fix(wix-manage): quote bundle-and-save flow description to fix YAML parse error - #829

pairwise judge: prod wins (high confidence) | Response B is substantially better because it successfully completed the core task (10/10 LLM judge score vs 0/10) by properly loading recipes and persisting recommendations via BatchCreate. While Response A was more efficient in terms of cost and tokens, the complete task failure means efficiency gains are irrelevant. Task completion and tool accuracy are critical for this ecommerce task. | dimensions - task completion: without, tool accuracy: without, error handling: without, efficiency: with | LLM judge fails in PR, passes in prod | LLM judge: 0/10 PR vs 10/10 prod | efficiency (PR vs prod) - cost: $0.6048 vs $2.1720, tokens: 296,915 vs 1,059,604, time: 48.8s vs 212.3s

View run (PR)

View run (prod)

Assertions (PR):

• ❌ Tool called with param
• ❌ Tool called with param
• ❌ LLM judge (0/10): The agent failed to load the required recipes and did not persist or track the proposed recommendations via BatchCreate, resulting in a complete failure.

Assertions (prod):

• ✅ Tool called with param
• ✅ Tool called with param
• ✅ LLM judge (10/10): The agent perfectly classified the goal, loaded the correct recipes, proposed multi-item quantity discounts, and persisted them via BatchCreate before presenting.

Compare result: Response B is substantially better because it successfully completed the core task (10/10 LLM judge score vs 0/10) by properly loading recipes and persisting recommendations via BatchCreate. While Response A was more efficient in terms of cost and tokens, the complete task failure means efficiency gains are irrelevant. Task completion and tool accuracy are critical for this ecommerce task.