Caveman Input Compression Skill#

Identity#

Compress natural language workspace files into caveman-speak (~45% fewer input tokens). Creates .original.md backup before overwriting.

Supported Files#

Always safe to compress:

• SOUL.md ✅
• IDENTITY.md ✅
• USER.md ✅

Feature-flagged (requires env var set):

• AGENTS.md 🔒 ( CAVEMAN_COMPRESS_AGENTS=1)
• HEARTBEAT.md 🔒 ( CAVEMAN_COMPRESS_HEARTBEAT=1)
• MEMORY.md 🔒 ( CAVEMAN_COMPRESS_MEMORY=1)

Security#

• No shell=True in subprocess calls
• No user content as shell arguments
• Files >500KB rejected
• .original.md backup created before overwrite
• Aborts if backup already exists (prevents accidental overwrite)

API Fallback Chain#

Models tried in order:

1. claude CLI (desktop auth) - Primary
2. MiniMax (OpenAI-compatible) - Requires MINIMAX_BASE_URL + MINIMAX_API_KEY
3. DeepSeek (OpenAI-compatible) - Requires DEEPSEEK_API_KEY

Uses subprocess + curl, no OpenAI SDK required.

Usage#

bash

# Compress a specific file
python3 -m skills.caveman_compress.scripts.compress /home/pi/.openclaw/workspace/SOUL.md

# Dry run (no write, just preview)
python3 -m skills.caveman_compress.scripts.compress --dry-run /home/pi/.openclaw/workspace/SOUL.md

# Feature flags (environment variables)
CAVEMAN_COMPRESS_AGENTS=1 /home/pi/.openclaw/workspace/scripts/caveman-enable.sh

Model Override#

Set COMPRESSION_MODEL env var to use specific model:

bash

COMPRESSION_MODEL=deepseek32 python3 -m skills.caveman_compress.scripts.compress SOUL.md

Default: deepseek32 (cheap, fast)

References#

• Upstream: https://github.com/JuliusBrussee/caveman
• Paper: arXiv 2604.00025 — "Brevity Constraints Reverse Performance Hierarchies"
• Project: memory/projects/caveman-token-reducer/