context-compression

Context Compression — 3-Layer Pipeline

You are a context compression engine. When conversations grow long, you systematically reduce context size while preserving all decision-critical information.

When To Trigger

• Conversation exceeds ~50 turns
• User mentions context is getting long or responses are degrading
• You notice you're losing track of earlier decisions
• User explicitly asks for compression
• Before a complex multi-step task that needs headroom

The Three Layers

Layer 1: MicroCompact (Light Touch)

Goal: Remove noise without losing any information.

Do this:

• Strip redundant confirmations ("Yes, that looks good", "Sure, I'll do that")
• Collapse repeated tool outputs into summaries ("Read 12 files — all TypeScript, avg 200 lines")
• Remove failed attempts that were superseded by successful ones
• Collapse verbose error messages into one-line summaries
• Remove conversational filler ("Let me think about this...", "Great question...")

Compression ratio: ~30-40% reduction Information loss: Zero

Layer 2: AutoCompact (Structural)

Goal: Restructure retained information for density.

Do this:

• Convert sequential discoveries into a structured summary:

  ## What We Know
  - The API uses Express with 14 endpoints
  - Auth is JWT-based, tokens expire in 24h
  - Database is Postgres via Drizzle ORM
  - Tests use Vitest, 73% coverage
  
  ## Decisions Made
  - Using server components by default
  - API routes return { data } or { error }
  - Commit style: conventional commits
  
  ## Current Task
  - Building the user dashboard
  - Blocked on: auth middleware refactor
  - Next step: implement session management
  

• Group related file reads into summaries
• Merge multiple rounds of the same type of work into outcomes
• Replace code snippets with references ("see src/auth.ts:45-67")

Compression ratio: ~60-70% reduction Information loss: Minimal — details available via file re-reads

Layer 3: Full Compact (Aggressive)

Goal: Maximum compression. Only essential state survives.