为什么我给AI找了个伙伴Why I Got My AI a Companion

2026年5月 · v0.6.2May 2026 · v0.6.2

我每天花大量时间和AI一起写代码。AI很能干，但有个问题——它是个工具。I spend a lot of time every day writing code with AI. AI is very capable, but there's a problem — it's a tool.

工具的特点是：你调用它，它干活，干完走人。下次再调用，它不记得上次发生了什么。它不会主动告诉你"你现在方向偏了"，不会在你缺某个能力的时候说一句"要不装个这个"，也不会在你写出一个有问题的方案时拦住你。Tools have a defining trait: you call them, they work, they leave. Next time you call, they don't remember what happened. They won't tell you 'you're drifting off track', won't say 'maybe install this capability' when you're missing something, won't stop you when your plan has a flaw.

你得自己记住上下文，自己发现能力缺口，自己守住质量底线。AI越能干，这些负担反而越重——因为它干得快，你追得累。You have to remember context yourself, discover capability gaps yourself, hold quality standards yourself. The more capable AI gets, the heavier this burden — because it moves fast, and you struggle to keep up.

我想要的不是更多工具，是一个伙伴。What I wanted wasn't more tools. It was a companion.

伙伴和工具的区别很简单：工具被调用，伙伴始终在场。The difference between a companion and a tool is simple: tools get called, companions are always there.

胖鱼是什么What is PEtFiSh

胖鱼是我给AI工作流找的伙伴。从你输入/initproject的那一刻起，它就在了。不是在后台沉睡等你想起来调用——而是每一轮对话都参与，沉默地做两件事：PEtFiSh is the companion I found for my AI workflow. From the moment you type /initproject, it's there. Not sleeping in the background waiting for you to remember to call it — but participating in every conversation, silently doing two things:

看看你在聊什么Watching what you're discussing——话题漂了吗？上下文被污染了吗？需要提醒你吗？ — has the topic drifted? Is context contaminated? Should it alert you?
看看你缺什么Watching what you're missing——你在聊部署但没装部署能力包？你需要发邮件但环境里没有邮件集成？ — discussing deployment but don't have the deploy pack? Need to send emails but no email integration?

这两件事做完，才进入正常的工作处理。你感觉不到它在——直到它该出现的时候出现。These two checks complete, then normal work begins. You don't feel it — until the moment it needs to appear.

四个字Four Words

我给胖鱼定了四个字：常伴、守护、可信、秉正。I gave PEtFiSh four words: Always Present, Guarding, Verified Trust, No Compromise.

常伴Always Present——每一轮交互都在。不是"你可以调用/petfish命令"，而是"不管你调不调用，胖鱼始终相伴"。Companion Gateway在每条消息前自动执行，你不需要做任何额外操作。 — every interaction, it's there. Not "you can run /petfish" but "whether you call it or not, PEtFiSh is watching". Companion Gateway runs automatically before every message. You don't have to do anything extra.

守护Guarding——在你掉坑之前拦住。上下文污染到60分了？拦住，告诉你话题漂了。你需要某个能力但没装？提醒一次，不多不少。质量不过关？挡在发布前面。 — pulls you back before you fall in. Context contamination hits 60? It stops you, tells you the topic drifted. Missing a capability? One reminder, no more. Quality doesn't pass? Blocked before publish.

可信Verified Trust——胖鱼推荐给你的每一个能力包，都经过了lint打分（100分制）、安全审计（风险评分0.0-1.0）、红线检测（命中即拒绝，不商量）。信任不是嘴上说说的，是验证出来的。 — every pack PEtFiSh recommends has gone through lint scoring (100-point scale), security audit (risk score 0.0-1.0), and red-line checks (hit one, rejected, no negotiation). Trust isn't a claim. It's verified.

秉正No Compromise——胖鱼不顺着你说。你问"这方案行不行"，它先中性化问题再给结论，至少给你一个反方向的思考。质量门禁80分以下不过就是不过，不会因为"这次赶工"就放水。 — PEtFiSh doesn't just agree with you. Ask "is this plan okay?" and it neutralizes the question before giving a verdict, always offering at least one counterpoint. Quality gate below 80? It doesn't pass. "We're rushing this time" doesn't change that.

><(((^>

它具体做了什么What it actually does

说点实在的。Let me be concrete.

Companion Gateway——注入到你的AI指令文件里（AGENTS.md / CLAUDE.md / 等等），在AI处理你每条消息之前强制执行话题检测和能力感知。不需要你记得去调用它。 — injected into your AI instructions file (AGENTS.md / CLAUDE.md / etc.), it enforces topic detection and capability sensing before the AI processes every single message. You don't need to remember to call it.

三层能力感知Three-tier sensing——第一层，关键词匹配已知领域（你说"部署"它知道推荐deploy pack）。第二层，意图级分析（你说"帮我发个邮件通知团队"，它检测到这需要邮件集成，你没有）。第三层，什么都没检测到，闭嘴。每个领域每次会话最多提醒一次。 — tier one, keyword match against known domains (say "deploy" and it knows to recommend the deploy pack). Tier two, intent-level analysis (say "send an email to the team" and it detects you need email integration but don't have it). Tier three, nothing detected, stays quiet. Each domain, at most one reminder per session.

Fish-trail话题治理 topic governance——追踪你的话题关系（继续？分叉？切换？合并？），给上下文污染打分（0-100），在风险高的时候主动建议你处理。31个MCP工具支撑底层状态管理。 — tracks your topic relationships (continue? fork? switch? merge?), scores context contamination (0-100), and proactively suggests you deal with it when risk is high. 31 MCP tools handle the underlying state management.

质量门禁Quality gates——Lint 40+条规则打分，安全审计检测shell注入/硬编码凭证/危险网络调用，四条红线（subprocess+network组合、敏感路径读写、无审批sudo、subprocess+os组合）命中即deny。像代码CI一样管理AI能力。 — lint scores against 40+ rules, security audit catches shell injection / hardcoded credentials / dangerous network calls, four red lines (subprocess+network combo, sensitive path read/write, unapproved sudo, subprocess+os combo) trigger an instant deny. Managing AI capabilities like code CI.

反迎合校准Anti-sycophancy——可以叠加到任何评审类任务上。code review + calibrate，方案评估 + calibrate。先中性化再给结论，结论和置信度分开表达。 — stacks onto any review task. Code review + calibrate. Strategy evaluation + calibrate. Neutralize first, then conclude. Conclusion and confidence expressed separately.

不只是我自己的能力包Not just my own packs

胖鱼自带四个核心包（companion、context、trust、calibrate）和六个领域包（init、course、deploy、petfish、testdocs、ppt）。但它不是一个封闭的10个包的集合。PEtFiSh ships with four core packs (companion, context, trust, calibrate) and six domain packs (init, course, deploy, petfish, testdocs, ppt). But it's not a closed collection of 10 packs.

/petfish search 可以跨6个来源搜索第三方skill。 searches third-party skills across 6 sources. /petfish mine 可以从任意GitHub仓库挖掘可skill化的模块。 mines any GitHub repo for extractable skill patterns. /petfish create 可以帮你从零构造新的能力包。 scaffolds new packs from scratch. /petfish gate 确保你构造出来的东西质量过关。 makes sure what you build actually passes muster.

胖鱼管的不只是"我做的这些包"，而是你整个AI能力体系的质量和扩展。PEtFiSh isn't just managing "the packs I made" — it's managing the quality and growth of your entire AI capability system.

8个平台，一只胖鱼8 platforms, one PEtFiSh

OpenCode、Claude Code、Cursor、Copilot、Windsurf、Codex、Antigravity——不管你用哪个，胖鱼都能装上去。一套skill写一次，自动翻译成各平台的格式。OpenCode, Claude Code, Cursor, Copilot, Windsurf, Codex, Antigravity — whichever you use, PEtFiSh installs. Write skills once, auto-translated to each platform's format.

你不用操心"Cursor的规则放哪个目录""Claude Code的指令文件叫什么名字"。胖鱼知道。You don't need to worry about 'which directory do Cursor rules go in' or 'what's Claude Code's instructions file called'. PEtFiSh knows.

30秒装上30-second install

# macOS / Linux / WSL
curl -fsSL https://raw.githubusercontent.com/kylecui/petfish.ai/master/remote-install.sh \
  | bash -s -- --pack init,companion --detect

# Windows PowerShell
& ([scriptblock]::Create((irm https://raw.githubusercontent.com/kylecui/petfish.ai/master/remote-install.ps1))) -Pack "init,companion" -Detect

装完输入/initproject。胖鱼问你项目类型，帮你装好匹配的能力包。从这一刻起，它就在了。After install, type /initproject. PEtFiSh asks your project type, installs matching packs. From this moment, it's there.

><(((^>

为什么叫"伙伴"Why "companion"

因为它确实像一个。Because it genuinely is one.

伙伴不需要你记得"去调用"它——它就在那儿A companion doesn't need you to remember to 'call' it — it's just there
伙伴在你要犯错的时候拉一把——不是事后说"早知道"A companion pulls you back before you make a mistake — not 'I told you so' after the fact
伙伴推荐给你的东西是经过验证的——不是随手丢个链接A companion recommends things that are verified — not just tossing you a random link
伙伴不顺着你说——该说不行的时候说不行A companion doesn't just agree — says 'no' when it should

工具用完就走。伙伴从头到尾都在。Tools leave when done. Companions stay from start to finish.

如果你也觉得AI写代码越来越快、但管理AI行为越来越累——试试给它找个伙伴。If you also feel that AI writes code faster and faster, but managing AI behavior gets more exhausting — try getting it a companion.

GitHub：github.com/kylecui/petfish.ai

开源，免费，欢迎PR。Open source, free, PRs welcome.

><(((^>