cwinvestments/compress

⚙️ Compress — Headroom Proxy Manager

Monitor and manage Headroom context compression for CC sessions.

Activation

When this skill activates, output:

⚙️ Compress — Checking Headroom status...

Then execute the protocol below.

• Keywords: headroom, compression stats, token savings, proxy status, check headroom
• Contextual: When user asks about token usage, context window limits, or session cost optimization
• Level: 2 (explicit trigger only)

Context Guard

| Context | Status | |---------|--------| | User says "headroom", "compression stats", "check proxy" | ACTIVE — run status check | | User asks about token savings or context window | ACTIVE — run session report | | Proxy errors or API connection failures appear | ACTIVE — run health diagnostics | | General discussion about CC features | DORMANT — do not activate | | User is actively coding (no proxy issues) | DORMANT — do not activate |

What It Does

Headroom is a transparent proxy between Claude Code and the Anthropic API that compresses tool outputs by removing redundant boilerplate. It extends effective context window by 30–40%.

This skill checks proxy health, reports compression stats, and troubleshoots connection issues.

Prerequisites

• Headroom installed: pip install headroom-ai[code]
  • The [code] extra installs tree-sitter for AST-based code compression. Without it, Code-Aware compression is disabled and CC sessions get 0% compression.
• Proxy running: headroom proxy --llmlingua-device cpu (defaults to localhost:8787)
• CC configured: ANTHROPIC_BASE_URL=http://127.0.0.1:8787

Recommended Startup

headroom proxy --llmlingua-device cpu

• --llmlingua-device cpu — Forces LLMLingua to use CPU (avoids silent CUDA failures on machines without GPU)
• Default port is 8787, no other flags needed
• Code-Aware and LLMLingua both load lazily when relevant content is detected

Workflow

1. Status Check

Run:

curl -s http://127.0.0.1:8787/stats | python -m json.tool

Report: proxy up/down, requests processed, compression ratio, tokens saved, estimated cost savings.

2. Health Diagnostics

If proxy is unreachable:

1. Check if process is running:

   # Windows
   tasklist | findstr headroom
   # Linux/macOS
   ps aux | grep headroom
   

2. Check port binding:

   netstat -ano | findstr 8787
   

3. Verify ANTHROPIC_BASE_URL is set:

   echo $ANTHROPIC_BASE_URL
   

4. Restart: headroom proxy in a separate terminal

3. Session Report

When triggered at session end or on request, report:

• Requests this session
• Tokens before/after compression
• Compression ratio (target: 30–40%)
• Estimated dollar savings (at $15/MTok input, $75/MTok output for Opus)

4. Configuration Reference

| Setting | Value | Notes | |---------|-------|-------| | Proxy URL | http://127.0.0.1:8787 | Default port | | Dashboard | AdminStack Infrastructure tab | Headroom monitoring panel | | Repo | github.com/chopratejas/headroom | Apache 2.0 | | Python | 3.14 compatible | Tested Feb 2026 |

Troubleshooting

| Symptom | Fix | |---------|-----| | 0% compression / 0.00x ratio | headroom-ai[code] is not installed. Run: pip install headroom-ai[code]. Restart proxy. | | "Code-Aware: NOT INSTALLED" in startup banner | Same fix — install the [code] extra and restart. | | Cost figures don't match Anthropic Console | Headroom estimates costs at list token prices without accounting for Anthropic's server-side prompt caching discounts. For actual costs, check console.anthropic.com. |

Output Format

⚙️ Headroom Status
├── Proxy: ✅ Running on :8787
├── Requests: 47 processed
├── Compression: 46.2% reduction
├── Tokens saved: ~18,500 tokens
└── Cost savings: ~$0.28 this session

Integration

• AdminStack: Infrastructure page has Headroom tab with live dashboard
• CC Sessions: Auto-routed when ANTHROPIC_BASE_URL is set
• Monitoring: Stats endpoint polled every 30s with visibility-aware polling

Level History

• Lv.1 — Base: Health check and stats reporting for Headroom proxy. (Origin: MemStack v3.0, Feb 2026)
• Lv.2 — Fixed: Added [code] extra for tree-sitter AST compression, updated startup flags (--llmlingua-device cpu), added troubleshooting. Compression 0% → 46%. (Feb 24, 2026)