whatsapp

WhatsApp Automation

Send messages via unified script with Green API (primary) + WAHA (fallback if configured).

Quick Start - Unified Script

cd ~/.claude/skills/whatsapp/scripts

# Text message
npx ts-node send-message.ts "Hello!" --to 972501234567

# Multiline text from shell: use $'...'
npx ts-node send-message.ts $'Line 1\n\nLine 2\n- Bullet A\n- Bullet B' --to 972501234567

# Image with caption
npx ts-node send-message.ts --to 972501234567 --image /path/to/image.jpg --caption "Check this out!"

# Video with caption
npx ts-node send-message.ts --to 972501234567 --video /path/to/video.mp4 --caption "Watch this!"

# Any media file (auto-detect)
npx ts-node send-message.ts --to 972501234567 --media /path/to/any-file --caption "Here!"

# Voice message
npx ts-node send-message.ts --to 972501234567 --voice /path/to/audio.ogg

# File/document
npx ts-node send-message.ts --to 972501234567 --file /path/to/doc.pdf --caption "Here's the doc"

# Send to group
npx ts-node send-message.ts "Group message" --group [email protected]

# Reply to a specific message (quoted reply)
npx ts-node send-message.ts "Reply!" --group [email protected] --quote BAE5367237E13A1B

# Voice as quoted reply
npx ts-node send-message.ts --group [email protected] --voice /tmp/voice.mp3 --quote BAE5367237E13A1B

# Dry run (preview)
npx ts-node send-message.ts "Test" --to 972501234567 --dry-run

Default Numbers

Alias	Number	Use Case
myself / me / test	$MY_PHONE (from config.sh)	Default recipient

Sending Video & Audio — encode to AAC

WhatsApp (and Facebook) play a video silently if its audio track is MP3 muxed into an MP4 — the file is valid and has sound in desktop players, but the mobile app drops it. Always ship video with AAC audio:

# Re-encode ONLY the audio to AAC, leave the video untouched; +faststart for streaming
ffmpeg -y -i in.mp4 -c:v copy -c:a aac -b:a 192k -movflags +faststart out.mp4

# Verify it isn't silent before sending
ffmpeg -i out.mp4 -af volumedetect -f null /dev/null 2>&1 | grep mean_volume

Voice notes: convert the source to OGG/Opus first — raw MP3 in a container is unreliable in the app.

Script Options

Option	Description
"message"	Text message (first positional argument)
--to, --phone <NUMBER>	Phone number (Israeli format accepted)
--group <ID>	Group ID (format: [email protected])
--image <PATH>	Attach image file
--video <PATH>	Attach video file (MP4, etc.)
--media <PATH>	Attach any media file (auto-detect)
--voice <PATH>	Send voice message (OGG/MP3)
--file <PATH>	Attach document/file
--caption <TEXT>	Caption for media files
--quote <MSG_ID>	Reply to a specific message (quotedMessageId)
--dry-run	Preview without sending

Multiline Formatting

When sending a WhatsApp message with line breaks from the shell, do not pass literal \n inside a normal quoted string if you want actual newlines in the final message.

Use shell ANSI-C quoting instead:

npx ts-node send-message.ts $'Hi\n\nLine 2\n- Item 1\n- Item 2' --group [email protected]

Why:

• "text with \n" sends the characters \ + n — but the script auto-converts these to real newlines
• $'text with \n' sends real line breaks natively

Either form works — the script normalizes literal \n to real newlines before sending.

Apostrophes inside $'...': Use \' directly — do NOT use the '\'' trick. The '\'' pattern breaks out of ANSI-C quoting, and everything after it becomes regular single-quoted (where \n is literal).

# CORRECT — apostrophe inside $'...'
$'I\'ve added a check.\n\nNext line.'

# WRONG — breaks ANSI-C quoting, \n becomes literal after the apostrophe
$'I'\''ve added a check.\n\nNext line.'

For structured WhatsApp messages, prefer:

• short opening line
• blank line
• short bullet list
• blank line
• closing line

Formatting Rule — Match Language

English messages: Use regular bullets/dashes. Never use Hebrew letters (א. ב. ג.) in English text.

- First item
- Second item
- Third item

Hebrew messages: Every line must start with a Hebrew character for correct RTL rendering. Use Hebrew letters for list items.

א. פריט ראשון
ב. פריט שני
ג. פריט שלישי

Don't mix: No Hebrew letters in English messages, no dashes/numbers in Hebrew messages.

Don't:

- פריט ראשון
1. פריט שני
* פריט שלישי

Lines starting with non-Hebrew characters (dashes, numbers, asterisks, English) break RTL alignment in WhatsApp. Always ensure the first visible character is Hebrew.

Provider Logic

Content Type	Primary	Fallback
Text	Green API	WAHA (if configured)
Image	Green API	-
Video	Green API	-
Voice	Green API	-
File	Green API	-

Note: WAHA is not currently installed. Green API handles all message types.

Configuration

Credentials in scripts/.env:

# Green API (primary provider)
GREEN_API_URL=https://7103.api.greenapi.com
GREEN_API_INSTANCE=your_instance
GREEN_API_TOKEN=your_token

# WAHA (optional fallback — uncomment if installed)
# WAHA_URL=http://localhost:3001
# WAHA_API_KEY=your_key
# WAHA_SESSION=default

Phone Number Format

Input	Normalized
972501234567	[email protected]
0501234567	[email protected]
501234567	[email protected]

Direct API (WAHA)

For text messages only:

source ~/.claude/skills/whatsapp/config.sh

curl -s -X POST "$WAHA_URL/api/sendText" \
  -H "X-Api-Key: $WAHA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "session": "default",
    "chatId": "[email protected]",
    "text": "Hello!"
  }'

Check Session Status

source ~/.claude/skills/whatsapp/config.sh
curl -s -H "X-Api-Key: $WAHA_API_KEY" "$WAHA_URL/api/sessions/default"

Finding Contact by Name

When user says "send WhatsApp to [name]" without phone number:

1. Use get-contact skill to search CRM
2. If not found → ask user for phone
3. If multiple → ask user to choose
4. If one → confirm before sending

Always confirm contact before sending.

Finding a Group by Name or Project

When user says "send to [project] group" — search Green API directly (WAHA may be down):

TOKEN=$(grep "GREEN_API_TOKEN" ~/.claude/skills/whatsapp/scripts/.env | cut -d'=' -f2)
INSTANCE=$(grep "GREEN_API_INSTANCE" ~/.claude/skills/whatsapp/scripts/.env | cut -d'=' -f2)
curl -s "https://7103.api.greenapi.com/waInstance$INSTANCE/getChats/$TOKEN" | python3 -c "
import json,sys
data=json.load(sys.stdin)
for c in (data if isinstance(data,list) else []):
    if '@g.us' in c.get('id',''):
        print(c.get('id'), '|', c.get('name',''))
"

Then send using send-message.ts --group <ID>.

Never check WAHA session before sending — the script handles fallback automatically.

Provider Notes

Green API is the primary provider. WAHA is not installed and only used as fallback if explicitly configured. Just run send-message.ts — it uses Green API directly.

Retrieving Messages (get-messages.ts)

Fetch incoming messages, voice recordings, and media from any contact:

# Get recent messages from a contact
npx ts-node get-messages.ts --chat 972526062921 --count 5

# Get messages from last 2 hours
npx ts-node get-messages.ts --chat 972526062921 --minutes 120

# Get all messages from a time range
npx ts-node get-messages.ts --start "2h ago" --end "now"

# Output as JSON (includes download URLs for media)
npx ts-node get-messages.ts --chat 972526062921 --json

What you get:

• Text messages with full content
• Voice messages with download URL (OGA/OGG format) — can be transcribed separately
• Images with download URL and metadata
• Timestamps, sender info, file names
• All message types (extended text, reactions, media)

Quoted/reply messages are shown inline — get-messages.ts now displays quoted context automatically with ↳ Reply to prefix. Both extendedTextMessage and quotedMessage types are handled.

Message types: textMessage, extendedTextMessage (with .text), quotedMessage (reply with .quotedMessage.textMessage), reactionMessage, imageMessage

Voice Message Workflow:

1. Get messages with get-messages.ts --json
2. Extract downloadUrl from audioMessage type
3. Download the OGA file
4. Convert to WAV: ffmpeg -i file.oga -ar 16000 file.wav
5. Transcribe: /transcribe --language he /path/to/file.wav

Waiting for Replies — Monitor Integration

Sometimes after sending a message you need the contact's reply before you can continue — an approval, a decision, feedback, a "how much do I owe you". Instead of stopping and asking the user to watch the chat, hand the waiting to the Monitor tool and keep working; each new reply arrives as a notification.

Use scripts/watch-replies.sh — it polls for genuinely new incoming messages (dedupes by id, handles text / reactions / media) and prints one line per reply, which is exactly what Monitor consumes. Exits only when killed (Monitor timeout / TaskStop).

# usage
scripts/watch-replies.sh --chat <number|groupId> [--since <epoch>] [--interval 30] [--label REPLY]

Launch it through the Monitor tool (not Bash), so replies notify you while you keep working:

• command: ~/.claude/skills/whatsapp/scripts/watch-replies.sh --chat 972501234567 --since "$(date +%s)" --label CLIENT
• timeout_ms: how long to wait (e.g. 3600000 for an hour).
• Always set --since to the moment you sent ($(date +%s)), or old history re-fires as if new.

⚠️ Do NOT hand-roll an inline poller in the Monitor command. Passing a group id or JSON body directly in the shell breaks under zsh — @g.us triggers bad math expression: illegal character: @, and the monitor exits 1. watch-replies.sh takes the chat as an argument (--chat <id>), so group ids are safe. Always reach for this script; never inline the group id + curl/python in the Monitor command.

Decide per situation whether to use it

Reach for it when the reply changes what you do next:

• you sent something that needs approval before the next step (a draft, a preview, a quote);
• you asked a question and will act on the answer;
• you're in an iterative loop (send fix → they review → send next);
• you started long work and want to hear back without blocking.

Skip it for fire-and-forget (a confirmation, an FYI, a "done") — no reply is expected, so there's nothing to watch.

Staying in the conversation

Often the right move is to remain engaged: when a reply lands, respond, then re-arm a fresh watcher (new --since) for the next turn. That's a real back-and-forth — send sample → they approve → send fix → wait again — until the thread naturally closes (final approval, thanks, payment settled, or the user says wrap up). Each round: send → watch → reply → re-watch. Use your judgement on when the exchange is genuinely done and it's time to stop watching.

Discipline

• Each watcher has its own dedupe file, so two watchers on the same chat double-fire — TaskStop the old one before arming a new one.
• Reactions surface as (reaction 👍) and media as [imageMessage]; when you need the image URL / caption / voice file, pull the full record with get-messages.ts --json.

When to Use

• Send notifications or alerts
• Share images/posters
• Voice messages
• Document sharing
• Automated reminders
• Workshop follow-ups
• Retrieve and transcribe voice messages from customers
• Check media attachments from contacts