ranxianglei/opencode-acp

This package is a plugin for OpenCode, an AI coding tool. The plugin's job is to keep the conversation between you and the AI from growing too long. As you chat back and forth with an AI assistant, every message and every file you share gets added to what is called the context window. Once that window fills up, the assistant either forgets the start of the conversation or has to be reset. ACP, which stands for Active Context Pruning, lets the model itself decide when older parts of the conversation are safe to summarize and which parts to summarize, so the chat can keep going much longer. ACP is a hardened fork of an earlier project called DCP, with 35 bug fixes applied. The README claims this turns sessions from roughly 200 messages before instability into 10,000 plus messages, drops the per-turn overhead from 20 to 50 seconds down to about 90 milliseconds, and lets the compression state survive a restart instead of being lost. The model is given a tool called compress that it can call whenever a chunk of the conversation is done with, and the tool replaces that chunk with a high-fidelity summary while the raw history is kept behind the scenes. The plugin offers two ways of compressing. Range mode shrinks a contiguous block of messages into one summary. Message mode is experimental and lets the model compress individual messages one at a time for finer control. On top of compression, ACP removes duplicate tool calls and trims the input from errored tool calls after a few turns, keeping the error message itself intact. Users interact with ACP through slash commands such as /acp context for a token breakdown, /acp stats for cumulative numbers, /acp sweep to prune recent tools, /acp compress to trigger one compression manually, and /acp decompress to restore a past summary. Configuration lives in an acp.jsonc file at a global, custom, or project level, with project settings winning. The README is clear that OpenCode's own built-in auto-compaction must be turned off when ACP is enabled, since the two systems conflict. Installation is a single command, or an entry in the OpenCode config file. The package is published on npm with English and Chinese READMEs.