harbor: Skill for Claude Code

Test Boost Module

Send a real prompt through a running Boost module via harbor launch + pi and validate the output yourself. This is a live integration test, not a unit test.

Prerequisites

• llamacpp running (harbor up llamacpp)
• boost running (harbor up boost)
• pi installed on the host

If services aren't running, start them. Wait for health checks before proceeding.

Picking a Test Model

Use a small-to-mid model already available in llamacpp. Check what's loaded:

curl -s -H "Authorization: Bearer sk-boost" http://localhost:$(docker port harbor.boost 8000/tcp | head -1 | cut -d: -f2)/v1/models | python3 -c "
import sys, json
for m in json.load(sys.stdin).get('data', []):
    if m.get('status', {}).get('value') == 'loaded':
        print(f\"  LOADED  {m['id']}\")
    else:
        print(f\"  avail   {m['id']}\")
" 2>/dev/null

Good defaults (if available): unsloth/Qwen3.6-35B-A3B-GGUF:Q4_K_XL, unsloth/Qwen3.5-4B-GGUF:Q4_K_M, or any loaded non-embedding model. Strip the module prefix from the model ID when passing to --model.

The Command

harbor launch --workflow <module_name> --model "<base_model_id>" pi \
  -p --no-tools --no-session "<prompt>"

• --workflow <module_name> routes through Boost with that module active
• --model is the base llamacpp model (no module prefix)
• -p makes pi print-and-exit (non-interactive)
• --no-tools disables tool use for a clean completion
• --no-session keeps it ephemeral

Argument order matters

Launch options (--workflow, --model, --backend) go before pi. Pi options (-p, --no-tools) and the prompt go after pi.

Choosing the Right Prompt

The prompt should make the module's effect obvious in the output. Pick a prompt that produces clearly different output with vs. without the module.

Prompt strategies by module type