PRs that claim tests passed are worthless without proof. Stephen Chin and Cassandra Chin used AI Engineer to introduce Tracetune, an open source writer that rebuilds pull requests from agent traces.
I think they nailed the failure mode.
Cassandra Chin said session logs compress and get more inaccurate as agents work. Stephen Chin pointed to a real PR that needed exact preview links and command attribution. It claimed tests passed while traces showed bun test and bun x tsc running together, which must run sequentially.
The fix adds Relate, a Rust embedded graph database called Archie that builds an evidence packet with zero LLM cost. Chin said it lets a cheaper Haiku reviewer check claims while a Sonnet writer drafts. PR Bench judges scored Tracetune plus Relate 74 to 76 and 82 to 86 on Claude and Codex while using about half as many tokens.
To be fair, simply swapping to cheaper writers cuts spend but Chin found quality drops. Strata takes a different local route with one-click install for Windows and Linux serving Qwen3.8-Flash-Next over localhost OpenAI and Anthropic APIs. Rune stays at the UI layer as a native framework, so neither checks whether tests actually ran.
Tracetune still made occasional mistakes before the graph brief, which is why the provenance step matters. If you ship AI PRs, stop summarizing diffs and ship the trace.
