name: opencode-cache-compact
repo: https://github.com/lennartschoch/opencode-cache-compact
tagline: Cache-friendly compaction for locally hosted models
description: Replaces OpenCode's built-in compaction, which rebuilds the prompt from scratch (no system prompt, no tools, history re-serialized) and therefore cold-prefills the whole conversation. This plugin appends the summary turn instead — a cache hit — and then rewrites only the outgoing message list down to that summary, so a locally hosted model reuses its KV cache instead of re-reading everything. Aimed at llama.cpp / LM Studio / vLLM setups where prefilling a large context is the expensive part.
kind: plugin
topics:
  - context
compat: v2
scope:
  - global
  - project
tags:
  - context
  - compaction
  - prompt-cache
  - local-models
  - llama.cpp
min_version: 1.17.9