The philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When running or manually monitoring agents locally, long waits are common. During this time, the screen may show only a progress bar or a blank waiting interface with little feedback. The dsh-crumbs plugin is designed to fill this gap by displaying short, verified facts relevant to the current task while long-running tasks are in progress.

Feature Overview

The plugin provides three primary usage modes. Its core goal is to “inform human observers” rather than interfere with agent execution.

  1. Auto-surfacing: When a long-running tool call (such as a shell command) exceeds minTaskMs, a crumb relevant to the current task is shown every intervalMs.
  2. Command-driven: Retrieve facts on demand using the /crumb command.
  3. Tool-driven: Models or workflows can retrieve facts directly through the crumb tool.

Installation

Add the plugin using the official installation command:

dsh plugin add dsh-crumbs

Configuration and Usage

The plugin is enabled by default, and all configuration options are optional.

Environment Variable

Set the following environment variable to disable automatic surfacing (commands and tools remain available):

DSH_CRUMBS_DISABLE=1

Configuration File

Specific parameters can be adjusted in .dsh/crumbs.config.json at the project root:

{
  "autoSurface": true,
  "minTaskMs": 8000,
  "intervalMs": 12000,
  "mode": "fact",
  "source": "auto",
  "longTools": ["bash", "shell", "exec", "run"]
}
  • autoSurface: Whether to automatically surface facts during long-running tasks.
  • minTaskMs: Minimum task duration (in milliseconds) required to trigger auto-surfacing.
  • intervalMs: Interval (in milliseconds) between auto-surfaced facts.
  • mode: Fact mode or quiz mode.
  • source: Fact source (pool, model, auto).
  • longTools: List of tool names treated as “long-running tools.”

Commands and Tools

  1. Commands
    * /crumb: Fetch a random fact.
    * /crumb <topic>: Fetch a fact for a specific topic, such as /crumb git or /crumb concrete.

  2. Tools
    * The model can invoke the crumb tool with parameters:

    • topic (optional): topic hint.
    • mode: "fact" or "quiz".

Fact Sources and Verification

The plugin supports three fact sources, each with a different verification status and presentation style:

  • pool: Curated facts sourced from data/crumbs.json. These facts are verified, marked with verified: true, and displayed with the 💡 symbol. They are offline and zero-cost.
  • model: Facts generated in real time by a side model. These facts are unverified, marked with verified: false, and displayed with the ✨ symbol. Generated trivia may be inaccurate.
  • auto: Prefer model-generated facts; if the model is unavailable or returns an empty result, fall back to pool.

Important Notes:
* Automatically generated crumbs are displayed via ctx.logger. They currently appear only as log lines and are not yet integrated into Web UI cards.
* Model-generated crumbs (marked with ✨) should not be cited or used as factual evidence. They are for entertainment only.
* The plugin does not touch agent context or results. It is intended solely for human observation.

  • GitHub Repository: https://github.com/MeghanBao/dsh-crumbs
  • Community Directory: https://www.skillhub.cn/plugins/MeghanBao/dsh-crumbs