Your PR Says the Tests Passed. Prove It. - Stephen Chin & Cassandra Chin
TraceTune combines coding-agent audit trails with Relate graph retrieval to produce evidence-backed PRs. Its early PR Bench comparison reports higher judge scores for Claude and Codex, but gives limited detail on the evaluation setup.
TraceTune uses agent and code history to draft PRs with command context, test evidence, and linked facts. Its Relate component is an open source, in-memory graph database written in Rust; the workflow uses a stronger model for drafting and a cheaper one for mechanical review.
If agents write your PRs, keep their audit trail and tie claims like “tests passed” to commands run against the submitted change. Graph retrieval can assemble related evidence before drafting, reducing repeated agent research; TraceTune also checks claims against that evidence.
TraceTune uses agent and code history to draft PRs with command context, test evidence, and linked facts. Its Relate component is an open source, in-memory graph database written in Rust; the workflow uses a stronger model for drafting and a cheaper one for mechanical review. If agents write your PRs, keep their audit trail and tie claims like “tests passed” to commands run against the submitted change. Graph retrieval can assemble related evidence before drafting, reducing repeated agent research; TraceTune also checks claims against that evidence. In the talk’s PR Bench comparison, judge scores rose from **74 to 76** for one of Claude or Codex and **82 to 86** for the other; the transcript does not clearly map each range to a model. Human review was also used for accuracy, but sample size and evaluation details are not given. Relate avoids LLM retrieval costs, while still requiring local compute and storage.
TraceTune makes PR claims more auditable by retaining command context and linking claims to evidence, while graph retrieval can reduce repeated research during drafting. Its judge-score gains are modest and not clearly mapped to models, so they support a promising workflow rather than a general performance claim. Evidence checks show what was run and what supports a claim; they do not establish that the underlying claim is true.