diff --git a/packages/agent/CHANGELOG.md b/packages/agent/CHANGELOG.md
index 87d91181474..77be2ac9c7c 100644
--- a/packages/agent/CHANGELOG.md
+++ b/packages/agent/CHANGELOG.md
@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed compaction summary inputs escaping Harmony control tokens so Copilot `gpt-5.6-*` models no longer reject serialized analysis-channel markers. ([#5184](https://github.com/can1357/oh-my-pi/issues/5184))
+
 ## [16.4.3] - 2026-07-11
 
 ### Fixed
diff --git a/packages/agent/src/compaction/branch-summarization.ts b/packages/agent/src/compaction/branch-summarization.ts
index 93901d67022..ec1e3bb7315 100644
--- a/packages/agent/src/compaction/branch-summarization.ts
+++ b/packages/agent/src/compaction/branch-summarization.ts
@@ -27,7 +27,7 @@ import {
 	extractFileOpsFromMessage,
 	type FileOperations,
 	SUMMARIZATION_SYSTEM_PROMPT,
-	serializeConversation,
+	serializeConversationForSummary,
 	stripReadSelector,
 	truncateToolResultForSummary,
 	upsertFileOperations,
@@ -320,7 +320,7 @@ export async function generateBranchSummary(
 	// Transform to LLM-compatible messages, then serialize to text
 	// Serialization prevents the model from treating it as a conversation to continue
 	const llmMessages = (options.convertToLlm ?? defaultConvertToLlm)(messages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 
 	// Build prompt
 	const instructions = customInstructions || BRANCH_SUMMARY_PROMPT;
diff --git a/packages/agent/src/compaction/compaction.ts b/packages/agent/src/compaction/compaction.ts
index 37dd360ca02..2b518601067 100644
--- a/packages/agent/src/compaction/compaction.ts
+++ b/packages/agent/src/compaction/compaction.ts
@@ -66,7 +66,7 @@ import {
 	extractFileOpsFromMessage,
 	type FileOperations,
 	SUMMARIZATION_SYSTEM_PROMPT,
-	serializeConversation,
+	serializeConversationForSummary,
 	stripReadSelector,
 	upsertFileOperations,
 } from "./utils";
@@ -816,7 +816,7 @@ export async function generateSummary(
 	// Serialize conversation to text so model doesn't try to continue it
 	// Convert to LLM messages first (handles custom app messages when caller provides a transformer).
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(currentMessages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 
 	// Build the prompt with conversation wrapped in tags
 	let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
@@ -1030,7 +1030,7 @@ async function generateShortSummary(
 ): Promise<string> {
 	const maxTokens = Math.min(512, Math.floor(0.2 * reserveTokens));
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(recentMessages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 
 	let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
 	if (historySummary) {
@@ -1578,7 +1578,7 @@ async function generateTurnPrefixSummary(
 	const maxTokens = Math.floor(0.5 * reserveTokens); // Smaller budget for turn prefix
 
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(messages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 	const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;
 	const summarizationMessages = [
 		{
diff --git a/packages/agent/src/compaction/utils.ts b/packages/agent/src/compaction/utils.ts
index fc3bb6965d7..286f4e7f951 100644
--- a/packages/agent/src/compaction/utils.ts
+++ b/packages/agent/src/compaction/utils.ts
@@ -207,9 +207,19 @@ export function truncateToolResultForSummary(text: string): string {
 	return `${text.slice(0, TOOL_RESULT_MAX_CHARS)}\n\n[... ${truncatedChars} more characters truncated]`;
 }
 
+const HARMONY_CONTROL_TOKEN_RE = /<\|(start|end|message|channel|constrain|return|call)\|>/g;
+
+/**
+ * Serialize LLM messages as plain summary input without provider control tokens.
+ */
+export function serializeConversationForSummary(messages: Message[], dialect?: Dialect): string {
+	const conversation = serializeConversation(messages, dialect);
+	if (dialect !== "harmony") return conversation;
+	return conversation.replace(HARMONY_CONTROL_TOKEN_RE, "<\\|$1\\|>");
+}
+
 /**
- * Serialize LLM messages to text for summarization.
- * This prevents the model from treating it as a conversation to continue.
+ * Serialize LLM messages to transcript text.
  * Call convertToLlm() first to handle custom message types.
  */
 export function serializeConversation(messages: Message[], dialect?: Dialect): string {
diff --git a/packages/agent/test/serialize-conversation.test.ts b/packages/agent/test/serialize-conversation.test.ts
index d91c059a0de..0c7b119e494 100644
--- a/packages/agent/test/serialize-conversation.test.ts
+++ b/packages/agent/test/serialize-conversation.test.ts
@@ -1,5 +1,5 @@
 import { describe, expect, test } from "bun:test";
-import { serializeConversation } from "@oh-my-pi/pi-agent-core/compaction";
+import { serializeConversation, serializeConversationForSummary } from "@oh-my-pi/pi-agent-core/compaction";
 import type { AssistantMessage, Message, ToolResultMessage, Usage } from "@oh-my-pi/pi-ai";
 
 const ZERO_USAGE: Usage = {
@@ -82,6 +82,41 @@ describe("serializeConversation — useless pairs", () => {
 		expect(out).not.toContain("[Assistant tool calls]:");
 	});
 
+	test("summary serialization escapes Harmony control tokens while preserving assistant thinking", () => {
+		const messages = [
+			assistantMessage([
+				{ type: "thinking", thinking: "Need to inspect the failing compaction path." },
+				{ type: "text", text: "The final answer stays visible." },
+			]),
+		];
+
+		const out = serializeConversationForSummary(messages, "harmony");
+
+		expect(out).not.toContain("<|channel|>analysis");
+		expect(out).not.toContain("<|message|>");
+		expect(out).toContain("<\\|channel\\|>analysis");
+		expect(out).toContain("<\\|channel\\|>final");
+		expect(out).toContain("Need to inspect the failing compaction path.");
+		expect(out).toContain("The final answer stays visible.");
+	});
+
+	test("native Harmony serialization keeps raw transcript markers", () => {
+		const out = serializeConversation(
+			[
+				assistantMessage([
+					{ type: "thinking", thinking: "Native transcript includes analysis." },
+					{ type: "text", text: "Native final text." },
+				]),
+			],
+			"harmony",
+		);
+
+		expect(out).toContain("<|channel|>analysis");
+		expect(out).toContain("<|message|>Native transcript includes analysis.");
+		expect(out).toContain("<|channel|>final");
+		expect(out).toContain("Native final text.");
+	});
+
 	test("native dialect serialization drops empty assistants left by useless calls", () => {
 		const out = serializeConversation(
 			[