Summary

Anthropic prompt caching is exact-prefix and cache-invalidating on prompt-prefix rewrites, so local prompt-mutating compaction can burn a freshly written cache even when lossless-claw is no longer stalling the foreground turn. OpenClaw already has a provider-native compaction precedent for OpenAI Responses; Anthropic should get the same first-class support, plus mixed prompt-cache TTLs that keep the expensive stable prefix warm without paying 1-hour write costs for high-churn conversation content.

Problem

• Anthropic-heavy coding sessions pay large prompt-cache write costs on Opus/Claude when long retention is applied too broadly.
• Local prompt rewrites after a turn can invalidate a still-hot conversation cache, reducing the value of lossless-claw compaction even after the no-stall/background-maintenance work lands.
• Right now OpenClaw does not expose Anthropic native compaction or a mixed TTL policy that matches Anthropic's cache model.

Design Goal

• Use Anthropic native active-context compaction when the provider supports it.
• Preserve stable system/tool/workspace prefixes with 1h cache retention.
• Keep high-churn conversation/trailing user prefixes at 5m retention by default.
• Round-trip Anthropic compaction blocks through the transport so follow-up requests can reuse the provider-native compacted prefix.
• Keep this separate from lossless-claw, which should stay the lossless searchable sidecar rather than the primary live prompt shaper.

Proposed Changes

• Add Anthropic request params:
  • anthropicServerCompaction: boolean
  • anthropicCompactThreshold: number
  • anthropicCompactPauseAfter?: boolean
  • anthropicCompactInstructions?: string
• In the Anthropic transport path:
  • inject context_management.edits with compact_20260112 when enabled
  • add the required Anthropic beta features
  • preserve compaction blocks in assistant history and streamed content
• In the prompt-cache policy:
  • keep system/tool/workspace prefix at 1h
  • keep trailing conversation / compaction / user prefix at 5m
  • avoid rewriting a still-hot cached conversation prefix unless needed for correctness

Why This Matters

• It reduces wasted cache-write spend on volatile conversation content.
• It lets provider-native compaction manage the active Claude prompt while lossless-claw continues to provide searchable durable memory.
• It preserves cache value for long coding sessions where follow-up turns are frequent.

Acceptance Criteria

• Anthropic requests can opt into native compaction via model/run params.
• Required beta headers are sent only when needed.
• Compaction blocks round-trip through request history and streaming output.
• Long Anthropic cache retention applies 1h only to the stable prefix and keeps trailing conversational content at 5m.
• Existing Anthropic behavior remains unchanged unless the new params are enabled.

Related Work

• Run context-engine turn maintenance as idle-aware background work #65233 (background turn maintenance for context engines)
• Proactive threshold compaction can stall foreground turns and hide maintenance state Martian-Engineering/lossless-claw#407
• Defer proactive compaction debt and surface LCM maintenance state Martian-Engineering/lossless-claw#408