13. Appendix: What We Don't Know
Transparency about gaps in our understanding:
Unknown or Unconfirmed
| Topic | What We Don't Know | Confidence in Current Understanding |
|---|---|---|
| Exact compaction threshold | Is it 75%? 85%? 92%? Varies by model? | 40% |
| System prompt contents | Full text not public, varies by model version | 30% |
| Token counting method | Exact tokenizer, overhead for tool schemas | 50% |
| Model fallback | Does Claude Code fallback if a model fails? | 20% |
| Internal caching | Is there result caching between sessions? | 20% |
| Rate limiting logic | How rate limits are applied per-tool | 40% |
Explicitly Undocumented
These are intentionally not documented by Anthropic:
- Session file format (internal implementation detail)
- System prompt variations between models
- Internal component names/architecture
- Token usage breakdown per component
- Exact permission evaluation order
How to Stay Updated
- Official changelog: Watch anthropic.com/changelog
- GitHub releases: github.com/anthropics/claude-code/releases
- Community Discord: Various Claude-focused servers
- This guide: Updated periodically based on verified information