What should context compression keep? I looked at how six agents handle it[D]

I use Claude Code, Codex CLI, OpenCode, Cline, Cursor, and Amp enough to notice a pattern in how they handle long context. They are all converging on layered progressive compression, but they disagree on what to protect. Most protect recent user messages as a first-class asset. That makes sense. The user said it, which is the source of truth. Most also protect tool outputs that carry state. What surprised me was how differently they treat old assistant messages. Artifacts keeps recent tool calls verbatim but drops older context aggressively. Cursor starts pruning earlier design decisions once the window gets full. Codex CLI lets the model itself decide what to keep in the summary tier. The other axis is transparency. Do you tell the model it was compressed? Some systems silently replace old tool results with a placeholder, which means the model is reasoning under the illusion that it never happened. Others make it explicit: "the previous 40 tool calls are summarized below." I lean explicit because the model needs to know its own context was degraded. Verdents agent loop uses a similar tiered approach: snip first, prune second, summarize last, and a hard red line that protects user messages, stateful tool outputs, and anything the user explicitly flagged. The tradeoff is cost vs accuracy. Aggressive compression saves tokens but degrades the plan. Under-compression hits the window and causes context rot.