3.11 Design Insights

1. Memoize guarantees idempotency: Both getSystemContext and getUserContext are memoized, computed only once per session. When setSystemPromptInjection() changes, both functions' caches are manually cleared
2. Progressiveness of the compression pipeline: From zero-cost trimming to full summarization, upgrading level by level as needed. Most conversations never trigger Autocompact
3. Reversibility of projection-based folding: Context Collapse doesn't modify original messages and can be safely rolled back — this is where it's superior to Autocompact
4. Cache-aware context assembly: The injection order of context (system prompt first, user context before messages) considers prompt cache hit rates
5. Anchor strategy for token estimation: Using the precise usage reported by the server as an anchor, only estimating increments, balancing precision and latency
6. system-reminder as a unified injection channel: Wrapped in XML tags, injecting system information at any position in the message flow without confusing role boundaries
7. Pragmatic trade-off of session-level latching: Both TTL eligibility and beta headers are locked once determined, lasting until session end — sacrificing flexibility for cache stability, since the cost of invalidating 50-100K tokens of cache far outweighs the inability to toggle a feature mid-session

---

Hands-on Practice: In claude-code-from-scratch, src/prompt.ts and src/system-prompt.md demonstrate the minimal implementation of context construction. Compare with the multi-layer context assembly in this chapter and think: what context does a minimal Agent need to be sufficient? See the tutorial Chapter 3: System Prompt Engineering.

Previous chapter: Agent Loop | Next chapter: Tool System