Slop Off

Apply browser edits

Process reports from the slop-off MCP server and apply them to the source. Everything happens in the MAIN thread — no subagents — so every report gives an immediate "✅ …" line and nothing sits between the queue and the work.

Mode

• Default (loop): say once "Waiting for browser edits — say 'stop' when you're done", then loop:
  1. Call wait_for_report with timeout_seconds: 60. Empty? Call it again, silently — empty waits are normal. After ~10 consecutive empty minutes, mention in one line that you're still waiting (don't ask, don't stop); repeat that at most once an hour.
  2. Report received → apply it per "Apply edits" below, then follow "Per processed report".
  3. Loop again immediately — more edits may already be queued. Stop only when the user says "stop" (or "done").
• Argument once: process exactly one report and stop.
• Argument latest: call get_latest_report (don't wait), process, stop.
• Argument list: call list_reports, show the queue, ask which one.
• Argument clear: call clear_reports and report in a single line how many reports were cleared. Process nothing, then stop.

Apply edits

The report contains per-URL sections with Before/After HTML pairs and/or Element/Instruction pairs, each with a CSS selector as a hint.

1. 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 says which route/page; the selector where in the DOM.
2. Replace the Before content with the After content. Preserve existing formatting, indentation and templating (translate HTML changes into the template/JSX/component if the source isn't plain HTML).
3. Not found? Report it explicitly with the closest match — never silently guess or skip. Then just continue the loop.
4. Placeholder, href and other attribute changes are attribute edits; change only that attribute.
5. Element/Instruction pairs are annotations: find the element (selector + Element snippet) and carry out the instruction on that element in the source. These are free-form tasks ("make this shorter", "different color") — carry them out to the best of your judgment and report what you did.

Per processed report

• Run a quick check if the project has one (typecheck/lint; no full build per report in loop mode).
• ALWAYS call the slop-off MCP tool notify_browser with a concrete 1-2 line summary of what changed (e.g. "Hero heading and CTA text updated in index.html") — also on failure, never with an empty message. The extension shows it as a toast on the page, and the call marks the report as done in the browser HUD: skip it and the report stays stuck on "applying" for the user.
• Report to the user in a single line: "✅ N change(s) applied — file1, file2 · M still queued" (N and M come from the report header; when M is 0 say "· queue empty, waiting"). Only if something failed a second line: "⚠️ not applicable: …". No further explanation.
• Go straight back to waiting.