wet trims stale Claude Code output

Hot take: this is a genuinely useful hack for Claude Code power users because it targets stale tool output, not the whole conversation, so long sessions stay usable without the blunt-force penalty of auto-compaction.

• –Sits as a reverse proxy in front of Anthropic, which makes it client-agnostic and easy to slot into existing Claude Code workflows.
• –Splitting deterministic compression from LLM rewrites is the right design: shell/test logs shrink cleanly, while agent replies and file reads need semantic preservation.
• –The control plane matters as much as the compressor; `pause`, `dry-run`, and rule tuning give users a way to audit what got trimmed.
• –It is best for heavy multi-agent workflows and likely overkill for casual users who only run short prompts.
• –The main tradeoff is trust: rewriting context can hide important breadcrumbs, so this wants a private, well-understood deployment environment.