DedrooM

Loop detection + context compression for AI coding agents.

DedrooM sits between your AI agent and the LLM provider to:

• Detect and block infinite loops — saves wasted API calls when tools repeat
• Compress context — reduces token usage by 60–95% without changing behavior
• Redact sensitive data — strip API keys, tokens, and secrets from tool outputs
• Track savings — attribution engine shows per-tool token usage

Quick Start

pip install dedroom

# Start the proxy daemon and route your agent through it
dedroom init
eval "$(dedroom init)"       # Sets ANTHROPIC_BASE_URL and OPENAI_BASE_URL

# Use your agent as normal
claude          # now routed through DedrooM
codex           # works immediately

# Check status and stop
dedroom status  # Running state, PID, uptime, tokens saved
dedroom stop    # Stop the daemon

One-shot alternative (no daemon)

dedroom wrap claude   # Starts proxy, launches agent, cleans up on exit

Commands

Use any LLM provider (not just OpenAI/Anthropic)

# DeepSeek API
dedroom wrap claude \
  --upstream-url https://api.deepseek.com \
  --api-key "sk-your-key"

# Local Ollama
dedroom wrap aider \
  --upstream-url http://localhost:11434/v1

Python API

from dedroom import DedrooM

pipeline = DedrooM("""
loop_detection:
  max_repeats: 3
""")

# Check for loops (0 = Allow)
verdict = pipeline.verify("write_file", '{"path": "/tmp/x.txt"}')

# Full pipeline
result = pipeline.process_tool("write_file", '{}', tool_result)
print(f"Blocked: {result['is_blocked']}")
print(f"Saved {result['original_tokens'] - result['compressed_tokens']} tokens")

Benchmarks

• Loop detection latency: ~1.3ms per tool call (negligible vs 2-10s LLM roundtrip)
• Pipeline throughput: 5.4µs (in-memory) / 260µs (SQLite)

Development

git clone https://github.com/Devaretanmay/dedroom
cd dedroom

# Build Rust binaries
cargo build -p dedroom-cli -p dedroom-proxy

# Install Python package in dev mode
pip install -e .

# Run tests
pytest python/tests/