diff --git a/pi-ralph-wiggum/CHANGELOG.md b/pi-ralph-wiggum/CHANGELOG.md
index 1cac967..c2fbbff 100644
--- a/pi-ralph-wiggum/CHANGELOG.md
+++ b/pi-ralph-wiggum/CHANGELOG.md
@@ -1,5 +1,12 @@
 # Changelog
 
+## Unreleased
+
+### Changed
+- Add a completion gate to Ralph prompts and skill guidance. Agents are now instructed to preserve required verification artifacts and record an exact monitor-rerunnable final command before emitting `<promise>COMPLETE</promise>`.
+- Queue Ralph follow-up messages with `streamingBehavior: "followUp"` to avoid runtime warnings when a loop tool schedules the next iteration while the agent is still processing.
+- Add a stale-prompt guard instructing agents to reload loop state and ignore already-completed loops instead of doing duplicate work.
+
 ## 0.2.0 - 2026-04-19
 
 ### Changed
diff --git a/pi-ralph-wiggum/README.md b/pi-ralph-wiggum/README.md
index 7eeeab9..ad7df84 100644
--- a/pi-ralph-wiggum/README.md
+++ b/pi-ralph-wiggum/README.md
@@ -46,11 +46,24 @@ You ask Pi to set up a ralph-wiggum loop.
   - It gets a prompt telling it to work on the task, update the task file, and call ralph_done when it finishes that iteration
   - When the iteration is done, it calls `ralph_done`, resending the same prompt*
 - Pi runs until either:
-  - All tasks are done (Pi sends `<promise>COMPLETE</promise>`)
+  - All tasks are done and final verification is externally rerunnable (Pi sends `<promise>COMPLETE</promise>`)
   - Max iterations (default 50)
   - You hit `esc` (pausing the loop)
 If you hit `esc`, you can run `/ralph-stop` to clear the loop. Alternatively, just tell Pi to continue to keep going.
 
+## Completion gate
+
+For build/test/refactor tasks, Ralph prompts the agent not to complete based only on checked checklist items. Before sending `<promise>COMPLETE</promise>`, the agent should:
+
+- Preserve any build artifacts, generated files, virtualenvs, or copied libraries required by final verification.
+- Record the exact final command, working directory, relevant environment variables, and output summary in the task file.
+- Ensure a separate monitor can rerun that command from the same worktree in a fresh shell.
+- Mark work blocked or deferred if the final command cannot be made externally rerunnable.
+
+## Stale prompt guard
+
+If an already-queued Ralph prompt arrives after a loop has completed, the agent should reload `.ralph/<name>.state.json` before doing work. If the loop state is `completed`, it should ignore the stale prompt, avoid file edits and task commands, and not call `ralph_done`.
+
 ## Commands
 
 | Command | Description |
diff --git a/pi-ralph-wiggum/SKILL.md b/pi-ralph-wiggum/SKILL.md
index a44568c..7b224ba 100644
--- a/pi-ralph-wiggum/SKILL.md
+++ b/pi-ralph-wiggum/SKILL.md
@@ -23,9 +23,25 @@ ralph_start({
 2. Work on the task and update the file each iteration.
 3. Record verification evidence (commands run, file paths, outputs) in the task file.
 4. Call `ralph_done` to proceed to the next iteration.
-5. Output `<promise>COMPLETE</promise>` when finished.
+5. Before outputting `<promise>COMPLETE</promise>`, run a final verification command that an external monitor can rerun from the same worktree.
 6. Stop when complete or when max iterations is reached (default 50).
 
+## Completion Gate
+
+For build/test/refactor tasks, do not mark complete based only on checked checklist items.
+
+Before emitting `<promise>COMPLETE</promise>`:
+
+- Preserve any build artifacts, generated files, virtualenvs, or environment setup required by the final verification command.
+- Record the exact final command, working directory, relevant environment variables, and output summary in the task file.
+- Ensure the command can be rerun by a separate monitor in a fresh shell from the same worktree.
+- If a test cannot be rerun externally, mark the item blocked or deferred instead of complete.
+- If cleanup removes required verification artifacts, recreate them or update the final command before completion.
+
+## Stale Prompt Guard
+
+Before doing any work from a Ralph prompt, reload `.ralph/<name>.state.json`. If the loop state says `"status": "completed"`, do not edit files, do not run task commands, and do not call `ralph_done`. Reply briefly that the stale prompt was ignored because the loop is already completed.
+
 ## User Commands
 
 - `/ralph start <name|path>` - Start a new loop.
@@ -58,7 +74,13 @@ Brief description.
 - [x] Completed item
 
 ## Verification
-- Evidence, commands run, or file paths
+- Commands run, working directories, relevant environment variables, outputs, and whether artifacts required for reruns were preserved
+
+## Final Verification
+- Exact monitor-rerunnable command: `<command>`
+- Working directory: `<path>`
+- Required preserved artifacts: `<paths>`
+- Result: `<output summary>`
 
 ## Notes
 (Update with progress, decisions, blockers)
@@ -70,4 +92,5 @@ Brief description.
 2. Update checklist and notes as you go.
 3. Capture verification evidence for completed items.
 4. Reflect when stuck to reassess approach.
-5. Output the completion marker only when truly done.
+5. Preserve the environment needed to rerun final verification.
+6. Output the completion marker only when truly done and externally rerunnable.
diff --git a/pi-ralph-wiggum/index.ts b/pi-ralph-wiggum/index.ts
index 8cce822..730dc7a 100644
--- a/pi-ralph-wiggum/index.ts
+++ b/pi-ralph-wiggum/index.ts
@@ -23,10 +23,34 @@ Describe your task here.
 - [ ] Item 1
 - [ ] Item 2
 
+## Verification
+- Commands run, working directories, relevant environment variables, outputs, and preserved artifacts
+
+## Final Verification
+- Exact monitor-rerunnable command: <command>
+- Working directory: <path>
+- Required preserved artifacts: <paths>
+- Result: <output summary>
+
 ## Notes
 (Update this as you work)
 `;
 
+const DEFAULT_COMPLETION_GATE = `COMPLETION GATE
+
+Do not output ${COMPLETE_MARKER} based only on checked checklist items.
+Before completion:
+1. Run a final verification command that an external monitor can rerun from the same worktree in a fresh shell.
+2. Record the exact command, working directory, relevant environment variables, and output summary in the task file.
+3. Preserve every artifact required by that command, including build directories, generated libraries, virtualenvs, caches, or copied dylibs.
+4. If cleanup removes required artifacts, recreate them or update the final command before completing.
+5. If the final command cannot be made externally rerunnable, mark the item blocked/deferred instead of complete.`;
+
+const DEFAULT_STALE_PROMPT_GUARD = `STALE PROMPT GUARD
+
+Before doing any work from a Ralph prompt, reload the loop state file named in the prompt (usually .ralph/<name>.state.json).
+If the state says \"status\": \"completed\", do not edit files, do not run task commands, and do not call ralph_done. Reply briefly that the stale prompt was ignored because the loop is already completed.`;
+
 const DEFAULT_REFLECT_INSTRUCTIONS = `REFLECTION CHECKPOINT
 
 Pause and reflect on your progress:
@@ -169,7 +193,10 @@ export default function (pi: ExtensionAPI) {
 		saveState(ctx, state);
 		currentLoop = null;
 		updateUI(ctx);
-		pi.sendUserMessage(banner);
+		pi.sendUserMessage(banner, {
+			deliverAs: "followUp",
+			streamingBehavior: "followUp",
+		});
 	}
 
 	function stopLoop(ctx: ExtensionContext, state: LoopState, message?: string): void {
@@ -235,6 +262,8 @@ export default function (pi: ExtensionAPI) {
 		if (isReflection) parts.push(state.reflectInstructions, "\n---\n");
 
 		parts.push(`## Current Task (from ${state.taskFile})\n\n${taskContent}\n\n---`);
+		parts.push(`\n## Stale Prompt Guard\n\n${DEFAULT_STALE_PROMPT_GUARD}\n`);
+		parts.push(`\n## Completion Gate\n\n${DEFAULT_COMPLETION_GATE}\n`);
 		parts.push(`\n## Instructions\n`);
 		parts.push("User controls: ESC pauses the assistant. Send a message to resume. Run /ralph-stop when idle to stop the loop.\n");
 		parts.push(
@@ -248,7 +277,7 @@ export default function (pi: ExtensionAPI) {
 			parts.push(`1. Continue working on the task`);
 		}
 		parts.push(`2. Update the task file (${state.taskFile}) with your progress`);
-		parts.push(`3. When FULLY COMPLETE, respond with: ${COMPLETE_MARKER}`);
+		parts.push(`3. When FULLY COMPLETE and the completion gate is satisfied, respond with: ${COMPLETE_MARKER}`);
 		parts.push(`4. Otherwise, call the ralph_done tool to proceed to next iteration`);
 
 		return parts.join("\n");
@@ -341,7 +370,10 @@ export default function (pi: ExtensionAPI) {
 				ctx.ui.notify(`Could not read task file: ${taskFile}`, "error");
 				return;
 			}
-			pi.sendUserMessage(buildPrompt(state, content, false));
+			pi.sendUserMessage(buildPrompt(state, content, false), {
+				deliverAs: "followUp",
+				streamingBehavior: "followUp",
+			});
 		},
 
 		stop(_rest, ctx) {
@@ -401,7 +433,10 @@ export default function (pi: ExtensionAPI) {
 
 			const needsReflection =
 				state.reflectEvery > 0 && state.iteration > 1 && (state.iteration - 1) % state.reflectEvery === 0;
-			pi.sendUserMessage(buildPrompt(state, content, needsReflection));
+			pi.sendUserMessage(buildPrompt(state, content, needsReflection), {
+				deliverAs: "followUp",
+				streamingBehavior: "followUp",
+			});
 		},
 
 		status(_rest, ctx) {
@@ -652,7 +687,10 @@ Examples:
 			currentLoop = loopName;
 			updateUI(ctx);
 
-			pi.sendUserMessage(buildPrompt(state, params.taskContent, false), { deliverAs: "followUp" });
+			pi.sendUserMessage(buildPrompt(state, params.taskContent, false), {
+				deliverAs: "followUp",
+				streamingBehavior: "followUp",
+			});
 
 			return {
 				content: [{ type: "text", text: `Started loop "${loopName}" (max ${state.maxIterations} iterations).` }],
@@ -717,7 +755,10 @@ Examples:
 			}
 
 			// Queue next iteration - use followUp so user can still interrupt
-			pi.sendUserMessage(buildPrompt(state, content, needsReflection), { deliverAs: "followUp" });
+			pi.sendUserMessage(buildPrompt(state, content, needsReflection), {
+				deliverAs: "followUp",
+				streamingBehavior: "followUp",
+			});
 
 			return {
 				content: [{ type: "text", text: `Iteration ${state.iteration - 1} complete. Next iteration queued.` }],
@@ -736,11 +777,14 @@ Examples:
 		const iterStr = `${state.iteration}${state.maxIterations > 0 ? `/${state.maxIterations}` : ""}`;
 
 		let instructions = `You are in a Ralph loop working on: ${state.taskFile}\n`;
+		instructions += `- Before doing work, reload .ralph/${state.name}.state.json; if status is completed, ignore this stale prompt and do not call ralph_done\n`;
 		if (state.itemsPerIteration > 0) {
 			instructions += `- Work on ~${state.itemsPerIteration} items this iteration\n`;
 		}
 		instructions += `- Update the task file as you progress\n`;
-		instructions += `- When FULLY COMPLETE: ${COMPLETE_MARKER}\n`;
+		instructions += `- Preserve artifacts needed by final verification\n`;
+		instructions += `- Record an exact monitor-rerunnable final command before completion\n`;
+		instructions += `- When FULLY COMPLETE and externally rerunnable: ${COMPLETE_MARKER}\n`;
 		instructions += `- Otherwise, call ralph_done tool to proceed to next iteration`;
 
 		return {