Pi Windows 工作流第二版：从第一版配置升级到现在有哪些改变？

xk128 2026-07-24 15:23 1

如果你还没看第一版，建议先回看同系列的《Pi 从零安装到入门配置：Windows 实战教程》；这版是在经历了 Pi 和 package 迭代，结合第一版的工作流深入体验后遇到的一些问题的迭代更新。不过 package 和 Pi 本体更新很快，升级前仍建议先看各自的 release 和 README。

编写时 Pi 版本：0.81.1；

更正模型窗口和成本

第一版里我直接搬了 GPT-5.6 API 的最大上下文，把 contextWindow 写成了 1050000。然而许多中转站能不能真按官方 API 全量提供百万窗口。而且现在无论是 OpenAI 和 Pi 都不推荐超过 272k ，所以后面经过迭代后把给 GPT-5.6 Luna、Terra、Sol 的信息都改成了 272000 上下文、128000 最大输出。

同时调整主压缩线到 217600，因为 272000 × 0.8 = 217600。

models.json

{

  "providers": {

    "openai": {

      "baseUrl": "<YOUR_OPENAI_COMPAT_BASE_URL>"

    },

    "anthropic": {

      "baseUrl": "<YOUR_ANTHROPIC_COMPAT_BASE_URL>"

    }

  }

}

上下文管理

现在我通过原生 + 三个额外 package 做上下文压缩管理

Pi 原生 compaction 负责保留最近对话和输出空间；

pi-observational-memory 负责从长过程里提炼可复用的事实和决策；

context-mode 尽量让无关内容别进窗口；

pi-ultra-compact 则在中高占用阶段做分级收缩。

pi-ultra-compact 是新增的 package，它会依据 Pi 当前模型 metadata 做分级压缩：中段优先丢弃低价值推理和工具输出，高占用时再做完整总结。它补的是长会话中的预压缩和缓存友好整理。

当前 compaction 与 observational-memory 配置

{

  "compaction": {

    "enabled": true,

    "reserveTokens": 32768,

    "keepRecentTokens": 48000

  },

  "observational-memory": {

    "observeAfterTokens": 48000,

    "reflectAfterTokens": 144000,

    "compactAfterTokens": 217600,

    "compactAfterTokensMode": "ratio",

    "compactAfterTokensRatio": 0.8,

    "observationsPoolMaxTokens": 32000,

    "observationsPoolTargetTokens": 16000,

    "agentMaxTurns": 16,

    "model": {

      "provider": "openai",

      "id": "gpt-5.6-luna",

      "thinking": "low"

    },

    "passive": false,

    "debugLog": false

  }

}

温馨提醒：模型 id、上下文、输出上限和价格随时可能变，升级前优先看 Pi 模型目录。

一些 package 改动

@gotgenes/pi-permission-system

第一版里我为了让 FFF 接管搜索，把原生 grep 和 find 直接设成 deny。后来发现 Plan Mode、常规只读审查和子agents都需要稳定（或者只支持）原生读工具，按原来设置会直接在plan mode和审核等子agents无法使用相关读工具。因此现在全局读工具默认 allow，FFF 作为额外高性能工具并存。

当前 pi-permission-system 脱敏配置

{

  "$schema": "https://raw.githubusercontent.com/gotgenes/pi-packages/main/packages/pi-permission-system/schemas/permissions.schema.json",

  "debugLog": false,

  "permissionReviewLog": false,

  "yoloMode": false,

  "piInfrastructureReadPaths": [

    "~/.pi/agent",

    "~/.pi/agent/*",

    "~/AppData/Roaming/npm/node_modules/@earendil-works",

    "~/AppData/Roaming/npm/node_modules/@earendil-works/*"

  ],

  "permission": {

    "*": "allow",

    "read": "allow",

    "grep": "allow",

    "find": "allow",

    "ls": "allow",

    "external_directory": {

      "*": "ask",

      "~/.pi/agent": "allow",

      "~/.pi/agent/*": "allow"

    },

    "path": {

      "*": "allow",

      "*.env": "deny",

      "*.env.*": "deny",

      "*.env.example": "allow",

      "~/.ssh/*": "deny",

      "$HOME/.ssh/*": "deny",

      "~/.pi/agent/auth.json": "deny",

      "~/.pi/agent/models.json": "deny",

      "~/.codex/auth.json": "deny"

    },

    "bash": {

      "*": "allow",

      "Remove-Item *": "ask",

      "rm *": "ask",

      "del *": "ask",

      "erase *": "ask",

      "rd *": "ask",

      "rmdir *": "ask",

      "ri *": "ask",

      "mv *": "ask",

      "Move-Item *": "ask",

      "Rename-Item *": "ask",

      "Clear-Content *": "ask",

      "Set-Content *": "ask",

      "Out-File *": "ask",

      "cmd /c del *": "ask",

      "cmd /c erase *": "ask",

      "cmd /c rd *": "ask",

      "cmd /c rmdir *": "ask",

      "cmd.exe /c del *": "ask",

      "cmd.exe /c erase *": "ask",

      "cmd.exe /c rd *": "ask",

      "cmd.exe /c rmdir *": "ask",

      "powershell *Remove-Item*": "ask",

      "powershell.exe *Remove-Item*": "ask",

      "pwsh *Remove-Item*": "ask",

      "pwsh.exe *Remove-Item*": "ask",

      "powershell *Move-Item*": "ask",

      "powershell.exe *Move-Item*": "ask",

      "pwsh *Move-Item*": "ask",

      "pwsh.exe *Move-Item*": "ask",

      "powershell *Rename-Item*": "ask",

      "powershell.exe *Rename-Item*": "ask",

      "pwsh *Rename-Item*": "ask",

      "pwsh.exe *Rename-Item*": "ask",

      "powershell *Clear-Content*": "ask",

      "powershell.exe *Clear-Content*": "ask",

      "pwsh *Clear-Content*": "ask",

      "powershell.exe *Clear-Content*": "ask",

      "pwsh *Set-Content*": "ask",

      "powershell.exe *Set-Content*": "ask",

      "pwsh *Set-Content*": "ask",

      "pwsh.exe *Set-Content*": "ask",

      "powershell *Out-File*": "ask",

      "powershell.exe *Out-File*": "ask",

      "pwsh *Out-File*": "ask",

      "pwsh.exe *Out-File*": "ask",

      "git status": "allow",

      "git status *": "allow",

      "git diff *": "allow",

      "git log *": "allow",

      "git show *": "allow",

      "git blame *": "allow",

      "git ls-files *": "allow",

      "git ls-tree *": "allow",

      "git rev-parse *": "allow",

      "git describe *": "allow",

      "git merge-base *": "allow",

      "git cat-file *": "allow",

      "git check-ignore *": "allow",

      "git check-attr *": "allow",

      "git branch": "allow",

      "git branch --show-current": "allow",

      "git tag --list": "allow",

      "git tag --list *": "allow",

      "git remote -v": "allow",

      "git remote get-url *": "allow",

      "git config --get *": "allow",

      "git config --get-all *": "allow",

      "git config --list": "allow",

      "git config --show-origin --list": "allow",

      "git stash list": "allow",

      "git stash show *": "allow",

      "git worktree list": "allow",

      "git submodule status *": "allow",

      "git clean *": "ask",

      "git reset --hard *": "ask",

      "git checkout -- *": "ask",

      "git restore *": "ask"

    }

  }

}

pi-subagents 和 @narumitw/pi-plan-mode

我最早装 pi-subagents 时，只是给每个内置角色分了模型和思考强度。但是会存在一个问题：我的工作流里子agents只负责读操作，给 Plan Mode 开放 subagents 工具以后，模型理论上也能叫到普通 worker；而 worker 为了正常开发本来就必须有 bash、edit 和 write，把全局权限收紧则会毁掉常规工作流，单靠“Plan 不要进行写操作”原来只能靠提示词。但后面随着 pi-plan-mode 迭代新增了 allowedPlanSubagents ，这个配置允许 Plan 里即便启用了 subagent，也只能启动我明确允许的只读角色。

因此现在我的解决方案变成了:

启用 Plan Mode 的角色白名单；

新增 plan-scout、plan-researcher、plan-reviewer 三个独立子 agent ；

通过每个 agent frontmatter 里的 permission: 默认给plan专属的三个独立子 agent 拒绝写操作。

这样 Plan 可以并行收集证据，但不能把任务偷换给普通可写角色。

pi-plan-mode.json

{

  "thinkingLevel": "inherit",

  "defaultPlanTools": [

    "read",

    "bash",

    "grep",

    "find",

    "ls",

    "subagent"

  ],

  "allowedPlanSubagents": [

    "plan-scout",

    "plan-researcher",

    "plan-reviewer"

  ],

  "safeSubcommands": {

    "git": [

      "rev-parse",

      "blame",

      "describe",

      "merge-base",

      "ls-tree",

      "cat-file"

    ]

  }

}

注意allowedPlanSubagents 是大小写敏感的精确名称匹配，Plan 会在 child process 启动前检查普通、并行、chain、fan-in 和 spawn 请求；不在名单里的角色会整次被拒绝。它不是 sandbox，所以白名单里的 agent 本身仍然必须保持只读。

~/.pi/agent/agents/plan-scout.md

---

name: plan-scout

description: Fast, read-only codebase reconnaissance for a planning session.

tools: read, grep, find, ls

extensions:

model: openai/gpt-5.6-luna

thinking: low

systemPromptMode: replace

inheritProjectContext: true

inheritSkills: false

defaultContext: fresh

acceptanceRole: read-only

permission:

  "*": deny

  read: allow

  grep: allow

  find: allow

  ls: allow

  bash: deny

  edit: deny

  write: deny

  mcp: deny

  skill: deny

  external_directory: deny

  special: deny

---



Treat the assigned workspace as read-only. Map the relevant files, control flow,

data contracts, existing behavior, and constraints for the parent planner.

Return exact paths, evidence, unresolved questions, and no filesystem changes.

~/.pi/agent/agents/plan-researcher.md

---

name: plan-researcher

description: Read-only technical investigation and evidence synthesis for planning.

tools: read, grep, find, ls

extensions:

model: openai/gpt-5.6-luna

thinking: medium

systemPromptMode: replace

inheritProjectContext: true

inheritSkills: false

defaultContext: fresh

acceptanceRole: read-only

permission:

  "*": deny

  read: allow

  grep: allow

  find: allow

  ls: allow

  bash: deny

  edit: deny

  write: deny

  mcp: deny

  skill: deny

  external_directory: deny

  special: deny

---



Treat the assigned workspace as read-only. Trace the requested technical

question through local code and configuration. Separate confirmed facts from

assumptions, then return evidence, risks, alternatives, and open questions.

~/.pi/agent/agents/plan-reviewer.md

---

name: plan-reviewer

description: Read-only design and change-risk review for a planning session.

tools: read, grep, find, ls

extensions:

model: openai/gpt-5.6-terra

thinking: medium

systemPromptMode: replace

inheritProjectContext: true

inheritSkills: false

defaultContext: fresh

acceptanceRole: read-only

permission:

  "*": deny

  read: allow

  grep: allow

  find: allow

  ls: allow

  bash: deny

  edit: deny

  write: deny

  mcp: deny

  skill: deny

  external_directory: deny

  special: deny

---



Treat the assigned workspace as read-only. Review the proposed direction or

existing implementation for correctness, regressions, missing evidence, and

scope risks. Return exact paths and actionable findings without changing state.

settings.json 中的 subagents 配置

{

  "subagents": {

    "defaultModel": "gpt-5.6-luna",

    "modelScope": {

      "allow": [

        "openai/gpt-5.5",

        "openai/gpt-5.6-*"

      ],

      "enforce": true

    },

    "agentOverrides": {

      "delegate": {

        "model": "gpt-5.6-luna",

        "thinking": "low",

        "fallbackModels": ["gpt-5.6-terra"],

        "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]

      },

      "scout": {

        "model": "gpt-5.6-luna",

        "thinking": "low",

        "fallbackModels": ["gpt-5.6-terra"],

        "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]

      },

      "context-builder": {

        "model": "gpt-5.6-luna",

        "thinking": "low",

        "fallbackModels": ["gpt-5.6-terra"],

        "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]

      },

      "researcher": {

        "model": "gpt-5.6-luna",

        "thinking": "low",

        "fallbackModels": ["gpt-5.6-terra"],

        "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]

      },

      "planner": {

        "model": "gpt-5.6-terra",

        "thinking": "medium",

        "inheritProjectContext": true,

        "fallbackModels": ["gpt-5.6-luna"],

        "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]

      },

      "worker": {

        "model": "gpt-5.6-terra",

        "thinking": "medium",

        "inheritProjectContext": true,

        "fallbackModels": ["gpt-5.6-luna"],

        "tools": ["read", "grep", "find", "ls", "bash", "edit", "write", "contact_supervisor"]

      },

      "reviewer": {

        "model": "gpt-5.6-luna",

        "thinking": "medium",

        "inheritProjectContext": true,

        "fallbackModels": ["gpt-5.6-terra"],

        "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]

      },

      "oracle": {

        "model": "gpt-5.6-terra",

        "thinking": "medium",

        "inheritProjectContext": true,

        "fallbackModels": ["gpt-5.6-luna"],

        "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]

      }

    }

  }

}

~/.pi/agent/extensions/subagent/config.json

{

  "toolDescriptionMode": "compact",

  "asyncByDefault": false,

  "waitTool": {

    "enabled": true

  },

  "forceTopLevelAsync": false,

  "globalConcurrencyLimit": 6,

  "parallel": {

    "maxTasks": 6,

    "concurrency": 6

  },

  "maxSubagentSpawnsPerSession": 24,

  "maxSubagentDepth": 1,

  "completionBatch": {

    "enabled": true,

    "debounceMs": 150,

    "maxWaitMs": 1000,

    "stragglerDebounceMs": 75,

    "stragglerMaxWaitMs": 400,

    "stragglerWindowMs": 2000

  },

  "worktreeBaseDir": "C:\\Users\\<YOUR_USER>\\.pi\\agent\\repos\\subagents-worktrees"

}

AGENTS.md 迭代

随着我从 codex 切到 Pi，AGENTS.md 就向着更加通用和精简的方向去迭代，完善了关于并行、子 agents 为什么只做只读调查、Skills/MCP 如何按任务主动调度、格式化噪声如何处理、什么时候只做静态核验的约束。

特别是使用 5.6 不约束清楚， 5.6 老是给我自作主张很烦人。

AGENTS.md.txt (34.0 KB)

参考了

5.6 Sol 必须知道的技巧 - 开发调优 - LINUX DO

【拯救 5.6 Sol（1）】开箱即用、快速高效、减少上下文腐烂的Codex子代理实践 - 开发调优 - LINUX DO

完整 settings.json

完整脱敏 settings.json

{

  "lastChangelogVersion": "0.81.1",

  "shellPath": "C:\\Program Files\\PowerShell\\7\\pwsh.exe",

  "externalEditor": "code --wait",

  "defaultProvider": "openai",

  "defaultModel": "gpt-5.6-terra",

  "subagents": {

    "defaultModel": "gpt-5.6-luna",

    "modelScope": {

      "allow": ["openai/gpt-5.5", "openai/gpt-5.6-*"],

      "enforce": true

    },

    "agentOverrides": {

      "delegate": {"model": "gpt-5.6-luna", "thinking": "low", "fallbackModels": ["gpt-5.6-terra"], "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]},

      "scout": {"model": "gpt-5.6-luna", "thinking": "low", "fallbackModels": ["gpt-5.6-terra"], "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]},

      "context-builder": {"model": "gpt-5.6-luna", "thinking": "low", "fallbackModels": ["gpt-5.6-terra"], "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]},

      "researcher": {"model": "gpt-5.6-luna", "thinking": "low", "fallbackModels": ["gpt-5.6-terra"], "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]},

      "planner": {"model": "gpt-5.6-terra", "thinking": "medium", "inheritProjectContext": true, "fallbackModels": ["gpt-5.6-luna"], "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]},

      "worker": {"model": "gpt-5.6-terra", "thinking": "medium", "inheritProjectContext": true, "fallbackModels": ["gpt-5.6-luna"], "tools": ["read", "grep", "find", "ls", "bash", "edit", "write", "contact_supervisor"]},

      "reviewer": {"model": "gpt-5.6-luna", "thinking": "medium", "inheritProjectContext": true, "fallbackModels": ["gpt-5.6-terra"], "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]},

      "oracle": {"model": "gpt-5.6-terra", "thinking": "medium", "inheritProjectContext": true, "fallbackModels": ["gpt-5.6-luna"], "tools": ["read", "grep", "find", "ls", "bash", "contact_supervisor"]}

    }

  },

  "defaultThinkingLevel": "high",

  "hideThinkingBlock": false,

  "theme": "codex-graphite",

  "quietStartup": true,

  "collapseChangelog": true,

  "outputPad": 1,

  "editorPaddingX": 1,

  "autocompleteMaxVisible": 8,

  "showHardwareCursor": true,

  "enabledModels": ["gpt-5.5", "gpt-5.6-luna", "gpt-5.6-terra", "gpt-5.6-sol"],

  "packages": [

    "npm:pi-mcp-adapter",

    "npm:context-mode",

    "npm:pi-until-done",

    "npm:@narumitw/pi-plan-mode",

    "git:github.com/daynin/nano-context",

    "npm:@juicesharp/rpiv-advisor",

    "npm:pi-simplify",

    "npm:pi-markdown-preview",

    "npm:@juicesharp/rpiv-ask-user-question",

    "npm:@juicesharp/rpiv-i18n",

    "npm:pi-rewind",

    "npm:@gotgenes/pi-permission-system",

    "npm:pi-ultra-compact",

    "npm:pi-observational-memory",

    "npm:pi-wtf",

    "npm:@ff-labs/pi-fff",

    "npm:pi-web-access",

    "npm:@juicesharp/rpiv-todo",

    "npm:pi-zentui",

    "npm:pi-subagents",

    "npm:pi-di18n",

    "npm:pi-lens"

  ],

  "compaction": {"enabled": true, "reserveTokens": 32768, "keepRecentTokens": 48000},

  "observational-memory": {"observeAfterTokens": 48000, "reflectAfterTokens": 144000, "compactAfterTokens": 217600, "compactAfterTokensMode": "ratio", "compactAfterTokensRatio": 0.8, "observationsPoolMaxTokens": 32000, "observationsPoolTargetTokens": 16000, "agentMaxTurns": 16, "model": {"provider": "openai", "id": "gpt-5.6-luna", "thinking": "low"}, "passive": false, "debugLog": false},

  "thinkingBudgets": {"minimal": 1024, "low": 4096, "medium": 12288, "high": 32768, "xhigh": 65536},

  "steeringMode": "one-at-a-time",

  "followUpMode": "one-at-a-time",

  "transport": "auto",

  "images": {"autoResize": true, "blockImages": false},

  "terminal": {"showImages": true, "imageWidthCells": 72, "clearOnShrink": false},

  "retry": {"enabled": true},

  "enableSkillCommands": true,

  "defaultProjectTrust": "ask",

  "treeFilterMode": "default",

  "doubleEscapeAction": "tree",

  "httpIdleTimeoutMs": 300000

}

最新回复 (7)

xk128 楼主 07-24 15:25

1楼

服了，怎么刚写完 Pi 就更新 0.82 了
aiia 07-24 15:26

2楼

随着我从 codex 切到 Pi，AGENTS.md 就向着更加通用和精简的方向去迭代，完善了关于并行、子 agents 为什么只做只读调查、Skills/MCP 如何按任务主动调度、格式化噪声如何处理、什么时候只做静态核验的约束。

特别是使用 5.6 不约束清楚， 5.6 老是给我自作主张很烦人。

最近看到pi出现变多了，真的有说法吗
xk128 楼主 07-24 15:29

3楼

Codex 、OpenCode Vibe coding 史山代码 16G 内存不够它吃，特别是一拉 subagents 更是卡的要命， Claude Code 后门前科。

同配置 Pi 开三个窗口并行都没问题不会爆内存，而且更青联/可控/ Token 消耗少更强不香吗？
miku 07-24 15:33

4楼

我也在用的pi，佬有没有遇到莫名奇妙对话跳到了置顶，还得按enter才能回来的情况
xk128 楼主 07-24 15:38

5楼

有这个问题，我发现是 subagents 容易触发，但是似乎是 Windows 终端的 BUG，暂时无解
三四 07-24 16:23

6楼

pi在windows下命令调用有问题，经常整出一些Linux下的东西。插件装多了bug也很多，好处是有什么问题都可以自己修，适合能折腾的人。
霁天 07-24 16:24

7楼

这里指的是具身算法的模型 $\pi$ 系列吗？

* 帖子来源Linux.do

附近帖子

↑Gemini APP 限制会话长度了？？
↑“男子献血后口吐白沫、神志不清”，当地通报：成立调查组
↑假设真的有一天你财富自由了会做什么
↑Axonhub 无法使用 OpenCode Go 的解决办法
↑GLM下午用的太快了
📍 Pi Windows 工作流第二版：从第一版配置升级到现在有哪些改变？
↓想要电路分析skill
↓GROK解决三十年前的数学难题
↓佬友们“她”那个叫三上悠亚日本的小姑娘你还记得吗？
↓重新升回三级，纪念一下
↓【富可敌国】昨天被攻击的太猛，CUN.AI 和L站都要科学上网了，继续给佬友们发兑换码！！888张！！

飞读

xk128

主题数
1

帖子数
1

注册排名
3

随机推荐