fix(sse): strip every <omniModel> tag, not just the first (#454) (#3248)

Strip ALL <omniModel> tags before forwarding to provider (global regex variant).

Integrated into release/v3.8.12. Thanks @MikeTuev.

open-sse/services/comboAgentMiddleware.ts

@@ -36,20 +36,27 @@ interface Message {
 
 // Handles both actual newlines (U+000A) and literal \n sequences injected
 // by combo.ts streaming around the <omniModel> tag (#531). Non-global so that
-// .exec() and .test() stay stateless; callers that need full replacement use
-// String.prototype.replace() which replaces all non-overlapping matches.
+// .exec() and .test() stay stateless (a global regex carries lastIndex between
+// calls and would skip matches).
 const CACHE_TAG_PATTERN = /(?:\\n|\n|\r)*<omniModel>([^<]+)<\/omniModel>(?:\\n|\n|\r)*/;
 
+// Global variant for .replace() callers that must strip EVERY tag. A non-global
+// regex only removes the first match, so a single message carrying more than one
+// <omniModel> tag (e.g. an Open WebUI follow-up/title request that inlines the
+// whole chat history) leaked the remaining tags to the provider — defeating the
+// cache-session protection stripModelTags exists to enforce (#454).
+const CACHE_TAG_PATTERN_GLOBAL = /(?:\\n|\n|\r)*<omniModel>([^<]+)<\/omniModel>(?:\\n|\n|\r)*/g;
+
 /**
  * Inject the model tag into the last assistant message (or append a new one).
  * Only modifies string content — does not touch array content to avoid breaking
  * Claude/Gemini multi-part message formats.
  */
 export function injectModelTag(messages: Message[], providerModel: string): Message[] {
-  // Remove any existing tag first to avoid duplication on context compaction
+  // Remove any existing tags first to avoid duplication on context compaction
   const cleaned = messages.map((msg) => {
     if (msg.role === "assistant" && typeof msg.content === "string") {
-      return { ...msg, content: msg.content.replace(CACHE_TAG_PATTERN, "").trimEnd() };
+      return { ...msg, content: msg.content.replace(CACHE_TAG_PATTERN_GLOBAL, "").trimEnd() };
     }
     return msg;
   });
@@ -147,7 +154,7 @@ export function applyToolFilter(
 export function stripModelTags(messages: Message[]): Message[] {
   return messages.map((msg) => {
     if (typeof msg.content === "string" && CACHE_TAG_PATTERN.test(msg.content)) {
-      return { ...msg, content: msg.content.replace(CACHE_TAG_PATTERN, "").trimEnd() };
+      return { ...msg, content: msg.content.replace(CACHE_TAG_PATTERN_GLOBAL, "").trimEnd() };
     }
     return msg;
   });

tests/unit/combo-omnimodel-tag-stripping.test.ts

@@ -0,0 +1,58 @@
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+import {
+  stripModelTags,
+  injectModelTag,
+} from "../../open-sse/services/comboAgentMiddleware.ts";
+
+// Regression for the non-global CACHE_TAG_PATTERN bug: `.replace()` with a
+// non-global regex only removes the FIRST match, so messages carrying more than
+// one <omniModel> tag (e.g. an Open WebUI follow-up/title request that inlines the
+// whole chat history) leaked the remaining tags straight to the provider — exactly
+// what stripModelTags is documented to prevent (#454).
+describe("comboAgentMiddleware — <omniModel> tag stripping (non-global regex bug)", () => {
+  test("stripModelTags removes ALL <omniModel> tags from a single message", () => {
+    const messages = [
+      {
+        role: "user",
+        content:
+          "first <omniModel>claude/claude-opus-4-8</omniModel> " +
+          "middle <omniModel>claude/claude-opus-4-8</omniModel> " +
+          "end <omniModel>claude/claude-opus-4-8</omniModel>",
+      },
+    ];
+
+    const stripped = stripModelTags(messages);
+    const remaining = (String(stripped[0].content).match(/<omniModel>/g) || []).length;
+
+    assert.equal(remaining, 0, "Provider must never see any <omniModel> tag (#454)");
+    assert.ok(
+      !String(stripped[0].content).includes("<omniModel>"),
+      "Stripped content should contain no tag fragments"
+    );
+  });
+
+  test("injectModelTag does not leave duplicate tags when the message already has several", () => {
+    const messages = [
+      { role: "user", content: "Continue" },
+      {
+        role: "assistant",
+        content:
+          "answer <omniModel>old/model-a</omniModel> more <omniModel>old/model-b</omniModel>",
+      },
+    ];
+
+    const result = injectModelTag(messages, "new/model");
+    const tagCount = (String(result[1].content).match(/<omniModel>/g) || []).length;
+
+    assert.equal(tagCount, 1, "Exactly one (the freshly pinned) tag should remain");
+    assert.ok(
+      String(result[1].content).includes("<omniModel>new/model</omniModel>"),
+      "The remaining tag must be the new pin"
+    );
+    assert.ok(
+      !String(result[1].content).includes("old/model"),
+      "All previous pins must be cleaned"
+    );
+  });
+});