How do I check code changes made by an agent?
Review the workspace patch, actual test evidence, and review report, and distinguish an agent's claim from a project check that ran.
Maintained by the 路墨 Lumoswitch product team
Inspect the patch, then confirm which tests and reviews actually ran. Do not treat an unrun check as a pass.
What should I inspect first?
1. Open this task's run history and file-change record; check every added, modified, deleted, or renamed path. 2. Read the patch against the request, configuration files, and any work that was already uncommitted. 3. Find the verification report and confirm its commands, results, and records belong to this run. If cited evidence is missing, check the task status before accepting. A diff preview may be truncated, so open the full file or diff in the project when needed.
How do I request a useful independent review?
Use an explicit read-only request, for example: “Review only the code changes from the current task; do not edit files. Check edge cases against the request and run the existing tests relevant to these files. Report the specific files and lines, commands actually run, pass or fail results, and risks that remain unverified.” 1. Choose Plan mode at the start if you want the workspace to plan, execute, and independently review the task. 2. If the change is already complete and you only want review, ask for no edits and inspect each reported check. 3. Model judgment does not replace tests; a command not listed in the report has no confirmed pass result.
What evidence shows that a check actually ran?
1. Read command output or the verification report in the run history, not just the final summary. 2. Compare the command with the repository's scripts; you can ask the agent to list failed test names instead of saying only “checks passed.” 3. For UI changes, run the project preview and inspect the target screen sizes yourself. The workspace can show captured file changes, test reports, and review references; the patch preview may be limited, and this is not a release sign-off.
What if a test fails or the result cannot be reviewed?
1. Note the failing command and first relevant error, then check whether the cause is the change or a missing environment dependency. 2. If the task is waiting for a decision or review, choose to continue adjusting and give the specific failure; do not complete the task just to bypass it. 3. If the runtime is unavailable, preserve the current result and run the check later in a local environment with the required dependencies. 4. Do not ask the agent to clean up out-of-scope files until you know whether they were your existing work.