常伴 · 守护 · 可信 · 秉正 Always Present · Always Guarding · Verified Trust · No Compromise
从项目第一行代码到最终交付，胖鱼始终相伴。
不是你想起来才调用的工具集，是每一轮交互都参与的AI伙伴。
From first commit to final delivery, PEtFiSh is always there.
Not a toolbox you remember to call — a companion in every interaction.
胖鱼不是工具。工具被调用，胖鱼始终相伴。PEtFiSh is not a tool. Tools get called. PEtFiSh is always there.
Companion Gateway在每条消息前自动执行话题检测和能力感知。不需要你主动调用，不需要特殊命令。从/initproject那一刻起，胖鱼就是你工作流的一部分——沉默地同行，在该出现的时刻出现。
Companion Gateway runs automatically before every message — topic detection and capability sensing. No commands needed. From the moment you run /initproject, PEtFiSh becomes part of your workflow — walking alongside silently, showing up when it matters.
检测到你需要部署能力但没装deploy pack？提醒一次。检测到上下文被跨话题污染、风险评分超过60？主动拦截。质量不过关？挡在发布前面。不是事后复盘，是实时守护。 Need deployment skills but don't have the deploy pack? Reminded once. Context contamination score above 60? Intercepted. Quality below threshold? Blocked before publish. Not a post-mortem — real-time protection.
胖鱼推荐的每个Skill都经过质量门禁：Lint 100分制打分、安全审计0.0–1.0风险评分、四条红线硬拒绝。TrustSkills治理引擎六维风险评估、五级动作裁决。信任不是宣称的，是审计出来的。 Every skill PEtFiSh recommends passes quality gates: Lint scoring (100-point scale), security audit (0.0-1.0 risk score), four red lines that trigger instant denial. TrustSkills engine: 6 risk dimensions, 5 action levels. Trust isn't claimed — it's audited.
反迎合校准：你问"这个方案行不行"，胖鱼先中性化问题再给结论，至少补一个反方。质量门禁不放水，红线命中即deny不走阈值。推你往更高标准走，而不是让你舒服。 Anti-sycophancy calibration: when you ask 'is this plan good?', PEtFiSh neutralizes the question first, then gives a conclusion — always with at least one counter-argument. Quality gates don't budge. Red lines mean denial, no score threshold negotiation. Pushes you toward higher standards, not comfort.
安装 init + companion pack。输入/initproject，胖鱼询问项目类型，自动安装匹配的能力包，完成配置。从此刻起，伙伴已就位。
Install init + companion pack. Run /initproject — PEtFiSh asks your project type, auto-installs matching capability packs. From this moment, your companion is in place.
Companion Gateway在每条消息前自动执行六步：读取项目模式 → 话题漂移检测 → 上轮失败信号捕获 → 能力缺口感知 → 反迎合校准 → 正常处理。你感觉不到它在，但它一直在——守护上下文、补齐能力、阻止迎合。 Companion Gateway runs six steps before every message: read project mode → detect topic drift → catch previous-turn failures → sense capability gaps → anti-sycophancy check → normal processing. You don't feel it — but it's always there, guarding context, filling gaps, blocking sycophancy.
三层感知模型：Tier 1 关键词白名单匹配已知领域 → Tier 2 意图级缺口检测（需要邮件服务？图表工具？监控集成？）→ Tier 3 无缺口，静默通过。每个领域每次会话最多提醒一次，不打扰。 Three-tier sensing: Tier 1 keyword whitelist for known domains → Tier 2 intent-level gap detection (need email service? chart tools? monitoring?) → Tier 3 no gap, stay silent. Each domain reminded at most once per session.
鱼迹Fish-trail话题治理器：7种话题关系类型、污染评分0-100、自动上下文隔离。当话题漂移风险超过阈值，主动提示你fork/switch/reset，而不是让混乱的上下文悄悄拖垮输出质量。 Fish-trail topic governance: 7 relationship types, contamination score 0-100, automatic context isolation. When drift risk exceeds threshold, proactively suggests fork/switch/reset instead of letting messy context silently degrade output quality.
Skill发布前必须通过：Lint ≥ 80分 → 安全审计 Risk ≤ 0.5 → 无CRITICAL → Quality Gate裁决PASS/CONDITIONAL/FAIL。四条红线命中即deny，不走分数阈值。像代码CI一样管理AI能力。 Before any skill publishes: Lint ≥ 80 → Security audit Risk ≤ 0.5 → No CRITICAL → Quality Gate verdict PASS/CONDITIONAL/FAIL. Four red lines trigger instant deny. Managing AI capabilities like code CI.
胖鱼有四个核心能力包，随仓库直接分发。此外提供十个领域扩展包，通过 petfish-market 分发——安装命令自动解析，体验与核心包一致。不够用？胖鱼可以搜索第三方市场、从任意仓库挖掘、甚至帮你直接构造新的能力包。 PEtFiSh has four core packs shipped directly. Plus ten domain packs distributed via petfish-market — install commands resolve automatically, same experience as core. Not enough? PEtFiSh can search third-party markets, mine any repo, or help you build new capability packs from scratch.
11 个领域包通过 petfish-market 分发。安装命令与核心包完全一致，用户无需关心仓库位置。支持离线回退和镜像加速。 11 domain packs distributed via petfish-market. Same install command as core packs — users never see the repo difference. Supports offline fallback and mirror acceleration.
PEtFiSh 现在有 ChatGPT GPT 版本 — 不需要本地安装也能用。不依赖 OpenCode、Codex、Cursor 或任何本地 IDE/CLI。 PEtFiSh now has a ChatGPT GPT version — no local installation needed. No dependency on OpenCode, Codex, Cursor, or any local IDE/CLI.
纯 GPT 模式 — 在 ChatGPT 内直接运行 Pure GPT mode — runs directly in ChatGPT
在线 API 模式 — 通过 API 查询 PEtFiSh 数据 Online API mode — queries PEtFiSh data via API
可选本地执行 — 可连接本地环境增强能力 Optional local execution — can connect to local environment for enhanced capabilities
注：需要本地安装或使用 Adapter 模式来访问这些能力 Note: local installation or Adapter mode required for these capabilities
安装脚本自动获取最新稳定版本。装完后输入 /initproject —— 胖鱼询问项目类型，自动安装匹配的Skill Pack，完成配置向导。从此刻起，伙伴已就位。
旧版Shell安装器（install.sh / install.ps1）已废弃。
Install script auto-resolves the latest stable version. After install, type /initproject — PEtFiSh asks your project type, auto-installs matching skill packs, completes setup wizard. From this moment, your companion is in place.
Legacy shell installers (install.sh / install.ps1) are deprecated.
PetFish Remote — 用 Telegram 遥控你的 AI 编程助手。随时随地查看进度、发送指令、接收通知。PetFish Remote — control your AI coding assistant from Telegram. Check progress, send commands, receive notifications anytime, anywhere.
PetFish RemoteBeta