fix(agents): skip thinking tags in code spans

src/agents/pi-embedded-subscribe.code-span-awareness.test.ts

@@ -0,0 +1,103 @@
+import { describe, expect, it, vi } from "vitest";
+import { subscribeEmbeddedPiSession } from "./pi-embedded-subscribe.js";
+
+type StubSession = {
+  subscribe: (fn: (evt: unknown) => void) => () => void;
+};
+
+describe("subscribeEmbeddedPiSession thinking tag code span awareness", () => {
+  it("does not strip thinking tags inside inline code backticks", () => {
+    let handler: ((evt: unknown) => void) | undefined;
+    const session: StubSession = {
+      subscribe: (fn) => {
+        handler = fn;
+        return () => {};
+      },
+    };
+
+    const onPartialReply = vi.fn();
+
+    subscribeEmbeddedPiSession({
+      session: session as unknown as Parameters<typeof subscribeEmbeddedPiSession>[0]["session"],
+      runId: "run",
+      onPartialReply,
+    });
+
+    handler?.({
+      type: "message_update",
+      message: { role: "assistant" },
+      assistantMessageEvent: {
+        type: "text_delta",
+        delta: "The fix strips leaked `<thinking>` tags from messages.",
+      },
+    });
+
+    expect(onPartialReply).toHaveBeenCalled();
+    const lastCall = onPartialReply.mock.calls[onPartialReply.mock.calls.length - 1];
+    expect(lastCall[0].text).toContain("`<thinking>`");
+  });
+
+  it("does not strip thinking tags inside fenced code blocks", () => {
+    let handler: ((evt: unknown) => void) | undefined;
+    const session: StubSession = {
+      subscribe: (fn) => {
+        handler = fn;
+        return () => {};
+      },
+    };
+
+    const onPartialReply = vi.fn();
+
+    subscribeEmbeddedPiSession({
+      session: session as unknown as Parameters<typeof subscribeEmbeddedPiSession>[0]["session"],
+      runId: "run",
+      onPartialReply,
+    });
+
+    handler?.({
+      type: "message_update",
+      message: { role: "assistant" },
+      assistantMessageEvent: {
+        type: "text_delta",
+        delta: "Example:\n  ````\n<thinking>code example</thinking>\n  ````\nDone.",
+      },
+    });
+
+    expect(onPartialReply).toHaveBeenCalled();
+    const lastCall = onPartialReply.mock.calls[onPartialReply.mock.calls.length - 1];
+    expect(lastCall[0].text).toContain("<thinking>code example</thinking>");
+  });
+
+  it("still strips actual thinking tags outside code spans", () => {
+    let handler: ((evt: unknown) => void) | undefined;
+    const session: StubSession = {
+      subscribe: (fn) => {
+        handler = fn;
+        return () => {};
+      },
+    };
+
+    const onPartialReply = vi.fn();
+
+    subscribeEmbeddedPiSession({
+      session: session as unknown as Parameters<typeof subscribeEmbeddedPiSession>[0]["session"],
+      runId: "run",
+      onPartialReply,
+    });
+
+    handler?.({
+      type: "message_update",
+      message: { role: "assistant" },
+      assistantMessageEvent: {
+        type: "text_delta",
+        delta: "Hello <thinking>internal thought</thinking> world",
+      },
+    });
+
+    expect(onPartialReply).toHaveBeenCalled();
+    const lastCall = onPartialReply.mock.calls[onPartialReply.mock.calls.length - 1];
+    expect(lastCall[0].text).not.toContain("internal thought");
+    expect(lastCall[0].text).toContain("Hello");
+    expect(lastCall[0].text).toContain("world");
+  });
+});