bobo-ai-cli 1.0.1 → 1.1.0

package/bundled-skills/context-compressor/SKILL.md

@@ -0,0 +1,75 @@
+---
+name: context-compressor
+description: 九段式上下文压缩 — 对话过长时生成结构化摘要，防止压缩后丢失关键上下文。解决"compact 后忘了之前在做什么"、"压缩后意图漂移"问题。适用于：(1) 对话接近 context 上限时 (2) 手动 /compact 时 (3) 需要生成对话摘要时。灵感源自 Claude Code Compact Prompt（prompt.ts 九段式模板）。
+---
+
+# Context Compressor
+
+## 核心理念
+
+> 压缩不是删除 — 是把 10 万 token 的对话浓缩成 5000 token 但不丢任何关键信息。
+
+## 压缩前准备
+
+1. **移除图片块** — 替换为 `[image was shared]` 文本标记
+2. **不调用任何工具** — 压缩是纯文本任务，tool call 会浪费 turn
+
+## 九段式摘要模板
+
+先用 `<analysis>` 整理思路（草稿，不进入最终上下文），再输出 `<summary>`：
+
+```
+<analysis>
+[按时间顺序分析每条消息，确保覆盖所有要点]
+</analysis>
+
+<summary>
+1. 用户请求和意图
+   [所有明确请求，完整详述]
+
+2. 关键技术概念
+   - [技术/框架/方案列表]
+
+3. 文件和代码
+   - [文件名]
+     - 为什么重要
+     - 做了什么修改
+     - 关键代码片段
+
+4. 错误和修复
+   - [错误描述]
+     - 如何修复
+     - 用户对此的反馈（如有）
+
+5. 问题解决
+   [已解决的问题 + 正在排查的问题]
+
+6. 所有用户消息
+   - [列出全部非工具结果的用户消息]
+   （⚠️ 这是捕捉用户意图变化的关键）
+
+7. 待办任务
+   - [明确被要求但未完成的任务]
+
+8. 当前工作
+   [压缩前正在做的事，附文件名和代码片段]
+
+9. 下一步（可选）
+   [必须与最近请求直接相关]
+   > 原文引用："用户说了什么" → 防止意图漂移
+</summary>
+```
+
+## 关键规则
+
+- **第 6 段（用户消息）是最重要的** — 不列用户消息 = 丢失意图变化
+- **第 8 段必须具体** — 不是"正在修复 bug"，而是"正在修复 src/auth.py:42 的空指针，已加了 null check 但测试还没跑"
+- **第 9 段必须附原文引用** — 防止压缩后做跑偏的事
+- **不要开始新的、切线的任务** — 只延续压缩前正在做的
+
+## 压缩后恢复
+
+压缩完成后自动恢复：
+- 最近修改的文件内容（最多 5 个文件）
+- 活跃 Skill 的指令重注入
+- 工具列表变更提示

package/bundled-skills/context-optimization-suite/SKILL.md

@@ -0,0 +1,162 @@
+---
+id: "context-optimization-suite"
+title: "Context Optimization Suite"
+category: "agent-engineering"
+tags: ["context optimization suite", "when to use", "compressed context", "combined workflow", "quick reference"]
+triggers: []
+dependencies: []
+source: "E:/Bobo's Coding cache/.claude/skills/context-optimization-suite"
+---
+
+---
+name: context-optimization-suite
+description: Unified context management - combines context-budget-analyzer (token budget tracking) + context-compressor (95% context compression). Use when context is tight, approaching limits, or when resuming long sessions.
+---
+
+# Context Optimization Suite
+
+> **Merged from**: context-budget-analyzer + context-compressor
+> **Token savings**: Up to 95% context reduction while preserving critical information
+
+## When to Use
+
+- Context window approaching 70%+ capacity
+- Resuming a long session with accumulated context
+- Before compaction to preserve key information
+- When `cacheCreationInputTokens` = 0 (cache miss)
+
+---
+
+## Module 1: Budget Analyzer
+
+### Budget Threshold System
+
+| Level | Usage | Action |
+|-------|-------|--------|
+| Green | <50% | Normal operation |
+| Yellow | 50-75% | Efficiency awareness, prefer targeted tools |
+| Orange | 75-90% | Consider compaction or session split |
+| Red | >90% | Immediate compaction required |
+
+### Token Category Tracking
+
+Monitor these categories:
+- **Human messages**: User input tokens
+- **Assistant messages**: AI response tokens
+- **Tool requests**: Tool call tokens
+- **Tool results**: Tool output tokens (often the biggest waste)
+- **Attachments**: File attachment tokens
+- **System/Other**: System message tokens
+
+### Optimization Rules
+
+1. **Cache file contents** after first read — never re-read the same file
+2. **Use Grep** instead of full file reads when searching
+3. **Compress images** before attachment
+4. **Limit bash output** with `head -N` / `tail -N`
+5. **Keep assistant responses concise** — no trailing summaries
+
+### Duplicate Detection
+
+Track file reads per session. If a file is read 3+ times, it's a waste signal:
+```
+File: src/index.ts — Read 4 times — ~8,000 tokens wasted
+Action: Cache content after first read
+```
+
+---
+
+## Module 2: Context Compressor
+
+### 9-Section Compression Template
+
+When compressing context, produce exactly this structure:
+
+```markdown
+## Compressed Context
+
+### 1. User Request
+[Original user prompt/goal in 1-2 sentences]
+
+### 2. Key Technical Concepts
+- [Concept 1]: [Brief explanation]
+- [Concept 2]: [Brief explanation]
+
+### 3. Key Files and Code
+- `path/to/file.ts:L42-L58` — [What it does]
+- `path/to/other.ts:L10-L25` — [Key function]
+
+### 4. Key Errors and Fixes
+- **Error**: [Error message] → **Fix**: [Solution applied]
+
+### 5. Problem-Solving Process
+1. [Step 1: What was tried]
+2. [Step 2: What worked]
+3. [Step 3: Current approach]
+
+### 6. User Messages (condensed)
+- [Message 1 summary]
+- [Message 2 summary]
+
+### 7. Todos and Tasks
+- [ ] [Remaining task 1]
+- [x] [Completed task 2]
+
+### 8. Current Work
+[What is being worked on RIGHT NOW — 2-3 sentences]
+
+### 9. Next Steps
+1. [Immediate next action]
+2. [Following action]
+3. [Final verification]
+```
+
+### Compression Ratios
+
+| Context Size | Compressed Size | Ratio |
+|-------------|----------------|-------|
+| 100k tokens | ~5k tokens | 95% |
+| 50k tokens | ~3k tokens | 94% |
+| 20k tokens | ~2k tokens | 90% |
+
+---
+
+## Combined Workflow
+
+```
+Session starts
+  ↓
+Monitor budget (Module 1)
+  ↓
+Budget > 75%?
+  ├─ No → Continue normally
+  └─ Yes → Apply compression (Module 2)
+       ↓
+     Resume with compressed context
+       ↓
+     Monitor budget again
+```
+
+### Cache-Aware Optimization (cc-cache-fix integration)
+
+With cc-cache-fix installed (`claude-patched`):
+- Cache TTL extended to 1 hour → fewer re-reads needed
+- Delta attachments preserved → resume works better
+- Hash stability → same content = same cache hit
+
+**Strategy shift**: With good caching, prioritize keeping referenced files in cache rather than aggressively compressing.
+
+---
+
+## Quick Reference
+
+```bash
+# Check context budget (mental model)
+# Green:  <50% — proceed normally
+# Yellow: 50-75% — be efficient
+# Orange: 75-90% — compress soon
+# Red:    >90% — compress NOW
+
+# Compress context using the 9-section template
+# Output only the compressed version, nothing else
+```

package/bundled-skills/context7/SKILL.md

@@ -0,0 +1,13 @@
+---
+id: "context7"
+title: "context7"
+category: "knowledge"
+tags: ["context7"]
+triggers: []
+dependencies: []
+source: "E:/Bobo's Coding cache/.claude/skills/context7"
+---
+
+# context7
+
+Skill directory: context7