Snip Compressor

Semantic conversation compressor — intelligently summarize long chat history while preserving key decisions, facts, and context.

Distilled from Claude Code's Snip compression engine. Analyzes conversation history and generates a compressed summary that preserves semantic continuity — not a simple truncation.

When to use

• Conversation exceeds 60% of context window
• User says "continue" / "刚才说到哪了" / "where were we"
• Before compaction to preserve key context
• Session handoff between agents

How it works

1. Parse — Split conversation into segments (user messages, assistant replies, tool calls, tool results)
2. Score — Each segment gets a retention score based on:
   • Contains decision/agreement (high)
   • Contains factual data/numbers (high)
   • Contains tool results (medium, summarized)
   • Contains greetings/small talk (low)
3. Compress — Low-score segments get dropped or condensed; high-score segments preserved verbatim or lightly summarized
4. Rebuild — Output a compressed transcript that reads naturally

Usage

bash

# Compress a conversation file
python3 {baseDir}/compressor.py --input conversation.json --output compressed.md

# Compress with custom token budget
python3 {baseDir}/compressor.py --input conversation.json --budget 2000

# Compress from stdin
cat conversation.json | python3 {baseDir}/compressor.py --budget 1500

Input format

JSON array of message objects:

json

[
  {
    "role": "user",
    "content": "帮我分析一下这个股票"
  },
  {
    "role": "assistant",
    "content": "好的，我来看看...",
    "tool_calls": [
      {"name": "exec", "result": "{...}"}
    ]
  }
]

Output

Markdown with sections:

• Summary: 1-2 paragraph overview
• Key Decisions: bullet list of decisions made
• Active Context: what's currently in progress
• Compressed Transcript: compressed conversation (token-optimized)

Algorithm reference

Based on Claude Code's src/services/snip/ module:

• Semantic boundary detection: topic shift → new segment
• Importance scoring: decision > data > tool_result > greeting
• Token-aware truncation: respects model-specific limits
• Cross-segment reference preservation: if segment B references segment A, both are retained