Scope
Repository, requested change, agent command, version, and the authority granted to the run.
Reproducible review workflow
Pair the final diff with the execution path that produced it.
Source review remains authoritative for the code. AgentSight adds the missing run context: what the agent executed, which tests failed or passed, what it touched, and where the evidence remains incomplete.
Artifact anatomy
Keep the review record compact enough to inspect.
The goal is not to attach a raw trace to every PR. It is to reduce the run into four evidence groups that support specific review questions.
Execution
Commands, child processes, test attempts, exit status, retries, and long-running phases.
Effects
Files read or changed, generated artifacts, network destinations, and activity outside the stated scope.
Gaps
Relevant tests not observed, unavailable evidence, ambiguous attribution, and claims requiring another run.
Capture
Record the same command that is allowed to edit the repository.
Use a bounded task and a clean worktree. Record the agent command, let the normal tests and review tools run, then export a sanitized report. Do not infer that a test ran from the final source tree; require execution evidence.
Clean repository baselineExplicit task and authoritySanitized export
Terminal
$ git status --short
$ sudo agentsight record -- codex
$ agentsight report audit --json
$ agentsight report export -o pr-evidence.jsonValidation
Which exact test commands ran, in what order, with which exit status and retries?
Scope
Which paths were read or written, including generated artifacts and activity outside the repository?
Dependencies
Did a remote response, package source, model endpoint, or external service influence the patch?
Evidence gaps
Which important review questions remain unsupported by the retained run?
Review standard
A clean artifact can say “not observed” without pretending that means “did not happen.”
Preserve the distinction between the final diff, recorded execution evidence, and reviewer inference.