zhangnan/think-like-codex
Codex-inspired decision and planning policy plugin for DeepSeek Harness
项目介绍Project Overview
think-like-codex 是一款面向 DeepSeek Harness (DSH) 的 Codex 风格行为策略插件。它通过注入协作、生成、校验与安全指导,引导编码代理遵循"计划→目标→实现"流程,在复杂任务中自动进入 Plan 模式并设置探索上限、目标校验与失败重试提醒等机制。适用于需要规范化决策、避免无效循环的复杂多步编码场景。需注意简单问题与只读任务不会自动启用 Plan 模式,且阈值为可配置项。
think-like-codex is a Codex-style behavior policy plugin for the DeepSeek Harness (DSH). It injects collaboration, code-generation, validation, and safety guidance so coding agents follow a plan → goal → implementation workflow. It auto-enters Plan mode for complex tasks, enforces exploration checkpoints, requires an approved plan plus goal before mutation tools, and reminds the agent to change approach after repeated failures. Use it for complex multi-step coding work that needs disciplined decisions. Simple questions and read-only tasks do not trigger planning, and all thresholds are configurable.
请帮我了解并安装插件:【think-like-codex】【https://github.com/zhangnan/think-like-codex】
把上面这条消息直接发给当前会话里的 DSH,让它帮你了解并安装。安装命令不一定准确,发给 DSH 更稳。Send this message to DSH in your current session. CLI install commands may not be accurate across systems — DSH will figure it out for you.
或使用命令行安装(适合开发者)Or use CLI install (for developers)
命令行安装CLI Install
dsh plugin --profile web add -w think-like-codex
把 zhangnan/think-like-codex 加入你的 DSH 配置(web profile)即可启用。
READMEREADME
think-like-codex
English | 简体中文
A Codex-inspired behavior policy and Profile Bundle for DeepSeek Harness (DSH).
It helps coding agents make deliberate decisions, avoid unproductive loops, ask the user for help when evidence is insufficient, and follow a structured plan → goal → implementation workflow for complex tasks.
Features
- Injects practical collaboration, code-generation, validation, and safety guidance.
- Applies equivalent task detection and policy guidance to both English and Chinese requests.
- Detects complex implementation requests and automatically enters DSH Plan mode.
- Adds an exploration checkpoint after 8 tools in Plan mode and enforces a hard limit after 12 tools.
- Requires an approved plan to be followed by a goal and todo list before mutation tools may run.
- Detects repeated tool failures and prompts the agent to change approach or ask the user for help.
- Adds periodic decision checkpoints during long turns.
- Distinguishes verified facts, inferences, and assumptions.
- Limits repeated attempts when an issue cannot be reproduced in the available environment.
Simple questions and explicitly read-only tasks do not automatically enter Plan mode. Users can also override automatic planning by clearly asking the agent not to plan. A direct user follow-up resets the exploration budget so a changed direction can be investigated normally.
The plugin reuses DSH's built-in planMode, create_goal, todo_write, ask_user_question, and repeat-tool-reminder capabilities. It does not register duplicate tools or modify session files directly.
Installation
dsh plugin --profile web add -w think-like-codex
dsh --profile web
This package is a DSH Profile Bundle. Its bundled cordis.patch.yml mounts the plugin automatically. Replace web with another profile name when needed. Restart the profile host after updating host-side code.
Configuration
The bundle installs the following defaults:
- 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
Override the complete config object for think-like-codex in your profile's cordis.patch.yml if you need different thresholds.
Uninstallation
dsh plugin --profile web remove think-like-codex
License
MIT
GanyuanRan/Aegis
Jayden-X-L/forkprobe
FeatherHunter/dsh-mattpocock-skills-deck
Soulize/hard-flash
dushaobindoudou/dsh-acp