fix: remove empty tagged messages after tool drop (#98, closes #97)

Persistence-boundary strip now removes whole §N§ pairs (not bare §), eliminating the digit-residue that stacked across turns into runaway 'multiplying numbers' overflow (#97). hasMeaningfulPart strips the tag prefix before the emptiness check so tag-only ghost messages are pruned. Shared tag-strip helpers in tag-content-primitives.ts are reused by both OpenCode and Pi.

Thanks to @Zireael for the fix.

Author: Zireael

packages/pi-plugin/src/strip-tag-prefix.test.ts

@@ -52,7 +52,7 @@ describe("stripTagPrefixFromAssistantMessage", () => {
 	});
 
 	describe("cargo-culted § mid-text (models mimicking MC notation)", () => {
-		it("strips § from mid-text §N§ pair (cargo-cult defense)", () => {
+		it("removes mid-text §N§ pair entirely (cargo-cult defense)", () => {
 			const msg = {
 				role: "assistant",
 				content: [
@@ -62,21 +62,20 @@ describe("stripTagPrefixFromAssistantMessage", () => {
 					},
 				],
 			};
-			// V7: mid-text § is stripped as cargo-cult defense; digits stay as plain text.
 			expect(stripTagPrefixFromAssistantMessage(msg)).toBe(true);
 			expect((msg.content[0] as { type: string; text: string }).text).toBe(
-				"Looking at 5 which references the earlier discussion",
+				"Looking at  which references the earlier discussion",
 			);
 		});
 
-		it('strips § from malformed §N"> hybrid mid-text', () => {
+		it('removes malformed §N"> hybrid mid-text', () => {
 			const msg = {
 				role: "assistant",
 				content: [{ type: "text", text: `Hello §40827">Oracle confirmed` }],
 			};
 			expect(stripTagPrefixFromAssistantMessage(msg)).toBe(true);
 			expect((msg.content[0] as { type: string; text: string }).text).toBe(
-				`Hello 40827">Oracle confirmed`,
+				"Hello Oracle confirmed",
 			);
 		});
 
@@ -103,7 +102,7 @@ describe("stripTagPrefixFromAssistantMessage", () => {
 			};
 			expect(stripTagPrefixFromAssistantMessage(msg)).toBe(true);
 			expect((msg.content[0] as { type: string; text: string }).text).toBe(
-				"The pattern 40827 appeared.",
+				"The pattern  appeared.",
 			);
 		});
 	});

packages/pi-plugin/src/strip-tag-prefix.ts

@@ -1,79 +1,36 @@
 /**
- * Strip injected `§N§` tag prefixes AND defensively strip any cargo-cult
- * `§` characters from assistant text before Pi persists the message.
+ * Strip injected `§N§` tag prefixes AND defensively strip cargo-cult MC tag
+ * notation from assistant text before Pi persists the message.
  *
- * # Why this exists
+ * Mirrors OpenCode's `text-complete.ts` via {@link stripPersistedAssistantText}:
+ * whole `§N§` pairs globally, malformed hybrids, then stray `§`. Does not strip
+ * bare leading digits on the transform path.
  *
- * Magic Context tags every visible message part with a `§N§` prefix in
- * the transform pipeline so the agent can reference parts by tag id
- * (`ctx_reduce(§3§)`). LLMs frequently mimic that prefix in their own
- * generated text — emitting `§4§ Yes...` at the start of an assistant
- * response. This is harmless for cache (the agent emitting the prefix
- * doesn't bust prefix cache; it's the same content shape we already
- * inject) and the next transform pass strips/re-injects with the
- * correct tag id.
+ * Pi persists raw assistant text from `message_end`; this hook mutates text parts
+ * before `agent-session.ts:appendMessage()` writes jsonl.
  *
- * BUT: Pi persists the raw assistant text from `message_end` events
- * directly into the session jsonl, and the Pi UI renders from that
- * stored text. So while OpenCode hides the prefix from its own UI via
- * `experimental.text.complete` (which mutates `output.text` before
- * persistence), Pi's UI shows the raw mimicked prefix to the user
- * because nothing scrubs it on the way to disk.
- *
- * This module mirrors OpenCode's `text-complete.ts` for Pi by hooking
- * `pi.on("message_end", ...)`. The event fires synchronously before
- * `agent-session.ts:appendMessage()` persists the message — the event
- * runner emits to extensions FIRST, then persists by reference, so
- * mutating `event.message.content[i].text` is visible to the
- * persistence call.
- *
- * # Two-step strip (matches OpenCode `text-complete.ts`)
- *
- * 1. `^(§N§\s*)+` removes well-formed leading prefix runs (the canonical
- *    case where the model correctly mimics MC's tag prefix at the start
- *    of a response). Digit-aware: removes the whole `§N§ ` pair cleanly,
- *    leaving no digit residue.
- *
- * 2. Global `§` strip removes ANY remaining `§` character anywhere in
- *    the text. Defends against cargo-cult patterns observed when execute
- *    passes drop most tool structure:
- *      - `§40827§` mid-text (well-formed cargo-cult pair)
- *      - `§40827">` (malformed partial — hybrid of MC tag and XML)
- *      - stray `§` anywhere
- *
- *    Only the MC transform layer is authorized to write `§N§` prefixes.
- *    Any `§` reaching this hook from the model is by definition wrong.
- *    Cost: legitimate `§5.1` section refs become `5.1`; models adapt
- *    naturally to alternative notation.
- *
- * # Scope
- *
- * Only `assistant` messages need stripping. User messages are
- * user-typed text (no LLM mimicking). Tool result messages keep their
- * tagger-injected prefix because that prefix is intentional context.
+ * Only `assistant` messages are stripped. User/tool messages keep intentional tags.
  */
 
-const LEADING_TAG_PREFIX_REGEX = /^(\u00a7\d+\u00a7\s*)+/;
-const SECTION_CHAR_REGEX = /\u00a7/g;
+import { stripPersistedAssistantText } from "../../plugin/src/hooks/magic-context/tag-content-primitives";
 
 /**
- * Mutate the given assistant message's text parts in place to strip
- * any leading `§N§` tag prefixes.
+ * Mutate the given assistant message's text parts in place to strip MC tag notation.
  *
- * Returns true if any text was modified, false otherwise. The return
- * value is informational; the actual mutation happens on the passed
- * message reference.
- *
- * Exported for testing. Production callers should use `registerStripTagPrefix`.
+ * Returns true if any text was modified. Production callers use `registerStripTagPrefix`.
  */
+
 export function stripTagPrefixFromAssistantMessage(message: {
 	role: string;
+
 	content: unknown;
 }): boolean {
 	if (message.role !== "assistant") return false;
+
 	if (!Array.isArray(message.content)) return false;
 
 	let mutated = false;
+
 	for (const part of message.content) {
 		if (
 			part === null ||
@@ -82,15 +39,19 @@ export function stripTagPrefixFromAssistantMessage(message: {
 		) {
 			continue;
 		}
+
 		const textPart = part as { type: "text"; text: unknown };
+
 		if (typeof textPart.text !== "string") continue;
-		const stripped = textPart.text
-			.replace(LEADING_TAG_PREFIX_REGEX, "")
-			.replace(SECTION_CHAR_REGEX, "");
+
+		const stripped = stripPersistedAssistantText(textPart.text);
+
 		if (stripped !== textPart.text) {
 			textPart.text = stripped;
+
 			mutated = true;
 		}
 	}
+
 	return mutated;
 }

packages/pi-plugin/src/temporal-awareness-pi.ts

@@ -27,6 +27,10 @@ import {
 	TEMPORAL_MARKER_PATTERN,
 	temporalMarkerPrefix,
 } from "@magic-context/core/hooks/magic-context/temporal-awareness";
+import {
+	peelLeadingMcTagNotation,
+	stripTagPrefix,
+} from "../../plugin/src/hooks/magic-context/tag-content-primitives";
 
 type PiTextContent = { type: "text"; text: string; textSignature?: string };
 type PiImageContent = { type: "image"; data: string; mimeType: string };
@@ -74,9 +78,9 @@ export function injectPiTemporalMarkers(messages: unknown[]): number {
 				const userMsg = msg as PiUserMessage;
 				if (typeof userMsg.content === "string") {
 					if (!TEMPORAL_MARKER_PATTERN.test(stripTagPrefix(userMsg.content))) {
-						const tagMatch = userMsg.content.match(/^(?:§\d+§\s*)+/);
-						const tagPrefix = tagMatch ? tagMatch[0] : "";
-						const body = userMsg.content.slice(tagPrefix.length);
+						const { tagPrefix, body } = peelLeadingMcTagNotation(
+							userMsg.content,
+						);
 						(messages as PiAgentMessage[])[i] = {
 							...userMsg,
 							content: tagPrefix + prefix + body,
@@ -92,9 +96,7 @@ export function injectPiTemporalMarkers(messages: unknown[]): number {
 					);
 					if (firstTextIndex >= 0) {
 						const existing = userMsg.content[firstTextIndex] as PiTextContent;
-						const tagMatch = existing.text.match(/^(?:§\d+§\s*)+/);
-						const tagPrefix = tagMatch ? tagMatch[0] : "";
-						const body = existing.text.slice(tagPrefix.length);
+						const { tagPrefix, body } = peelLeadingMcTagNotation(existing.text);
 						if (!TEMPORAL_MARKER_PATTERN.test(body)) {
 							const newContent = userMsg.content.slice();
 							newContent[firstTextIndex] = {
@@ -122,14 +124,3 @@ export function injectPiTemporalMarkers(messages: unknown[]): number {
 
 	return injected;
 }
-
-/**
- * Strip the leading `§N§` tag prefix(es) so we test the marker pattern
- * against the body, not against the tagged form. Matches OpenCode's
- * approach in `injectTemporalMarkers`.
- */
-function stripTagPrefix(text: string): string {
-	const match = text.match(/^(?:§\d+§\s*)+/);
-	if (!match) return text;
-	return text.slice(match[0].length);
-}