Skip to main content

Kaggle Demo

This repo includes a demo package that behaves like a small research project. It is the fastest way to see goalseek produce runs, logs, and git-backed iteration history.

Fresh project walkthrough​

1. Install the package​

uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"

2. Create a new project​

uv run goalseek project init demo --provider claude_code --model claude-haiku-4-5-20251001

3. Copy the included sample assets​

./move-testpackage.sh --overwrite ./demo

4. Validate the manifest​

For the irrigation example, the metric should be Balanced accuracy Score.

metric:
  name: Balanced accuracy Score
  direction: maximize

Then validate:

uv run goalseek manifest validate ./demo

5. Run setup​

uv run goalseek setup ./demo

6. Clean the git tree​

uv run goalseek gittreeclean --message "clean repo" ./demo

7. Run the baseline​

uv run goalseek baseline ./demo

8. Run three iterations​

uv run goalseek run ./demo --iterations 3

9. Review the result​

uv run goalseek status ./demo
uv run goalseek summary ./demo

Helpful files during the demo​

  • demo/logs/state.json
  • demo/logs/results.jsonl
  • demo/runs/0001/prompt.md
  • demo/runs/0001/provider_output.md
  • demo/runs/0001/result.json

Common fixes​

Dirty working tree before iteration​

Check the tree:

git -C demo status --short

Keep the edits:

git -C demo add --all
git -C demo commit -m "save local changes"

Discard the edits you do not want:

git -C demo restore <path>

Missing Python packages during setup​

Install the required dependencies in your project environment and rerun goalseek setup ./demo.

Verification surprises​

Inspect demo/runs/<iteration>/verifier.log and confirm the metric extractor matches the verification output.