In the DeepSeek Harness (DSH) plugin ecosystem, code review tools usually focus on logic bugs or syntax errors. dsh-mbti-jury offers a different perspective: it feeds the latest commit to a review jury composed of 16 MBTI personalities. Based on different cognitive functions, the personalities produce completely different review results for the same change.

This is a DSH plugin maintained by SleepEggTart. Its core is to analyze the latest commit and the historical commit cadence, simulating 16 MBTI personalities (based on cognitive functions: Si/Ni/Fe/Ne/Te/Ti/Se/Fi) to run parallel reviews. The output format is determined by the dominant cognitive function rather than merely changing tone.

Installation

Install the plugin with the following command:

dsh plugin --profile web add dsh-mbti-jury

Core Features

  1. 16-type personality review: Supports 16 MBTI personality types and generates review reports in different formats based on dominant and auxiliary cognitive functions.
  2. Historical behavior analysis: In addition to analyzing the latest commit, it reads the summaries and messages of the most recent 5 commits, identifies files that have been modified repeatedly, and even determines which commit of the week it is (focusing on cadence).
  3. Parallel execution: Each personality independently initiates a request and cannot see the others, ensuring that the differences represent “parallel worlds” rather than mutual influence.
  4. Fault tolerance: A failed call for one personality does not block the overall flow; a placeholder message is displayed.
  5. Roast mode: After the review, each personality can rebut the statement it disagrees with most.

Usage

In any git repository session, use the following commands:

  • /mbti: Review the latest commit with the default classic four-person panel (ISTJ/INTJ/ESFJ/ENTP).
  • /mbti random: Randomly select a four-person panel from the 16 types.
  • /mbti ENFP ISTP INTP ISFP: Specify the lineup (supports space, comma, or enumeration-comma separators).
  • /mbti-fight: Specified lineup + roast segment.
  • /mbti-fight random: Random four-person panel + roast.

Technical Implementation

The plugin relies on the host’s ctx.llm model. It uses parallel execution to ensure that the review processes of different personalities do not interfere with one another. If a request for one personality fails, it does not interrupt the other personalities’ reviews; instead, a placeholder message is displayed. Commands support cancellation, and the abort signal propagates to every model call.

Development and Debugging

The plugin is developed in TypeScript and depends on @deepseek-ai/cordis. For local debugging, after modifying the code, bump the version number in package.json, then install it with dsh plugin --profile web add "file:<本目录路径>".

Summary

By observing how different personalities diverge on the same code commit, developers can more intuitively identify potential blind spots or risks across different dimensions in the code. This is a code review tool that combines psychology with engineering practice.