pi-advisor:给 Pi 编码智能体加可配置的顾问/执行双模型工作流
开源插件 pi-advisor 让 Pi 编码智能体在关键决策、卡死与收尾时呼叫第二模型顾问,执行权仍留在 Executor,MIT 许可。
开发者 Dan Kornas 发布面向 Pi 编码智能体的开源插件 pi-advisor,把「顾问模型」和「执行模型」拆开:实现继续由 Executor 负责,需要拍板、反复失败或宣告完成前,可再请 Advisor 做第二意见。
帖文列出的能力包括:通过 ask_advisor 或 /advisor-manual 按需咨询;可在出计划前、连续失败后或完成前配置审查门禁;对重复工具调用做环路检测,并给出继续、修订或阻断;Executor、Advisor 与可选回退模型可分开配置,同模型咨询默认跳过;还可控制对话历史、仓库上下文、受信任的 AGENTS.md、文件/图片交接与密钥脱敏。项目为 MIT 许可,安装与文档链接放在帖子回复中。
对已在用 Pi 或类似双模型编排的团队,这把「贵模型只在决策点介入」产品化,有助于控制成本并减少执行模型独自钻牛角尖;落地时仍需验证与本机 Pi 版本的兼容性,以及审查门禁会不会过度打断流畅改码。若团队已有人工代码评审节奏,也可把 Advisor 输出当作预审意见,再决定哪些结论需要人类确认。
pi-advisor Advisor/Executor plugin for Pi agents
When your coding agent gets stuck, get a second opinion without handing over the work.
pi-advisor is a configurable Advisor and Executor workflow plugin for Pi coding agents.
It helps you keep implementation with the Executor while bringing a second model into consequential decisions, stalled work, and final reviews.
Key features:
• On-demand advice – request a second opinion through ask_advisor or /advisor-manual
• Configurable review gates – run reviews before plans, after repeated failures, or before declaring completion
• Loop detection – evaluate repeated tool calls with explicit proceed, revise, or blocked decisions
• Separate model controls – configure Executor, Advisor, and optional fallback models; same-model consultations are skipped by default
• Context controls – configure conversation history, repository context, trusted AGENTS.md rules, file/image handoff, and secret redaction
It’s open-source (MIT license).
Link in the reply 👇