胖鱼 PEtFiSh — 产品介绍PEtFiSh — Product Overview

v0.6.2 · 2026-05

问题不在Skill，在协作方式The problem isn't skills — it's how you collaborate

你用AI写代码。AI很能干。但你注意到几个问题：You write code with AI. AI is capable. But you notice a few problems:

• 它没有边界感No boundaries——话题漂移了它不知道，上下文污染了它不管，你得自己记住"现在在做什么"。 — topics drift without notice, context gets contaminated, you have to remember 'what am I working on right now'.
• 它不会主动补位Won't proactively fill gaps——你开始搞部署，它不会说"你还没装部署能力包"。你缺什么得你自己发现。 — you start working on deployment, it won't say 'you don't have deployment skills installed'. You discover what's missing on your own.
• 它容易顺着你说Tends to agree with you——你提出一个有问题的方案，它说"好的，这是个很好的方向"。没人拦你。 — you propose a flawed plan, it says 'great direction!' Nobody stops you.
• 它没有质量底线No quality floor——推荐的Skill有没有安全问题？触发准确率多少？不知道，也不检查。 — are the recommended skills secure? What's the trigger accuracy? Unknown, unchecked.

这些不是"Skill管理"问题。这是协作质量问题。你需要的不是更多工具，是一个靠谱的伙伴。These aren't 'skill management' problems. They're collaboration quality problems. You don't need more tools — you need a reliable companion.

四个承诺Four Promises

常伴Always Present — 每一轮交互都在 — In every interaction

Companion Gateway在每条消息前自动执行两步检查：Companion Gateway automatically runs two checks before every message:

1. 话题检测Topic detection：当前消息和活跃话题什么关系？漂移了吗？风险多高？需要提示吗？: What's the relationship between this message and the active topic? Has it drifted? How risky? Need a prompt?
2. 能力感知Capability sensing：你正在做的事，需要什么能力？你有吗？没有的话要不要装？: What does your current task need? Do you have it? Should you install it?

然后才进入正常处理。你感觉不到它在，但它每一轮都在。Then normal processing begins. You don't feel it — but it's there every round.

守护Guarding — 在你掉坑之前拦住 — Catches you before you fall

三层感知模型：Three-tier sensing model:

• Tier 1：关键词白名单匹配已知领域（部署→deploy、课程→course、PPT→ppt...）: Keyword whitelist matching known domains (deployment→deploy, course→course, PPT→ppt...)
• Tier 2：意图级缺口检测——需要外部集成（邮件服务、图表工具、监控）且agent原生能力不覆盖时触发: Intent-level gap detection — triggered when external integration needed (email, charts, monitoring) and agent native capabilities don't cover it
• Tier 3：无缺口，静默通过。不打扰。: No gap, stay silent. Don't interrupt.

每个领域每次会话最多提醒一次。上下文守护同理：Fish-trail话题治理器持续监测话题关系和污染风险，高于阈值才出声。Each domain reminded at most once per session. Context guard works the same: Fish-trail monitors topic relationships and contamination risk, only speaks up above threshold.

可信Verified Trust — 信任是验证出来的 — Trust is earned through verification

胖鱼推荐的每个Skill都经过完整门禁：Every skill PEtFiSh recommends passes a full gate:

Lint (100分制, 40+规则)
  → Score ≥ 80? Continue
Security Audit (0.0-1.0风险分)
  → Risk ≤ 0.5? No CRITICAL? Continue
四条红线检测
  → subprocess+network combo? deny.
  → 敏感路径读写? deny.
  → sudo without approval? deny.
  → subprocess+os combo? deny.
Quality Gate → PASS / CONDITIONAL / FAIL

TrustSkills治理引擎：6维风险（shell/network/file_write/sensitive_data/script_risk/persistence）、5级动作（allow → deny）、4条红线硬拒绝。命中红线不走分数阈值，直接deny。TrustSkills governance engine: 6 risk dimensions (shell/network/file_write/sensitive_data/script_risk/persistence), 5 action levels (allow → deny), 4 red lines that trigger hard denial. Red lines bypass score thresholds — instant deny.

秉正No Compromise — 标准不因便利而降格 — Standards don't bend for convenience

反迎合校准机制：Anti-sycophancy calibration:

• 先中性化问题，再给结论——不顺着用户预设表态Neutralize the question first, then conclude — don't default to agreeing with the user's premise
• 先给评分维度，再做判断——至少补一个反方或替代方案Score on dimensions first, then judge — always include at least one counter-argument or alternative
• 结论与置信度分开表达——证据不足时明确降级Separate conclusion from confidence — explicitly downgrade when evidence is insufficient
• 可叠加到任何评审类Skill：Composable with any review skill: outline-design + calibrate、code-review + calibrate

质量门禁同理：不会因为"这个Skill看起来有用"就放水。80分以下不过，0.5以上不过，红线不商量。Quality gates work the same: won't pass because 'this skill looks useful'. Below 80 is below 80. Above 0.5 is above 0.5. Red lines are non-negotiable.

它具体做什么What it actually does

四个承诺落地为具体能力：Four promises, materialized as concrete capabilities:

10个能力包10 Capability Packs

为什么是伙伴，不是工具Why companion, not tool

工具是你想起来才用的。伙伴是一直在的。Tools are things you remember to use. Companions are always there.

• 工具等你调用——胖鱼每一轮都自动执行Gateway检查Tools wait for you to call them — PEtFiSh runs Gateway checks every round automatically
• 工具不管上下文——胖鱼追踪话题关系，防止污染累积Tools ignore context — PEtFiSh tracks topic relationships, prevents contamination buildup
• 工具不会说"不"——胖鱼在红线面前硬拒绝，在评审中不迎合Tools never say 'no' — PEtFiSh hard-rejects at red lines, doesn't sycophant in reviews
• 工具用完就走——胖鱼从init到交付全程在场，跨会话保持记忆Tools leave when done — PEtFiSh stays from init to delivery, maintains memory across sessions

Skill管理是手段。让你的AI始终处于正确的工作状态，才是目的。Skill management is the means. Keeping your AI in the right working state is the purpose.

安装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。胖鱼询问项目类型，自动装上匹配的能力包。从此刻起，伙伴已就位。After install, type /initproject. PEtFiSh asks your project type, installs matching packs. From this moment, your companion is in place.

GitHub：github.com/kylecui/petfish.ai