fix: split telegram reasoning and answer draft streams (#20774)

Merged via /review-pr -> /prepare-pr -> /merge-pr. Prepared head SHA: 7458444144 Co-authored-by: obviyus <22031114+obviyus@users.noreply.github.com> Co-authored-by: obviyus <22031114+obviyus@users.noreply.github.com> Reviewed-by: @obviyus
2026-03-12 07:20:45 +00:00 · 2026-02-20 11:14:39 +05:30
parent beb2b74b5b
commit ab256b8ec7
7 changed files with 1194 additions and 279 deletions
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -34,6 +34,7 @@ Docs: https://docs.openclaw.ai
 - Config/Memory: restore schema help/label metadata for hybrid `mmr` and `temporalDecay` settings so configuration surfaces show correct names and guidance. (#18786) Thanks @rodrigouroz.
 - Tools/web_search: handle xAI Responses API payloads that emit top-level `output_text` blocks (without a `message` wrapper) so Grok web_search no longer returns `No response` for those results. (#20508) Thanks @echoVic.
 - Telegram/Streaming: always clean up draft previews even when dispatch throws before fallback handling, preventing orphaned preview messages during failed runs. (#19041) thanks @mudrii.
+- Telegram/Streaming: split reasoning and answer draft preview lanes to prevent cross-lane overwrites, and ignore literal `<think>` tags inside inline/fenced code snippets so sample markup is not misrouted as reasoning. (#20774) Thanks @obviyus.

 - Discord/Gateway: handle close code 4014 (missing privileged gateway intents) without crashing the gateway. Thanks @thewilloftheshadow.
 - Security/Net: strip sensitive headers (`Authorization`, `Proxy-Authorization`, `Cookie`, `Cookie2`) on cross-origin redirects in `fetchWithSsrFGuard` to prevent credential forwarding across origin boundaries. (#20313) Thanks @afurm.
--- a/src/telegram/bot-message-dispatch.test.ts
+++ b/src/telegram/bot-message-dispatch.test.ts
@@ -7,6 +7,8 @@ const createTelegramDraftStream = vi.hoisted(() => vi.fn());
 const dispatchReplyWithBufferedBlockDispatcher = vi.hoisted(() => vi.fn());
 const deliverReplies = vi.hoisted(() => vi.fn());
 const editMessageTelegram = vi.hoisted(() => vi.fn());
+const loadSessionStore = vi.hoisted(() => vi.fn());
+const resolveStorePath = vi.hoisted(() => vi.fn(() => "/tmp/sessions.json"));

 vi.mock("./draft-stream.js", () => ({
  createTelegramDraftStream,
@@ -24,6 +26,11 @@ vi.mock("./send.js", () => ({
  editMessageTelegram,
 }));

+vi.mock("../config/sessions.js", async () => ({
+  loadSessionStore,
+  resolveStorePath,
+}));
+
 vi.mock("./sticker-cache.js", () => ({
  cacheSticker: vi.fn(),
  describeStickerImage: vi.fn(),
@@ -39,6 +46,10 @@ describe("dispatchTelegramMessage draft streaming", () => {
    dispatchReplyWithBufferedBlockDispatcher.mockReset();
    deliverReplies.mockReset();
    editMessageTelegram.mockReset();
+    loadSessionStore.mockReset();
+    resolveStorePath.mockReset();
+    resolveStorePath.mockReturnValue("/tmp/sessions.json");
+    loadSessionStore.mockReturnValue({});
  });

  function createDraftStream(messageId?: number) {
@@ -52,6 +63,15 @@ describe("dispatchTelegramMessage draft streaming", () => {
    };
  }

+  function setupDraftStreams(params?: { answerMessageId?: number; reasoningMessageId?: number }) {
+    const answerDraftStream = createDraftStream(params?.answerMessageId);
+    const reasoningDraftStream = createDraftStream(params?.reasoningMessageId);
+    createTelegramDraftStream
+      .mockImplementationOnce(() => answerDraftStream)
+      .mockImplementationOnce(() => reasoningDraftStream);
+    return { answerDraftStream, reasoningDraftStream };
+  }
+
  function createContext(overrides?: Partial<TelegramMessageContext>): TelegramMessageContext {
    const base = {
      ctxPayload: {},
@@ -152,6 +172,7 @@ describe("dispatchTelegramMessage draft streaming", () => {
      expect.objectContaining({
        chatId: 123,
        thread: { id: 777, scope: "dm" },
+        minInitialChars: 1,
      }),
    );
    expect(draftStream.update).toHaveBeenCalledWith("Hello");
@@ -172,6 +193,27 @@ describe("dispatchTelegramMessage draft streaming", () => {
    expect(draftStream.clear).toHaveBeenCalledTimes(1);
  });

+  it("keeps a higher initial debounce threshold in block stream mode", async () => {
+    const draftStream = createDraftStream();
+    createTelegramDraftStream.mockReturnValue(draftStream);
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onPartialReply?.({ text: "Hello" });
+        await dispatcherOptions.deliver({ text: "Hello" }, { kind: "final" });
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "block" });
+
+    expect(createTelegramDraftStream).toHaveBeenCalledWith(
+      expect.objectContaining({
+        minInitialChars: 30,
+      }),
+    );
+  });
+
  it("keeps block streaming enabled when account config enables it", async () => {
    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(async ({ dispatcherOptions }) => {
      await dispatcherOptions.deliver({ text: "Hello" }, { kind: "final" });
@@ -195,6 +237,66 @@ describe("dispatchTelegramMessage draft streaming", () => {
    );
  });

+  it("keeps block streaming enabled when session reasoning level is on", async () => {
+    loadSessionStore.mockReturnValue({
+      s1: { reasoningLevel: "on" },
+    });
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(async ({ dispatcherOptions }) => {
+      await dispatcherOptions.deliver({ text: "Reasoning:\n_step_" }, { kind: "block" });
+      await dispatcherOptions.deliver({ text: "Hello" }, { kind: "final" });
+      return { queuedFinal: true };
+    });
+    deliverReplies.mockResolvedValue({ delivered: true });
+
+    await dispatchWithContext({
+      context: createContext({
+        ctxPayload: { SessionKey: "s1" } as unknown as TelegramMessageContext["ctxPayload"],
+      }),
+    });
+
+    expect(createTelegramDraftStream).not.toHaveBeenCalled();
+    expect(dispatchReplyWithBufferedBlockDispatcher).toHaveBeenCalledWith(
+      expect.objectContaining({
+        replyOptions: expect.objectContaining({
+          disableBlockStreaming: false,
+        }),
+      }),
+    );
+    expect(loadSessionStore).toHaveBeenCalledWith("/tmp/sessions.json", { skipCache: true });
+    expect(deliverReplies).toHaveBeenCalledWith(
+      expect.objectContaining({
+        replies: [expect.objectContaining({ text: "Reasoning:\n_step_" })],
+      }),
+    );
+  });
+
+  it("streams reasoning draft updates even when answer stream mode is off", async () => {
+    loadSessionStore.mockReturnValue({
+      s1: { reasoningLevel: "stream" },
+    });
+    const reasoningDraftStream = createDraftStream(111);
+    createTelegramDraftStream.mockImplementationOnce(() => reasoningDraftStream);
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onReasoningStream?.({ text: "Reasoning:\n_step_" });
+        await dispatcherOptions.deliver({ text: "Hello" }, { kind: "final" });
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+
+    await dispatchWithContext({
+      context: createContext({
+        ctxPayload: { SessionKey: "s1" } as unknown as TelegramMessageContext["ctxPayload"],
+      }),
+      streamMode: "off",
+    });
+
+    expect(createTelegramDraftStream).toHaveBeenCalledTimes(1);
+    expect(reasoningDraftStream.update).toHaveBeenCalledWith("Reasoning:\n_step_");
+    expect(loadSessionStore).toHaveBeenCalledWith("/tmp/sessions.json", { skipCache: true });
+  });
+
  it("finalizes text-only replies by editing the preview message in place", async () => {
    const draftStream = createDraftStream(999);
    createTelegramDraftStream.mockReturnValue(draftStream);
@@ -407,71 +509,398 @@ describe("dispatchTelegramMessage draft streaming", () => {
    expect(draftStream.forceNewMessage).not.toHaveBeenCalled();
  });

-  it("forces new message when reasoning ends after previous output", async () => {
-    const draftStream = createDraftStream(999);
-    createTelegramDraftStream.mockReturnValue(draftStream);
+  it.each(["block", "partial"] as const)(
+    "splits reasoning lane only when a later reasoning block starts (%s mode)",
+    async (streamMode) => {
+      const { reasoningDraftStream } = setupDraftStreams({
+        answerMessageId: 999,
+        reasoningMessageId: 111,
+      });
+      dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+        async ({ dispatcherOptions, replyOptions }) => {
+          await replyOptions?.onReasoningStream?.({ text: "Reasoning:\n_first block_" });
+          await replyOptions?.onReasoningEnd?.();
+          expect(reasoningDraftStream.forceNewMessage).not.toHaveBeenCalled();
+          await replyOptions?.onPartialReply?.({ text: "checking files..." });
+          await replyOptions?.onReasoningStream?.({ text: "Reasoning:\n_second block_" });
+          await dispatcherOptions.deliver({ text: "Done" }, { kind: "final" });
+          return { queuedFinal: true };
+        },
+      );
+      deliverReplies.mockResolvedValue({ delivered: true });
+      editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+      await dispatchWithContext({ context: createContext(), streamMode });
+
+      expect(reasoningDraftStream.forceNewMessage).toHaveBeenCalledTimes(1);
+    },
+  );
+
+  it.each(["block", "partial"] as const)(
+    "does not split reasoning lane on reasoning end without a later reasoning block (%s mode)",
+    async (streamMode) => {
+      const { reasoningDraftStream } = setupDraftStreams({
+        answerMessageId: 999,
+        reasoningMessageId: 111,
+      });
+      dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+        async ({ dispatcherOptions, replyOptions }) => {
+          await replyOptions?.onReasoningStream?.({ text: "Reasoning:\n_first block_" });
+          await replyOptions?.onReasoningEnd?.();
+          await replyOptions?.onPartialReply?.({ text: "Here's the answer" });
+          await dispatcherOptions.deliver({ text: "Here's the answer" }, { kind: "final" });
+          return { queuedFinal: true };
+        },
+      );
+      deliverReplies.mockResolvedValue({ delivered: true });
+
+      await dispatchWithContext({ context: createContext(), streamMode });
+
+      expect(reasoningDraftStream.forceNewMessage).not.toHaveBeenCalled();
+    },
+  );
+
+  it("does not finalize preview with reasoning payloads before answer payloads", async () => {
+    setupDraftStreams({ answerMessageId: 999 });
    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
      async ({ dispatcherOptions, replyOptions }) => {
-        // First partial: text before thinking
-        await replyOptions?.onPartialReply?.({ text: "Let me check" });
-        // Reasoning stream (thinking block)
-        await replyOptions?.onReasoningStream?.({ text: "Analyzing..." });
-        // Reasoning ends
-        await replyOptions?.onReasoningEnd?.();
-        // Second partial: text after thinking
-        await replyOptions?.onPartialReply?.({ text: "Here's the answer" });
-        await dispatcherOptions.deliver({ text: "Here's the answer" }, { kind: "final" });
-        return { queuedFinal: true };
-      },
-    );
-    deliverReplies.mockResolvedValue({ delivered: true });
-
-    await dispatchWithContext({ context: createContext(), streamMode: "block" });
-
-    // Should force new message when reasoning ends
-    expect(draftStream.forceNewMessage).toHaveBeenCalled();
-  });
-
-  it("does not force new message in partial mode when reasoning ends", async () => {
-    const draftStream = createDraftStream(999);
-    createTelegramDraftStream.mockReturnValue(draftStream);
-    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
-      async ({ dispatcherOptions, replyOptions }) => {
-        await replyOptions?.onPartialReply?.({ text: "Let me check" });
-        await replyOptions?.onReasoningEnd?.();
-        await replyOptions?.onPartialReply?.({ text: "Here's the answer" });
-        await dispatcherOptions.deliver({ text: "Here's the answer" }, { kind: "final" });
+        await replyOptions?.onPartialReply?.({ text: "Hi, I did what you asked and..." });
+        await dispatcherOptions.deliver({ text: "Reasoning:\n_step one_" }, { kind: "final" });
+        await dispatcherOptions.deliver(
+          { text: "Hi, I did what you asked and..." },
+          { kind: "final" },
+        );
        return { queuedFinal: true };
      },
    );
    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });

    await dispatchWithContext({ context: createContext(), streamMode: "partial" });

-    expect(draftStream.forceNewMessage).not.toHaveBeenCalled();
+    // Keep reasoning as its own message.
+    expect(deliverReplies).toHaveBeenCalledTimes(1);
+    expect(deliverReplies).toHaveBeenCalledWith(
+      expect.objectContaining({
+        replies: [expect.objectContaining({ text: "Reasoning:\n_step one_" })],
+      }),
+    );
+    // Finalize preview with the actual answer instead of overwriting with reasoning.
+    expect(editMessageTelegram).toHaveBeenCalledTimes(1);
+    expect(editMessageTelegram).toHaveBeenCalledWith(
+      123,
+      999,
+      "Hi, I did what you asked and...",
+      expect.any(Object),
+    );
  });

-  it("does not force new message on reasoning end without previous output", async () => {
-    const draftStream = createDraftStream(999);
-    createTelegramDraftStream.mockReturnValue(draftStream);
+  it("keeps reasoning and answer streaming in separate preview lanes", async () => {
+    const { answerDraftStream, reasoningDraftStream } = setupDraftStreams({
+      answerMessageId: 999,
+      reasoningMessageId: 111,
+    });
    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
      async ({ dispatcherOptions, replyOptions }) => {
-        // Reasoning starts immediately (no previous text output)
-        await replyOptions?.onReasoningStream?.({ text: "Thinking..." });
-        // Reasoning ends
-        await replyOptions?.onReasoningEnd?.();
-        // First actual text output
-        await replyOptions?.onPartialReply?.({ text: "Here's my answer" });
-        await dispatcherOptions.deliver({ text: "Here's my answer" }, { kind: "final" });
+        await replyOptions?.onReasoningStream?.({ text: "Reasoning:\n_Working on it..._" });
+        await replyOptions?.onPartialReply?.({ text: "Checking the directory..." });
+        await dispatcherOptions.deliver({ text: "Checking the directory..." }, { kind: "final" });
        return { queuedFinal: true };
      },
    );
    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "partial" });
+
+    expect(reasoningDraftStream.update).toHaveBeenCalledWith("Reasoning:\n_Working on it..._");
+    expect(answerDraftStream.update).toHaveBeenCalledWith("Checking the directory...");
+    expect(answerDraftStream.forceNewMessage).not.toHaveBeenCalled();
+    expect(reasoningDraftStream.forceNewMessage).not.toHaveBeenCalled();
+  });
+
+  it("does not edit reasoning preview bubble with final answer when no assistant partial arrived yet", async () => {
+    setupDraftStreams({ reasoningMessageId: 999 });
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onReasoningStream?.({ text: "Reasoning:\n_Working on it..._" });
+        await dispatcherOptions.deliver({ text: "Here's what I found." }, { kind: "final" });
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "partial" });
+
+    expect(editMessageTelegram).not.toHaveBeenCalled();
+    expect(deliverReplies).toHaveBeenCalledWith(
+      expect.objectContaining({
+        replies: [expect.objectContaining({ text: "Here's what I found." })],
+      }),
+    );
+  });
+
+  it.each(["partial", "block"] as const)(
+    "does not duplicate reasoning final after reasoning end (%s mode)",
+    async (streamMode) => {
+      let reasoningMessageId: number | undefined = 111;
+      const reasoningDraftStream = {
+        update: vi.fn(),
+        flush: vi.fn().mockResolvedValue(undefined),
+        messageId: vi.fn().mockImplementation(() => reasoningMessageId),
+        clear: vi.fn().mockResolvedValue(undefined),
+        stop: vi.fn().mockResolvedValue(undefined),
+        forceNewMessage: vi.fn().mockImplementation(() => {
+          reasoningMessageId = undefined;
+        }),
+      };
+      const answerDraftStream = createDraftStream(999);
+      createTelegramDraftStream
+        .mockImplementationOnce(() => answerDraftStream)
+        .mockImplementationOnce(() => reasoningDraftStream);
+      dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+        async ({ dispatcherOptions, replyOptions }) => {
+          await replyOptions?.onReasoningStream?.({ text: "Reasoning:\n_step one_" });
+          await replyOptions?.onReasoningEnd?.();
+          await dispatcherOptions.deliver(
+            { text: "Reasoning:\n_step one expanded_" },
+            { kind: "final" },
+          );
+          return { queuedFinal: true };
+        },
+      );
+      deliverReplies.mockResolvedValue({ delivered: true });
+      editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "111" });
+
+      await dispatchWithContext({ context: createContext(), streamMode });
+
+      expect(reasoningDraftStream.forceNewMessage).not.toHaveBeenCalled();
+      expect(editMessageTelegram).toHaveBeenCalledWith(
+        123,
+        111,
+        "Reasoning:\n_step one expanded_",
+        expect.any(Object),
+      );
+      expect(deliverReplies).not.toHaveBeenCalled();
+    },
+  );
+
+  it("updates reasoning preview for reasoning block payloads instead of sending duplicates", async () => {
+    setupDraftStreams({ answerMessageId: 999, reasoningMessageId: 111 });
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onReasoningStream?.({
+          text: "Reasoning:\nIf I count r in strawberry, I see positions 3, 8, and",
+        });
+        await replyOptions?.onReasoningEnd?.();
+        await replyOptions?.onPartialReply?.({ text: "3" });
+        await dispatcherOptions.deliver({ text: "3" }, { kind: "final" });
+        await dispatcherOptions.deliver(
+          {
+            text: "Reasoning:\nIf I count r in strawberry, I see positions 3, 8, and 9. So the total is 3.",
+          },
+          { kind: "block" },
+        );
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "partial" });
+
+    expect(editMessageTelegram).toHaveBeenNthCalledWith(1, 123, 999, "3", expect.any(Object));
+    expect(editMessageTelegram).toHaveBeenNthCalledWith(
+      2,
+      123,
+      111,
+      "Reasoning:\nIf I count r in strawberry, I see positions 3, 8, and 9. So the total is 3.",
+      expect.any(Object),
+    );
+    expect(deliverReplies).not.toHaveBeenCalledWith(
+      expect.objectContaining({
+        replies: [
+          expect.objectContaining({
+            text: expect.stringContaining("Reasoning:\nIf I count r in strawberry"),
+          }),
+        ],
+      }),
+    );
+  });
+
+  it("routes think-tag partials to reasoning lane and keeps answer lane clean", async () => {
+    const { answerDraftStream, reasoningDraftStream } = setupDraftStreams({
+      answerMessageId: 999,
+      reasoningMessageId: 111,
+    });
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onPartialReply?.({
+          text: "<think>Counting letters in strawberry</think>3",
+        });
+        await dispatcherOptions.deliver({ text: "3" }, { kind: "final" });
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "partial" });
+
+    expect(reasoningDraftStream.update).toHaveBeenCalledWith(
+      "Reasoning:\n_Counting letters in strawberry_",
+    );
+    expect(answerDraftStream.update).toHaveBeenCalledWith("3");
+    expect(
+      answerDraftStream.update.mock.calls.some((call) => String(call[0] ?? "").includes("<think>")),
+    ).toBe(false);
+    expect(editMessageTelegram).toHaveBeenCalledWith(123, 999, "3", expect.any(Object));
+  });
+
+  it("routes unmatched think partials to reasoning lane without leaking answer lane", async () => {
+    const { answerDraftStream, reasoningDraftStream } = setupDraftStreams({
+      answerMessageId: 999,
+      reasoningMessageId: 111,
+    });
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onPartialReply?.({
+          text: "<think>Counting letters in strawberry",
+        });
+        await dispatcherOptions.deliver(
+          { text: "There are 3 r's in strawberry." },
+          { kind: "final" },
+        );
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "partial" });
+
+    expect(reasoningDraftStream.update).toHaveBeenCalledWith(
+      "Reasoning:\n_Counting letters in strawberry_",
+    );
+    expect(
+      answerDraftStream.update.mock.calls.some((call) => String(call[0] ?? "").includes("<")),
+    ).toBe(false);
+    expect(editMessageTelegram).toHaveBeenCalledWith(
+      123,
+      999,
+      "There are 3 r's in strawberry.",
+      expect.any(Object),
+    );
+  });
+
+  it("keeps reasoning preview message when reasoning is streamed but final is answer-only", async () => {
+    const { reasoningDraftStream } = setupDraftStreams({
+      answerMessageId: 999,
+      reasoningMessageId: 111,
+    });
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onPartialReply?.({
+          text: "<think>Word: strawberry. r appears at 3, 8, 9.</think>",
+        });
+        await dispatcherOptions.deliver(
+          { text: "There are 3 r's in strawberry." },
+          { kind: "final" },
+        );
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "partial" });
+
+    expect(reasoningDraftStream.update).toHaveBeenCalledWith(
+      "Reasoning:\n_Word: strawberry. r appears at 3, 8, 9._",
+    );
+    expect(reasoningDraftStream.clear).not.toHaveBeenCalled();
+    expect(editMessageTelegram).toHaveBeenCalledWith(
+      123,
+      999,
+      "There are 3 r's in strawberry.",
+      expect.any(Object),
+    );
+  });
+
+  it("splits think-tag final payload into reasoning and answer lanes", async () => {
+    setupDraftStreams({
+      answerMessageId: 999,
+      reasoningMessageId: 111,
+    });
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(async ({ dispatcherOptions }) => {
+      await dispatcherOptions.deliver(
+        {
+          text: "<think>Word: strawberry. r appears at 3, 8, 9.</think>There are 3 r's in strawberry.",
+        },
+        { kind: "final" },
+      );
+      return { queuedFinal: true };
+    });
+    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });
+
+    await dispatchWithContext({ context: createContext(), streamMode: "partial" });
+
+    expect(editMessageTelegram).toHaveBeenNthCalledWith(
+      1,
+      123,
+      111,
+      "Reasoning:\n_Word: strawberry. r appears at 3, 8, 9._",
+      expect.any(Object),
+    );
+    expect(editMessageTelegram).toHaveBeenNthCalledWith(
+      2,
+      123,
+      999,
+      "There are 3 r's in strawberry.",
+      expect.any(Object),
+    );
+    expect(deliverReplies).not.toHaveBeenCalled();
+  });
+
+  it("edits stop-created preview when final text is shorter than buffered draft", async () => {
+    let answerMessageId: number | undefined;
+    const answerDraftStream = {
+      update: vi.fn(),
+      flush: vi.fn().mockResolvedValue(undefined),
+      messageId: vi.fn().mockImplementation(() => answerMessageId),
+      clear: vi.fn().mockResolvedValue(undefined),
+      stop: vi.fn().mockImplementation(async () => {
+        answerMessageId = 999;
+      }),
+      forceNewMessage: vi.fn(),
+    };
+    const reasoningDraftStream = createDraftStream();
+    createTelegramDraftStream
+      .mockImplementationOnce(() => answerDraftStream)
+      .mockImplementationOnce(() => reasoningDraftStream);
+    dispatchReplyWithBufferedBlockDispatcher.mockImplementation(
+      async ({ dispatcherOptions, replyOptions }) => {
+        await replyOptions?.onPartialReply?.({
+          text: "Let me check that file and confirm details for you.",
+        });
+        await dispatcherOptions.deliver({ text: "Let me check that file." }, { kind: "final" });
+        return { queuedFinal: true };
+      },
+    );
+    deliverReplies.mockResolvedValue({ delivered: true });
+    editMessageTelegram.mockResolvedValue({ ok: true, chatId: "123", messageId: "999" });

    await dispatchWithContext({ context: createContext(), streamMode: "block" });

-    // No previous text output, so no forceNewMessage needed
-    expect(draftStream.forceNewMessage).not.toHaveBeenCalled();
+    expect(editMessageTelegram).toHaveBeenCalledWith(
+      123,
+      999,
+      "Let me check that file.",
+      expect.any(Object),
+    );
+    expect(deliverReplies).not.toHaveBeenCalled();
  });

  it("does not edit preview message when final payload is an error", async () => {
--- a/src/telegram/bot-message-dispatch.ts
+++ b/src/telegram/bot-message-dispatch.ts
@@ -10,11 +10,13 @@ import { EmbeddedBlockChunker } from "../agents/pi-embedded-block-chunker.js";
 import { resolveChunkMode } from "../auto-reply/chunk.js";
 import { clearHistoryEntriesIfEnabled } from "../auto-reply/reply/history.js";
 import { dispatchReplyWithBufferedBlockDispatcher } from "../auto-reply/reply/provider-dispatcher.js";
+import type { ReplyPayload } from "../auto-reply/types.js";
 import { removeAckReactionAfterReply } from "../channels/ack-reactions.js";
 import { logAckFailure, logTypingFailure } from "../channels/logging.js";
 import { createReplyPrefixOptions } from "../channels/reply-prefix.js";
 import { createTypingCallbacks } from "../channels/typing.js";
 import { resolveMarkdownTableMode } from "../config/markdown-tables.js";
+import { loadSessionStore, resolveStorePath } from "../config/sessions.js";
 import type { OpenClawConfig, ReplyToMode, TelegramAccountConfig } from "../config/types.js";
 import { danger, logVerbose } from "../globals.js";
 import { getAgentScopedMediaLocalRoots } from "../media/local-roots.js";
@@ -26,6 +28,11 @@ import type { TelegramStreamMode } from "./bot/types.js";
 import type { TelegramInlineButtons } from "./button-types.js";
 import { resolveTelegramDraftStreamingChunking } from "./draft-chunking.js";
 import { createTelegramDraftStream } from "./draft-stream.js";
+import { renderTelegramHtmlText } from "./format.js";
+import {
+  createTelegramReasoningStepState,
+  splitTelegramReasoningText,
+} from "./reasoning-lane-coordinator.js";
 import { editMessageTelegram } from "./send.js";
 import { cacheSticker, describeStickerImage } from "./sticker-cache.js";

@@ -60,6 +67,31 @@ type DispatchTelegramMessageParams = {
  opts: Pick<TelegramBotOptions, "token">;
 };

+type TelegramReasoningLevel = "off" | "on" | "stream";
+
+function resolveTelegramReasoningLevel(params: {
+  cfg: OpenClawConfig;
+  sessionKey?: string;
+  agentId: string;
+}): TelegramReasoningLevel {
+  const { cfg, sessionKey, agentId } = params;
+  if (!sessionKey) {
+    return "off";
+  }
+  try {
+    const storePath = resolveStorePath(cfg.session?.store, { agentId });
+    const store = loadSessionStore(storePath, { skipCache: true });
+    const entry = store[sessionKey.toLowerCase()] ?? store[sessionKey];
+    const level = entry?.reasoningLevel;
+    if (level === "on" || level === "stream") {
+      return level;
+    }
+  } catch {
+    // Fall through to default.
+  }
+  return "off";
+}
+
 export const dispatchTelegramMessage = async ({
  context,
  bot,
@@ -90,112 +122,183 @@ export const dispatchTelegramMessage = async ({
  } = context;

  const draftMaxChars = Math.min(textLimit, 4096);
+  const tableMode = resolveMarkdownTableMode({
+    cfg,
+    channel: "telegram",
+    accountId: route.accountId,
+  });
+  const renderDraftPreview = (text: string) => ({
+    text: renderTelegramHtmlText(text, { tableMode }),
+    parseMode: "HTML" as const,
+  });
  const accountBlockStreamingEnabled =
    typeof telegramCfg.blockStreaming === "boolean"
      ? telegramCfg.blockStreaming
      : cfg.agents?.defaults?.blockStreamingDefault === "on";
-  const canStreamDraft = streamMode !== "off" && !accountBlockStreamingEnabled;
+  const resolvedReasoningLevel = resolveTelegramReasoningLevel({
+    cfg,
+    sessionKey: ctxPayload.SessionKey,
+    agentId: route.agentId,
+  });
+  const forceBlockStreamingForReasoning = resolvedReasoningLevel === "on";
+  const streamReasoningDraft = resolvedReasoningLevel === "stream";
+  const canStreamAnswerDraft =
+    streamMode !== "off" && !accountBlockStreamingEnabled && !forceBlockStreamingForReasoning;
+  const canStreamReasoningDraft = canStreamAnswerDraft || streamReasoningDraft;
  const draftReplyToMessageId =
    replyToMode !== "off" && typeof msg.message_id === "number" ? msg.message_id : undefined;
-  const draftStream = canStreamDraft
-    ? createTelegramDraftStream({
-        api: bot.api,
-        chatId,
-        maxChars: draftMaxChars,
-        thread: threadSpec,
-        replyToMessageId: draftReplyToMessageId,
-        minInitialChars: DRAFT_MIN_INITIAL_CHARS,
-        log: logVerbose,
-        warn: logVerbose,
-      })
-    : undefined;
-  const draftChunking =
-    draftStream && streamMode === "block"
-      ? resolveTelegramDraftStreamingChunking(cfg, route.accountId)
-      : undefined;
-  const shouldSplitPreviewMessages = streamMode === "block";
-  const draftChunker = draftChunking ? new EmbeddedBlockChunker(draftChunking) : undefined;
+  const draftMinInitialChars =
+    streamMode === "partial" || streamReasoningDraft ? 1 : DRAFT_MIN_INITIAL_CHARS;
  const mediaLocalRoots = getAgentScopedMediaLocalRoots(cfg, route.agentId);
-  let lastPartialText = "";
-  let draftText = "";
-  let hasStreamedMessage = false;
-  const updateDraftFromPartial = (text?: string) => {
-    if (!draftStream || !text) {
+  type LaneName = "answer" | "reasoning";
+  type DraftLaneState = {
+    stream: ReturnType<typeof createTelegramDraftStream> | undefined;
+    lastPartialText: string;
+    draftText: string;
+    hasStreamedMessage: boolean;
+    chunker: EmbeddedBlockChunker | undefined;
+  };
+  const createDraftLane = (enabled: boolean): DraftLaneState => {
+    const stream = enabled
+      ? createTelegramDraftStream({
+          api: bot.api,
+          chatId,
+          maxChars: draftMaxChars,
+          thread: threadSpec,
+          replyToMessageId: draftReplyToMessageId,
+          minInitialChars: draftMinInitialChars,
+          renderText: renderDraftPreview,
+          log: logVerbose,
+          warn: logVerbose,
+        })
+      : undefined;
+    const chunker =
+      stream && streamMode === "block"
+        ? new EmbeddedBlockChunker(resolveTelegramDraftStreamingChunking(cfg, route.accountId))
+        : undefined;
+    return {
+      stream,
+      lastPartialText: "",
+      draftText: "",
+      hasStreamedMessage: false,
+      chunker,
+    };
+  };
+  const lanes: Record<LaneName, DraftLaneState> = {
+    answer: createDraftLane(canStreamAnswerDraft),
+    reasoning: createDraftLane(canStreamReasoningDraft),
+  };
+  const answerLane = lanes.answer;
+  const reasoningLane = lanes.reasoning;
+  let splitReasoningOnNextStream = false;
+  const reasoningStepState = createTelegramReasoningStepState();
+  type SplitLaneSegment = { lane: LaneName; text: string };
+  const splitTextIntoLaneSegments = (text?: string): SplitLaneSegment[] => {
+    const split = splitTelegramReasoningText(text);
+    const segments: SplitLaneSegment[] = [];
+    if (split.reasoningText) {
+      segments.push({ lane: "reasoning", text: split.reasoningText });
+    }
+    if (split.answerText) {
+      segments.push({ lane: "answer", text: split.answerText });
+    }
+    return segments;
+  };
+  const resetDraftLaneState = (lane: DraftLaneState) => {
+    lane.lastPartialText = "";
+    lane.draftText = "";
+    lane.hasStreamedMessage = false;
+    lane.chunker?.reset();
+  };
+  const updateDraftFromPartial = (lane: DraftLaneState, text: string | undefined) => {
+    const laneStream = lane.stream;
+    if (!laneStream || !text) {
      return;
    }
-    if (text === lastPartialText) {
+    if (text === lane.lastPartialText) {
      return;
    }
    // Mark that we've received streaming content (for forceNewMessage decision).
-    hasStreamedMessage = true;
+    lane.hasStreamedMessage = true;
    if (streamMode === "partial") {
      // Some providers briefly emit a shorter prefix snapshot (for example
      // "Sure." -> "Sure" -> "Sure."). Keep the longer preview to avoid
      // visible punctuation flicker.
      if (
-        lastPartialText &&
-        lastPartialText.startsWith(text) &&
-        text.length < lastPartialText.length
+        lane.lastPartialText &&
+        lane.lastPartialText.startsWith(text) &&
+        text.length < lane.lastPartialText.length
      ) {
        return;
      }
-      lastPartialText = text;
-      draftStream.update(text);
+      lane.lastPartialText = text;
+      laneStream.update(text);
      return;
    }
    let delta = text;
-    if (text.startsWith(lastPartialText)) {
-      delta = text.slice(lastPartialText.length);
+    if (text.startsWith(lane.lastPartialText)) {
+      delta = text.slice(lane.lastPartialText.length);
    } else {
      // Streaming buffer reset (or non-monotonic stream). Start fresh.
-      draftChunker?.reset();
-      draftText = "";
+      lane.chunker?.reset();
+      lane.draftText = "";
    }
-    lastPartialText = text;
+    lane.lastPartialText = text;
    if (!delta) {
      return;
    }
-    if (!draftChunker) {
-      draftText = text;
-      draftStream.update(draftText);
+    if (!lane.chunker) {
+      lane.draftText = text;
+      laneStream.update(lane.draftText);
      return;
    }
-    draftChunker.append(delta);
-    draftChunker.drain({
+    lane.chunker.append(delta);
+    lane.chunker.drain({
      force: false,
      emit: (chunk) => {
-        draftText += chunk;
-        draftStream.update(draftText);
+        lane.draftText += chunk;
+        laneStream.update(lane.draftText);
      },
    });
  };
-  const flushDraft = async () => {
-    if (!draftStream) {
+  const ingestDraftLaneSegments = (text: string | undefined) => {
+    for (const segment of splitTextIntoLaneSegments(text)) {
+      if (segment.lane === "reasoning") {
+        reasoningStepState.noteReasoningHint();
+        reasoningStepState.noteReasoningDelivered();
+      }
+      updateDraftFromPartial(lanes[segment.lane], segment.text);
+    }
+  };
+  const flushDraftLane = async (lane: DraftLaneState) => {
+    if (!lane.stream) {
      return;
    }
-    if (draftChunker?.hasBuffered()) {
-      draftChunker.drain({
+    if (lane.chunker?.hasBuffered()) {
+      lane.chunker.drain({
        force: true,
        emit: (chunk) => {
-          draftText += chunk;
+          lane.draftText += chunk;
        },
      });
-      draftChunker.reset();
-      if (draftText) {
-        draftStream.update(draftText);
+      lane.chunker.reset();
+      if (lane.draftText) {
+        lane.stream.update(lane.draftText);
      }
    }
-    await draftStream.flush();
+    await lane.stream.flush();
  };

  const disableBlockStreaming =
    streamMode === "off"
-      ? true // off mode must always disable block streaming
-      : typeof telegramCfg.blockStreaming === "boolean"
-        ? !telegramCfg.blockStreaming
-        : draftStream
-          ? true
-          : undefined;
+      ? true
+      : forceBlockStreamingForReasoning
+        ? false
+        : typeof telegramCfg.blockStreaming === "boolean"
+          ? !telegramCfg.blockStreaming
+          : canStreamAnswerDraft
+            ? true
+            : undefined;

  const { onModelSelected, ...prefixOptions } = createReplyPrefixOptions({
    cfg,
@@ -203,11 +306,6 @@ export const dispatchTelegramMessage = async ({
    channel: "telegram",
    accountId: route.accountId,
  });
-  const tableMode = resolveMarkdownTableMode({
-    cfg,
-    channel: "telegram",
-    accountId: route.accountId,
-  });
  const chunkMode = resolveChunkMode(cfg, "telegram", route.accountId);

  // Handle uncached stickers: get a dedicated vision description before dispatch
@@ -271,26 +369,12 @@ export const dispatchTelegramMessage = async ({
  const deliveryState = {
    delivered: false,
    skippedNonSilent: 0,
-    failedDeliveries: 0,
+    failedNonSilent: 0,
  };
-  let finalizedViaPreviewMessage = false;
-
-  /**
-   * Clean up the draft preview message.  The preview must be removed in every
-   * case EXCEPT when it was successfully finalized as the actual response via
-   * an in-place edit (`finalizedViaPreviewMessage === true`).
-   */
-  const clearDraftPreviewIfNeeded = async () => {
-    if (finalizedViaPreviewMessage) {
-      return;
-    }
-    try {
-      await draftStream?.clear();
-    } catch (err) {
-      logVerbose(`telegram: draft preview cleanup failed: ${String(err)}`);
-    }
+  const finalizedPreviewByLane: Record<LaneName, boolean> = {
+    answer: false,
+    reasoning: false,
  };
-
  const clearGroupHistory = () => {
    if (isGroup && historyKey) {
      clearHistoryEntriesIfEnabled({ historyMap: groupHistories, historyKey, limit: historyLimit });
@@ -310,9 +394,157 @@ export const dispatchTelegramMessage = async ({
    linkPreview: telegramCfg.linkPreview,
    replyQuoteText,
  };
+  const getLanePreviewText = (lane: DraftLaneState) =>
+    streamMode === "block" ? lane.draftText : lane.lastPartialText;
+  const tryUpdatePreviewForLane = async (params: {
+    lane: DraftLaneState;
+    laneName: LaneName;
+    text: string;
+    previewButtons?: TelegramInlineButtons;
+    stopBeforeEdit?: boolean;
+    updateLaneSnapshot?: boolean;
+    skipRegressive: "always" | "existingOnly";
+    context: "final" | "update";
+  }): Promise<boolean> => {
+    const {
+      lane,
+      laneName,
+      text,
+      previewButtons,
+      stopBeforeEdit = false,
+      updateLaneSnapshot = false,
+      skipRegressive,
+      context,
+    } = params;
+    if (!lane.stream) {
+      return false;
+    }
+    const hadPreviewMessage = typeof lane.stream.messageId() === "number";
+    if (stopBeforeEdit) {
+      await lane.stream.stop();
+    }
+    const previewMessageId = lane.stream.messageId();
+    if (typeof previewMessageId !== "number") {
+      return false;
+    }
+    const currentPreviewText = getLanePreviewText(lane);
+    const shouldSkipRegressive =
+      Boolean(currentPreviewText) &&
+      currentPreviewText.startsWith(text) &&
+      text.length < currentPreviewText.length &&
+      (skipRegressive === "always" || hadPreviewMessage);
+    if (shouldSkipRegressive) {
+      // Avoid regressive punctuation/wording flicker from occasional shorter finals.
+      deliveryState.delivered = true;
+      return true;
+    }
+    try {
+      await editMessageTelegram(chatId, previewMessageId, text, {
+        api: bot.api,
+        cfg,
+        accountId: route.accountId,
+        linkPreview: telegramCfg.linkPreview,
+        buttons: previewButtons,
+      });
+      if (updateLaneSnapshot) {
+        lane.lastPartialText = text;
+        lane.draftText = text;
+      }
+      deliveryState.delivered = true;
+      return true;
+    } catch (err) {
+      logVerbose(
+        `telegram: ${laneName} preview ${context} edit failed; falling back to standard send (${String(err)})`,
+      );
+      return false;
+    }
+  };
+  const applyTextToPayload = (payload: ReplyPayload, text: string): ReplyPayload => {
+    if (payload.text === text) {
+      return payload;
+    }
+    return { ...payload, text };
+  };
+  const sendPayload = async (payload: ReplyPayload) => {
+    const result = await deliverReplies({
+      ...deliveryBaseOptions,
+      replies: [payload],
+      onVoiceRecording: sendRecordVoice,
+    });
+    if (result.delivered) {
+      deliveryState.delivered = true;
+    }
+    return result.delivered;
+  };
+  type LaneDeliveryResult = "preview-finalized" | "preview-updated" | "sent" | "skipped";
+  const deliverLaneText = async (params: {
+    laneName: LaneName;
+    text: string;
+    payload: ReplyPayload;
+    infoKind: string;
+    previewButtons?: TelegramInlineButtons;
+    allowPreviewUpdateForNonFinal?: boolean;
+  }): Promise<LaneDeliveryResult> => {
+    const {
+      laneName,
+      text,
+      payload,
+      infoKind,
+      previewButtons,
+      allowPreviewUpdateForNonFinal = false,
+    } = params;
+    const lane = lanes[laneName];
+    const hasMedia = Boolean(payload.mediaUrl) || (payload.mediaUrls?.length ?? 0) > 0;
+    const canEditViaPreview =
+      !hasMedia && text.length > 0 && text.length <= draftMaxChars && !payload.isError;
+
+    if (infoKind === "final") {
+      if (canEditViaPreview && !finalizedPreviewByLane[laneName]) {
+        await flushDraftLane(lane);
+        const finalized = await tryUpdatePreviewForLane({
+          lane,
+          laneName,
+          text,
+          previewButtons,
+          stopBeforeEdit: true,
+          skipRegressive: "existingOnly",
+          context: "final",
+        });
+        if (finalized) {
+          finalizedPreviewByLane[laneName] = true;
+          return "preview-finalized";
+        }
+      } else if (!hasMedia && !payload.isError && text.length > draftMaxChars) {
+        logVerbose(
+          `telegram: preview final too long for edit (${text.length} > ${draftMaxChars}); falling back to standard send`,
+        );
+      }
+      await lane.stream?.stop();
+      const delivered = await sendPayload(applyTextToPayload(payload, text));
+      return delivered ? "sent" : "skipped";
+    }
+
+    if (allowPreviewUpdateForNonFinal && canEditViaPreview) {
+      const updated = await tryUpdatePreviewForLane({
+        lane,
+        laneName,
+        text,
+        previewButtons,
+        stopBeforeEdit: false,
+        updateLaneSnapshot: true,
+        skipRegressive: "always",
+        context: "update",
+      });
+      if (updated) {
+        return "preview-updated";
+      }
+    }
+
+    const delivered = await sendPayload(applyTextToPayload(payload, text));
+    return delivered ? "sent" : "skipped";
+  };

  let queuedFinal = false;
-  let dispatchError: unknown;
  try {
    ({ queuedFinal } = await dispatchReplyWithBufferedBlockDispatcher({
      ctx: ctxPayload,
@@ -320,117 +552,86 @@ export const dispatchTelegramMessage = async ({
      dispatcherOptions: {
        ...prefixOptions,
        deliver: async (payload, info) => {
-          if (info.kind === "final") {
-            await flushDraft();
-            const hasMedia = Boolean(payload.mediaUrl) || (payload.mediaUrls?.length ?? 0) > 0;
-            const previewMessageId = draftStream?.messageId();
-            const finalText = payload.text;
-            const currentPreviewText = streamMode === "block" ? draftText : lastPartialText;
-            const previewButtons = (
-              payload.channelData?.telegram as { buttons?: TelegramInlineButtons } | undefined
+          const previewButtons = (
+            payload.channelData?.telegram as { buttons?: TelegramInlineButtons } | undefined
+          )?.buttons;
+          const segments = splitTextIntoLaneSegments(payload.text);
+          const hasMedia = Boolean(payload.mediaUrl) || (payload.mediaUrls?.length ?? 0) > 0;
+
+          const flushBufferedFinalAnswer = async () => {
+            const buffered = reasoningStepState.takeBufferedFinalAnswer();
+            if (!buffered) {
+              return;
+            }
+            const bufferedButtons = (
+              buffered.payload.channelData?.telegram as
+                | { buttons?: TelegramInlineButtons }
+                | undefined
            )?.buttons;
-            let draftStoppedForPreviewEdit = false;
-            // Skip preview edit for error payloads to avoid overwriting previous content
-            const canFinalizeViaPreviewEdit =
-              !finalizedViaPreviewMessage &&
-              !hasMedia &&
-              typeof finalText === "string" &&
-              finalText.length > 0 &&
-              typeof previewMessageId === "number" &&
-              finalText.length <= draftMaxChars &&
-              !payload.isError;
-            if (canFinalizeViaPreviewEdit) {
-              await draftStream?.stop();
-              draftStoppedForPreviewEdit = true;
-              if (
-                currentPreviewText &&
-                currentPreviewText.startsWith(finalText) &&
-                finalText.length < currentPreviewText.length
-              ) {
-                // Ignore regressive final edits (e.g., "Okay." -> "Ok"), which
-                // can appear transiently in some provider streams.
-                return;
-              }
-              try {
-                await editMessageTelegram(chatId, previewMessageId, finalText, {
-                  api: bot.api,
-                  cfg,
-                  accountId: route.accountId,
-                  linkPreview: telegramCfg.linkPreview,
-                  buttons: previewButtons,
-                });
-                finalizedViaPreviewMessage = true;
-                deliveryState.delivered = true;
-                logVerbose(
-                  `telegram: finalized response via preview edit (messageId=${previewMessageId})`,
-                );
-                return;
-              } catch (err) {
-                logVerbose(
-                  `telegram: preview final edit failed; falling back to standard send (${String(err)})`,
-                );
-              }
-            }
+            await deliverLaneText({
+              laneName: "answer",
+              text: buffered.text,
+              payload: buffered.payload,
+              infoKind: "final",
+              previewButtons: bufferedButtons,
+            });
+            reasoningStepState.resetForNextStep();
+          };
+
+          for (const segment of segments) {
            if (
-              !hasMedia &&
-              !payload.isError &&
-              typeof finalText === "string" &&
-              finalText.length > draftMaxChars
+              segment.lane === "answer" &&
+              info.kind === "final" &&
+              reasoningStepState.shouldBufferFinalAnswer()
            ) {
-              logVerbose(
-                `telegram: preview final too long for edit (${finalText.length} > ${draftMaxChars}); falling back to standard send`,
-              );
+              reasoningStepState.bufferFinalAnswer({ payload, text: segment.text });
+              continue;
            }
-            if (!draftStoppedForPreviewEdit) {
-              await draftStream?.stop();
+            if (segment.lane === "reasoning") {
+              reasoningStepState.noteReasoningHint();
            }
-            // Check if stop() sent a message (debounce released on isFinal)
-            // If so, edit that message instead of sending a new one
-            const messageIdAfterStop = draftStream?.messageId();
-            if (
-              !finalizedViaPreviewMessage &&
-              typeof messageIdAfterStop === "number" &&
-              typeof finalText === "string" &&
-              finalText.length > 0 &&
-              finalText.length <= draftMaxChars &&
-              !hasMedia &&
-              !payload.isError
-            ) {
-              try {
-                await editMessageTelegram(chatId, messageIdAfterStop, finalText, {
-                  api: bot.api,
-                  cfg,
-                  accountId: route.accountId,
-                  linkPreview: telegramCfg.linkPreview,
-                  buttons: previewButtons,
-                });
-                finalizedViaPreviewMessage = true;
-                deliveryState.delivered = true;
-                logVerbose(
-                  `telegram: finalized response via post-stop preview edit (messageId=${messageIdAfterStop})`,
-                );
-                return;
-              } catch (err) {
-                logVerbose(
-                  `telegram: post-stop preview edit failed; falling back to standard send (${String(err)})`,
-                );
+            const result = await deliverLaneText({
+              laneName: segment.lane,
+              text: segment.text,
+              payload,
+              infoKind: info.kind,
+              previewButtons,
+              allowPreviewUpdateForNonFinal: segment.lane === "reasoning",
+            });
+            if (segment.lane === "reasoning") {
+              if (result !== "skipped") {
+                reasoningStepState.noteReasoningDelivered();
+                await flushBufferedFinalAnswer();
              }
+              continue;
+            }
+            if (info.kind === "final") {
+              if (reasoningLane.hasStreamedMessage) {
+                finalizedPreviewByLane.reasoning = true;
+              }
+              reasoningStepState.resetForNextStep();
            }
          }
-          const result = await deliverReplies({
-            ...deliveryBaseOptions,
-            replies: [payload],
-            onVoiceRecording: sendRecordVoice,
-          });
-          if (result.delivered) {
-            deliveryState.delivered = true;
-            logVerbose(
-              `telegram: ${info.kind} reply delivered to chat ${chatId}${payload.isError ? " (error payload)" : ""}`,
-            );
-          } else {
-            logVerbose(
-              `telegram: ${info.kind} reply delivery returned not-delivered for chat ${chatId}`,
-            );
+          if (segments.length > 0) {
+            return;
+          }
+
+          if (info.kind === "final") {
+            await answerLane.stream?.stop();
+            await reasoningLane.stream?.stop();
+            reasoningStepState.resetForNextStep();
+          }
+          const canSendAsIs =
+            hasMedia || typeof payload.text !== "string" || payload.text.length > 0;
+          if (!canSendAsIs) {
+            if (info.kind === "final") {
+              await flushBufferedFinalAnswer();
+            }
+            return;
+          }
+          await sendPayload(payload);
+          if (info.kind === "final") {
+            await flushBufferedFinalAnswer();
          }
        },
        onSkip: (_payload, info) => {
@@ -439,7 +640,7 @@ export const dispatchTelegramMessage = async ({
          }
        },
        onError: (err, info) => {
-          deliveryState.failedDeliveries += 1;
+          deliveryState.failedNonSilent += 1;
          runtime.error?.(danger(`telegram ${info.kind} reply failed: ${String(err)}`));
        },
        onReplyStart: createTypingCallbacks({
@@ -457,60 +658,82 @@ export const dispatchTelegramMessage = async ({
      replyOptions: {
        skillFilter,
        disableBlockStreaming,
-        onPartialReply: draftStream ? (payload) => updateDraftFromPartial(payload.text) : undefined,
-        onAssistantMessageStart: draftStream
-          ? () => {
-              // Only split preview bubbles in block mode. In partial mode, keep
-              // editing one preview message to avoid flooding the chat.
-              logVerbose(
-                `telegram: onAssistantMessageStart called, hasStreamedMessage=${hasStreamedMessage}`,
-              );
-              if (shouldSplitPreviewMessages && hasStreamedMessage) {
-                logVerbose(`telegram: calling forceNewMessage()`);
-                draftStream.forceNewMessage();
+        onPartialReply:
+          answerLane.stream || reasoningLane.stream
+            ? (payload) => ingestDraftLaneSegments(payload.text)
+            : undefined,
+        onReasoningStream: reasoningLane.stream
+          ? (payload) => {
+              // Split between reasoning blocks only when the next reasoning
+              // stream starts. Splitting at reasoning-end can orphan the active
+              // preview and cause duplicate reasoning sends on reasoning final.
+              if (splitReasoningOnNextStream) {
+                reasoningLane.stream?.forceNewMessage();
+                resetDraftLaneState(reasoningLane);
+                splitReasoningOnNextStream = false;
              }
-              lastPartialText = "";
-              draftText = "";
-              draftChunker?.reset();
+              ingestDraftLaneSegments(payload.text);
            }
          : undefined,
-        onReasoningEnd: draftStream
+        onAssistantMessageStart: answerLane.stream
          ? () => {
-              // Same policy as assistant-message boundaries: split only in block mode.
-              if (shouldSplitPreviewMessages && hasStreamedMessage) {
-                draftStream.forceNewMessage();
+              reasoningStepState.resetForNextStep();
+              // Keep answer blocks separated in block mode; partial mode keeps one answer lane.
+              if (streamMode === "block" && answerLane.hasStreamedMessage) {
+                answerLane.stream?.forceNewMessage();
              }
-              lastPartialText = "";
-              draftText = "";
-              draftChunker?.reset();
+              resetDraftLaneState(answerLane);
+            }
+          : undefined,
+        onReasoningEnd: reasoningLane.stream
+          ? () => {
+              // Split when/if a later reasoning block begins.
+              splitReasoningOnNextStream = reasoningLane.hasStreamedMessage;
            }
          : undefined,
        onModelSelected,
      },
    }));
-  } catch (err) {
-    dispatchError = err;
  } finally {
-    await draftStream?.stop();
+    // Must stop() first to flush debounced content before clear() wipes state.
+    const streamCleanupStates = new Map<
+      NonNullable<DraftLaneState["stream"]>,
+      { shouldClear: boolean }
+    >();
+    const lanesToCleanup: Array<{ laneName: LaneName; lane: DraftLaneState }> = [
+      { laneName: "answer", lane: answerLane },
+      { laneName: "reasoning", lane: reasoningLane },
+    ];
+    for (const laneState of lanesToCleanup) {
+      const stream = laneState.lane.stream;
+      if (!stream) {
+        continue;
+      }
+      const shouldClear = !finalizedPreviewByLane[laneState.laneName];
+      const existing = streamCleanupStates.get(stream);
+      if (!existing) {
+        streamCleanupStates.set(stream, { shouldClear });
+        continue;
+      }
+      existing.shouldClear = existing.shouldClear && shouldClear;
+    }
+    for (const [stream, cleanupState] of streamCleanupStates) {
+      await stream.stop();
+      if (cleanupState.shouldClear) {
+        await stream.clear();
+      }
+    }
  }
  let sentFallback = false;
-  try {
-    if (
-      !dispatchError &&
-      !deliveryState.delivered &&
-      (deliveryState.skippedNonSilent > 0 || deliveryState.failedDeliveries > 0)
-    ) {
-      const result = await deliverReplies({
-        replies: [{ text: EMPTY_RESPONSE_FALLBACK }],
-        ...deliveryBaseOptions,
-      });
-      sentFallback = result.delivered;
-    }
-  } finally {
-    await clearDraftPreviewIfNeeded();
-  }
-  if (dispatchError) {
-    throw dispatchError;
+  if (
+    !deliveryState.delivered &&
+    (deliveryState.skippedNonSilent > 0 || deliveryState.failedNonSilent > 0)
+  ) {
+    const result = await deliverReplies({
+      replies: [{ text: EMPTY_RESPONSE_FALLBACK }],
+      ...deliveryBaseOptions,
+    });
+    sentFallback = result.delivered;
  }

  const hasFinalResponse = queuedFinal || sentFallback;
--- a/src/telegram/draft-stream.test.ts
+++ b/src/telegram/draft-stream.test.ts
@@ -133,6 +133,48 @@ describe("createTelegramDraftStream", () => {
    expect(api.sendMessage).toHaveBeenCalledTimes(2);
    expect(api.sendMessage).toHaveBeenLastCalledWith(123, "After thinking", undefined);
  });
+
+  it("supports rendered previews with parse_mode", async () => {
+    const api = createMockDraftApi();
+    const stream = createTelegramDraftStream({
+      // oxlint-disable-next-line typescript/no-explicit-any
+      api: api as any,
+      chatId: 123,
+      renderText: (text) => ({ text: `<i>${text}</i>`, parseMode: "HTML" }),
+    });
+
+    stream.update("hello");
+    await stream.flush();
+    expect(api.sendMessage).toHaveBeenCalledWith(123, "<i>hello</i>", { parse_mode: "HTML" });
+
+    stream.update("hello again");
+    await stream.flush();
+    expect(api.editMessageText).toHaveBeenCalledWith(123, 17, "<i>hello again</i>", {
+      parse_mode: "HTML",
+    });
+  });
+
+  it("enforces maxChars after renderText expansion", async () => {
+    const api = createMockDraftApi();
+    const warn = vi.fn();
+    const stream = createTelegramDraftStream({
+      // oxlint-disable-next-line typescript/no-explicit-any
+      api: api as any,
+      chatId: 123,
+      maxChars: 100,
+      renderText: () => ({ text: `<b>${"<".repeat(120)}</b>`, parseMode: "HTML" }),
+      warn,
+    });
+
+    stream.update("short raw text");
+    await stream.flush();
+
+    expect(api.sendMessage).not.toHaveBeenCalled();
+    expect(api.editMessageText).not.toHaveBeenCalled();
+    expect(warn).toHaveBeenCalledWith(
+      expect.stringContaining("telegram stream preview stopped (text length 127 > 100)"),
+    );
+  });
 });

 describe("draft stream initial message debounce", () => {
--- a/src/telegram/draft-stream.ts
+++ b/src/telegram/draft-stream.ts
@@ -15,6 +15,11 @@ export type TelegramDraftStream = {
  forceNewMessage: () => void;
 };

+type TelegramDraftPreview = {
+  text: string;
+  parseMode?: "HTML";
+};
+
 export function createTelegramDraftStream(params: {
  api: Bot["api"];
  chatId: number;
@@ -24,6 +29,8 @@ export function createTelegramDraftStream(params: {
  throttleMs?: number;
  /** Minimum chars before sending first message (debounce for push notifications) */
  minInitialChars?: number;
+  /** Optional preview renderer (e.g. markdown -> HTML + parse mode). */
+  renderText?: (text: string) => TelegramDraftPreview;
  log?: (message: string) => void;
  warn?: (message: string) => void;
 }): TelegramDraftStream {
@@ -42,6 +49,7 @@ export function createTelegramDraftStream(params: {

  let streamMessageId: number | undefined;
  let lastSentText = "";
+  let lastSentParseMode: "HTML" | undefined;
  let stopped = false;
  let isFinal = false;

@@ -54,33 +62,52 @@ export function createTelegramDraftStream(params: {
    if (!trimmed) {
      return false;
    }
-    if (trimmed.length > maxChars) {
+    const rendered = params.renderText?.(trimmed) ?? { text: trimmed };
+    const renderedText = rendered.text.trimEnd();
+    const renderedParseMode = rendered.parseMode;
+    if (!renderedText) {
+      return false;
+    }
+    if (renderedText.length > maxChars) {
      // Telegram text messages/edits cap at 4096 chars.
      // Stop streaming once we exceed the cap to avoid repeated API failures.
      stopped = true;
      params.warn?.(
-        `telegram stream preview stopped (text length ${trimmed.length} > ${maxChars})`,
+        `telegram stream preview stopped (text length ${renderedText.length} > ${maxChars})`,
      );
      return false;
    }
-    if (trimmed === lastSentText) {
+    if (renderedText === lastSentText && renderedParseMode === lastSentParseMode) {
      return true;
    }

    // Debounce first preview send for better push notification quality.
    if (typeof streamMessageId !== "number" && minInitialChars != null && !isFinal) {
-      if (trimmed.length < minInitialChars) {
+      if (renderedText.length < minInitialChars) {
        return false;
      }
    }

-    lastSentText = trimmed;
+    lastSentText = renderedText;
+    lastSentParseMode = renderedParseMode;
    try {
      if (typeof streamMessageId === "number") {
-        await params.api.editMessageText(chatId, streamMessageId, trimmed);
+        if (renderedParseMode) {
+          await params.api.editMessageText(chatId, streamMessageId, renderedText, {
+            parse_mode: renderedParseMode,
+          });
+        } else {
+          await params.api.editMessageText(chatId, streamMessageId, renderedText);
+        }
        return true;
      }
-      const sent = await params.api.sendMessage(chatId, trimmed, replyParams);
+      const sendParams = renderedParseMode
+        ? {
+            ...replyParams,
+            parse_mode: renderedParseMode,
+          }
+        : replyParams;
+      const sent = await params.api.sendMessage(chatId, renderedText, sendParams);
      const sentMessageId = sent?.message_id;
      if (typeof sentMessageId !== "number" || !Number.isFinite(sentMessageId)) {
        stopped = true;
@@ -138,6 +165,7 @@ export function createTelegramDraftStream(params: {
  const forceNewMessage = () => {
    streamMessageId = undefined;
    lastSentText = "";
+    lastSentParseMode = undefined;
    loop.resetPending();
  };

--- a/src/telegram/reasoning-lane-coordinator.test.ts
+++ b/src/telegram/reasoning-lane-coordinator.test.ts
@@ -0,0 +1,25 @@
+import { describe, expect, it } from "vitest";
+import { splitTelegramReasoningText } from "./reasoning-lane-coordinator.js";
+
+describe("splitTelegramReasoningText", () => {
+  it("splits real tagged reasoning and answer", () => {
+    expect(splitTelegramReasoningText("<think>example</think>Done")).toEqual({
+      reasoningText: "Reasoning:\n_example_",
+      answerText: "Done",
+    });
+  });
+
+  it("ignores literal think tags inside inline code", () => {
+    const text = "Use `<think>example</think>` literally.";
+    expect(splitTelegramReasoningText(text)).toEqual({
+      answerText: text,
+    });
+  });
+
+  it("ignores literal think tags inside fenced code", () => {
+    const text = "```xml\n<think>example</think>\n```";
+    expect(splitTelegramReasoningText(text)).toEqual({
+      answerText: text,
+    });
+  });
+});
--- a/src/telegram/reasoning-lane-coordinator.ts
+++ b/src/telegram/reasoning-lane-coordinator.ts
@@ -0,0 +1,167 @@
+import { formatReasoningMessage } from "../agents/pi-embedded-utils.js";
+import type { ReplyPayload } from "../auto-reply/types.js";
+import { stripReasoningTagsFromText } from "../shared/text/reasoning-tags.js";
+
+const REASONING_MESSAGE_PREFIX = "Reasoning:\n";
+const REASONING_TAG_PREFIXES = [
+  "<think",
+  "<thinking",
+  "<thought",
+  "<antthinking",
+  "</think",
+  "</thinking",
+  "</thought",
+  "</antthinking",
+];
+const THINKING_TAG_RE = /<\s*(\/?)\s*(?:think(?:ing)?|thought|antthinking)\b[^<>]*>/gi;
+
+interface CodeRegion {
+  start: number;
+  end: number;
+}
+
+function findCodeRegions(text: string): CodeRegion[] {
+  const regions: CodeRegion[] = [];
+
+  const fencedRe = /(^|\n)(```|~~~)[^\n]*\n[\s\S]*?(?:\n\2(?:\n|$)|$)/g;
+  for (const match of text.matchAll(fencedRe)) {
+    const start = (match.index ?? 0) + match[1].length;
+    regions.push({ start, end: start + match[0].length - match[1].length });
+  }
+
+  const inlineRe = /`+[^`]+`+/g;
+  for (const match of text.matchAll(inlineRe)) {
+    const start = match.index ?? 0;
+    const end = start + match[0].length;
+    const insideFenced = regions.some((r) => start >= r.start && end <= r.end);
+    if (!insideFenced) {
+      regions.push({ start, end });
+    }
+  }
+
+  regions.sort((a, b) => a.start - b.start);
+  return regions;
+}
+
+function isInsideCode(pos: number, regions: CodeRegion[]): boolean {
+  return regions.some((r) => pos >= r.start && pos < r.end);
+}
+
+function extractThinkingFromTaggedStreamOutsideCode(text: string): string {
+  if (!text) {
+    return "";
+  }
+  const codeRegions = findCodeRegions(text);
+  let result = "";
+  let lastIndex = 0;
+  let inThinking = false;
+  THINKING_TAG_RE.lastIndex = 0;
+  for (const match of text.matchAll(THINKING_TAG_RE)) {
+    const idx = match.index ?? 0;
+    if (isInsideCode(idx, codeRegions)) {
+      continue;
+    }
+    if (inThinking) {
+      result += text.slice(lastIndex, idx);
+    }
+    const isClose = match[1] === "/";
+    inThinking = !isClose;
+    lastIndex = idx + match[0].length;
+  }
+  if (inThinking) {
+    result += text.slice(lastIndex);
+  }
+  return result.trim();
+}
+
+function isPartialReasoningTagPrefix(text: string): boolean {
+  const trimmed = text.trimStart().toLowerCase();
+  if (!trimmed.startsWith("<")) {
+    return false;
+  }
+  if (trimmed.includes(">")) {
+    return false;
+  }
+  return REASONING_TAG_PREFIXES.some((prefix) => prefix.startsWith(trimmed));
+}
+
+export type TelegramReasoningSplit = {
+  reasoningText?: string;
+  answerText?: string;
+};
+
+export function splitTelegramReasoningText(text?: string): TelegramReasoningSplit {
+  if (typeof text !== "string") {
+    return {};
+  }
+
+  const trimmed = text.trim();
+  if (isPartialReasoningTagPrefix(trimmed)) {
+    return {};
+  }
+  if (
+    trimmed.startsWith(REASONING_MESSAGE_PREFIX) &&
+    trimmed.length > REASONING_MESSAGE_PREFIX.length
+  ) {
+    return { reasoningText: trimmed };
+  }
+
+  const taggedReasoning = extractThinkingFromTaggedStreamOutsideCode(text);
+  const strippedAnswer = stripReasoningTagsFromText(text, { mode: "strict", trim: "both" });
+
+  if (!taggedReasoning && strippedAnswer === text) {
+    return { answerText: text };
+  }
+
+  const reasoningText = taggedReasoning ? formatReasoningMessage(taggedReasoning) : undefined;
+  const answerText = strippedAnswer || undefined;
+  return { reasoningText, answerText };
+}
+
+export type BufferedFinalAnswer = {
+  payload: ReplyPayload;
+  text: string;
+};
+
+export function createTelegramReasoningStepState() {
+  let reasoningStatus: "none" | "hinted" | "delivered" = "none";
+  let bufferedFinalAnswer: BufferedFinalAnswer | undefined;
+
+  const noteReasoningHint = () => {
+    if (reasoningStatus === "none") {
+      reasoningStatus = "hinted";
+    }
+  };
+
+  const noteReasoningDelivered = () => {
+    reasoningStatus = "delivered";
+  };
+
+  const shouldBufferFinalAnswer = () => {
+    return reasoningStatus === "hinted" && !bufferedFinalAnswer;
+  };
+
+  const bufferFinalAnswer = (value: BufferedFinalAnswer) => {
+    bufferedFinalAnswer = value;
+  };
+
+  const takeBufferedFinalAnswer = (): BufferedFinalAnswer | undefined => {
+    const value = bufferedFinalAnswer;
+    bufferedFinalAnswer = undefined;
+    return value;
+  };
+
+  const resetForNextStep = () => {
+    reasoningStatus = "none";
+    bufferedFinalAnswer = undefined;
+  };
+
+  return {
+    noteReasoningHint,
+    noteReasoningDelivered,
+    shouldBufferFinalAnswer,
+    bufferFinalAnswer,
+    takeBufferedFinalAnswer,
+    resetForNextStep,
+  };
+}