When you and a coding agent work in the same repository, the final diff can contain both sets of edits. A file may include a line you wrote, a line the agent wrote, and a later agent change that overwrote your work. Reviewing only the final file can hide that sequence.
The safest review starts with the current working tree and the task you gave the agent. Then it treats origin as evidence to check, not as a reason to accept or reject a change automatically.
Preserve a reference point
Before a long agent run, inspect git status --short. If you already have local edits, note which files they touch. A small commit or a separate worktree can make the boundary clearer when that fits your workflow, but neither is required to review the final diff.
After the run, inspect git status --short again. New and modified files show where to focus. Compare the result with your starting state, especially in files you edited yourself.
Review overlapping hunks closely
An overlap matters when the agent changes nearby lines or replaces code you wrote during the session. Look at the surrounding function, not only the colored lines of the diff. Ask whether the final behavior still includes your intended change.
A clean test run may miss a lost edit if there is no test for that behavior. If you find an overwrite, restore the needed behavior and run the relevant checks again.
Avoid treating origin as certainty
Attribution can be incomplete. Formatting tools, generators, and edits made between snapshots can blur the boundary between human and agent work. An unknown label is useful information: it tells you to inspect the change without making an assumption about who wrote it.
The decision to keep a hunk should rest on what it does, whether it belongs to the request, and whether it has been verified. Its likely author helps you decide where to look first.
End with one coherent diff
When the review is done, read git diff from start to finish. Every remaining change should have a reason to be there. If a useful but unrelated edit remains, consider separating it from the agent's requested work.
Diffward records changes during a local agent session and marks each hunk as agent, human, or unknown when it can. It also warns when the agent appears to overwrite lines you edited. Those signals are still being evaluated, so confirm important decisions in the actual code and test results.