fix: suppress <think> leakage + split reasoning output (#614) (thanks @zknicker)

src/agents/pi-embedded-runner.ts

@@ -1604,23 +1604,21 @@ export async function runEmbeddedPiAgent(params: {
             }
           }
 
-          const fallbackText = lastAssistant
-            ? (() => {
-                const base = extractAssistantText(lastAssistant);
-                if (params.reasoningLevel !== "on") return base;
-                const thinking = extractAssistantThinking(lastAssistant);
-                const formatted = thinking
-                  ? formatReasoningMarkdown(thinking)
-                  : "";
-                if (!formatted) return base;
-                return base ? `${formatted}\n\n${base}` : formatted;
-              })()
+          const reasoningText =
+            lastAssistant && params.reasoningLevel === "on"
+              ? formatReasoningMarkdown(extractAssistantThinking(lastAssistant))
+              : "";
+          if (reasoningText) replyItems.push({ text: reasoningText });
+
+          const fallbackAnswerText = lastAssistant
+            ? extractAssistantText(lastAssistant)
             : "";
-          for (const text of assistantTexts.length
+          const answerTexts = assistantTexts.length
             ? assistantTexts
-            : fallbackText
-              ? [fallbackText]
-              : []) {
+            : fallbackAnswerText
+              ? [fallbackAnswerText]
+              : [];
+          for (const text of answerTexts) {
             const { text: cleanedText, mediaUrls } = splitMediaFromOutput(text);
             if (!cleanedText && (!mediaUrls || mediaUrls.length === 0))
               continue;

src/agents/pi-embedded-subscribe.test.ts

@@ -129,7 +129,7 @@ describe("subscribeEmbeddedPiSession", () => {
     expect(payload.text).toBe("Hello block");
   });
 
-  it("prepends reasoning before text when enabled", () => {
+  it("emits reasoning as a separate message when enabled", () => {
     let handler: ((evt: unknown) => void) | undefined;
     const session: StubSession = {
       subscribe: (fn) => {
@@ -160,11 +160,11 @@ describe("subscribeEmbeddedPiSession", () => {
 
     handler?.({ type: "message_end", message: assistantMessage });
 
-    expect(onBlockReply).toHaveBeenCalledTimes(1);
-    const payload = onBlockReply.mock.calls[0][0];
-    expect(payload.text).toBe(
-      "_Reasoning:_\n_Because it helps_\n\nFinal answer",
+    expect(onBlockReply).toHaveBeenCalledTimes(2);
+    expect(onBlockReply.mock.calls[0][0].text).toBe(
+      "Reasoning:\nBecause it helps",
     );
+    expect(onBlockReply.mock.calls[1][0].text).toBe("Final answer");
   });
 
   it("promotes <think> tags to thinking blocks at write-time", () => {
@@ -200,10 +200,11 @@ describe("subscribeEmbeddedPiSession", () => {
 
     handler?.({ type: "message_end", message: assistantMessage });
 
-    expect(onBlockReply).toHaveBeenCalledTimes(1);
+    expect(onBlockReply).toHaveBeenCalledTimes(2);
     expect(onBlockReply.mock.calls[0][0].text).toBe(
-      "_Reasoning:_\n_Because it helps_\n\nFinal answer",
+      "Reasoning:\nBecause it helps",
     );
+    expect(onBlockReply.mock.calls[1][0].text).toBe("Final answer");
 
     expect(assistantMessage.content).toEqual([
       { type: "thinking", thinking: "Because it helps" },

src/agents/pi-embedded-subscribe.ts

@@ -936,11 +936,7 @@ export function subscribeEmbeddedPiSession(params: {
           const formattedReasoning = rawThinking
             ? formatReasoningMarkdown(rawThinking)
             : "";
-          const text = includeReasoning
-            ? baseText && formattedReasoning
-              ? `${formattedReasoning}\n\n${baseText}`
-              : formattedReasoning || baseText
-            : baseText;
+          const text = baseText;
 
           const addedDuringMessage =
             assistantTexts.length > assistantTextBaseline;
@@ -953,13 +949,28 @@ export function subscribeEmbeddedPiSession(params: {
           }
           assistantTextBaseline = assistantTexts.length;
 
+          const onBlockReply = params.onBlockReply;
+          const shouldEmitReasoning =
+            includeReasoning &&
+            Boolean(formattedReasoning) &&
+            Boolean(onBlockReply) &&
+            formattedReasoning !== lastReasoningSent;
+          const shouldEmitReasoningBeforeAnswer =
+            shouldEmitReasoning &&
+            blockReplyBreak === "message_end" &&
+            !addedDuringMessage;
+          if (shouldEmitReasoningBeforeAnswer && formattedReasoning) {
+            lastReasoningSent = formattedReasoning;
+            void onBlockReply?.({ text: formattedReasoning });
+          }
+
           if (
             (blockReplyBreak === "message_end" ||
               (blockChunker
                 ? blockChunker.hasBuffered()
                 : blockBuffer.length > 0)) &&
             text &&
-            params.onBlockReply
+            onBlockReply
           ) {
             if (blockChunker?.hasBuffered()) {
               blockChunker.drain({ force: true, emit: emitBlockChunk });
@@ -975,7 +986,7 @@ export function subscribeEmbeddedPiSession(params: {
                 const { text: cleanedText, mediaUrls } =
                   splitMediaFromOutput(text);
                 if (cleanedText || (mediaUrls && mediaUrls.length > 0)) {
-                  void params.onBlockReply({
+                  void onBlockReply({
                     text: cleanedText,
                     mediaUrls: mediaUrls?.length ? mediaUrls : undefined,
                   });
@@ -983,16 +994,13 @@ export function subscribeEmbeddedPiSession(params: {
               }
             }
           }
-          const onBlockReply = params.onBlockReply;
-          const shouldEmitReasoningBlock =
-            includeReasoning &&
-            Boolean(formattedReasoning) &&
-            Boolean(onBlockReply) &&
-            formattedReasoning !== lastReasoningSent &&
-            (blockReplyBreak === "text_end" || Boolean(blockChunker));
-          if (shouldEmitReasoningBlock && formattedReasoning && onBlockReply) {
+          if (
+            shouldEmitReasoning &&
+            !shouldEmitReasoningBeforeAnswer &&
+            formattedReasoning
+          ) {
             lastReasoningSent = formattedReasoning;
-            void onBlockReply({ text: formattedReasoning });
+            void onBlockReply?.({ text: formattedReasoning });
           }
           if (streamReasoning && rawThinking) {
             emitReasoningStream(rawThinking);

src/agents/pi-embedded-utils.ts

@@ -37,12 +37,7 @@ export function extractAssistantThinking(msg: AssistantMessage): string {
 export function formatReasoningMarkdown(text: string): string {
   const trimmed = text.trim();
   if (!trimmed) return "";
-  const lines = trimmed.split(/\r?\n/);
-  const wrapped = lines
-    .map((line) => line.trim())
-    .map((line) => (line ? `_${line}_` : ""))
-    .filter((line) => line.length > 0);
-  return wrapped.length > 0 ? [`_Reasoning:_`, ...wrapped].join("\n") : "";
+  return `Reasoning:\n${trimmed}`;
 }
 
 export function inferToolMetaFromArgs(