context-viz

TotalClaw 作者 totalclaw

可视化当前上下文窗口的使用情况 - 每个组件的令牌估计（系统提示、工具、工作区文件、消息、可用空间）。当用户询问上下文大小、令牌使用情况、上下文细分、“上下文有多满”或想要查看正在消耗上下文窗口的内容时使用。

安装 / 下载方式

TotalClaw CLI推荐

totalclaw skills install shop:totalclaw:totalclaw~furukama-context-viz

cURL直接下载，无需登录

curl -fsSL https://skills.taituai.com/api/skills/totalclaw%3Atotalclaw~furukama-context-viz/file -o furukama-context-viz.md

## 概述（中文）

可视化当前上下文窗口的使用情况 - 每个组件的令牌估计（系统提示、工具、工作区文件、消息、可用空间）。当用户询问上下文大小、令牌使用情况、上下文细分、“上下文有多满”或想要查看正在消耗上下文窗口的内容时使用。

## 原文

# Context Visualization

Estimate and display a breakdown of the current context window usage.

## How It Works

Run the bundled script to estimate token counts for workspace files:

```bash
python3 scripts/estimate_tokens.py /path/to/workspace
```

The script counts characters in known workspace files and estimates tokens (~4 chars/token).

Then call `session_status` to get the actual context usage from OpenClaw.

## Generating the Visualization

1. Run `session_status` to get: model, context used/total, compactions
2. Run `scripts/estimate_tokens.py <workspace_path>` to estimate file token sizes
3. Estimate message tokens: `context_used - system_overhead - file_tokens`
4. Present the breakdown using the format below

## Output Format

Use a monospace block with bar chart. Adapt the bar lengths proportionally.

```
📊 Context Usage
<model> • <used>k/<total>k tokens (<pct>%)

Component                    Tokens    %     
─────────────────────────────────────────────
⚙️  System prompt + tools    ~Xk      X%    ░░
📋  AGENTS.md                ~Xk      X%    ░
👻  SOUL.md                  ~Xk      X%    
👤  USER.md                  ~Xk      X%    
🔧  TOOLS.md                 ~Xk      X%    ░
💓  HEARTBEAT.md             ~Xk      X%    
🧠  MEMORY.md                ~Xk      X%    ░
🪪  IDENTITY.md              ~Xk      X%    
💬  Messages                 ~Xk      X%    ░░░░░░░░░░░░
📭  Free space               ~Xk      X%    ░░░░░
─────────────────────────────────────────────
```

Use ░ blocks: 1 block per ~2% of total context. Round to nearest block.

## Memory Inventory (not in context)

Below the context chart, add a **Memory on Disk** section showing what's stored in `memory/` — grouped by category. These files are NOT loaded into context but represent the agent's total knowledge base.

```
💾 Memory on Disk (not in context)
Category                     Files  Tokens   Size
──────────────────────────────────────────────────
📰  chinese-ai-digests        12    ~23k     92KB
📁  other                     11    ~12k     46KB
📅  daily-notes                9    ~5k      17KB
🗃️  zettelkasten               8    ~4k      15KB
💼  linkedin                   2    ~1k       5KB
──────────────────────────────────────────────────
     Total:                   42    ~44k    177KB
```

The script auto-categorizes files by directory or filename pattern.

## Notes

- Token estimates use ~4 chars/token (rough average for English/mixed content)
- System prompt + tools overhead is estimated at ~8-10k tokens for a typical OpenClaw setup
- Message tokens are the remainder after subtracting files + system overhead
- Memory files are informational only — they show what the agent has accumulated
- For Discord/WhatsApp: skip markdown tables, use the block format above