上下文与压缩
Claude Code 主动管理其上下文窗口。当 token 不足时，它通过总结历史自动压缩。微压缩在整个会话中在行内清理旧的工具结果。
200K 默认上下文窗口 1M Sonnet 4.x / Opus 4.6 ~13K 剩余 token → 自动压缩 50K 压缩后 token 预算
上下文窗口大小
| Context | Size | Available for |
|---|---|---|
| Default | 200,000 tokens | All users |
| 1M context | 1,000,000 tokens | Sonnet 4.x [1m] and Opus 4.6 [1m] |
| Override | Custom | Set via CLAUDE_CODE_MAX_CONTEXT_TOKENS |
自动压缩流程
当接近上下文限制时，Claude 自动总结对话历史以释放空间。
1
~13,000 tokens remaining trigger
Compaction threshold reached. Auto-compact triggers automatically. 2
Fork agent to summarize
A background agent reads the full history and produces a compressed summary. 3
Replace history with summary
The conversation history is replaced with the summary. The context window is freed. 4
Post-compact budget: 50,000 tokens 50K budget
Claude gets 50K tokens to restore important files and skills. Max 5 files, 5,000 tokens per file. 5
Circuit breaker: 3 failures
If compaction fails 3 consecutive times, it stops trying to avoid infinite loops. 微压缩
微压缩与自动压缩不同。它在整个会话中在行内发生，专门针对旧的工具结果。
When a tool result (Read, Bash, Grep, Glob, WebSearch, WebFetch, Edit, Write) becomes old enough, its content is replaced with:
[Old tool result content cleared]
Images are estimated at a fixed 2,000 tokens regardless of actual size, a conservative estimate that helps budgeting.
输出 token 限制
| Constant | Value | Used when |
|---|---|---|
| Default max output | 32,000 tokens | Standard responses |
| Capped (slot-reservation) | 8,000 tokens | When slots are reserved for other operations |
| Escalated (recovery) | 64,000 tokens | Recovery mode after errors |
| Compaction output | 20,000 tokens | The summary written by the compaction agent |
→ 在达到限制之前使用 /compact
自动压缩会自动触发，但你可以在达到阈值之前手动运行
/compact。
手动压缩往往比在最后一刻触发的自动压缩产生更清晰、更有意图的总结。