| name | run-propeller-optimizer |
| description | Run the quadcopter-propeller AutoResearch optimization loop (the multi-agent search for the quietest, most efficient, highest-thrust propeller). Use when the user says "run the optimizer", "run the propeller loop", "start the autoresearch", "go to work on the propeller", "resume the run", or "kick off the overnight run". Handles fresh runs, crash-resume, the local-LLM swarm, and the unattended overnight runner.
|
Run the Propeller Optimizer
This skill launches the AutoResearch loop in src/. The loop
proposes propeller designs, scores them with trusted physics (Figure of Merit,
thrust, noise), keeps the Pareto-best, and writes everything to the SQLite store
of record at data/research.db. Full design: implementation_plan.md.
TL;DR — just run it
cd src
python -m autoresearch.researcher --no-llm --budget 60
That runs a fast, dependency-light optimization (no Ollama needed) for 60
seconds and prints the best designs. Outputs land in data/ and docs/.
Pick the mode
| Goal | Command (from src/) |
|---|
| Fast deterministic run (no AI, just the GA — best first try) | python -m autoresearch.researcher --no-llm --budget 60 |
| Full swarm run (local LLMs propose designs — needs Ollama) | python -m autoresearch.researcher --budget 1800 |
| Resume the last run after a crash/stop | python -m autoresearch.researcher --resume --budget 600 |
| Unattended overnight (auto-restart, caps, logging) | powershell -ExecutionPolicy Bypass -File ..\run_overnight.ps1 |
--budget is wall-clock seconds. Increase it for a longer search.
Before a full swarm run (only --no-llm needs nothing)
- Ollama running with:
ollama pull qwen2.5-coder:7b and ollama pull phi4-mini.
Check with python -m autoresearch.local_llm --check. If Ollama is down the
loop automatically falls back to the deterministic GA, so it never hard-fails.
- Python deps:
pip install -r ../requirements.txt.
What you get after a run
data/research.db — every design + score (the source of truth; query it with any SQLite tool).data/journal.md — one human-readable line per generation.docs/pareto.png — the trade-off plot (FM vs noise, colored by thrust).data/optimization/pareto_front.json + all_candidates.csv — exports.
Monitor / stop
- Watch progress live: open
data/journal.md, or query data/research.db
(WAL mode lets you read it while the loop runs).
- Stop an overnight run cleanly: drop a file named
STOP in the project root.
- Morning verdict of an overnight run:
data/RUN_STATUS.txt.
How it resumes (why a crash is safe)
Each generation is committed to data/research.db in one transaction, and the
runs table tracks last_gen. --resume adopts the latest unfinished run,
reloads its designs, and continues from the next generation — it does not
start over. The overnight runner passes --resume automatically on every restart.
