Research

Current compaction triggers at compaction_threshold = 0.80. Industry best practices recommend:

• Soft compaction at 70%: apply deferred tool summaries, compress old tool outputs (10:1+ ratio)
• Hard compaction at 90%: aggressive summarization, strip recoverable content
• Different compression ratios per content type: conversation (3:1-5:1) vs tool outputs (10:1-20:1)

Current state: deferred_apply_threshold = 0.70 already exists, but compaction itself is single-threshold and not type-aware.

Proposal

1. Add per-category compression targets (conversation vs tool output vs retrieved context)
2. Strip recoverable information during compaction (file contents agent can re-fetch)
3. Make summarization aggressiveness increase with message age

Sources

• https://www.getmaxim.ai/articles/context-engineering-for-ai-agents-production-optimization-strategies/
• https://www.philschmid.de/context-engineering-part-2