@tarquinen/opencode-dynamic-context-pruning

Dynamic Context Pruning Plugin

OpenCode plugin that optimizes token usage by analyzing conversation history and pruning obsolete tool outputs from requests.

Features

• Zero Configuration: Uses the free opencode/big-pickle model - no API keys required
• Automatic Optimization: Runs in the background when sessions become idle
• Smart Analysis: Uses AI to identify truly obsolete context
• Debug Logging: Optional file-based logging for troubleshooting

How It Works

• When a session becomes idle, the Janitor analyzes the conversation using shadow inference
• It identifies tool call outputs that are no longer relevant to the conversation
• These IDs are stored in memory for that session
• On subsequent requests, a custom fetch function filters out the obsolete tool responses
• Result: Fewer tokens sent to the LLM, lower costs, faster responses

Installation

This plugin is in development. Once complete, it will be installable via:

opencode plugin install dynamic-context-pruning

Configuration

Edit lib/config.ts to enable debug logging:

const defaultConfig: PluginConfig = {
  debug: true, // Set to true to enable debug logging
}

Debug logs are written to logs/YYYY-MM-DD.log files.

Architecture

• Janitor: Background process that analyzes sessions using opencode/big-pickle model
• State Manager: In-memory store for pruned tool call IDs per session
• Fetch Injector: Injects custom fetch via chat.params hook
• Pruning Fetch Wrapper: Filters request bodies based on session state
• Logger: File-based debug logging system

Development Status

Current implementation status: Step 1 - Project Scaffolding ✅

See @notes/dynamic-context-pruning/IMPLEMENTATION-PLAN.md for full implementation details.

License

MIT