AI Agent Hub
Back to plugins
⚙️

think-like-codex

Workflow Updated 2026.08.21

Run the following command in DeepSeek Harness:

dsh plugin install zhangnan/think-like-codex

Paste the following prompt into your AI chat to install this plugin:

To install this plugin in DeepSeek Harness, execute the command dsh plugin install zhangnan/think-like-codex with the source code from https://github.com/zhangnan/think-like-codex.

About this plugin

In the development of code agents, blind loops and a lack of structured planning are common pain points. think-like-codex aims to address this by introducing Codex-inspired behavioral strategies that help agents escape inefficient exploration cycles. It not only automatically identifies complex implementation requests to force the agent into a structured plan mode but also proactively seeks user help when evidence is insufficient, ensuring every step is deliberate.

The plugin's core capability lies in its fine-grained control over agent behavior. It injects practical guidance for collaboration, code generation, validation, and safety, while enforcing strict exploration budgets. In plan mode, agents are required to pass a checkpoint after 8 tools and face a hard limit after 12. Additionally, it detects repeated tool failures and distinguishes facts from assumptions, guiding the agent to adjust strategies or seek user input.

This is an essential auxiliary tool for DeepSeek Harness users tackling complex engineering tasks. By standardizing the "Plan → Goal → Implementation" workflow, it effectively eliminates meaningless attempts and improves the quality and safety of code generation. Whether building large features or debugging complex logic, the plugin provides a clear behavioral compass for your agents.

Use Cases

  • Prevents agents from getting stuck in meaningless loops
  • Handles complex coding tasks requiring multi-step planning
  • Validates code logic safety before execution

Best For

  • Developers orchestrating and optimizing DeepSeek Harness agents
  • Engineering teams seeking improved code generation quality and safety
  • Researchers or engineers tackling complex logical tasks