feat(tui): Compact Mode Phase 2 - Compression Engine Integration #9018

Summary

Implement Phase 2 of Compact Mode: Compression engine for actual token reduction.

Background

Phase 1 provides UI compression (display only). Phase 2 adds actual token reduction sent to LLM APIs.

Phase 2 Scope (4 weeks)

Compression Technologies (Vendor-Agnostic Only)

Quality vs Compression Trade-off

Deliverables

• LLMLingua-2 Python sidecar integration
• Sidecar lifecycle management (lazy loading, timeout cleanup)
• Conversation pruning logic (keep recent N, compress older)
• TOON format serializer for structured data
• Compression threshold configuration
• Quality preservation settings (protect entities, numbers)
• Compression metrics display (tokens saved)

Technical Notes

• LLM Understanding Impact: 2-5% at recommended settings
• Vendor Agnostic: ✅ All techniques work with any LLM
• Excluded: Anthropic Token-Efficient Tools (Claude-only)
• Excluded: Provider-specific Prompt Caching (non-interoperable)

Best Practices

1. Preserve named entities and numbers explicitly
2. Use extractive (LLMLingua-2) over abstractive methods
3. Never compress recent 3 messages
4. Set maximum compression threshold at 80%

Files to Create/Modify

• packages/opencode/script/llmlingua-sidecar.py - Python compression service
• packages/opencode/src/session/compression.ts - Compression client
• packages/opencode/src/session/compaction.ts - Enhance existing pruning

Dependencies

• Requires: feat(tui): Compact Mode Phase 1 - Display Hierarchy Implementation #9017 (Phase 1)

Success Criteria

• 50-70% token reduction achieved
• Quality retention >95% on standard tasks
• Sidecar starts/stops correctly
• Metrics displayed in TUI

Related

• Phase 1: feat(tui): Compact Mode Phase 1 - Display Hierarchy Implementation #9017
• Phase 3: #TBD (Intelligence Layer)

Generated by Trinitas Full Mode Analysis