Skip to content

feat: add Atlas series — goal-integrity skills for long-running agent tasks - #8

Open
wede-wx wants to merge 1 commit into
XiaomiMiMo:mainfrom
wede-wx:feat/atlas-series
Open

feat: add Atlas series — goal-integrity skills for long-running agent tasks#8
wede-wx wants to merge 1 commit into
XiaomiMiMo:mainfrom
wede-wx:feat/atlas-series

Conversation

@wede-wx

@wede-wx wede-wx commented Jun 11, 2026

Copy link
Copy Markdown

新增两个 Skill

atlas-contract

防止 AI agent 在执行过程中静默偏离用户目标。核心机制:

  • 根据任务复杂度自动选择治理力度(Light / Medium / Heavy)
  • Heavy 任务强制生成目标合同 + 阶段账本(最多 4 个阶段),每阶段停止确认
  • 内置硬性 Heavy 锚点,识别多步骤请求、rework、端到端任务等高漂移场景
  • Final Audit 对所有合同项进行对抗性自检

atlas-ledger

atlas-contract 的记忆层,将捕获的偏离编译为项目级规则:

  • Final Audit 发现硬偏离后自动触发,无需用户手动召唤
  • 将偏离蒸馏为 WHEN / DON'T / INSTEAD 条款,写入 Atlas.md
  • 四道验收闸确保条款可触发、可泛化、不误伤
  • 每次捕获都加厚项目防线,复利效应

兼容性

  • description ≤ 230 字符,兼容 Claude Code(250 上限)和 Codex(1024 上限)
  • 纯 prompt 无依赖,适配所有 SKILL.md 兼容的 agent
  • MIT 协议

相关项目

https://github.com/wede-wx/atlas

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant