolidation.

Maintainer: Classicoke
License: MIT (see package.json)
Current Version: 1.2.0
Category: Workflow
GitHub: https://github.com/Classicoke/cleverer-dsh
Community Directory Page: https://www.skillhub.cn/plugins/Classicoke/cleverer-dsh

Core Features¶

Below is an introduction based on the architectural layers described in the README, listing only verified capabilities.

Discipline Layer (8 Plugins + 1 Hub)¶

Discipline plugins share failure logs and reminder pipelines via discipline-hub, avoiding isolated operations.

Tool Layer (2 Plugins)¶

Skill Layer (6 Built-in Skills)¶

The README lists built-in Skills including: six-step error handling, error quick reference, fast file location, root cause debugging, local-first approach, and plan-before-execute. At runtime, dsh-skill-provider resolves and registers them within the package, and dsh-skill-loader handles on-demand loading.

Real-World Data (README self-reported, limited samples)¶

The README records a comparative experiment: same task (analyzing software packaging logs) with the suite installed vs. bare DSH:

The README explicitly notes: each group has n=1, tokens are character estimates (±20%), not yet widely verified, and specific figures await more samples and real API bill data. This table is cited to illustrate a trend, not as a universal conclusion.

Installation and Activation¶

Prerequisites: DSH is installed and initialized; pnpm is available locally (required by DSH’s plugin manager).

Only one of the two installation methods can be chosen. The README warns: installing both will cause each plugin to load twice, leading to duplicated behavior or early rejection.

Method 1: DSH Official Plugin Manager (Recommended)¶

Suitable for scenarios where a single command completes installation. Plugins are written as inline patches to cordis.patch.yml.

dsh plugin --profile web add github:Classicoke/cleverer-dsh

For headless environments, replace web with headless:

dsh plugin --profile headless add github:Classicoke/cleverer-dsh

After the command completes, no build is needed; plugins and Skills are immediately available. To uninstall:

dsh plugin --profile web remove cleverer-dsh

Method 2: PowerShell Script Installation¶

Suitable for scenarios requiring a board-grouped architecture. The script downloads the v1.2 zip from GitHub Release, extracts it, and runs install.ps1, dividing plugins into discipline-board (discipline group) and tools-board (tool group). This enhances structured collaboration between discipline plugins and the Hub.

Paste and execute in PowerShell 7+:

$u = 'https://github.com/Classicoke/cleverer-dsh/archive/refs/tags/v1.2.zip'
$z = "$env:TEMP\cleverer-dsh.zip"; $d = "$env:TEMP\cleverer-dsh-install"
Invoke-WebRequest $u -OutFile $z
Expand-Archive $z $d -Force
pwsh -File "$d\cleverer-dsh-1.2\install.ps1"
Remove-Item $z, $d -Recurse -Force

Uninstallation requires restoring the pre-installation backup of cordis.patch.yml, then deleting each plugin and Skill file as listed in the README. A complete uninstall script is in the GitHub README under “Uninstall (Option 2)”.

Typical Usage¶

After installation, the suite loads automatically with the DSH process, requiring no additional startup commands. Behavior manifests during task execution:

1. Multi-step tasks: dsh-plan-discipline reminds to create a plan first, reducing unstructured progression.
2. Command failure: anti-stuck intercepts identical parameter retries; dsh-env-triage requests a stop and report after multiple approaches fail.
3. File search: Uses dsh-fast-locate to scan multiple directories in parallel, replacing sequential guessing.
4. Environment troubleshooting: Invokes dsh-env-check-tool to perform 9 health checks, quickly locating environment issues.
5. Experience consolidation: skill-evolver distills failures and solutions into reusable Skills for subsequent tasks.

Observable differences from the README comparative experiment: after installing the suite, the agent asks users at key decision points, automatically switches failed approaches, and consolidates available commands into scripts; bare DSH retried the same problem 13 times, distilled no experience, and did not actively consult.

Applicable Scenarios and Considerations¶

Who is this for?

• Those who regularly run multi-step development tasks with DSH and want to reduce “deadlock retries” and ineffective token consumption.
• Those needing structured discipline constraints (planning, reflection, memory deduplication, Skill auto-loading) without modifying DSH’s source code.
• Those willing to use the official dsh plugin add on Linux/macOS or use the board architecture script installation on Windows.

Considerations

1. Permissions: Plugins run with the current DSH process’s permissions. Before installation, review the source code and confirm the MIT license to assess acceptance of its file access and tool call scope.
2. Installation Exclusivity: Method 1 and Method 2 cannot coexist; uninstall the existing installation before switching.
3. Real-World Data: The README performance comparison is a single-sample experiment and should not be directly extrapolated to all task types.
4. Community Directory: SkillHub (https://www.skillhub.cn) is an independent community site with no official affiliation to DeepSeek or High-Flyer; plugin distribution is based on the GitHub repository.

Conclusion¶

cleverer-dsh packages the execution discipline, practical tools, and Skill library that DSH lacks by default into a zero-dependency solution: 11 plugins with clear division of labor, 6 built-in Skills loaded on demand, and a single dsh plugin add command to enable it. If you are using DSH for daily agent workflows and are troubled by repeated failures, lack of planning, or idle Skills, you can start by installing via the official plugin manager, refer to the architecture table in the README to understand each plugin’s responsibilities, and then observe the actual effects of the discipline and tool layers by task type.

• Community Directory: https://www.skillhub.cn/plugins/Classicoke/cleverer-dsh
• GitHub: https://github.com/Classicoke/cleverer-dsh