Introduction

In the use of DeepSeek Harness (DSH), developers often need to read logs to determine the current state of an agent. This approach is not only time-consuming but can also easily miss subtle behavioral anomalies. The dsh-mood plugin aims to address this pain point. It does not interfere with model behavior but instead observes tool calls and their results within the session to provide an intuitive “status light.”

Plugin Overview

dsh-mood is a micro behavioral observer designed for AI coding agents. It collapses complex session behavior into four verifiable states: GOOD, CONFUSED, FRUSTRATED, and OVERWHELMED. The plugin runs in pure observer mode and does not inject prompts or block tool execution. The project is maintained by Raphaelutumn and is licensed under the MIT license.

Core Features

The plugin primarily provides the following capabilities:
* Four-state indicator: Displays a low-friction status light next to the session title. State changes are based on consecutive failures, repeated operations, or high activity levels.
* Explainable reasons: Each state corresponds to verifiable trigger conditions (for example, “3 consecutive failures”).
* Session journey: Records state changes throughout the session, viewable on hover.
* Pure observation mode: Does not modify model requests, does not block tool calls, and does not interrupt the workflow.

Installation and Enabling

After downloading or building the plugin package, install it into the Web Profile using the following command:

dsh plugin --profile web add .\dsh-mood-0.1.0.tgz

After installation is complete, refresh the dsh web interface, and the status light will automatically appear next to the session title.

Configuration and Tuning

The plugin supports adjusting behavioral thresholds through a configuration file. Override the following fields in cordis.patch.yml (default values):

- id: mood
  config:
    frustratedFailureThreshold: 3        # 触发 FRUSTRATED 的连续失败次数
    stableSuccessesToRecover: 2          # 恢复到 GOOD 所需的连续成功次数
    confusedRepeatThreshold: 3           # 触发 CONFUSED 的重复操作次数
    changeCooldownMs: 60000              # 原因去重冷却时间
    repetitionWindow: 12                 # 用于检测重复的最近工具调用窗口
    highActivityThreshold: 4             # 触发高活动的窗口长度

Note: All configuration values must be positive integers.

Use Cases and Limitations

  • Use cases: Suitable for developers who need to quickly understand an agent’s current health and identify potential stalls during development.
  • Limitations: This is not a scientific measurement but an explainable behavioral classification. The plugin only provides the session header UI in the Web Profile.

Conclusion

dsh-mood provides a low-friction way for developers to understand an agent’s behavioral tendencies without digging deep into logs. By adjusting the thresholds in cordis.patch.yml, it can be adapted to different project requirements.