Introduction

The following introduces dsh-adhd-copilot. It is part of the DeepSeek Harness (DSH) plugin ecosystem, maintained by zimai233, and aims to guide ADHD users from “wanting to do it but unable to start” to “taking away a minimal action.”

For DSH Agent developers, a specific problem is: when users get stuck, the Agent tends to give lengthy plans, yet the user still cannot start. dsh-adhd-copilot does not continue to expand the plan; instead, it compresses the current request into an executable next step and controls the length of the response and the granularity of the action.

What is this

dsh-adhd-copilot is a DSH behavioral coaching skill plugin, with a focus on behavioral/mental health assistance. It provides the adhd_breakdown tool to scan tasks and status signals, processing them according to the routing priority of “Emotion > Structure > Initiation,” and finally returning a single next-step action breakdown.

The plugin license is MIT. The complete skill content includes skills/adhd-copilot/SKILL.md and references/*.md, totaling 11 files, packaged as-is.

Core Features

adhd_breakdown Tool

adhd_breakdown scans tasks and status signals, then routes to the M1-M8 modules. It does not require completing the entire task but returns a structured next-step breakdown.

The breakdown result contains the following fields:

  • module
  • moduleName
  • nextAction
  • whyNow
  • timeboxEstimate
  • energyRequired
  • steps

Where nextAction points to a next step that can be done within 30 seconds; steps contains only 1 to 3 steps, not throwing the entire task list at the user all at once.

Pre-flight Gate

The plugin has a Pre-flight Gate:

  1. Remain silent with no task signals, zero interference.
  2. Answer directly for very small requests, without going through the workflow.

This prevents the Agent from forcibly entering the coaching workflow during casual chit-chat or very small requests.

Self-help Menu for Ambiguous Signals

When signals are ambiguous and it is impossible to judge what category of problem the user is currently stuck on, the plugin provides an A-G self-help menu, allowing the user to choose from available options. This transfers the decision-making power back to the user, reducing misjudgments.

Output Discipline

The plugin has clear constraints on Agent responses:

  • The first line is an action.
  • Numbered steps do not exceed 3.
  • End with a next step that can be done within 30 seconds.
  • Praise is tied to specific victories.
  • Never pretend to be a timer.

The focus of these rules is to control output length, avoiding the use of a full checklist, vague encouragement, or fake status updates to replace an executable action.

Zero-build Run

The plugin is zero-build, zero runtime dependencies, pure ESM, run-as-published. package.json specifies type as module, main as src/index.js, and includes the dsh.bundle.patch configuration for ./cordis.patch.yml.

Optional Parameters

The following optional parameters can be supported during invocation:

  • battery: Battery level, used by M3.
  • narrative: Narrative style, used for M5 opt-in.
  • estimatedMinutes: User’s own time estimate, used by M6.
  • parked / dropped: Later / abandoned list, used by M2.

Installation and Activation

The README provides the following installation command:

dsh plugin --profile myprofile add dsh-adhd-copilot

If installing from GitHub and locking a commit, the README also provides:

dsh plugin --profile myprofile add github:zimai233/dsh-adhd-copilot#<sha>

It is recommended to check the README, source code, and MIT license before installation. DSH plugins run with the permissions of the current dsh process; therefore, do not rely solely on the name and description. You should also verify that the tools actually invoked by the plugin, output rules, and optional parameters match your Agent configuration.

Typical Usage

Simply state the currently stuck task in natural language to the Agent, for example:

I’ve been dragging my feet on writing the weekly report for two weeks, help me start.

There are so many things piled up, my head is a mess, I don’t know where to start first.

I didn’t finish it again, why am I so bad.

The Agent will invoke adhd_breakdown and return a structured breakdown. For example, in the weekly report scenario, you might get a result similar to this:

{
  "module": "M1",
  "moduleName": "Task Breaker",
  "nextAction": "Open the file or page related to 'Weekly Report' (within 30 seconds)",
  "whyNow": "Initiation is the hardest: the first action must be obvious, tiny, and doable right now. Only give 1-3 steps at a time, each step < 60 seconds, ask for the next after finishing.",
  "timeboxEstimate": "First step < 60 seconds; 2 steps for this round, ask for the next step after finishing",
  "energyRequired": "Low (just need to open the file)",
  "steps": [
    "Open the file or page related to 'Weekly Report' (within 30 seconds)",
    "Write down this line as the title: 'Weekly Report'"
  ]
}

After the above breakdown, the Agent can render it into a response that conforms to the output discipline:

Weekly report, don’t think about finishing it for now. Just do the first step:

  1. Open the file or page related to ‘Weekly Report’ (within 30 seconds)
  2. Write down this line as the title: ‘Weekly Report’

Each step < 60 seconds. Reply “Done” when finished (optional, non-blocking). You have started, and this step is done.

The key in this example is not “finishing the weekly report,” but first completing two actions that take less than 60 seconds.

Skill Files and Package Structure

The plugin places the complete skill content under skills/adhd-copilot/:

  • SKILL.md
  • references/*.md

The README explains that these files are packaged as-is and are part of the plugin delivery. For developers, this means you can check how the skill content constrains routing, output, and state handling, rather than relying solely on invisible runtime behavior.

Suitable Scenarios and Notes

The suitable scenario is when the user already has a task or state but needs to reduce the cost of starting. It can assist in task breakdown, reduce overload, provide optional checkpoints, and shrink the goal to a 30-second action when the user is self-blaming.

However, there are a few things to clarify:

  1. This is an execution assistance tool, not a medical diagnostic or treatment tool.
  2. When a user shows signs of emotional crisis / self-harm, gently suggest seeking professional help.
  3. There is no timer. The AI cannot appear during execution; all mid-process checkpoints are optional.
  4. No cross-session memory is promised. It is only persisted when the platform has long-term memory.
  5. Completely silent when there are no task signals, preventing normal requests from being forcibly forced into the coaching workflow.

Links

GitHub:

https://github.com/zimai233/dsh-adhd-copilot

Directory Page Link:

https://www.skillhub.cn/plugins/zimai233/dsh-adhd-copilot