---
title: "coach"
url: "https://claudemods.dev/zh-cn/builds/coach-d18572"
source: "https://github.com/eshin087/claude-code-mods/tree/main/coach"
source_type: "github"
repo: "https://github.com/eshin087/claude-code-mods/tree/main/coach"
author: "eshin087"
author_handle: "eshin087"
stars: 0
category: "其他"
categories: ["other"]
tags: []
permission_level: 2
scan_complete: true
install: "claude plugin marketplace add eshin087/claude-code-mods\nclaude plugin install coach"
added: "2026-10-05T01:57:25.163Z"
updated: "2026-10-25T21:14:22Z"
locale: "zh-cn"
content_locale: "zh-cn"
---

# coach

/coach：实时成本、缓存、上下文和子代理统计、SDLC 习惯检查和架构师课程；关键时刻，转弯后给出提示。

## Description

# 教练

**它的作用：** 显示您的会话背后推动成本和质量的数字，根据正常软件生命周期 (SDLC) 检查您的习惯，并一次教授一个与您刚所做的事情相关的 architect-level 概念。目的是将vibe-coding技能运用到实际的工程工作中。

**您在哪里看到它：**
- 扩展坞（或 `/coach`）中的 **C** 切换面板：
```
  ● Cache    92%   ━━━━━━━━━━━  reusing context well
  ● Context  41%   ━━━━━━━━━━━  room to spare
  ● Cost     $3.20             $0.27 per turn
  Habits   ✓ Plan  ✓ Branch  ✗ Tests  ✓ Commits  · PR
  Do next  Run the tests: 8 edits since the last run
  Tokens   120k fresh  4.1M cached  300k cache-write  52k out
  Model    claude-opus-5-5  effort high  subagents 4 runs · 31% of tokens
  Lesson   The test pyramid
           Many fast unit tests at the base, fewer integration tests…
  
```-
 **颜色：** 绿色表示良好，黄色表示观看，红色表示行动。在令牌行中，缓存的令牌为绿色（廉价），cache-writes 为黄色，输出为青色。
  - **更多**添加了课程的“下一步尝试”和“工作中”注释（带有**另一节课**）和 Claude Code 专业提示（带有**下一步提示**和链接）。
- **将 C** 悬停在扩展坞中，查看紫罗兰卡，其中包含缓存和上下文栏、成本、习惯和“下一步”。
- **一轮后一个提示，仅当规则触发时**，作为聊天中暗淡的 `🎓 Coach · …` 行：
  - 直接致力于主要
  - 在大转弯时 50% 下的缓存命中（可能的原因）
  - 75% 完整上下文
  - 自上次运行测试以来 5+ 编辑
  - 一轮费用超过~$2
  - 10+ 没有计划的建筑分钟数
  - 子代理使用超过一半的代币

  相同的刀尖在重复之前至少等待 6 圈。

**术语表：**
- **提示缓存：**应用程序每次都会 re-sends 整个对话。缓存命中意味着前缀被重用，并按新输入的 ~10% 计费。模型切换、CLAUDE.md 或内存编辑、/compact 或空闲间隙（缓存过期）上的命中断。
- **上下文填充：**对话使用模型窗口的多少。富勒意味着速度更慢、价格更高、质量接近极限。
- **成本：** API 价格。在订阅中，这是您使用的价值，而不是账单，但它会跟踪您燃烧 5-hour 窗口的速度。

**它是如何工作的：**一切都来自应用程序已经报告的内容（每回合的令牌使用情况、会话成本和上下文、编辑、测试、提交和 `gh pr create` 等工具调用）。没有模型调用。

**成本：**无。

**保存的数据：**磁盘上没有数据。会议统计数据实时显示。

## Install

```
claude plugin marketplace add eshin087/claude-code-mods
claude plugin install coach
```

## Original README

# Coach

**What it does:** shows the numbers behind your sessions that drive cost and quality, checks your habits against a normal software lifecycle (SDLC), and teaches one architect-level concept at a time, tied to what you just did. The aim is to carry vibe-coding skill into real engineering work.

**Where you see it:**
- **C** in the dock (or `/coach`) toggles the panel:
  ```
  ● Cache    92%   ━━━━━━━━━━━  reusing context well
  ● Context  41%   ━━━━━━━━━━━  room to spare
  ● Cost     $3.20             $0.27 per turn
  Habits   ✓ Plan  ✓ Branch  ✗ Tests  ✓ Commits  · PR
  Do next  Run the tests: 8 edits since the last run
  Tokens   120k fresh  4.1M cached  300k cache-write  52k out
  Model    claude-opus-5-5  effort high  subagents 4 runs · 31% of tokens
  Lesson   The test pyramid
           Many fast unit tests at the base, fewer integration tests…
  ```
  - **Colors:** green is good, yellow means watch it, red means act. In the token row, cached tokens are green (cheap), cache-writes yellow and output cyan.
  - **More** adds the lesson's "try next" and "at work" notes (with **Another lesson**) and a Claude Code pro tip (with **Next tip** and links).
- **Hover C** in the dock for a violet card with the cache and context bars, cost, habits and "do next".
- **One tip after a turn, only when a rule fires,** as a dim `🎓 Coach · …` row in the chat:
  - committed straight to main
  - cache hit under 50% on a big turn (with the likely cause)
  - context over 75% full
  - 5+ edits since tests last ran
  - a turn costing over ~$2
  - 10+ minutes of building with no plan
  - subagents using over half the tokens

  The same tip waits at least 6 turns before repeating.

**Glossary:**
- **Prompt cache:** the app re-sends the whole conversation every turn. A cache hit means that prefix was reused and billed at ~10% of fresh input. Hits break on model switches, CLAUDE.md or memory edits, /compact, or idle gaps (the cache expires).
- **Context fill:** how much of the model's window the conversation uses. Fuller means slower, pricier, and lower quality near the limit.
- **Cost:** API prices. On a subscription this is the value you used, not a bill, but it tracks how fast you burn your 5-hour window.

**How it works:** everything comes from what the app already reports (token usage per turn, session cost and context, tool calls such as edits, tests, commits and `gh pr create`). No model calls.

**Cost:** none.

**Data saved:** none on disk. Stats live for the session.
