Review AI-generated pull requests with run evidence

A source diff shows the final tree. It does not show which tests failed first, which files were read outside the repository, or which remote services influenced the result.

Pair the diff with the generating run.

See commands, exits, retries, and touched paths.

Focus human review on unexplained effects and missing validation.

Treat the trace as review evidence

Use the run profile to reconstruct how the agent reached the patch. Keep the source diff authoritative for code review and use AgentSight to expose the execution context the diff cannot contain.

Ask bounded questions

Did the agent run the relevant test? Did it retry after a failure? Did it read or write outside the stated scope? Did a network response or generated artifact affect the final patch? Each answer should point to recorded evidence.

Terminal

$ sudo agentsight record -- codex
$ agentsight report audit --json

Primary sources

AgentSight reports and audit commands