codestable/CodeStable

CodeStable 是一套面向严肃软件工程、坚持人在环的 AI 编码工作流。它不以编排 Agent 为中心，而是组织需求、架构、特性、问题与历史决策，让 Codex 和 Claude 驱动的开发过程可控、可追溯、可持续演进。

What it solves

CodeStable provides a lightweight set of "skill contracts" for AI-assisted software development in long-term projects. It prevents AI agents from losing context or boundaries over time, ensuring that AI actions are grounded in evidence, respect project boundaries, and integrate knowledge back into the existing project documentation rather than creating redundant, parallel systems.

How it works

Instead of using complex agent orchestration or rigid scripts, CodeStable uses a "thin harness, thick context" approach. It defines specific skills (like cs-feat for features or cs-issue for bug fixes) that constrain goals and boundaries while letting the AI model choose the execution path based on repository facts.

Key mechanisms include:

  • Evidence-based delivery: Requiring verification (e.g., "red-to-green" for bug fixes) before concluding a task.
  • Project Memory: A minimal skeleton (.codestable/) that tracks essential project facts (attention.md), validated lessons learned (lessons/), and active task cursors (work/).
  • Canonical Ownership: It avoids duplicating documentation, pushing stable conclusions back to the project's original source of truth (ADRs, code, or existing docs).

Who it’s for

  • Teams working on software projects that iterate over months or years.
  • Developers who want AI to execute efficiently while humans maintain control over product boundaries and final acceptance.
  • Teams that prioritize verifiable results, independent reviews, and knowledge reuse across different AI sessions or models.

Highlights

  • Skill-based Routing: Offers 8 specialized skills for routing actions, onboarding, feature implementation, issue diagnosis, refactoring, epic management, review, and memory management.
  • Minimalist Memory: Uses a small, structured directory to store high-value project facts and lessons without bloating the project.
  • Epic Dual-Layer Model: Separates long-term contracts (goals/decisions) from temporary execution states (progress/evidence).
  • Plugin Ecosystem: Integrates with Codex and Claude plugin marketplaces.

Related

  • Project
    zgsm-ai/costrict

    An open-source AI coding assistant for enterprises that provides standardized workflows and supports private deployment for secure development.

  • Project
    modiqo/skillspec

    A CLI and contract system that converts prose-based agent skills into structured, auditable specifications to ensure AI agents follow safety rules and provide proof of completion.

  • Dispatch
    Combatting AI Brain Drain: Deliberate Skill Development with Claude Code

    Explore how to use AI 'skills' to transform coding assistants from passive generators into active tutors, preventing the loss of codebase knowledge and cognitive atrophy.

  • Project
    jsmastery-pro/skills

    A set of modular AI agent skills that implement a professional engineering workflow, moving project state from chat sessions into repository files for persistence and team collaboration.

  • Project
    mattpocock/skills

    A collection of composable AI agent skills that apply software engineering fundamentals like TDD and domain modeling to prevent common AI coding failures.