简体中文 · English

--- # Context Management & Prefix Caching DeepSeek's prefix cache mechanism: on each request, the API compares `messages[0]` onward against the previous request, finding the longest common prefix. The cached portion is billed at the cache-hit rate; the remainder at the standard rate. **The price gap between cache-hit and cache-miss is massive** — for V4-Flash and V4-Pro, cache-hit price is just **1/30** of the cache-miss price (after the Aug 2026 price adjustment; previously 1/50 ~ 1/120). Waveloom systematically optimizes for this: 1. **System prompt fixed as `messages[0]`**: The first message never changes, no matter how long the conversation — ensuring the prefix starting point is always stable. 2. **Message history accumulated across turns**: Each turn appends to the end rather than resetting. The first N-1 turns become the prefix for turn N's request. 3. **Four-tier watermark compaction (Tier 0–3)**: As context utilization rises, history is compressed in stages. The key insight — **compacted byte content never changes again**. Once a message is truncated or replaced with a placeholder, it keeps the exact same byte representation in all future turns, so the prefix cache keeps hitting. 4. **Monotonic boundary guarantee**: The decision table (`compactionDecisionSet`) + triple cursor mechanism ensures each message is compacted exactly once — never modified repeatedly, which would invalidate the cache. ```mermaid flowchart LR t0["Tier 0 idle < 60%"] t1["Tier 1 · Snip tool output truncation 60-80%"] t2["Tier 2 · Prune clear reasoning 80-95%"] t3["Tier 3 · Summarize LLM incremental summary ≥ 95%"] stop["Hard limit block further LLM calls ≥ 98%"] t0 --> t1 --> t2 --> t3 --> stop style t0 fill:#2d8,stroke:#333,color:#fff style t1 fill:#5b5,stroke:#333,color:#fff style t2 fill:#da5,stroke:#333,color:#000 style t3 fill:#e73,stroke:#333,color:#fff style stop fill:#c22,stroke:#333,color:#fff ``` Cache hit rates are typically **95–99%**, meaning in a 1M-token context window, only 10K–50K tokens are billed at the standard rate. This is not luck — it's by architectural design. > See [`specs/compaction.md`](../specs/compaction.md) — complete design of context compaction.