Reconstructs a failed agent trace, isolates root causes, proposes a minimal patch, and generates one regression test per failure. I used Codex with GPT-5.6 to implement trace normalization, causal analysis, tests, and the product experience. A future live path can use GPT-5.6 to explain unstructured traces; the public demo is explicitly a precomputed, tested fixture, and deterministic checks enforce regression coverage.

Built With

Share this project:

Updates