[wusijian007]

Open a new iteration track: rebuild context compaction from today's single-box compactMessagesTiered + chars/4 trigger into a Claude Code-style five-layer pipeline that runs before every model call, cheapest-first, short-circuiting as soon as it is under budget, with semantic auto-compact as the guarded last resort.

The doc is the spec of record for the track. It locks:

• the five layers + usage anchor (Budget / Snip / Microcompact / Context collapse / Auto-compact; budget judged against last API usage, not chars/4),
• an audit of what myagent already has (L1 spill is non-destructive — keep it) vs the real gaps (heuristic trigger, middle three layers mashed into one in-place rewrite, no guardrails on the semantic layer),
• the load-bearing constraint: the gateway does not expose Anthropic cache_edits, so L3's hot path is a native cache-preserving adaptation (defer rewrite / append after the breakpoint; warmth approximated by elapsed time since last call via an injectable clock), not server-side reference deletion,
• five cross-cutting invariants (v3's three + per-layer freed-token short-circuit
  • reversible-before-irreversible),
• the build order (foundation-first: M4.0 cascade spine + usage anchor, then L1..L5 cheapest-first, one PR each) and per-milestone design + decisions,
• honest fidelity trade-offs adapted from CC (cache_edits, real TTL, plan state, the view-overlay architecture change).

No code yet — M4.0 (cascade spine + usage anchor) follows in its own PR and references this file.