Introduction

When using DeepSeek Harness (DSH), unclear prompts often cause the agent to “get stuck” or misunderstand. Repeatedly correcting the agent’s behavior in conversation is time-consuming and makes it difficult to establish stable conventions. dsh-tacit aims to solve this problem. It is a workflow plugin that automatically analyzes your interaction history with the agent, distills your unstated habits and preferences, and injects them directly into future conversations to reduce repetitive corrections.

What Is It?

dsh-tacit is a DeepSeek Harness plugin maintained by developer hackernotfound. By observing your conversation turns (especially unsuccessful turns and your corrections), it learns the contextual information you often omit and converts these rules into instructions the agent can execute.

Installation and Prerequisites

Before installing, ensure the following environment requirements are met:
* DeepSeek Harness version >= 0.1.1-rc.1
* Node.js version >= 22

Run the following command in the terminal to install:

npx @deepseek-ai/dsh plugin --profile web add dsh-tacit

After installation, start or restart the DSH Web interface and refresh the page.

Core Capabilities

  1. Zero-Click Learning
    The plugin automatically analyzes conversations in the background. When a turn ends imperfectly or you send a correction such as “no, I meant…”, the system extracts the missing information. By default, the daily cap for automatically analyzed entries is 30.

  2. Instructions That Prove Their Worth
    The extracted rules are injected as part of the system prompt. These instructions can be edited, toggled, or deleted. New instructions enter a “trial period” (default: 10 turns). If you start correcting the agent more frequently during this period, the instruction is retired.

  3. Improvement
    It provides a rewrite button that uses the learned rules to optimize your current draft. It offers a before-and-after preview and supports thumbs-up or thumbs-down feedback.

  4. Measurable
    Charts show actual interaction trends, including how often you correct the agent, the chaotic turn rate, and token consumption per turn. All costs are calculated and displayed using the official pricing list.

How It Works

The workflow of dsh-tacit consists of four steps:
1. Maintains a bounded-size conversation summary that records prompts, tool calls, error messages, and end states.
2. When a turn is chaotic or a correction is detected, it triggers a deepseek-v4-flash analysis to record the missing information.
3. It periodically distills findings into 1-4 single-sentence instructions.
4. It injects these instructions into the system prompt of new conversations (about 300 tokens). Only fragments of the conversation summary leave the local machine.

Privacy and Cost

  • Privacy: The plugin never directly reads your API Key, and all model calls are routed through Harness’s service. Sensitive information (such as API Keys and JWTs) is masked before it is stored or sent. Data is stored in ~/.dsh/storages/tacit/.
  • Cost: By estimate, the learning cost per instruction is about \(0.001–\)0.003. The plugin itself follows the MIT license.

Typical Usage

你: "make the login page better"
智能体: ...卡住了...

你: "no, I meant the Next.js app under apps/web"

Tacit: learns → "The user often omits which app they mean — check apps/web first."

The plugin automatically injects the rule learned above into the system prompt of subsequent conversations.

Conclusion

dsh-tacit is a tool that helps you fine-tune agent behavior through actual interactions. It makes implicit knowledge explicit, and is suitable for DSH users who want to reduce the cost of repetitive communication. For more details, see the plugin directory or the source code.