fix: strip leaked model control tokens from user-facing text (#42173)

Models like GLM-5 and DeepSeek sometimes emit internal delimiter tokens in their responses. Uses generic pattern in the text extraction pipeline, following the same architecture as stripMinimaxToolCallXml.

Closes #40020
Supersedes #40573

Co-authored-by: imwyvern <100903837+imwyvern@users.noreply.github.com>

src/agents/pi-embedded-utils.strip-model-special-tokens.test.ts

@@ -0,0 +1,25 @@
+import { describe, expect, it } from "vitest";
+import { stripModelSpecialTokens } from "./pi-embedded-utils.js";
+
+/**
+ * @see https://github.com/openclaw/openclaw/issues/40020
+ */
+describe("stripModelSpecialTokens", () => {
+  it("strips tokens and inserts space between adjacent words", () => {
+    expect(stripModelSpecialTokens("<|user|>Question<|assistant|>Answer")).toBe("Question Answer");
+  });
+
+  it("strips full-width pipe variants (DeepSeek U+FF5C)", () => {
+    expect(stripModelSpecialTokens("<｜begin▁of▁sentence｜>Hello there")).toBe("Hello there");
+  });
+
+  it("does not strip normal angle brackets or HTML", () => {
+    expect(stripModelSpecialTokens("a < b && c > d")).toBe("a < b && c > d");
+    expect(stripModelSpecialTokens("<div>hello</div>")).toBe("<div>hello</div>");
+  });
+
+  it("passes through text without tokens unchanged", () => {
+    const text = "Just a normal response.";
+    expect(stripModelSpecialTokens(text)).toBe(text);
+  });
+});

src/agents/pi-embedded-utils.ts

@@ -33,6 +33,32 @@ export function stripMinimaxToolCallXml(text: string): string {
   return cleaned;
 }
 
+/**
+ * Strip model control tokens leaked into assistant text output.
+ *
+ * Models like GLM-5 and DeepSeek sometimes emit internal delimiter tokens
+ * (e.g. `<|assistant|>`, `<|tool_call_result_begin|>`, `<｜begin▁of▁sentence｜>`)
+ * in their responses. These use the universal `<|...|>` convention (ASCII or
+ * full-width pipe variants) and should never reach end users.
+ *
+ * This is a provider bug — no upstream fix tracked yet.
+ * Remove this function when upstream providers stop leaking tokens.
+ * @see https://github.com/openclaw/openclaw/issues/40020
+ */
+// Match both ASCII pipe <|...|> and full-width pipe <｜...｜> (U+FF5C) variants.
+const MODEL_SPECIAL_TOKEN_RE = /<[|｜][^|｜]*[|｜]>/g;
+
+export function stripModelSpecialTokens(text: string): string {
+  if (!text) {
+    return text;
+  }
+  if (!MODEL_SPECIAL_TOKEN_RE.test(text)) {
+    return text;
+  }
+  MODEL_SPECIAL_TOKEN_RE.lastIndex = 0;
+  return text.replace(MODEL_SPECIAL_TOKEN_RE, " ").replace(/  +/g, " ").trim();
+}
+
 /**
  * Strip downgraded tool call text representations that leak into text content.
  * When replaying history to Gemini, tool calls without `thought_signature` are
@@ -212,7 +238,7 @@ export function extractAssistantText(msg: AssistantMessage): string {
     extractTextFromChatContent(msg.content, {
       sanitizeText: (text) =>
         stripThinkingTagsFromText(
-          stripDowngradedToolCallText(stripMinimaxToolCallXml(text)),
+          stripDowngradedToolCallText(stripModelSpecialTokens(stripMinimaxToolCallXml(text))),
         ).trim(),
       joinWith: "\n",
       normalizeText: (text) => text.trim(),

src/agents/tools/sessions-helpers.ts

@@ -32,6 +32,7 @@ import { sanitizeUserFacingText } from "../pi-embedded-helpers.js";
 import {
   stripDowngradedToolCallText,
   stripMinimaxToolCallXml,
+  stripModelSpecialTokens,
   stripThinkingTagsFromText,
 } from "../pi-embedded-utils.js";
 
@@ -142,7 +143,9 @@ export function sanitizeTextContent(text: string): string {
   if (!text) {
     return text;
   }
-  return stripThinkingTagsFromText(stripDowngradedToolCallText(stripMinimaxToolCallXml(text)));
+  return stripThinkingTagsFromText(
+    stripDowngradedToolCallText(stripModelSpecialTokens(stripMinimaxToolCallXml(text))),
+  );
 }
 
 export function extractAssistantText(message: unknown): string | undefined {