Analysis and re-analysis

ReviewArc makes the pull request’s complete current diff available against its base revision. The harness reads relevant per-file pages through bounded local tools, and ReviewArc records full, partial, and unread diff coverage. Re-analysis is intentionally end to end: it evaluates the current pull request again rather than analyzing only commits added since the previous run.

That choice keeps the recommendation, judgments, evidence, blind spots, and cross-file behavior coherent after a revision changes. It also means re-analysis cost and time generally scale with the full pull request, not only its newest commit.

Read-only execution

ReviewArc invokes the selected compatible harness with a strict JSON output schema and controlled access to an immutable local analysis bundle. The initial prompt does not embed the full diff. The analysis process is read-only. ReviewArc does not run arbitrary project commands, package scripts, tests, hooks, or generated instructions from the repository.

Progress is streamed into the app while the structured agent review is built. The result includes an approval recommendation, explicit human judgments, the observations behind them, verified signals, and remaining blind spots. ReviewArc does not merge separate findings in the interface based on similar wording; any grouping must be explicit in the agent’s structured result and explain the shared root cause.

Evidence validation

AI output is treated as a proposal, not ground truth. Before displaying evidence, ReviewArc validates that each referenced file belongs to the reviewed revision and that each line range exists in that file or diff. This validates the citation, not the agent’s interpretation of it.

Claims without valid supporting evidence remain explicitly unverified. ReviewArc preserves the distinction between:

• Facts supported directly by the reviewed source or diff.
• Inferences reasoned from available evidence.
• Unknowns that cannot be resolved from the pull request alone.

Stale analysis

ReviewArc records the head revision used by an analysis. When GitHub reports a different pull request head, the existing analysis is marked stale. Your previous analysis and notes remain available, but the app makes the revision mismatch visible and offers re-analysis against the new head.