Plan ordering decides success in 9-week codegen experiment

Ok_Astronomer_526 · reddit · 2026-08-31

A user ran a repetitive code generation task for 9 weeks using the same repo, prompt, and Claude model to generate an OpenAPI client. The only variable was saving each run's execution plan. Comparisons reveal near-identical plans except for one step: whether shared test fixtures are updated before or after call sites. In 6 runs where fixtures came first, the type check passed immediately. In the 3 where they came last, it failed and required a retry. Despite minimal spec changes and no prompt mentions of fixtures, this single random variation dictated the outcome.

Original post →

More from coding & agent

coding & agent channel →