Process reports from the slop-off MCP server and apply them to the source.
Mode
- Default (loop): call
wait_for_report(timeout_seconds: 60), apply the report, briefly report what was done, and then immediately callwait_for_reportagain. Keep repeating this until the user says to stop ("stop", "done", or gives another command).- Timeout without a report? Just call
wait_for_reportagain, without comment. After ~10 empty timeouts in a row: ask the user once whether you should keep waiting. - At the start, announce once: "I'm waiting for browser edits — say 'stop' when you're done."
- Timeout without a report? Just call
- Argument
once: process exactly one report and stop. - Argument
latest: callget_latest_report(do not wait), apply, stop. - Argument
list: calllist_reports, show the queue, ask which one.
Apply edits
The report contains per-URL sections with Before/After HTML pairs and a CSS selector as a hint.
- For each edit, find the source file that renders that page/HTML: search for distinctive text from the Before block (literal strings first, then fuzzy). The URL indicates which route/page; the selector where in the DOM.
- Replace the Before content with the After content. Preserve existing formatting, indentation and templating (translate HTML changes to the template/JSX/component if the source is not plain HTML).
- Not found? Report it explicitly with the closest match — never silently guess or skip. Then just continue with the loop.
- Placeholder, href and other attribute changes are attribute edits; adjust only that attribute.
Per processed report
- Run a quick check if the project has one (typecheck/lint; no full build per report in loop mode).
- Summarize in 1-3 lines: which files, which edits, what failed.
- Go straight back to waiting.