opencode-dynamic-context-pruning

Why we included this project

Long OpenCode sessions tend to balloon in token usage, and this plugin gives the model a way to trim the conversation without losing the important parts. It exposes a compress tool that replaces stale, closed messages with technical summaries, either across contiguous ranges or on individual messages, so the model can decide what no longer needs to be verbatim instead of waiting for a blunt full-session compaction. It also deduplicates repeated tool calls and prunes large inputs from errored calls, which keeps prompt caches and context windows under control. The config is granular enough to protect subagent outputs, specific file patterns, and your own user messages from compression, and you can switch to manual mode if you prefer to trigger cleanup yourself. One caveat: active development has moved to the Sleev proxy, so treat this as a stable, working plugin for existing OpenCode users rather than a fast-moving project.