Three-layer compression

Three layers, each handling a different scale: Layer 1 (micro): Every turn, silently replace old tool results with short placeholders. Keeps the last 3 results intact. Layer 2 (auto): When token estimate exceeds threshold, save the full transcript to disk and ask the LLM to summarize. Fresh context with continuity. Layer 3 (manual): The model can trigger compression itself via the compact tool when it knows it needs a fresh start.