Cache-Friendly Context Compacting Plugin for OpenCode Cuts Local Model Time from 10min to 1-2min
schennardo · reddit · 2026-09-23
The author released the open-source plugin opencode-cache-compact to fix a pain point in OpenCode's default context compacting: it strips tokens from the start of the conversation (system prompt, tools), which is fine for hosted models but forces a full prefill of the already-cached conversation on local models.
The plugin keeps the conversation intact, has the model write a summary, then rewrites the conversation to just system prompt + tools + summary. Since the full conversation remains cached, compacting drops from >10 minutes to 1-2 minutes on the author's Strix Halo. It's their first open-source project in the local LLM space, and feedback is welcome.
More from coding & agent
- Claude Opus 5.5 Hits GitHub Copilot: Opus-5-Level Results With Far Fewer Tokens — film_girl · 2026-09-23
- debt-gate: an open-source CI gate that blocks new tech debt in AI-built code — This_Cell_1829 · 2026-09-23
- Claude Opus 5.5 rolls out on Claude, API, AWS, Azure and Google Cloud, scoring 1846 on GDPval-AA v2.1 — minchoi · 2026-09-23
- Opus 5.5 audits a 200k-line codebase in under 3 hours using 2.5x fewer tokens than Opus 5 — minchoi · 2026-09-23
- Google paper proposes RRSI: regularized recursive self-improvement of agent harnesses — _akhaliq · 2026-09-23
- Claude Code Desktop bug: MCP elicitation declared but never wired, servers silently degrade — round · 2026-09-23