diff --git a/src/agents/pi-embedded-runner.ts b/src/agents/pi-embedded-runner.ts
index 7a416c60c..5ef32ea65 100644
--- a/src/agents/pi-embedded-runner.ts
+++ b/src/agents/pi-embedded-runner.ts
@@ -1044,6 +1044,7 @@ export async function runEmbeddedPiAgent(params: {
   onBlockReply?: (payload: {
     text?: string;
     mediaUrls?: string[];
+    audioAsVoice?: boolean;
   }) => void | Promise<void>;
   blockReplyBreak?: "text_end" | "message_end";
   blockReplyChunking?: BlockReplyChunking;
@@ -1640,6 +1641,7 @@ export async function runEmbeddedPiAgent(params: {
             text: string;
             media?: string[];
             isError?: boolean;
+            audioAsVoice?: boolean;
           }> = [];
 
           const errorText = lastAssistant
@@ -1656,10 +1658,10 @@ export async function runEmbeddedPiAgent(params: {
           if (inlineToolResults) {
             for (const { toolName, meta } of toolMetas) {
               const agg = formatToolAggregate(toolName, meta ? [meta] : []);
-              const { text: cleanedText, mediaUrls } =
+              const { text: cleanedText, mediaUrls, audioAsVoice } =
                 splitMediaFromOutput(agg);
               if (cleanedText)
-                replyItems.push({ text: cleanedText, media: mediaUrls });
+                replyItems.push({ text: cleanedText, media: mediaUrls, audioAsVoice });
             }
           }
 
@@ -1678,18 +1680,31 @@ export async function runEmbeddedPiAgent(params: {
               ? [fallbackAnswerText]
               : [];
           for (const text of answerTexts) {
-            const { text: cleanedText, mediaUrls } = splitMediaFromOutput(text);
-            if (!cleanedText && (!mediaUrls || mediaUrls.length === 0))
+            const { text: cleanedText, mediaUrls, audioAsVoice } =
+              splitMediaFromOutput(text);
+            if (
+              !cleanedText &&
+              (!mediaUrls || mediaUrls.length === 0) &&
+              !audioAsVoice
+            )
               continue;
-            replyItems.push({ text: cleanedText, media: mediaUrls });
+            replyItems.push({ text: cleanedText, media: mediaUrls, audioAsVoice });
           }
 
+          // Check if any replyItem has audioAsVoice tag - if so, apply to all media payloads
+          const hasAudioAsVoiceTag = replyItems.some(
+            (item) => item.audioAsVoice,
+          );
+
           const payloads = replyItems
             .map((item) => ({
               text: item.text?.trim() ? item.text.trim() : undefined,
               mediaUrls: item.media?.length ? item.media : undefined,
               mediaUrl: item.media?.[0],
               isError: item.isError,
+              // Apply audioAsVoice to media payloads if tag was found anywhere in response
+              audioAsVoice:
+                item.audioAsVoice || (hasAudioAsVoiceTag && item.media?.length),
             }))
             .filter(
               (p) =>
diff --git a/src/agents/pi-embedded-subscribe.ts b/src/agents/pi-embedded-subscribe.ts
index 4c407a483..6dc5957cd 100644
--- a/src/agents/pi-embedded-subscribe.ts
+++ b/src/agents/pi-embedded-subscribe.ts
@@ -262,6 +262,7 @@ export function subscribeEmbeddedPiSession(params: {
   onBlockReply?: (payload: {
     text?: string;
     mediaUrls?: string[];
+    audioAsVoice?: boolean;
   }) => void | Promise<void>;
   blockReplyBreak?: "text_end" | "message_end";
   blockReplyChunking?: BlockReplyChunking;
@@ -436,11 +437,13 @@ export function subscribeEmbeddedPiSession(params: {
     lastBlockReplyText = chunk;
     assistantTexts.push(chunk);
     if (!params.onBlockReply) return;
-    const { text: cleanedText, mediaUrls } = splitMediaFromOutput(chunk);
-    if (!cleanedText && (!mediaUrls || mediaUrls.length === 0)) return;
+    const { text: cleanedText, mediaUrls, audioAsVoice } = splitMediaFromOutput(chunk);
+    // Skip empty payloads, but always emit if audioAsVoice is set (to propagate the flag)
+    if (!cleanedText && (!mediaUrls || mediaUrls.length === 0) && !audioAsVoice) return;
     void params.onBlockReply({
       text: cleanedText,
       mediaUrls: mediaUrls?.length ? mediaUrls : undefined,
+      audioAsVoice,
     });
   };
 
@@ -859,12 +862,18 @@ export function subscribeEmbeddedPiSession(params: {
                 );
               } else {
                 lastBlockReplyText = text;
-                const { text: cleanedText, mediaUrls } =
+                const { text: cleanedText, mediaUrls, audioAsVoice } =
                   splitMediaFromOutput(text);
-                if (cleanedText || (mediaUrls && mediaUrls.length > 0)) {
+                // Emit if there's content OR audioAsVoice flag (to propagate the flag)
+                if (
+                  cleanedText ||
+                  (mediaUrls && mediaUrls.length > 0) ||
+                  audioAsVoice
+                ) {
                   void onBlockReply({
                     text: cleanedText,
                     mediaUrls: mediaUrls?.length ? mediaUrls : undefined,
+                    audioAsVoice,
                   });
                 }
               }
diff --git a/src/auto-reply/reply/agent-runner.ts b/src/auto-reply/reply/agent-runner.ts
index 56a04b5b2..1063dcdd7 100644
--- a/src/auto-reply/reply/agent-runner.ts
+++ b/src/auto-reply/reply/agent-runner.ts
@@ -260,6 +260,14 @@ export async function runReplyAgent(params: {
   const pendingToolTasks = new Set<Promise<void>>();
   const blockReplyTimeoutMs =
     opts?.blockReplyTimeoutMs ?? BLOCK_REPLY_SEND_TIMEOUT_MS;
+
+  // Buffer audio blocks to apply [[audio_as_voice]] tag that may come later
+  const bufferedAudioBlocks: ReplyPayload[] = [];
+  let seenAudioAsVoice = false;
+
+  const AUDIO_EXTENSIONS = /\.(opus|mp3|m4a|wav|ogg|aac|flac)$/i;
+  const hasAudioMedia = (urls?: string[]): boolean =>
+    Boolean(urls?.some((u) => AUDIO_EXTENSIONS.test(u)));
   const replyToChannel =
     sessionCtx.OriginatingChannel ??
     ((sessionCtx.Surface ?? sessionCtx.Provider)?.toLowerCase() as
@@ -532,23 +540,34 @@ export async function runReplyAgent(params: {
                       },
                       sessionCtx.MessageSid,
                     );
-                    if (!isRenderablePayload(taggedPayload)) return;
+                    // Let through payloads with audioAsVoice flag even if empty (need to track it)
+                    if (!isRenderablePayload(taggedPayload) && !payload.audioAsVoice)
+                      return;
                     const audioTagResult = extractAudioTag(taggedPayload.text);
                     const cleaned = audioTagResult.cleaned || undefined;
                     const hasMedia =
                       Boolean(taggedPayload.mediaUrl) ||
                       (taggedPayload.mediaUrls?.length ?? 0) > 0;
-                    if (!cleaned && !hasMedia) return;
+                    // Skip empty payloads unless they have audioAsVoice flag (need to track it)
+                    if (!cleaned && !hasMedia && !payload.audioAsVoice) return;
                     if (
                       isSilentReplyText(cleaned, SILENT_REPLY_TOKEN) &&
                       !hasMedia
                     )
                       return;
+
+                    // Track if we've seen [[audio_as_voice]] from payload or text extraction
+                    if (payload.audioAsVoice || audioTagResult.audioAsVoice) {
+                      seenAudioAsVoice = true;
+                    }
+
                     const blockPayload: ReplyPayload = applyReplyToMode({
                       ...taggedPayload,
                       text: cleaned,
-                      audioAsVoice: audioTagResult.audioAsVoice,
+                      audioAsVoice:
+                        audioTagResult.audioAsVoice || payload.audioAsVoice,
                     });
+
                     void typingSignals
                       .signalTextDelta(taggedPayload.text)
                       .catch((err) => {
@@ -556,6 +575,14 @@ export async function runReplyAgent(params: {
                           `block reply typing signal failed: ${String(err)}`,
                         );
                       });
+
+                    // Buffer audio blocks to apply [[audio_as_voice]] that may come later
+                    const isAudioBlock = hasAudioMedia(taggedPayload.mediaUrls);
+                    if (isAudioBlock) {
+                      bufferedAudioBlocks.push(blockPayload);
+                      return; // Don't send immediately - wait for potential [[audio_as_voice]] tag
+                    }
+
                     blockReplyPipeline?.enqueue(blockPayload);
                   }
                 : undefined,
@@ -670,6 +697,17 @@ export async function runReplyAgent(params: {
     }
 
     const payloadArray = runResult.payloads ?? [];
+
+    if (bufferedAudioBlocks.length > 0 && blockReplyPipeline) {
+      for (const audioPayload of bufferedAudioBlocks) {
+        const finalPayload = seenAudioAsVoice
+          ? { ...audioPayload, audioAsVoice: true }
+          : audioPayload;
+        blockReplyPipeline.enqueue(finalPayload);
+      }
+      bufferedAudioBlocks.length = 0;
+    }
+
     if (blockReplyPipeline) {
       await blockReplyPipeline.flush({ force: true });
       blockReplyPipeline.stop();
@@ -677,6 +715,7 @@ export async function runReplyAgent(params: {
     if (pendingToolTasks.size > 0) {
       await Promise.allSettled(pendingToolTasks);
     }
+
     // Drain any late tool/block deliveries before deciding there's "nothing to send".
     // Otherwise, a late typing trigger (e.g. from a tool callback) can outlive the run and
     // keep the typing indicator stuck.
diff --git a/src/media/parse.ts b/src/media/parse.ts
index 2a01156c4..08f6d492e 100644
--- a/src/media/parse.ts
+++ b/src/media/parse.ts
@@ -1,5 +1,7 @@
 // Shared helpers for parsing MEDIA tokens from command/stdout text.
 
+import { parseFenceSpans } from "../markdown/fences.js";
+
 // Allow optional wrapping backticks and punctuation after the token; capture the core token.
 export const MEDIA_TOKEN_RE = /\bMEDIA:\s*`?([^\n]+)`?/gi;
 
@@ -22,10 +24,22 @@ function isValidMedia(candidate: string) {
   );
 }
 
+// Check if a character offset is inside any fenced code block
+function isInsideFence(
+  fenceSpans: Array<{ start: number; end: number }>,
+  offset: number,
+): boolean {
+  return fenceSpans.some((span) => offset >= span.start && offset < span.end);
+}
+
+// Regex to detect [[audio_as_voice]] tag
+const AUDIO_AS_VOICE_RE = /\[\[audio_as_voice\]\]/gi;
+
 export function splitMediaFromOutput(raw: string): {
   text: string;
   mediaUrls?: string[];
   mediaUrl?: string; // legacy first item for backward compatibility
+  audioAsVoice?: boolean; // true if [[audio_as_voice]] tag was found
 } {
   // KNOWN: Leading whitespace is semantically meaningful in Markdown (lists, indented fences).
   // We only trim the end; token cleanup below handles removing `MEDIA:` lines.
@@ -35,14 +49,26 @@ export function splitMediaFromOutput(raw: string): {
   const media: string[] = [];
   let foundMediaToken = false;
 
+  // Parse fenced code blocks to avoid extracting MEDIA tokens from inside them
+  const fenceSpans = parseFenceSpans(trimmedRaw);
+
   // Collect tokens line by line so we can strip them cleanly.
   const lines = trimmedRaw.split("\n");
   const keptLines: string[] = [];
 
+  let lineOffset = 0; // Track character offset for fence checking
   for (const line of lines) {
+    // Skip MEDIA extraction if this line is inside a fenced code block
+    if (isInsideFence(fenceSpans, lineOffset)) {
+      keptLines.push(line);
+      lineOffset += line.length + 1; // +1 for newline
+      continue;
+    }
+
     const matches = Array.from(line.matchAll(MEDIA_TOKEN_RE));
     if (matches.length === 0) {
       keptLines.push(line);
+      lineOffset += line.length + 1; // +1 for newline
       continue;
     }
 
@@ -86,18 +112,39 @@ export function splitMediaFromOutput(raw: string): {
     if (cleanedLine) {
       keptLines.push(cleanedLine);
     }
+    lineOffset += line.length + 1; // +1 for newline
   }
 
-  const cleanedText = keptLines
+  let cleanedText = keptLines
     .join("\n")
     .replace(/[ \t]+\n/g, "\n")
     .replace(/[ \t]{2,}/g, " ")
     .replace(/\n{2,}/g, "\n")
     .trim();
 
-  if (media.length === 0) {
-    return { text: foundMediaToken ? cleanedText : trimmedRaw };
+  // Detect and strip [[audio_as_voice]] tag
+  const hasAudioAsVoice = AUDIO_AS_VOICE_RE.test(cleanedText);
+  if (hasAudioAsVoice) {
+    cleanedText = cleanedText
+      .replace(AUDIO_AS_VOICE_RE, "")
+      .replace(/[ \t]+/g, " ")
+      .replace(/\n{2,}/g, "\n")
+      .trim();
   }
 
-  return { text: cleanedText, mediaUrls: media, mediaUrl: media[0] };
+  if (media.length === 0) {
+    const result: ReturnType<typeof splitMediaFromOutput> = {
+      // Return cleaned text if we found a media token OR audio tag, otherwise original
+      text: (foundMediaToken || hasAudioAsVoice) ? cleanedText : trimmedRaw,
+    };
+    if (hasAudioAsVoice) result.audioAsVoice = true;
+    return result;
+  }
+
+  return {
+    text: cleanedText,
+    mediaUrls: media,
+    mediaUrl: media[0],
+    ...(hasAudioAsVoice ? { audioAsVoice: true } : {}),
+  };
 }
diff --git a/src/telegram/bot.ts b/src/telegram/bot.ts
index aaba18307..80404a185 100644
--- a/src/telegram/bot.ts
+++ b/src/telegram/bot.ts
@@ -1388,6 +1388,7 @@ async function deliverReplies(params: {
         });
       } else if (kind === "audio") {
         const useVoice = reply.audioAsVoice === true; // default false (backward compatible)
+        log.warn(`[DEBUG] Audio media: audioAsVoice=${reply.audioAsVoice}, useVoice=${useVoice}`);
         if (useVoice) {
           // Voice message - displays as round playable bubble (opt-in via [[audio_as_voice]])
           await bot.api.sendVoice(chatId, file, {