cline/docs/model-config/context-windows.mdx

---
title: "Context Window Guide"
description: "Understanding and managing AI model context windows"
---

## What is a Context Window?

A context window is the maximum amount of text an AI model can process at once. Think of it as the model's "working memory" - it determines how much of your conversation and code the model can consider when generating responses.

<Note>
**Key Point**: Larger context windows allow the model to understand more of your codebase at once, but may increase costs and response times.
</Note>

## Context Window Sizes

### Quick Reference

| Size | Tokens | Approximate Words | Use Case |
|------|--------|------------------|----------|
| **Small** | 8K-32K | 6,000-24,000 | Single files, quick fixes |
| **Medium** | 128K | ~96,000 | Most coding projects |
| **Large** | 200K | ~150,000 | Complex codebases |
| **Extra Large** | 400K+ | ~300,000+ | Entire applications |
| **Massive** | 1M+ | ~750,000+ | Multi-project analysis |

### Model Context Windows

| Model | Context Window | Effective Window* | Notes |
|-------|---------------|------------------|-------|
| **Claude Sonnet 4.5** | 1M tokens | ~500K tokens | Best quality at high context |
| **GPT-5** | 400K tokens | ~300K tokens | Three modes affect performance |
| **Gemini 2.5 Pro** | 1M+ tokens | ~600K tokens | Excellent for documents |
| **DeepSeek V3** | 128K tokens | ~100K tokens | Optimal for most tasks |
| **Qwen3 Coder** | 256K tokens | ~200K tokens | Good balance |

*Effective window is where model maintains high quality

## Managing Context Efficiently

### What Counts Toward Context

1. **Your current conversation** - All messages in the chat
2. **File contents** - Any files you've shared or Cline has read
3. **Tool outputs** - Results from executed commands
4. **System prompts** - Cline's instructions (minimal impact)

### Optimization Strategies

#### 1. Start Fresh for New Features
```
/new - Creates a new task with clean context
```
Benefits:
- Maximum context available
- No irrelevant history
- Better model focus

#### 2. Use @ Mentions Strategically
Instead of including entire files:
- `@filename.ts` - Include only when needed
- Use search instead of reading large files
- Reference specific functions rather than whole files

#### 3. Enable Auto-compact
Cline can automatically summarize long conversations:
- Settings → Features → Auto-compact
- Preserves important context
- Reduces token usage

## Context Window Warnings

### Signs You're Hitting Limits

| Warning Sign | What It Means | Solution |
|-------------|---------------|----------|
| **"Context window exceeded"** | Hard limit reached | Start new task or enable auto-compact |
| **Slower responses** | Model struggling with context | Reduce included files |
| **Repetitive suggestions** | Context fragmentation | Summarize and start fresh |
| **Missing recent changes** | Context overflow | Use checkpoints to track changes |

### Best Practices by Project Size

#### Small Projects (< 50 files)
- Any model works well
- Include relevant files freely
- No special optimization needed

#### Medium Projects (50-500 files)
- Use 128K+ context models
- Include only working set of files
- Clear context between features

#### Large Projects (500+ files)
- Use 200K+ context models
- Focus on specific modules
- Use search instead of reading many files
- Break work into smaller tasks

## Advanced Context Management

### Plan/Act Mode Optimization

Leverage Plan/Act mode for better context usage:
- **Plan Mode**: Use smaller context for discussion
- **Act Mode**: Include necessary files for implementation

Configuration:
```
Plan Mode: DeepSeek V3 (128K) - Lower cost planning
Act Mode: Claude Sonnet (1M) - Maximum context for coding
```

### Context Pruning Strategies

1. **Temporal Pruning**: Remove old conversation parts
2. **Semantic Pruning**: Keep only relevant code sections
3. **Hierarchical Pruning**: Maintain high-level structure, prune details

### Token Counting Tips

#### Rough Estimates
- **1 token ≈ 0.75 words**
- **1 token ≈ 4 characters**
- **100 lines of code ≈ 500-1000 tokens**

#### File Size Guidelines
| File Type | Tokens per KB |
|-----------|---------------|
| **Code** | ~250-400 |
| **JSON** | ~300-500 |
| **Markdown** | ~200-300 |
| **Plain text** | ~200-250 |

## Context Window FAQ

### Q: Why do responses get worse with very long conversations?
**A:** Models can lose focus with too much context. The "effective window" is typically 50-70% of the advertised limit.

### Q: Should I use the largest context window available?
**A:** Not always. Larger contexts increase cost and can reduce response quality. Match the context to your task size.

### Q: How can I tell how much context I'm using?
**A:** Cline shows token usage in the interface. Watch for the context meter approaching limits.

### Q: What happens when I exceed the context limit?
**A:** Cline will either:
- Automatically compact the conversation (if enabled)
- Show an error and suggest starting a new task
- Truncate older messages (with warning)

## Recommendations by Use Case

| Use Case | Recommended Context | Model Suggestion |
|----------|-------------------|------------------|
| **Quick fixes** | 32K-128K | DeepSeek V3 |
| **Feature development** | 128K-200K | Qwen3 Coder |
| **Large refactoring** | 400K+ | Claude Sonnet 4.5 |
| **Code review** | 200K-400K | GPT-5 |
| **Documentation** | 128K | Any budget model |