feat: add voice memo responses via TTS (#394)

.env.example

@@ -141,6 +141,21 @@ TELEGRAM_BOT_TOKEN=your_telegram_bot_token
 # GMAIL_REFRESH_TOKEN=your_refresh_token
 # GMAIL_TELEGRAM_USER=123456789
 
+# ============================================
+# Voice Memos / TTS (optional)
+# ============================================
+# TTS provider: "elevenlabs" (default) or "openai"
+# TTS_PROVIDER=elevenlabs
+
+# ElevenLabs (default provider)
+# ELEVENLABS_API_KEY=sk_your_elevenlabs_key
+# ELEVENLABS_VOICE_ID=21m00Tcm4TlvDq8ikWAM
+# ELEVENLABS_MODEL_ID=eleven_multilingual_v2
+
+# OpenAI TTS (uses OPENAI_API_KEY from above)
+# OPENAI_TTS_VOICE=alloy
+# OPENAI_TTS_MODEL=tts-1
+
 # ============================================
 # API Server (for Docker/CLI integration)
 # ============================================

docs/cli-tools.md

@@ -11,6 +11,7 @@ Send a message to the most recent chat, or target a specific channel/chat.
 lettabot-message send --text "Hello from a background task"
 lettabot-message send --text "Hello" --channel slack --chat C123456
 lettabot-message send --file /tmp/report.pdf --text "Report attached" --channel discord --chat 123456789
+lettabot-message send --file /tmp/voice.ogg --voice    # Send as native voice note
 ```
 
 ## lettabot-react

docs/configuration.md

@@ -675,6 +675,36 @@ transcription:
   model: whisper-1     # Default
 ```
 
+## Text-to-Speech (TTS) Configuration
+
+Voice memo generation via the `<voice>` directive. The agent can reply with voice notes on Telegram and WhatsApp:
+
+```yaml
+tts:
+  provider: elevenlabs    # "elevenlabs" (default) or "openai"
+  apiKey: sk_475a...      # Provider API key
+  voiceId: 21m00Tcm4TlvDq8ikWAM  # Voice selection (see below)
+  model: eleven_multilingual_v2   # Optional model override
+```
+
+**ElevenLabs** (default):
+- `voiceId` is an ElevenLabs voice ID. Default: `21m00Tcm4TlvDq8ikWAM` (Rachel). Browse voices at [elevenlabs.io/voice-library](https://elevenlabs.io/voice-library).
+- `model` defaults to `eleven_multilingual_v2`.
+
+**OpenAI**:
+- `voiceId` is one of: `alloy`, `echo`, `fable`, `onyx`, `nova`, `shimmer`. Default: `alloy`.
+- `model` defaults to `tts-1`. Use `tts-1-hd` for higher quality.
+
+The agent uses the `<voice>` directive in responses:
+
+```xml
+<actions>
+  <voice>Hey, here's a quick voice reply!</voice>
+</actions>
+```
+
+The `lettabot-tts` CLI tool is also available for background tasks (heartbeats, cron).
+
 ## Attachments Configuration
 
 ```yaml
@@ -807,5 +837,11 @@ Environment variables override config file values:
 | `LOG_LEVEL` | `server.logLevel` (fatal/error/warn/info/debug/trace). Overrides config. |
 | `LETTABOT_LOG_LEVEL` | Alias for `LOG_LEVEL` |
 | `LOG_FORMAT` | Set to `json` for structured JSON output (recommended for Railway/Docker) |
+| `TTS_PROVIDER` | TTS backend: `elevenlabs` (default) or `openai` |
+| `ELEVENLABS_API_KEY` | API key for ElevenLabs TTS |
+| `ELEVENLABS_VOICE_ID` | ElevenLabs voice ID (default: `21m00Tcm4TlvDq8ikWAM` / Rachel) |
+| `ELEVENLABS_MODEL_ID` | ElevenLabs model (default: `eleven_multilingual_v2`) |
+| `OPENAI_TTS_VOICE` | OpenAI TTS voice (default: `alloy`) |
+| `OPENAI_TTS_MODEL` | OpenAI TTS model (default: `tts-1`) |
 
 See [SKILL.md](../SKILL.md) for complete environment variable reference.

docs/directives.md

@@ -48,13 +48,14 @@ Sends a file or image to the same channel/chat as the triggering message.
 ```xml
 <send-file path="/tmp/report.pdf" caption="Report attached" />
 <send-file path="/tmp/photo.png" kind="image" caption="Look!" />
+<send-file path="/tmp/voice.ogg" kind="audio" cleanup="true" />
 <send-file path="/tmp/temp-export.csv" cleanup="true" />
 ```
 
 **Attributes:**
 - `path` / `file` (required) -- Local file path on the LettaBot server
 - `caption` / `text` (optional) -- Caption text for the file
-- `kind` (optional) -- `image` or `file` (defaults to auto-detect based on extension)
+- `kind` (optional) -- `image`, `file`, or `audio` (defaults to auto-detect based on extension). Audio files (.ogg, .opus, .mp3, .m4a, .wav, .aac, .flac) are auto-detected as `audio`.
 - `cleanup` (optional) -- `true` to delete the file after sending (default: false)
 
 **Security:**
@@ -63,6 +64,22 @@ Sends a file or image to the same channel/chat as the triggering message.
 - File size is limited to `sendFileMaxSize` (default: 50MB).
 - The `cleanup` attribute only works when `sendFileCleanup: true` is set in the agent's features config (disabled by default).
 
+### `<voice>`
+
+Generates speech from text via TTS and sends it as a native voice note. No tool calls needed.
+
+```xml
+<voice>Hey, here's a quick voice reply!</voice>
+```
+
+The text content is sent to the configured TTS provider (see [TTS Configuration](./configuration.md#text-to-speech-tts-configuration)), converted to audio, and delivered as a voice note. Audio is automatically cleaned up after sending.
+
+- Requires `tts` to be configured in `lettabot.yaml`
+- Renders as native voice bubbles on Telegram and WhatsApp
+- Discord and Slack receive a playable audio attachment
+- On Telegram, falls back to audio file if voice messages are restricted by Premium privacy settings
+- Can be combined with text: any text after the `</actions>` block is sent as a normal message alongside the voice note
+
 ### `<no-reply/>`
 
 Suppresses response delivery entirely. The agent's text is discarded.
@@ -88,13 +105,13 @@ Backslash-escaped quotes (common when LLMs generate XML inside a JSON context) a
 
 ## Channel Support
 
-| Channel   | `addReaction` | `send-file` | Notes |
+| Channel   | `addReaction` | `send-file` | `kind="audio"` | Notes |
-|-----------|:---:|:---:|-------|
+|-----------|:---:|:---:|:---:|-------|
-| Telegram  | Yes | Yes | Reactions limited to Telegram's [allowed reaction set](https://core.telegram.org/bots/api#reactiontype). |
+| Telegram  | Yes | Yes | Voice note (`sendVoice`) | Falls back to `sendAudio` if voice messages are restricted by Telegram Premium privacy settings. |
-| Slack     | Yes | Yes | Reactions use Slack emoji names (`:thumbsup:` style). |
+| Slack     | Yes | Yes | Audio attachment | Reactions use Slack emoji names (`:thumbsup:` style). |
-| Discord   | Yes | Yes | Custom server emoji not yet supported. |
+| Discord   | Yes | Yes | Audio attachment | Custom server emoji not yet supported. |
-| WhatsApp  | No  | Yes | Reactions skipped with a warning. |
+| WhatsApp  | No  | Yes | Voice note (PTT) | Sent with `ptt: true` for native voice bubble. |
-| Signal    | No  | No  | Directive skipped with a warning. |
+| Signal    | No  | No  | No | Directive skipped with a warning. |
 
 When a channel doesn't implement `addReaction`, the directive is silently skipped and a warning is logged. This never blocks message delivery.
 

skills/voice-memo/SKILL.md

@@ -0,0 +1,58 @@
+---
+name: voice-memo
+description: Reply with voice memos using text-to-speech. Use when the user sends a voice message, asks for an audio reply, or when a voice response would be more natural.
+---
+
+# Voice Memo Responses
+
+Generate voice memos using TTS and send them as native voice notes.
+
+## Usage
+
+Use the `<voice>` directive to send voice memos. No tool calls needed:
+
+```
+<actions>
+  <voice>Hey, here's a quick update on that thing we discussed.</voice>
+</actions>
+```
+
+With accompanying text:
+
+```
+<actions>
+  <voice>Here's the summary as audio.</voice>
+</actions>
+And here it is in text form too!
+```
+
+### Silent mode (heartbeats, cron)
+
+For background tasks that need to send voice without a user message context:
+
+```bash
+OUTPUT=$(lettabot-tts "Your message here") || exit 1
+lettabot-message send --file "$OUTPUT" --voice
+```
+
+## When to Use Voice
+
+- User sent a voice message and a voice reply feels natural
+- User explicitly asks for a voice/audio response
+- Short, conversational responses (voice is awkward for long technical content)
+
+## When NOT to Use Voice
+
+- Code snippets, file paths, URLs, or structured data (these should be text)
+- Long responses -- keep voice memos under ~30 seconds of speech
+- When the user has indicated a preference for text
+- When `ELEVENLABS_API_KEY` is not set
+
+## Notes
+
+- Audio format is OGG Opus, which renders as native voice bubbles on Telegram and WhatsApp
+- Discord and Slack will show it as a playable audio attachment
+- Use `cleanup="true"` to delete the audio file after sending
+- The `data/outbound/` directory is the default allowed path for send-file directives
+- The script uses `$LETTABOT_WORKING_DIR` to output files to the correct directory
+- On Telegram, if the user has voice message privacy enabled (Telegram Premium), the bot falls back to sending as an audio file instead of a voice bubble. Users can allow voice messages via Settings > Privacy and Security > Voice Messages.

skills/voice-memo/lettabot-tts

@@ -0,0 +1,130 @@
+#!/usr/bin/env bash
+# lettabot-tts - Generate speech audio via configurable TTS provider
+#
+# Usage: lettabot-tts <text> [output_path]
+#
+# Environment:
+#   TTS_PROVIDER         - Optional. "elevenlabs" (default) or "openai".
+#
+#   ElevenLabs:
+#     ELEVENLABS_API_KEY   - Required. API key.
+#     ELEVENLABS_VOICE_ID  - Optional. Voice ID (default: 21m00Tcm4TlvDq8ikWAM / Rachel).
+#     ELEVENLABS_MODEL_ID  - Optional. Model ID (default: eleven_multilingual_v2).
+#
+#   OpenAI:
+#     OPENAI_API_KEY       - Required. API key.
+#     OPENAI_TTS_VOICE     - Optional. Voice name (default: alloy).
+#     OPENAI_TTS_MODEL     - Optional. Model (default: tts-1).
+
+set -euo pipefail
+
+TEXT="${1:?Usage: lettabot-tts <text> [output_path]}"
+
+# The session subprocess CWD is set to workingDir (bot.ts:642), which is the
+# same base directory that <send-file> directives resolve from. This means
+# $(pwd) and LETTABOT_WORKING_DIR produce paths in the correct coordinate space.
+OUTBOUND_DIR="${LETTABOT_WORKING_DIR:-$(pwd)}/data/outbound"
+
+PROVIDER="${TTS_PROVIDER:-elevenlabs}"
+
+# Ensure output directory exists
+mkdir -p "$OUTBOUND_DIR"
+
+# Use collision-safe random filenames when output path is not explicitly provided.
+if [ -n "${2:-}" ]; then
+  OUTPUT="$2"
+else
+  # Clean stale voice files older than 1 hour
+  find "$OUTBOUND_DIR" -name 'voice-*.ogg' -mmin +60 -delete 2>/dev/null || true
+  OUTPUT=$(mktemp "${OUTBOUND_DIR}/voice-XXXXXXXXXX.ogg")
+fi
+
+# ---------------------------------------------------------------------------
+# Provider: ElevenLabs
+# ---------------------------------------------------------------------------
+tts_elevenlabs() {
+  if [ -z "${ELEVENLABS_API_KEY:-}" ]; then
+    echo "Error: ELEVENLABS_API_KEY is not set" >&2
+    exit 1
+  fi
+
+  local voice_id="${ELEVENLABS_VOICE_ID:-onwK4e9ZLuTAKqWW03F9}"
+  local model_id="${ELEVENLABS_MODEL_ID:-eleven_multilingual_v2}"
+
+  local http_code
+  http_code=$(curl -s -w "%{http_code}" -o "$OUTPUT" \
+    "https://api.elevenlabs.io/v1/text-to-speech/${voice_id}" \
+    -H "xi-api-key: ${ELEVENLABS_API_KEY}" \
+    -H "Content-Type: application/json" \
+    -d "$(jq -n \
+      --arg text "$TEXT" \
+      --arg model "$model_id" \
+      '{
+        text: $text,
+        model_id: $model,
+        output_format: "ogg_opus"
+      }'
+    )")
+
+  if [ "$http_code" -lt 200 ] || [ "$http_code" -ge 300 ]; then
+    echo "Error: ElevenLabs API returned HTTP $http_code" >&2
+    if file "$OUTPUT" | grep -q "text\|JSON\|ASCII"; then
+      cat "$OUTPUT" >&2
+    fi
+    rm -f "$OUTPUT"
+    exit 1
+  fi
+}
+
+# ---------------------------------------------------------------------------
+# Provider: OpenAI
+# ---------------------------------------------------------------------------
+tts_openai() {
+  if [ -z "${OPENAI_API_KEY:-}" ]; then
+    echo "Error: OPENAI_API_KEY is not set" >&2
+    exit 1
+  fi
+
+  local voice="${OPENAI_TTS_VOICE:-alloy}"
+  local model="${OPENAI_TTS_MODEL:-tts-1}"
+
+  local http_code
+  http_code=$(curl -s -w "%{http_code}" -o "$OUTPUT" \
+    "https://api.openai.com/v1/audio/speech" \
+    -H "Authorization: Bearer ${OPENAI_API_KEY}" \
+    -H "Content-Type: application/json" \
+    -d "$(jq -n \
+      --arg text "$TEXT" \
+      --arg model "$model" \
+      --arg voice "$voice" \
+      '{
+        model: $model,
+        input: $text,
+        voice: $voice,
+        response_format: "opus"
+      }'
+    )")
+
+  if [ "$http_code" -lt 200 ] || [ "$http_code" -ge 300 ]; then
+    echo "Error: OpenAI TTS API returned HTTP $http_code" >&2
+    if file "$OUTPUT" | grep -q "text\|JSON\|ASCII"; then
+      cat "$OUTPUT" >&2
+    fi
+    rm -f "$OUTPUT"
+    exit 1
+  fi
+}
+
+# ---------------------------------------------------------------------------
+# Dispatch
+# ---------------------------------------------------------------------------
+case "$PROVIDER" in
+  elevenlabs) tts_elevenlabs ;;
+  openai)     tts_openai ;;
+  *)
+    echo "Error: Unknown TTS_PROVIDER: $PROVIDER (supported: elevenlabs, openai)" >&2
+    exit 1
+    ;;
+esac
+
+echo "$OUTPUT"

src/api/server.ts

@@ -104,7 +104,7 @@ export function createApiServer(deliverer: AgentRouter, options: ServerOptions):
           {
             text: fields.text,
             filePath: file?.tempPath,
-            kind: fields.kind as 'image' | 'file' | undefined,
+            kind: fields.kind as 'image' | 'file' | 'audio' | undefined,
           }
         );
 

src/api/types.ts

@@ -23,7 +23,7 @@ export interface SendFileRequest {
   chatId: string;
   filePath: string;  // Temporary file path on server
   caption?: string;
-  kind?: 'image' | 'file';
+  kind?: 'image' | 'file' | 'audio';
   threadId?: string;
 }
 

src/channels/telegram.ts

@@ -565,6 +565,21 @@ export class TelegramAdapter implements ChannelAdapter {
       return { messageId: String(result.message_id) };
     }
 
+    if (file.kind === 'audio') {
+      try {
+        const result = await this.bot.api.sendVoice(file.chatId, input, { caption });
+        return { messageId: String(result.message_id) };
+      } catch (err: any) {
+        // Fall back to sendAudio if voice messages are restricted (Telegram Premium privacy setting)
+        if (err?.description?.includes('VOICE_MESSAGES_FORBIDDEN')) {
+          log.warn('sendVoice forbidden, falling back to sendAudio');
+          const result = await this.bot.api.sendAudio(file.chatId, new InputFile(file.filePath), { caption });
+          return { messageId: String(result.message_id) };
+        }
+        throw err;
+      }
+    }
+
     const result = await this.bot.api.sendDocument(file.chatId, input, { caption });
     return { messageId: String(result.message_id) };
   }

src/channels/whatsapp/outbound.test.ts

@@ -0,0 +1,36 @@
+import { describe, it, expect, vi } from 'vitest';
+import { sendWhatsAppFile, type LidMapper } from './outbound.js';
+
+describe('sendWhatsAppFile', () => {
+  it('sends audio as native voice note payload', async () => {
+    const sock = {
+      sendMessage: vi.fn(async () => ({ key: { id: '' } })),
+    } as any;
+
+    const lidMapper: LidMapper = {
+      selfChatLid: '',
+      myNumber: '',
+      lidToJid: new Map(),
+    };
+
+    await sendWhatsAppFile(
+      sock,
+      {
+        chatId: '12345@s.whatsapp.net',
+        filePath: '/tmp/voice.ogg',
+        caption: 'hello',
+        kind: 'audio',
+      },
+      lidMapper,
+      new Set<string>(),
+    );
+
+    expect(sock.sendMessage).toHaveBeenCalledWith(
+      '12345@s.whatsapp.net',
+      {
+        audio: { url: '/tmp/voice.ogg' },
+        ptt: true,
+      },
+    );
+  });
+});

src/channels/whatsapp/outbound.ts

@@ -244,10 +244,14 @@ export async function sendWhatsAppFile(
   const caption = file.caption || undefined;
   const fileName = basename(file.filePath);
 
-  const payload =
+  let payload;
-    file.kind === "image"
+  if (file.kind === "image") {
-      ? { image: { url: file.filePath }, caption }
+    payload = { image: { url: file.filePath }, caption };
-      : { document: { url: file.filePath }, mimetype: "application/octet-stream", caption, fileName };
+  } else if (file.kind === "audio") {
+    payload = { audio: { url: file.filePath }, ptt: true };
+  } else {
+    payload = { document: { url: file.filePath }, mimetype: "application/octet-stream", caption, fileName };
+  }
 
   try {
     // Send file

src/cli/message.ts

@@ -147,7 +147,7 @@ async function sendViaApi(
   options: {
     text?: string;
     filePath?: string;
-    kind?: 'image' | 'file';
+    kind?: 'image' | 'file' | 'audio';
   }
 ): Promise<void> {
   const apiUrl = process.env.LETTABOT_API_URL || 'http://localhost:8080';
@@ -249,7 +249,7 @@ async function sendToChannel(channel: string, chatId: string, text: string): Pro
 async function sendCommand(args: string[]): Promise<void> {
   let text = '';
   let filePath = '';
-  let kind: 'image' | 'file' | undefined = undefined;
+  let kind: 'image' | 'file' | 'audio' | undefined = undefined;
   let channel = '';
   let chatId = '';
   const fileCapableChannels = new Set(['telegram', 'slack', 'discord', 'whatsapp']);
@@ -267,6 +267,8 @@ async function sendCommand(args: string[]): Promise<void> {
       i++;
     } else if (arg === '--image') {
       kind = 'image';
+    } else if (arg === '--voice') {
+      kind = 'audio';
     } else if ((arg === '--channel' || arg === '-c' || arg === '-C') && next) {
       channel = next;
       i++;
@@ -332,6 +334,7 @@ Send options:
   --text, -t <text>       Message text (or caption when used with --file)
   --file, -f <path>       File path (optional, for file messages)
   --image                 Treat file as image (vs document)
+  --voice                 Treat file as voice note (sends as native voice memo)
   --channel, -c <name>    Channel: telegram, slack, whatsapp, discord (default: last used)
   --chat, --to <id>       Chat/conversation ID (default: last messaged)
 
@@ -348,6 +351,9 @@ Examples:
   # Send to specific WhatsApp chat
   lettabot-message send --file report.pdf --text "Report attached" --channel whatsapp --chat "+1555@s.whatsapp.net"
 
+  # Send voice note
+  lettabot-message send --file voice.ogg --voice
+
   # Short form
   lettabot-message send -t "Done!" -f doc.pdf -c telegram
 

src/config/io.ts

@@ -352,6 +352,36 @@ export function configToEnv(config: LettaBotConfig): Record<string, string> {
     env.ATTACHMENTS_MAX_AGE_DAYS = String(config.attachments.maxAgeDays);
   }
 
+  // TTS (text-to-speech for voice memos)
+  if (config.tts?.provider) {
+    env.TTS_PROVIDER = config.tts.provider;
+  }
+  if (config.tts?.apiKey) {
+    // Set the provider-specific key based on provider
+    const provider = config.tts.provider || 'elevenlabs';
+    if (provider === 'elevenlabs') {
+      env.ELEVENLABS_API_KEY = config.tts.apiKey;
+    } else if (provider === 'openai') {
+      env.OPENAI_API_KEY = config.tts.apiKey;
+    }
+  }
+  if (config.tts?.voiceId) {
+    const provider = config.tts.provider || 'elevenlabs';
+    if (provider === 'elevenlabs') {
+      env.ELEVENLABS_VOICE_ID = config.tts.voiceId;
+    } else if (provider === 'openai') {
+      env.OPENAI_TTS_VOICE = config.tts.voiceId;
+    }
+  }
+  if (config.tts?.model) {
+    const provider = config.tts.provider || 'elevenlabs';
+    if (provider === 'elevenlabs') {
+      env.ELEVENLABS_MODEL_ID = config.tts.model;
+    } else if (provider === 'openai') {
+      env.OPENAI_TTS_MODEL = config.tts.model;
+    }
+  }
+
   // API server (server.api is canonical, top-level api is deprecated fallback)
   const apiConfig = config.server.api ?? config.api;
   if (apiConfig?.port !== undefined) {

src/config/types.ts

@@ -177,9 +177,12 @@ export interface LettaBotConfig {
     google?: GoogleConfig;
   };
 
-  // Transcription (voice messages)
+  // Transcription (inbound voice messages)
   transcription?: TranscriptionConfig;
 
+  // Text-to-speech (outbound voice memos)
+  tts?: TtsConfig;
+
   // Attachment handling
   attachments?: {
     maxMB?: number;
@@ -195,6 +198,13 @@ export interface LettaBotConfig {
   };
 }
 
+export interface TtsConfig {
+  provider?: 'elevenlabs' | 'openai';  // Default: 'elevenlabs'
+  apiKey?: string;                      // Falls back to ELEVENLABS_API_KEY or OPENAI_API_KEY env var
+  voiceId?: string;                     // ElevenLabs voice ID or OpenAI voice name
+  model?: string;                       // Model ID (provider-specific defaults)
+}
+
 export interface TranscriptionConfig {
   provider: 'openai' | 'mistral';
   apiKey?: string;     // Falls back to OPENAI_API_KEY or MISTRAL_API_KEY env var

src/core/bot.ts

@@ -5,15 +5,16 @@
  */
 
 import { createAgent, createSession, resumeSession, imageFromFile, imageFromURL, type Session, type MessageContentItem, type SendMessage, type CanUseToolCallback } from '@letta-ai/letta-code-sdk';
-import { mkdirSync } from 'node:fs';
+import { mkdirSync, existsSync } from 'node:fs';
 import { access, unlink, realpath, stat, constants } from 'node:fs/promises';
+import { execFile } from 'node:child_process';
 import { extname, resolve, join } from 'node:path';
 import type { ChannelAdapter } from '../channels/types.js';
 import type { BotConfig, InboundMessage, TriggerContext } from './types.js';
 import type { AgentSession } from './interfaces.js';
 import { Store } from './store.js';
 import { updateAgentName, getPendingApprovals, rejectApproval, cancelRuns, recoverOrphanedConversationApproval, getLatestRunError } from '../tools/letta-api.js';
-import { installSkillsToAgent } from '../skills/loader.js';
+import { installSkillsToAgent, withAgentSkillsOnPath, getAgentSkillExecutableDirs, isVoiceMemoConfigured } from '../skills/loader.js';
 import { formatMessageEnvelope, formatGroupBatchEnvelope, type SessionContextOptions } from './formatter.js';
 import type { GroupBatcher } from './group-batcher.js';
 import { loadMemoryBlocks } from './memory.js';
@@ -117,10 +118,16 @@ const IMAGE_FILE_EXTENSIONS = new Set([
   '.png', '.jpg', '.jpeg', '.gif', '.webp', '.bmp', '.tiff',
 ]);
 
-/** Infer whether a file is an image or generic file based on extension. */
+const AUDIO_FILE_EXTENSIONS = new Set([
-export function inferFileKind(filePath: string): 'image' | 'file' {
+  '.ogg', '.opus', '.mp3', '.m4a', '.wav', '.aac', '.flac',
+]);
+
+/** Infer whether a file is an image, audio, or generic file based on extension. */
+export function inferFileKind(filePath: string): 'image' | 'file' | 'audio' {
   const ext = extname(filePath).toLowerCase();
-  return IMAGE_FILE_EXTENSIONS.has(ext) ? 'image' : 'file';
+  if (IMAGE_FILE_EXTENSIONS.has(ext)) return 'image';
+  if (AUDIO_FILE_EXTENSIONS.has(ext)) return 'audio';
+  return 'file';
 }
 
 /**
@@ -776,6 +783,59 @@ export class LettaBot implements AgentSession {
           console.warn('[Bot] Directive send-file failed:', err instanceof Error ? err.message : err);
         }
       }
+
+      if (directive.type === 'voice') {
+        if (!isVoiceMemoConfigured()) {
+          log.warn('Directive voice skipped: no TTS credentials configured');
+          continue;
+        }
+        if (typeof adapter.sendFile !== 'function') {
+          log.warn(`Directive voice skipped: ${adapter.name} does not support sendFile`);
+          continue;
+        }
+
+        // Find lettabot-tts in agent's skill dirs
+        const agentId = this.store.agentId;
+        const skillDirs = agentId ? getAgentSkillExecutableDirs(agentId) : [];
+        const ttsPath = skillDirs
+          .map(dir => join(dir, 'lettabot-tts'))
+          .find(p => existsSync(p));
+
+        if (!ttsPath) {
+          log.warn('Directive voice skipped: lettabot-tts not found in skill dirs');
+          continue;
+        }
+
+        try {
+          const outputPath = await new Promise<string>((resolve, reject) => {
+            execFile(ttsPath, [directive.text], {
+              cwd: this.config.workingDir,
+              env: { ...process.env, LETTABOT_WORKING_DIR: this.config.workingDir },
+              timeout: 30_000,
+            }, (err, stdout, stderr) => {
+              if (err) {
+                reject(new Error(stderr?.trim() || err.message));
+              } else {
+                resolve(stdout.trim());
+              }
+            });
+          });
+
+          await adapter.sendFile({
+            chatId,
+            filePath: outputPath,
+            kind: 'audio',
+            threadId,
+          });
+          acted = true;
+          log.info(`Directive: sent voice memo (${directive.text.length} chars)`);
+
+          // Clean up generated file
+          try { await unlink(outputPath); } catch {}
+        } catch (err) {
+          log.warn('Directive voice failed:', err instanceof Error ? err.message : err);
+        }
+      }
     }
     return acted;
   }
@@ -864,6 +924,7 @@ export class LettaBot implements AgentSession {
 
     const opts = this.baseSessionOptions(this.sessionCanUseTool);
     let session: Session;
+    let sessionAgentId: string | undefined;
 
     // In per-channel mode, look up per-key conversation ID.
     // In shared mode (key === "shared"), use the legacy single conversationId.
@@ -873,9 +934,15 @@ export class LettaBot implements AgentSession {
 
     if (convId) {
       process.env.LETTA_AGENT_ID = this.store.agentId || undefined;
+      if (this.store.agentId) {
+        installSkillsToAgent(this.store.agentId, this.config.skills);
+        sessionAgentId = this.store.agentId;
+      }
       session = resumeSession(convId, opts);
     } else if (this.store.agentId) {
       process.env.LETTA_AGENT_ID = this.store.agentId;
+      installSkillsToAgent(this.store.agentId, this.config.skills);
+      sessionAgentId = this.store.agentId;
       session = createSession(this.store.agentId, opts);
     } else {
       // Create new agent -- persist immediately so we don't orphan it on later failures
@@ -893,6 +960,7 @@ export class LettaBot implements AgentSession {
         updateAgentName(newAgentId, this.config.agentName).catch(() => {});
       }
       installSkillsToAgent(newAgentId, this.config.skills);
+      sessionAgentId = newAgentId;
 
       session = createSession(newAgentId, opts);
     }
@@ -900,7 +968,14 @@ export class LettaBot implements AgentSession {
     // Initialize eagerly so the subprocess is ready before the first send()
     log.info(`Initializing session subprocess (key=${key})...`);
     try {
-      await this.withSessionTimeout(session.initialize(), `Session initialize (key=${key})`);
+      if (sessionAgentId) {
+        await withAgentSkillsOnPath(
+          sessionAgentId,
+          () => this.withSessionTimeout(session.initialize(), `Session initialize (key=${key})`),
+        );
+      } else {
+        await this.withSessionTimeout(session.initialize(), `Session initialize (key=${key})`);
+      }
       log.info(`Session subprocess ready (key=${key})`);
     } catch (error) {
       // Close immediately so failed initialization cannot leak a subprocess.
@@ -1680,7 +1755,11 @@ export class LettaBot implements AgentSession {
           msgTypeCounts[streamMsg.type] = (msgTypeCounts[streamMsg.type] || 0) + 1;
           
           const preview = JSON.stringify(streamMsg).slice(0, 300);
-          log.info(`type=${streamMsg.type} ${preview}`);
+          if (streamMsg.type === 'reasoning' || streamMsg.type === 'assistant') {
+            log.debug(`type=${streamMsg.type} ${preview}`);
+          } else {
+            log.info(`type=${streamMsg.type} ${preview}`);
+          }
           
           // stream_event is a low-level streaming primitive (partial deltas), not a
           // semantic type change. Skip it for type-transition logic so it doesn't
@@ -1694,6 +1773,7 @@ export class LettaBot implements AgentSession {
 
           // Flush reasoning buffer when type changes away from reasoning
           if (isSemanticType && lastMsgType === 'reasoning' && streamMsg.type !== 'reasoning' && reasoningBuffer.trim()) {
+            log.info(`Reasoning: ${reasoningBuffer.trim()}`);
             if (this.config.display?.showReasoning && !suppressDelivery) {
               try {
                 const reasoning = this.formatReasoningDisplay(reasoningBuffer, adapter.id);
@@ -2180,7 +2260,7 @@ export class LettaBot implements AgentSession {
     options: {
       text?: string;
       filePath?: string;
-      kind?: 'image' | 'file';
+      kind?: 'image' | 'file' | 'audio';
     }
   ): Promise<string | undefined> {
     const adapter = this.channels.get(channelId);

src/core/directives.test.ts

@@ -77,6 +77,14 @@ describe('parseDirectives', () => {
     ]);
   });
 
+  it('parses send-file directive with audio kind', () => {
+    const result = parseDirectives('<actions><send-file path="voice.ogg" kind="audio" /></actions>');
+    expect(result.cleanText).toBe('');
+    expect(result.directives).toEqual([
+      { type: 'send-file', path: 'voice.ogg', kind: 'audio' },
+    ]);
+  });
+
   it('parses send-file directive with cleanup attribute', () => {
     const result = parseDirectives('<actions><send-file path="/tmp/report.pdf" cleanup="true" /></actions>');
     expect(result.cleanText).toBe('');
@@ -149,6 +157,57 @@ describe('parseDirectives', () => {
     expect(result.cleanText).toBe('');
     expect(result.directives).toEqual([]);
   });
+
+  it('parses voice directive with text content', () => {
+    const result = parseDirectives('<actions><voice>Hello from a voice memo</voice></actions>');
+    expect(result.cleanText).toBe('');
+    expect(result.directives).toEqual([{ type: 'voice', text: 'Hello from a voice memo' }]);
+  });
+
+  it('parses voice directive with text after actions block', () => {
+    const result = parseDirectives('<actions><voice>Here is a voice note</voice></actions>\nHere\'s the audio!');
+    expect(result.cleanText).toBe("Here's the audio!");
+    expect(result.directives).toEqual([{ type: 'voice', text: 'Here is a voice note' }]);
+  });
+
+  it('parses voice directive with multiline text', () => {
+    const result = parseDirectives('<actions><voice>Line one.\nLine two.</voice></actions>');
+    expect(result.cleanText).toBe('');
+    expect(result.directives).toEqual([{ type: 'voice', text: 'Line one.\nLine two.' }]);
+  });
+
+  it('ignores empty voice directive', () => {
+    const result = parseDirectives('<actions><voice>  </voice></actions>');
+    expect(result.cleanText).toBe('');
+    expect(result.directives).toEqual([]);
+  });
+
+  it('parses voice and react directives together', () => {
+    const result = parseDirectives('<actions><react emoji="🎤" /><voice>Check this out</voice></actions>');
+    expect(result.directives).toHaveLength(2);
+    expect(result.directives[0]).toEqual({ type: 'react', emoji: '🎤' });
+    expect(result.directives[1]).toEqual({ type: 'voice', text: 'Check this out' });
+  });
+
+  it('preserves order when voice appears before react', () => {
+    const result = parseDirectives('<actions><voice>First</voice><react emoji="🎤" /></actions>');
+    expect(result.directives).toEqual([
+      { type: 'voice', text: 'First' },
+      { type: 'react', emoji: '🎤' },
+    ]);
+  });
+
+  it('preserves mixed directive order across voice and self-closing tags', () => {
+    const result = parseDirectives(
+      '<actions><send-file path="a.pdf" /><voice>One</voice><react emoji="👍" /><voice>Two</voice></actions>',
+    );
+    expect(result.directives).toEqual([
+      { type: 'send-file', path: 'a.pdf' },
+      { type: 'voice', text: 'One' },
+      { type: 'react', emoji: '👍' },
+      { type: 'voice', text: 'Two' },
+    ]);
+  });
 });
 
 describe('stripActionsBlock', () => {

src/core/directives.ts

@@ -26,12 +26,17 @@ export interface SendFileDirective {
   type: 'send-file';
   path: string;
   caption?: string;
-  kind?: 'image' | 'file';
+  kind?: 'image' | 'file' | 'audio';
   cleanup?: boolean;
 }
 
+export interface VoiceDirective {
+  type: 'voice';
+  text: string;
+}
+
 // Union type — extend with more directive types later
-export type Directive = ReactDirective | SendFileDirective;
+export type Directive = ReactDirective | SendFileDirective | VoiceDirective;
 
 export interface ParseResult {
   cleanText: string;
@@ -45,10 +50,11 @@ export interface ParseResult {
 const ACTIONS_BLOCK_REGEX = /^\s*<actions>([\s\S]*?)<\/actions>/;
 
 /**
- * Match self-closing child directive tags inside the actions block.
+ * Match supported directive tags inside the actions block in source order.
- * Captures the tag name and the full attributes string.
+ * - Self-closing: <react ... />, <send-file ... />
+ * - Content-bearing: <voice>...</voice>
  */
-const CHILD_DIRECTIVE_REGEX = /<(react|send-file)\b([^>]*)\/>/g;
+const DIRECTIVE_TOKEN_REGEX = /<(react|send-file)\b([^>]*)\/>|<voice>([\s\S]*?)<\/voice>/g;
 
 /**
  * Parse a single attribute string like: emoji="eyes" message="123"
@@ -73,13 +79,21 @@ function parseChildDirectives(block: string): Directive[] {
   const normalizedBlock = block.replace(/\\(['"])/g, '$1');
 
   // Reset regex state (global flag)
-  CHILD_DIRECTIVE_REGEX.lastIndex = 0;
+  DIRECTIVE_TOKEN_REGEX.lastIndex = 0;
 
-  while ((match = CHILD_DIRECTIVE_REGEX.exec(normalizedBlock)) !== null) {
+  while ((match = DIRECTIVE_TOKEN_REGEX.exec(normalizedBlock)) !== null) {
-    const [, tagName, attrString] = match;
+    const [, tagName, attrString, voiceText] = match;
+
+    if (voiceText !== undefined) {
+      const text = voiceText.trim();
+      if (text) {
+        directives.push({ type: 'voice', text });
+      }
+      continue;
+    }
 
     if (tagName === 'react') {
-      const attrs = parseAttributes(attrString);
+      const attrs = parseAttributes(attrString || '');
       if (attrs.emoji) {
         directives.push({
           type: 'react',
@@ -91,11 +105,11 @@ function parseChildDirectives(block: string): Directive[] {
     }
 
     if (tagName === 'send-file') {
-      const attrs = parseAttributes(attrString);
+      const attrs = parseAttributes(attrString || '');
       const path = attrs.path || attrs.file;
       if (!path) continue;
       const caption = attrs.caption || attrs.text;
-      const kind = attrs.kind === 'image' || attrs.kind === 'file'
+      const kind = attrs.kind === 'image' || attrs.kind === 'file' || attrs.kind === 'audio'
         ? attrs.kind
         : undefined;
       const cleanup = attrs.cleanup === 'true';

src/core/formatter.ts

@@ -356,6 +356,7 @@ export function formatMessageEnvelope(
     `- To skip replying: \`<no-reply/>\``,
     `- To perform actions: wrap in \`<actions>\` at the start of your response`,
     `  Example: \`<actions><react emoji="thumbsup" /></actions>Your text here\``,
+    `- To send a voice memo: \`<actions><voice>Your message here</voice></actions>\``,
   ];
   sections.push(`## Response Directives\n${directiveLines.join('\n')}`);
 

src/core/gateway.ts

@@ -114,7 +114,7 @@ export class LettaGateway implements AgentRouter {
   async deliverToChannel(
     channelId: string,
     chatId: string,
-    options: { text?: string; filePath?: string; kind?: 'image' | 'file' }
+    options: { text?: string; filePath?: string; kind?: 'image' | 'file' | 'audio' }
   ): Promise<string | undefined> {
     // Try each agent until one owns the channel
     for (const [name, session] of this.agents) {

src/core/interfaces.ts

@@ -37,7 +37,7 @@ export interface AgentSession {
   deliverToChannel(channelId: string, chatId: string, options: {
     text?: string;
     filePath?: string;
-    kind?: 'image' | 'file';
+    kind?: 'image' | 'file' | 'audio';
   }): Promise<string | undefined>;
 
   /** Get agent status */
@@ -67,7 +67,7 @@ export interface MessageDeliverer {
   deliverToChannel(channelId: string, chatId: string, options: {
     text?: string;
     filePath?: string;
-    kind?: 'image' | 'file';
+    kind?: 'image' | 'file' | 'audio';
   }): Promise<string | undefined>;
 }
 

src/core/send-file.test.ts

@@ -23,10 +23,21 @@ describe('inferFileKind', () => {
     expect(inferFileKind('/tmp/script.ts')).toBe('file');
   });
 
+  it('returns audio for common audio extensions', () => {
+    expect(inferFileKind('/tmp/voice.ogg')).toBe('audio');
+    expect(inferFileKind('/tmp/voice.opus')).toBe('audio');
+    expect(inferFileKind('/tmp/voice.mp3')).toBe('audio');
+    expect(inferFileKind('/tmp/voice.m4a')).toBe('audio');
+    expect(inferFileKind('/tmp/voice.wav')).toBe('audio');
+    expect(inferFileKind('/tmp/voice.aac')).toBe('audio');
+    expect(inferFileKind('/tmp/voice.flac')).toBe('audio');
+  });
+
   it('is case insensitive', () => {
     expect(inferFileKind('/tmp/PHOTO.PNG')).toBe('image');
     expect(inferFileKind('/tmp/photo.JPG')).toBe('image');
     expect(inferFileKind('/tmp/photo.Jpeg')).toBe('image');
+    expect(inferFileKind('/tmp/VOICE.OGG')).toBe('audio');
   });
 
   it('returns file for extensionless paths', () => {

src/core/system-prompt.ts

@@ -36,6 +36,9 @@ lettabot-message send --file /path/to/image.jpg --text "Check this out!"
 # Send file without text (treated as image)
 lettabot-message send --file photo.png --image
 
+# Send voice note
+lettabot-message send --file voice.ogg --voice
+
 # Send to specific channel and chat
 lettabot-message send --text "Hello!" --channel telegram --chat 123456789
 
@@ -104,6 +107,8 @@ This sends "Great idea!" and reacts with thumbsup.
 - \`<react emoji="👀" />\` -- react to the message you are responding to. Use the actual emoji character (👀, 👍, ❤️, 🔥, 🎉, 👏).
 - \`<react emoji="🔥" message="123" />\` -- react to a specific message by ID.
 - \`<send-file path="/path/to/file.png" kind="image" caption="..." />\` -- send a file or image to the same channel/chat. File paths are restricted to the configured send-file directory (default: \`data/outbound/\` in the working directory). Paths outside this directory are blocked.
+- \`<send-file path="/path/to/voice.ogg" kind="audio" cleanup="true" />\` -- send a voice note. Audio files (.ogg, .mp3, etc.) are sent as native voice memos on Telegram and WhatsApp. Use \`cleanup="true"\` to delete the file after sending.
+- \`<voice>Your message here</voice>\` -- generate and send a voice memo. The text is converted to speech via TTS and sent as a native voice note. No tool calls needed. Use for short conversational replies, responding to voice messages, or when the user asks for audio.
 
 ### Actions-only response
 

src/core/types.ts

@@ -108,7 +108,7 @@ export interface OutboundFile {
   filePath: string;
   caption?: string;
   threadId?: string;
-  kind?: 'image' | 'file';
+  kind?: 'image' | 'file' | 'audio';
 }
 
 /**
@@ -117,6 +117,7 @@ export interface OutboundFile {
 export interface SkillsConfig {
   cronEnabled?: boolean;
   googleEnabled?: boolean;
+  ttsEnabled?: boolean;
   additionalSkills?: string[];
 }
 

src/main.ts

@@ -178,6 +178,7 @@ import { CronService } from './cron/service.js';
 import { HeartbeatService } from './cron/heartbeat.js';
 import { PollingService, parseGmailAccounts } from './polling/service.js';
 import { agentExists, findAgentByName, ensureNoToolApprovals } from './tools/letta-api.js';
+import { isVoiceMemoConfigured } from './skills/loader.js';
 // Skills are now installed to agent-scoped location after agent creation (see bot.ts)
 
 // Check if config exists (skip in Railway/Docker where env vars are used directly)
@@ -523,6 +524,7 @@ async function main() {
   }
   log.info(`Data directory: ${dataDir}`);
   log.info(`Working directory: ${globalConfig.workingDir}`);
+  process.env.LETTABOT_WORKING_DIR = globalConfig.workingDir;
   
   // Normalize config to agents array
   const agents = normalizeAgents(yamlConfig);
@@ -551,6 +553,7 @@ async function main() {
   }
   
   const gateway = new LettaGateway();
+  const voiceMemoEnabled = isVoiceMemoConfigured();
   const services: { 
     cronServices: CronService[], 
     heartbeatServices: HeartbeatService[], 
@@ -590,6 +593,7 @@ async function main() {
       skills: {
         cronEnabled: agentConfig.features?.cron ?? globalConfig.cronEnabled,
         googleEnabled: !!agentConfig.integrations?.google?.enabled || !!agentConfig.polling?.gmail?.enabled,
+        ttsEnabled: voiceMemoEnabled,
       },
     });
     
@@ -771,7 +775,9 @@ async function main() {
       },
     };
   });
-  printStartupBanner(bannerAgents);
+  if (!process.env.LETTABOT_NO_BANNER) {
+    printStartupBanner(bannerAgents);
+  }
   
   // Shutdown
   const shutdown = async () => {

src/onboard.ts

@@ -1528,7 +1528,7 @@ export async function onboard(options?: { nonInteractive?: boolean }): Promise<v
   if (config.agentChoice === 'new' && !config.agentId) {
     const { createAgent } = await import('@letta-ai/letta-code-sdk');
     const { updateAgentName, ensureNoToolApprovals } = await import('./tools/letta-api.js');
-    const { installSkillsToAgent } = await import('./skills/loader.js');
+    const { installSkillsToAgent, isVoiceMemoConfigured } = await import('./skills/loader.js');
     const { loadMemoryBlocks } = await import('./core/memory.js');
     const { SYSTEM_PROMPT } = await import('./core/system-prompt.js');
     
@@ -1545,9 +1545,11 @@ export async function onboard(options?: { nonInteractive?: boolean }): Promise<v
       if (config.agentName) {
         await updateAgentName(agentId, config.agentName).catch(() => {});
       }
+      const ttsEnv = { ...process.env, ...env };
       installSkillsToAgent(agentId, {
         cronEnabled: config.cron,
         googleEnabled: config.google.enabled,
+        ttsEnabled: isVoiceMemoConfigured(ttsEnv),
       });
       
       // Disable tool approvals

src/skills/loader.test.ts

@@ -2,13 +2,14 @@
  * Skills Loader Tests
  */
 
-import { describe, it, expect, beforeEach, afterEach } from 'vitest';
+import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest';
 import { mkdtempSync, rmSync, mkdirSync, writeFileSync, existsSync, readdirSync } from 'node:fs';
 import { tmpdir } from 'node:os';
 import { join } from 'node:path';
 import {
   getAgentSkillsDir,
   FEATURE_SKILLS,
+  isVoiceMemoConfigured,
 } from './loader.js';
 
 describe('skills loader', () => {
@@ -52,6 +53,23 @@ describe('skills loader', () => {
       expect(FEATURE_SKILLS.google).toContain('gog');
       expect(FEATURE_SKILLS.google).toContain('google');
     });
+
+    it('has tts feature with voice-memo skill', () => {
+      expect(FEATURE_SKILLS.tts).toBeDefined();
+      expect(FEATURE_SKILLS.tts).toContain('voice-memo');
+    });
+  });
+
+  describe('isVoiceMemoConfigured', () => {
+    it('defaults to elevenlabs and requires ELEVENLABS_API_KEY', () => {
+      expect(isVoiceMemoConfigured({})).toBe(false);
+      expect(isVoiceMemoConfigured({ ELEVENLABS_API_KEY: 'test' })).toBe(true);
+    });
+
+    it('supports openai provider and requires OPENAI_API_KEY', () => {
+      expect(isVoiceMemoConfigured({ TTS_PROVIDER: 'openai' })).toBe(false);
+      expect(isVoiceMemoConfigured({ TTS_PROVIDER: 'openai', OPENAI_API_KEY: 'test' })).toBe(true);
+    });
   });
 
   describe('installSkillsToAgent', () => {
@@ -145,4 +163,53 @@ describe('skills loader', () => {
       expect(content).toBe('target version');
     });
   });
+
+  describe('loadAllSkills precedence', () => {
+    it('prefers global skills over bundled skills for the same name', async () => {
+      const originalHome = process.env.HOME;
+      const originalUserProfile = process.env.USERPROFILE;
+      const originalCwd = process.cwd();
+      const tempHome = mkdtempSync(join(tmpdir(), 'lettabot-home-test-'));
+      const tempProject = mkdtempSync(join(tmpdir(), 'lettabot-project-test-'));
+
+      try {
+        process.env.HOME = tempHome;
+        process.env.USERPROFILE = tempHome;
+        process.chdir(tempProject);
+
+        const globalVoiceMemoDir = join(tempHome, '.letta', 'skills', 'voice-memo');
+        mkdirSync(globalVoiceMemoDir, { recursive: true });
+        writeFileSync(
+          join(globalVoiceMemoDir, 'SKILL.md'),
+          [
+            '---',
+            'name: voice-memo',
+            'description: global override',
+            '---',
+            '',
+            '# Global override',
+            '',
+          ].join('\n'),
+        );
+
+        vi.resetModules();
+        const mod = await import('./loader.js');
+        const skills = mod.loadAllSkills();
+        const voiceMemo = skills.find((skill: any) => skill.name === 'voice-memo');
+        const expectedPath = join(tempHome, '.letta', 'skills', 'voice-memo', 'SKILL.md');
+
+        expect(voiceMemo).toBeDefined();
+        expect(voiceMemo!.description).toBe('global override');
+        expect(voiceMemo!.filePath).toContain(expectedPath);
+      } finally {
+        process.chdir(originalCwd);
+        if (originalHome === undefined) delete process.env.HOME;
+        else process.env.HOME = originalHome;
+        if (originalUserProfile === undefined) delete process.env.USERPROFILE;
+        else process.env.USERPROFILE = originalUserProfile;
+        rmSync(tempHome, { recursive: true, force: true });
+        rmSync(tempProject, { recursive: true, force: true });
+      }
+    });
+  });
 });

src/skills/loader.ts

@@ -4,7 +4,7 @@
 
 import { existsSync, readdirSync, readFileSync, mkdirSync, cpSync } from 'node:fs';
 import { execSync } from 'node:child_process';
-import { join, resolve } from 'node:path';
+import { join, resolve, delimiter } from 'node:path';
 import matter from 'gray-matter';
 import type { SkillEntry, ClawdbotMetadata } from './types.js';
 
@@ -30,6 +30,91 @@ export function getAgentSkillsDir(agentId: string): string {
   return join(HOME, '.letta', 'agents', agentId, 'skills');
 }
 
+/**
+ * Resolve subdirectories that contain executable skill files.
+ */
+function resolveSkillExecutableDirs(skillsDir: string): string[] {
+  // Only add dirs that contain at least one executable (non-.md) file
+  return readdirSync(skillsDir, { withFileTypes: true })
+    .filter(d => d.isDirectory())
+    .map(d => join(skillsDir, d.name))
+    .filter(dir => {
+      try {
+        return readdirSync(dir).some(f => !f.endsWith('.md'));
+      } catch { return false; }
+    });
+}
+
+/**
+ * Get executable skill directories for a specific agent.
+ */
+export function getAgentSkillExecutableDirs(agentId: string): string[] {
+  const skillsDir = getAgentSkillsDir(agentId);
+  if (!existsSync(skillsDir)) return [];
+  return resolveSkillExecutableDirs(skillsDir);
+}
+
+/**
+ * Temporarily prepend agent skill directories to PATH for one async operation.
+ *
+ * PATH is process-global, so serialize PATH mutations to avoid races when
+ * multiple sessions initialize concurrently.
+ */
+let _pathMutationQueue: Promise<void> = Promise.resolve();
+async function withPathMutationLock<T>(fn: () => Promise<T>): Promise<T> {
+  const previous = _pathMutationQueue;
+  let release!: () => void;
+  _pathMutationQueue = new Promise<void>((resolve) => {
+    release = resolve;
+  });
+
+  await previous;
+  try {
+    return await fn();
+  } finally {
+    release();
+  }
+}
+
+export async function withAgentSkillsOnPath<T>(agentId: string, fn: () => Promise<T>): Promise<T> {
+  const skillDirs = getAgentSkillExecutableDirs(agentId);
+  if (skillDirs.length === 0) {
+    return fn();
+  }
+
+  return withPathMutationLock(async () => {
+    const originalPath = process.env.PATH || '';
+    const originalParts = originalPath.split(delimiter).filter(Boolean);
+    const existing = new Set(originalParts);
+    const prepend = skillDirs.filter((dir) => !existing.has(dir));
+
+    if (prepend.length > 0) {
+      process.env.PATH = [...prepend, ...originalParts].join(delimiter);
+      log.info(`Added ${prepend.length} skill dir(s) to PATH: ${prepend.join(', ')}`);
+    }
+
+    try {
+      return await fn();
+    } finally {
+      process.env.PATH = originalPath;
+    }
+  });
+}
+
+/**
+ * Whether TTS credentials are configured enough to use voice-memo skill.
+ */
+export function isVoiceMemoConfigured(env: NodeJS.ProcessEnv = process.env): boolean {
+  const provider = (env.TTS_PROVIDER || 'elevenlabs').toLowerCase();
+  if (provider === 'openai') {
+    return !!env.OPENAI_API_KEY;
+  }
+  if (provider === 'elevenlabs') {
+    return !!env.ELEVENLABS_API_KEY;
+  }
+  return false;
+}
+
 /**
  * Check if a binary exists on PATH
  */
@@ -158,7 +243,10 @@ export function loadAllSkills(agentId?: string | null): SkillEntry[] {
   // skills.sh global installs (lowest priority)
   dirs.push(SKILLS_SH_DIR);
   
-  // Global skills
+  // Bundled skills (ship with the project in skills/)
+  dirs.push(BUNDLED_SKILLS_DIR);
+
+  // Global skills (override bundled defaults)
   dirs.push(GLOBAL_SKILLS_DIR);
   
   // Agent-scoped skills (middle priority)
@@ -208,6 +296,7 @@ function installSkillsFromDir(sourceDir: string, targetDir: string): string[] {
 export const FEATURE_SKILLS: Record<string, string[]> = {
   cron: ['scheduling'],      // Scheduling handles both one-off reminders and recurring cron jobs
   google: ['gog', 'google'], // Installed when Google/Gmail is configured
+  tts: ['voice-memo'],       // Voice memo replies via lettabot-tts helper
 };
 
 /**
@@ -242,6 +331,7 @@ function installSpecificSkills(
 export interface SkillsInstallConfig {
   cronEnabled?: boolean;
   googleEnabled?: boolean;  // Gmail polling or Google integration
+  ttsEnabled?: boolean;     // Voice memo replies via TTS providers
   additionalSkills?: string[]; // Explicitly enabled skills
 }
 
@@ -261,22 +351,29 @@ export function installSkillsToWorkingDir(workingDir: string, config: SkillsInst
   mkdirSync(targetDir, { recursive: true });
   
   // Collect skills to install based on enabled features
-  const skillsToInstall: string[] = [];
+  const requestedSkills: string[] = [];
   
   // Cron skills (always if cron is enabled)
   if (config.cronEnabled) {
-    skillsToInstall.push(...FEATURE_SKILLS.cron);
+    requestedSkills.push(...FEATURE_SKILLS.cron);
   }
   
   // Google skills (if Gmail polling or Google is configured)
   if (config.googleEnabled) {
-    skillsToInstall.push(...FEATURE_SKILLS.google);
+    requestedSkills.push(...FEATURE_SKILLS.google);
+  }
+
+  // Voice memo skill (if TTS is configured)
+  if (config.ttsEnabled) {
+    requestedSkills.push(...FEATURE_SKILLS.tts);
   }
   
   // Additional explicitly enabled skills
   if (config.additionalSkills?.length) {
-    skillsToInstall.push(...config.additionalSkills);
+    requestedSkills.push(...config.additionalSkills);
   }
+
+  const skillsToInstall = Array.from(new Set(requestedSkills));
   
   if (skillsToInstall.length === 0) {
     log.info('No feature-gated skills to install');
@@ -310,22 +407,29 @@ export function installSkillsToAgent(agentId: string, config: SkillsInstallConfi
   mkdirSync(targetDir, { recursive: true });
   
   // Collect skills to install based on enabled features
-  const skillsToInstall: string[] = [];
+  const requestedSkills: string[] = [];
   
   // Cron skills (always if cron is enabled)
   if (config.cronEnabled) {
-    skillsToInstall.push(...FEATURE_SKILLS.cron);
+    requestedSkills.push(...FEATURE_SKILLS.cron);
   }
   
   // Google skills (if Gmail polling or Google is configured)
   if (config.googleEnabled) {
-    skillsToInstall.push(...FEATURE_SKILLS.google);
+    requestedSkills.push(...FEATURE_SKILLS.google);
+  }
+
+  // Voice memo skill (if TTS is configured)
+  if (config.ttsEnabled) {
+    requestedSkills.push(...FEATURE_SKILLS.tts);
   }
   
   // Additional explicitly enabled skills
   if (config.additionalSkills?.length) {
-    skillsToInstall.push(...config.additionalSkills);
+    requestedSkills.push(...config.additionalSkills);
   }
+
+  const skillsToInstall = Array.from(new Set(requestedSkills));
   
   if (skillsToInstall.length === 0) {
     return; // No skills to install - silent return