Product evidence
Real open-source findings, regression replays, and proposed Fix PRs, with supporting evidence and review status.
Early identified a saved-state bug in Pydantic AI, then generated a repair and test after user initiation. Follow the product flow into draft Fix PR #3.
Read the Case FileFailed turns could bypass output handling and persistence behavior configured for the stream.
Read the Case FileA valid no-op replacement could raise a false pattern-not-found exception and stop a Conan recipe build.
Read the Case FileA streaming handler that returned early could leave an agent run with an incomplete or empty response.
Read the Case FileProduction optimization could remove a valid CommonJS default export, causing consumers' default imports to resolve to undefined at runtime.
Read the Case FileA valid form request could fail with HTTP 500 or return HTTP 200 with corrupted fields.
Read the Case FileHistorical regression replays follow the path below. Finding-to-fix cases follow a bug into a proposed repair.
Start with a confirmed regression and compare the last known-good release with the affected release.
Analyze the repository changes without providing the later issue or fix as the expected answer.
Execute a focused reproduction across the known-good, affected, and fixed versions.
Compare Early's finding with the public issue, introducing change, and upstream fix.
