potatoupdate-ralph-loop - SKILL.md Agent Skill

name: potato:update-ralph-loop description: "Use this to update the status of a ralph loop status for any iterative workflow. Controls whether to continue iterating, complete successfully, or escalate when blocked."

Updating Ralph Loop

The Rule

When your review is complete, call the ralph_loop_dock immediately. Not after cleanup. Not after one more check. Immediately.

Review done → ralph_loop_dock() → Nothing else

The Tool

ralph_loop_dock({
  approved: boolean,      // true = work is good, false = needs revision
  feedback?: string       // REQUIRED when approved=false
})

When to Approve

Call ralph_loop_dock({ approved: true }) when:

All requirements are met
Tests pass
No blocking issues remain

When to Reject

Call ralph_loop_dock({ approved: false, feedback: "..." }) when work needs revision.

Your feedback is injected into the next iteration. The builder will see exactly what you wrote. Be specific. Be actionable.

Bad Feedback	Good Feedback
"Needs work"	"Missing error handling for network timeouts in fetchUser()"
"Tests incomplete"	"No tests for edge case: empty input array in processItems()"
"Not quite right"	"Button placement violates design spec - should be right-aligned per mockup"

When to Block

If you cannot approve AND revision won't help (missing requirements, need user input, fundamental design question), still call:

ralph_loop_dock({
  approved: false,
  feedback:
    "BLOCKED: Cannot proceed without user clarification on auth strategy",
});

The system handles escalation. Your job is to signal.

Red Flags - STOP Immediately

These thoughts mean you're avoiding the tool:

Thought	Reality
"I'll just note this in the code"	Notes don't signal the orchestrator. Call the tool.
"The exit code will handle it"	Exit codes lose your feedback. Call the tool.
"I need to do one more thing first"	Review is done when review is done. Call the tool now.
"This is obvious, no feedback needed"	If rejected, feedback is required. Always explain why.

The Contract

One call per review - Call ralph_loop_dock exactly once at the end of your review
Feedback on rejection - If approved: false, explain what's wrong
No file updates - The tool handles storage. Don't write to ralph loop files.
Immediate call - When your verdict is clear, call immediately

What Happens Next

Approved: Loop completes, phase advances
Rejected + attempts remain: Builder gets your feedback, tries again
Rejected + max attempts: Ticket blocks, user notified

Your feedback directly shapes the next iteration. Make it count.