Showcase: Automated our prompt testing with Freeplay and GitHub Actions

We were wasting hours manually testing LLM prompt changes across our RAG pipelines. Freeplay's SDK and API gave us the hooks to automate it.

Key setup:
* Freeplay project for each major workflow (e.g., `support_agent`, `doc_summarizer`)
* GitHub Actions workflow triggered on PRs with prompt changes

The workflow:
1. On PR, action extracts the changed prompt files.
2. Runs a test suite that uses Freeplay's SDK to log `test_scenario` runs against the new prompts.
3. Fetches evaluation results (latency, cost, scored outputs) via Freeplay's API.
4. Posts a comment on the PR with a summary table.

Example GitHub Actions step (simplified):

```yaml
- name: Run Freeplay Scenario Tests
env:
FREEPLAY_API_KEY: ${{ secrets.FREEPLAY_API_KEY }}
run: |
python scripts/test_prompts.py
--scenario "support_triage"
--prompt-version "${{ github.head_ref }}"
--test-file "test_data/triage_cases.json"
```

Output posted to PR:

```markdown
## Freeplay Test Summary | support_triage
| Metric | Baseline | This PR | Delta |
|--------|----------|---------|-------|
| Avg Latency | 1.2s | 0.9s | -25% |
| Avg Cost/Req | $0.003 | $0.002 | -33% |
| Accuracy Score | 0.92 | 0.94 | +0.02 |
```

Results:
* Catch performance regressions before merge.
* Quantify cost impact of prompt changes.
* Stopped guessing about "better" prompts. We have data.