Research Finding

Anthropic introduced a server-side Compaction API (launched Feb 5, 2026, beta header compact-2026-01-12) for Opus 4.6 and Sonnet 4.6. When enabled, the API automatically summarizes conversation history when input tokens approach a threshold, inserting a compaction content block in the response. The client does not need to maintain summarization logic — just append the response as usual and the API handles history pruning on the next turn.

Applicability

Zeph currently implements client-side chunked compaction in zeph-memory and zeph-core context management. The server-side API could:

• Replace the custom compaction pipeline for Claude sessions (apply_deferred_summaries, compact_context)
• Eliminate token-counting heuristics and summarization LLM calls
• Reduce code complexity in crates/zeph-core/src/context_manager.rs

This is a complement/alternative to #1338 (tiered context compaction).

Design Decision Needed

Evaluate whether to:

1. Use server-side compaction as the primary path for Claude sessions, keeping client-side as fallback for non-Claude providers
2. Or keep existing pipeline and only adopt the server-side API for Sonnet 4.6/Opus 4.6 sessions

Action

1. Review the Compaction API spec at https://platform.claude.com/docs/en/build-with-claude/context-management
2. Prototype adding context_management: {auto_truncate: true} (or equivalent) to the Claude request builder
3. Parse the compaction content block type in response deserialization
4. Add feature-flag or config option to opt-in per session

Source

Research session 2026-03-13. Anthropic changelog Feb 5, 2026.