Introduction

DSH’s philosophy is “Everything is a plugin”; the community directory is an independent site with no official affiliation with DeepSeek / Huaxuan. For people using DSH to build coding agents, a common issue is: in complex tasks, agents might call tools sequentially but lack clear plans, evidence boundaries, and moments to ask for help.

think-like-codex is a behavioral strategy plugin for DeepSeek Harness (DSH). It helps coding agents make more cautious decisions, avoid inefficient loops, ask users when evidence is insufficient, and proceed with complex tasks according to the plan → goal → implementation flow.

What is this

One-sentence positioning: think-like-codex is a Codex-inspired planning, evidence, goal, and anti-loop policy plugin for DeepSeek Harness.

The maintainer is zhangnan, and the license is MIT. The repository address is:

https://github.com/zhangnan/think-like-codex

It is a DSH Profile Bundle; the bundled cordis.patch.yml will automatically mount the plugin. The plugin does not register duplicate tools and does not directly modify session files.

Core Features

Below are the verified capabilities:

  • Injecting practical guidance on collaboration, code generation, verification, and security.
  • Applying equal task detection and strategy guidance to both English and Chinese requests.
  • Detecting complex implementation requests and automatically entering DSH Plan mode.
  • In Plan mode, an exploration checkpoint is added after calling 8 tools; a hard limit is triggered after calling 12 tools.
  • Before running mutation-type tools, requiring a goal and todo list after an approved plan.
  • Detecting duplicate tool failures and prompting the agent to switch methods or ask the user for help.
  • Adding periodic decision checkpoints in longer turns.
  • Distinguishing verified facts, inferences, and assumptions.
  • Limiting repeated attempts when the problem cannot be reproduced in the available environment.
  • Reusing DSH built-in capabilities like planMode, create_goal, todo_write, ask_user_question, and repeat-tool-reminder.

Installation and Activation

The installation commands are as follows:

dsh plugin --profile web add -w think-like-codex
dsh --profile web

web is the profile name and can be replaced with other profile names as needed. The plugin is a DSH Profile Bundle, and the bundled cordis.patch.yml will automatically mount it. If the host-side code is updated, the profile host needs to be restarted.

The uninstallation command is as follows:

dsh plugin --profile web remove think-like-codex

Typical Usage and Configuration

After installation, the plugin installs the following default configurations:

- id: think-like-codex
  name: think-like-codex
  config:
    autoPlan: true
    complexityMinChars: 120
    stepReminderAt: 12
    stepReminderEvery: 8
    planCheckpointToolAt: 8
    planHardToolLimit: 12
    enforceGoalAfterPlan: true
    failureReminderCount: 2

These thresholds are used to control automatic planning, complex task determination, step reminders, checkpoints in Plan mode and hard tool limits, goal constraints after plan, and failure reminder counts.

If different thresholds are needed, you can override the complete config object of think-like-codex in the cordis.patch.yml of the corresponding profile.

In terms of behavior, there are a few points to note:

  1. Simple questions and clearly read-only tasks will not automatically enter Plan mode.
  2. Users can explicitly tell the agent not to plan, thereby overriding automatic planning.
  3. When users directly press for an answer or change direction, the exploration budget is reset, allowing the new investigation direction to unfold normally.
  4. After a complex implementation request is detected, it will first enter Plan mode and proceed in the order of plan, goal, and todo list.

Use Cases and Notes

Suitable scenarios for using DSH profiles to manage coding agents: those who want complex tasks to be planned before execution, wish to reduce meaningless tool loops, clearly ask for help when evidence is insufficient, and preserve decision checkpoints in longer tasks.

The plugin runs with the current dsh process permissions. Before installation, it is recommended to check the source code, configuration options, and MIT license. After the above installation and configuration steps, the plugin will work through DSH built-in capabilities rather than directly modifying session files or registering duplicate tools.

Conclusion

The value of think-like-codex lies in putting “planning, evidence, goals, and anti-looping” into DSH’s daily workflow. The plugin name is think-like-codex and can be found in the DSH plugin directory by name. The GitHub address is:

https://github.com/zhangnan/think-like-codex