cli-zoo：otter 一条命令拉起 tmux 工位，wren 两行实时显示 token 消耗

模型大家都在用同一批，差距在工位上。

GitHub 上有个叫 Spoon94 的用户（头像是只狗），最近维护了一个叫 cli-zoo 的仓库，攒了几只用于 AI 编程流的 Shell 小工具，每个工具以动物命名，一个目录一个脚本，symlink 安装到全局。目前收了两只：otter 和 wren。

仓库地址：github.com/Spoon94/cli-zoo，MIT 协议，Shell 写成。

---

otter：一条命令拉起整个 AI 开发会话

你用 Claude Code、Opencode 或者 Qoder CLI 写代码，通常要手动打开 tmux，分个窗格给终端，再开文件管理器，再开 git 面板……每次都要重复这套操作。

otter 解决的就是这件事：

otter -c claude          # 启动或复用 session，默认布局
otter -c opencode        # 换成 Opencode
otter -ks my-session     # 杀掉指定 session

布局固定为三格：

• 左：你选的 AI CLI 工具（claude / qodercli / opencode）
• 右上：yazi 文件管理器
• 右下：空白 shell（备用）

检测到 nvim 时额外开一个 window 给编辑器，检测到 lazygit 时再开一个 window 给 git 面板。软依赖缺失自动降级，不会因为没装 yazi 就崩。

-c 的白名单是 ALLOWED_TOOLS，当前接受 claude、qodercli、opencode 三个值。想加新的，改这个变量就行。

---

wren：两行状态栏，实时看 token 和缓存

wren 是个安装器，把一条两行状态栏接进三个宿主：Claude Code（cc）、pi 和 Qoder CLI（qc）。

安装：

./cli-zoo-install.sh wren        # 先把 wren 安装到 /usr/local/bin
wren install cc                  # 装进 Claude Code
wren install all                 # 三个宿主一起装

装完 statusline 长这样（Claude Code 侧，Dracula 配色）：

~/Code/cli-zoo | main ↑0↓0 +4 ✱2 | wC:t1:p1 | cc
↑12K ↓3K | R1.2M CH57.14% CP2 | 8.40%/200K | claude-opus-5 · high · 1h5m

两行各显示什么：

第一行：当前目录（长路径自动折叠）+ git 分支/状态 + header 在哪个 tmux pane + 宿主徽标（cc / pi / qc）

第二行：

• ↑12K ↓3K：本次会话累计发出/收到的 token
• R1.2M：缓存读取量
• CH57.14%：缓存命中率（cache hit rate）
• CP2：上下文压缩次数（context compression count）
• 8.40%/200K：当前上下文占比
• claude-opus-5 · high · 1h5m：模型 + 思考档位 + 运行时长

Qoder CLI 侧同构，数据源不同：

~/Code/cli-zoo | feat/x ↑0↓0 | wW:t1:p2 | qc
↑4.5M ↓65K | R4.2M CH98.21% | 15.00%/1M | Qwen3.8-Max · xhigh · 16m

两级安装设计值得注意：cli-zoo-install.sh 装的是 symlink（跟随仓库，改脚本即时生效）；wren install 装到宿主的是文件副本（仓库删了、移了，状态栏照常工作）。所以卸载顺序有讲究：先 wren uninstall 拆宿主接线，再 cli-zoo-uninstall.sh wren 摘本体，别搞反。

pi 里装完后用 /footer 命令切换开关。

---

工程细节

安装脚本：

git clone https://github.com/Spoon94/cli-zoo.git
cd cli-zoo
./cli-zoo-install.sh otter       # 或 wren
# 默认写 /usr/local/bin，不想动就：
PREFIX=$HOME/bin ./cli-zoo-install.sh otter

目录结构：每只动物一个目录 zoo-scripts/<name>/，入口脚本直接可执行，安装器用 ln -s 软链到 $PREFIX（wren 因为是多文件工具，软链的是入口脚本 zoo-scripts/wren/wren）。

测试：docs/testing.md 有专门的测试用例说明，覆盖安装/卸载幂等性（目标不存在时直接 exit 0）。

---

几点局限

• 星数低（1 star）：仓库新，2026-09-19 才更新，代码质量需要自己评估后再用于生产工作流。
• otter 工具白名单写死在脚本里：加新工具需改源码，没有配置文件抽象。
• wren 不同步更新：装到宿主是副本，作者改脚本后你要重跑 wren install 才会更新，不会自动拉新版。
• 只测了三个宿主：cc / pi / qc 以外的 AI CLI 工具（如 Gemini CLI）目前没有官方支持。

---

同作者的另一个仓库

顺带看了一下，Spoon94 还有个 skill-vault（3 stars，MIT，Python）：给 Claude Code、Codex CLI、Copilot CLI、Gemini CLI、OpenCode 整理了一批社区 Agent Skill，每个 skill 是自包含的 SKILL.md，遵循 Agent Skills 规范。两个仓库放在一起，方向一致——把 AI 编程工作流的”工位层”标准化。

开源代码仅供学习参考，正式使用前请自行测试评估。

---

仓库：github.com/Spoon94/cli-zoo
License：MIT | 语言：Shell
工具：otter（tmux 工位布局）、wren（AI CLI token 状态栏）

Everyone’s running the same models. The difference is the workstation.

A GitHub user called Spoon94 (dog avatar) maintains cli-zoo — a small collection of shell tools for AI coding workflows, one tool per directory, symlink-installed globally. Each tool is named after an animal. Two so far: otter and wren.

Repository: github.com/Spoon94/cli-zoo — MIT license, Shell.

---

otter: One command to spin up a full AI dev session

When you use Claude Code, Opencode, or Qoder CLI, you typically have to manually open tmux, split panes for the terminal, add a file manager, open a git panel… same routine every time.

otter solves this:

otter -c claude          # start or reuse session with default layout
otter -c opencode        # switch to Opencode instead
otter -ks my-session     # kill a named session

Fixed three-pane layout:

• Left: your chosen AI CLI tool (claude / qodercli / opencode)
• Top-right: yazi file manager
• Bottom-right: empty shell (spare)

Detects nvim → opens a dedicated window. Detects lazygit → opens another. Soft dependencies that aren’t installed are gracefully skipped — missing yazi won’t crash the launch.

-c accepts the ALLOWED_TOOLS whitelist: claude, qodercli, opencode. Add more by editing that variable.

---

wren: Two-line statusline with live token and cache data

wren is an installer that attaches a two-line statusline to three hosts: Claude Code (cc), pi, and Qoder CLI (qc).

Install:

./cli-zoo-install.sh wren        # install wren to /usr/local/bin
wren install cc                  # wire it into Claude Code
wren install all                 # wire all three hosts at once

After install, the statusline looks like this (Claude Code side, Dracula palette):

~/Code/cli-zoo | main ↑0↓0 +4 ✱2 | wC:t1:p1 | cc
↑12K ↓3K | R1.2M CH57.14% CP2 | 8.40%/200K | claude-opus-5 · high · 1h5m

Line 1: current directory (auto-truncated for long paths) + git branch/status + which tmux pane holds the header + host badge (cc / pi / qc)

Line 2:

• ↑12K ↓3K: cumulative tokens sent/received this session
• R1.2M: cache read volume
• CH57.14%: cache hit rate
• CP2: context compression count
• 8.40%/200K: context occupancy
• claude-opus-5 · high · 1h5m: model + thinking tier + session duration

Qoder CLI side is structurally identical, different data source:

~/Code/cli-zoo | feat/x ↑0↓0 | wW:t1:p2 | qc
↑4.5M ↓65K | R4.2M CH98.21% | 15.00%/1M | Qwen3.8-Max · xhigh · 16m

The two-level installation is intentional: cli-zoo-install.sh creates a symlink (tracks the repo — edits take effect instantly); wren install copies files into each host (works even if the repo is deleted or moved). Uninstall order matters: run wren uninstall first to unwire the hosts, then cli-zoo-uninstall.sh wren to remove the binary. In pi, toggle the footer with /footer.

---

Limitations

• 1 star, new repo: updated 2026-09-19. Evaluate the code yourself before adopting it in a real workflow.
• otter’s tool whitelist is hardcoded: no config file abstraction; adding new tools means editing the script.
• wren installs file copies, not live links: the author’s upstream changes don’t reach your hosts automatically — you have to re-run wren install to update.
• Only three hosts supported: cc / pi / qc. No official support yet for Gemini CLI or other AI CLIs.

---

Same author, related repo

Spoon94 also has skill-vault (3 stars, MIT, Python): a curated collection of community Agent Skills for Claude Code, Codex CLI, Copilot CLI, Gemini CLI, and OpenCode — each skill a self-contained SKILL.md following the Agent Skills spec. Same direction as cli-zoo: standardizing the “workstation layer” of AI coding workflows.

Open-source code is for learning and reference. Evaluate before use in production.

---

Repository: github.com/Spoon94/cli-zoo
License: MIT | Language: Shell
Tools: otter (tmux workspace layout), wren (AI CLI token statusline)