Skip to content

l9-context.en.md

Latest commit

 

History

History
43 lines (28 loc) · 1.08 KB

l9-context.en.md

File metadata and controls

43 lines (28 loc) · 1.08 KB

Layer 9: Context Window Management And Prompt Cache

中文 | English

Core Question

How does Claude Code control context budget in long conversations while protecting prompt-cache hit rate?

Run First

python examples/l9_context_mgmt.py

Read First

  • utils/context.ts
  • constants/prompts.ts
  • utils/api.ts
  • memdir/memdir.ts

Search Anchors

  • SYSTEM_PROMPT_DYNAMIC_BOUNDARY
  • context_1m_beta
  • MEMORY.md
  • ENTRYPOINT_NAME

What To Notice

  • why the static/dynamic boundary matters
  • how snapshotting and sticky latches protect cache stability
  • what auto-compact solves versus what memory indexing solves

Demo vs Real Source

The demo only illustrates token budgeting and compaction. The real source also handles model window capability, cache boundaries, system-prompt assembly, and memory injection rules.

Questions To Answer

  1. Why does prompt-cache hit rate end up shaping architecture?
  2. Why must MEMORY.md stay short and index-like?
  3. Auto-compact solves “history is too long” — what problem does memory solve?