Reasoning Pruner

Prunes historical reasoning parts from the message history before it is re-sent to the LLM, keeping only the last turn's reasoning. Saves context window and tokens on reasoning models (DeepSeek, GLM, Kimi, Qwen, ...) without touching what the model sees in the current turn.

• Model-agnostic: strips any part.type === "reasoning", regardless of provider key naming (reasoning_content vs reasoning / reasoning_details).
• Zero regex, zero event-stream manipulation, zero system-prompt injection.
• Reasoning is still visible in the TUI (/thinking) and still consumed during generation — only the historical copies are pruned.

Install

npm i @ghilteras/opencode-reasoning-pruner
# or: bun add @ghilteras/opencode-reasoning-pruner

Then add it to the plugin array of opencode.json:

{
  "plugin": ["@ghilteras/opencode-reasoning-pruner"]
}

How it works

Hooks experimental.chat.messages.transform: finds the index of the last KEEP_TURNS user message(s) and removes every reasoning part before that index. Default KEEP_TURNS = 1 (keeps the reasoning of the current turn, prunes everything older).

Config

Edit the KEEP_TURNS constant in plugin.js (default 1).

Notes

• opencode's built-in compaction.prune only trims tool-output bytes — it does NOT touch reasoning parts. This plugin complements it.
• Adjacent plugins exist for tool-output pruning (DCP/ACP) and for the opposite direction (preserving reasoning to fix provider 400s); none strip reasoning parts from history.

License

MIT © 2026 Angelo Pantano