[P1] Selective context compaction (upgrade summarize strategy)

[JackChen-me]

Problem

Our current contextStrategy: 'summarize' compresses the entire conversation history into a single summary. This is a blunt instrument — it loses tool call decisions, action sequences, and key reasoning while retaining generic summaries.

Industry comparison: Anthropic's Compaction API and ACON framework use selective compression — preserve decisions and action outcomes, discard intermediate reasoning and raw tool outputs. This achieves better compression ratios with higher fidelity.

Proposal

Upgrade the summarize strategy to a compaction-style approach:

1. Preserve: tool call decisions (which tool, why), final outcomes, errors encountered, key user instructions
2. Discard: thinking/reasoning blocks, raw tool output (replace with one-line summary), redundant assistant acknowledgments
3. Structure: output as a structured context block (not free-form prose) so the model can parse it reliably

This could be implemented as:

• A new contextStrategy: 'compact' option alongside the existing summarize
• Or an upgrade to the existing summarize with a smarter default prompt

Source

• Competitive analysis: Anthropic Compaction API, ACON framework (arXiv 2510.00615, 26-54% token reduction), Mem0 semantic distillation
• Triggered by [Feature] Provide Custom Tools. #108 discussion and context engineering landscape research