feat: add preemptive compaction with multi-phase context reduction

[harshav167]

Summary

Adds an opt-in preemptive compaction feature that proactively manages context before it exceeds model limits. When enabled, it monitors token usage after each message and triggers compaction when usage exceeds a configurable threshold (default: 85%).

Multi-Phase Compaction Flow

1. DCP Strategies — Runs all enabled DCP strategies (deduplication, supersede writes, purge errors)
2. Tool Truncation — If still over threshold, truncates large tool outputs (preserving recent messages)
3. Summarization — If still over threshold, triggers OpenCode's built-in summarization

Features

• Model-aware context limits (Claude, GPT-4, Gemini, o1/o3, etc.)
• Configurable threshold, cooldown, and minimum tokens
• Protected recent messages from truncation
• Cooldown to prevent rapid re-compaction
• Event hook integration via message.updated

Configuration

Disabled by default (opt-in):

{
    "preemptiveCompaction": {
        "enabled": false,           // opt-in
        "threshold": 0.85,          // 85% context usage triggers compaction
        "cooldownMs": 60000,        // 1 minute between compactions
        "minTokens": 50000,         // minimum tokens before compaction can trigger
        "truncation": {
            "enabled": true,
            "protectedMessages": 3  // protect last 3 messages from truncation
        }
    }
}

Files Changed

• lib/preemptive/ - New preemptive compaction module
  • types.ts - Type definitions
  • constants.ts - Constants and model context limits
  • storage.ts - Truncation utilities
  • model-limits.ts - Context limit inference
  • index.ts - Main handler
• lib/config.ts - Added PreemptiveCompaction config interface and defaults
• dcp.schema.json - Added preemptiveCompaction schema
• index.ts - Registered event hook
• README.md - Documentation

Testing

• Build passes: bun run build ✓
• No TypeScript errors

[Tarquinen]

Hey, I don't think I can accept this for a few reasons:

1. Compaction is complicated and outside the scope of this project. It's possible that when this plugin matures, compaction may not be necessary with it. There's probably a lot that can be done to improve the default compaction, and if you're interested in that I would encourage you to make a plugin for it seperately.
2. The way you're detecting model context sizes in this PR is very fragile and will require lots of maintenance. I don't think hard coding values the way you've done is a good idea, you should at least use models.dev in a similar way to opencode. This implementation also doesn't account for custom models created in the opencode config.

Let me know what you think