Prompt Optimizer v2.5.4

Summary

• Added feedback-driven evaluation flows so users can guide analysis with explicit evaluation notes.
• Updated evaluation templates to pay more attention to user feedback and improved result parsing robustness.
• Polished the analyze interaction flow and aligned more provider environment mappings.

Highlights

• Evaluation became more iterative: users can now add feedback directly into the evaluation flow instead of only reading a score afterward.
• The evaluation stack now treats user feedback as a stronger signal, which improves the usefulness of follow-up analysis.
• Parsing and provider configuration became more resilient, reducing avoidable friction in evaluation-heavy sessions.

Product Updates

Web

• Added an evaluation feedback editor and feedback-aware analyze flow across the main workspace surfaces.
• Merged focus input more cleanly into the analyze action so evaluation guidance can be supplied at the right moment.
• Kept the evaluation popover open during apply actions, making iterative review less disruptive.

Core/Infra

• Added support for user feedback across evaluation templates and made that feedback a higher-priority signal during evaluation.
• Improved evaluation result parsing robustness so structured output is interpreted more consistently.
• Added missing provider environment mappings and aligned provider handling across desktop logs, MCP selection, and docs.

Fixes

• Fixed evaluation result parsing edge cases that could weaken downstream analysis.
• Fixed missing provider environment mappings that affected some env-driven model setups.
• Fixed the evaluation popover behavior so apply actions no longer interrupt iterative review as easily.

Breaking Changes / Upgrade Notes

• None. Existing evaluation workflows continue to work, now with stronger feedback support.

Developer Notes

• v2.5.4 is the release where evaluation shifts from a mostly one-way score display toward a feedback-guided loop.
• If you maintain custom evaluation prompts or result consumers, this is the right baseline for the stronger feedback signal and more defensive parsing behavior.