AGENTS/skills/voice-notify/SKILL.md

name, description, compatibility

name	description	compatibility
voice-notify	Use when: (1) A long-running task (build, test, migration, cruncher job) completes, (2) The user asks to be notified audibly, (3) An agent finishes delegated work and should announce results, (4) Important errors or completions need immediate attention. Triggers: talk, voice, notify, audio, abbrechen, notify me, let me know when done, cruncher, job complete.	opencode

voice-notify

ElevenLabs TTS voice notifications with a dismissable desktop popup. Audio playback and notification appear simultaneously — click ⛔ Abbrechen to stop playback.

Prerequisites

• talk binary installed (via m3ta-nixpkgs pkgs.talk)
• ElevenLabs API key at /run/agenix/elevenlabs-key (preferred) or ELEVENLABS_API_KEY env var (fallback)
• Desktop notification daemon (DMS, dunst, mako, etc.)
• Audio playback (mpv) and PipeWire/PulseAudio running

When to Use

• Task completed — after builds, tests, migrations, deployments (>30s runtime)
• Cruncher job finished — long data processing pipelines
• Agent delegated work done — when a subagent finishes and the user should know
• User requested audio feedback — "notify me when done", "let me know"
• Critical errors — build failures, test crashes, missing dependencies

When NOT to Use

• Quick responses (< 30s of work)
• VOICE_NOTIFY=0 is set — respect the toggle, stay silent
• Headless server without audio/desktop
• User is in a meeting/streaming — audio would be disruptive

Usage

Basic

talk "Build abgeschlossen. Alle 247 Tests erfolgreich."

Stdin (Pipe)

echo "Deployment fertig" | talk
nixos-rebuild build 2>&1 | tail -3 | talk

# AI agent print mode → talk
pi -p "Review diesen Code" | talk
opencode run "Erkläre diese Funktion" -q | talk

Argument has priority; if no argument is given, talk reads from stdin. This enables piping from any command or AI agent.

After Task Completion

# Am Ende eines langen Jobs:
talk "Cruncher Job komplett. 3 Dateien verarbeitet, 0 Fehler."

Disable Temporarily

export VOICE_NOTIFY=0   # stumm schalten
# ... agent work ...
export VOICE_NOTIFY=1   # wieder an

Integration Patterns

Pattern 1: Cruncher Job Wrapper

For any long-running job, use the wrapper script at scripts/cruncher-notify.sh:

# Statt direkt:
python3 process_data.py --input data.csv

# Mit Voice Notification:
cruncher-notify.sh "Data Import" -- python3 process_data.py --input data.csv
# → 🔊 "Data Import erfolgreich. Dauer: 3 Minuten 42 Sekunden."

On failure: ❌ "Data Import fehlgeschlagen, Code 1".

Pattern 2: Agent Hook

Add to AGENTS.md in any project:

## Voice Notifications
When `VOICE_NOTIFY != 0` and the task took longer than 30 seconds,
call at the end: `talk "[brief summary]"`

Pattern 3: Post-Build / Post-Test

# Makefile oder CI script:
test:
    pytest tests/
    talk "Tests durchgelaufen: $$(pytest tests/ -q --tb=no 2>&1 | tail -1)"

Toggle Behaviour

Value	Effect
VOICE_NOTIFY=1 (default)	Notifications active
VOICE_NOTIFY=0	talk exits immediately, no audio, no popup
unset	Defaults to enabled

Set per-shell, per-user, or system-wide via Home-Manager home.sessionVariables.

Pitfalls

• No audio? Check /run/agenix/elevenlabs-key exists and is readable (or ELEVENLABS_API_KEY env var). HTTP 402 = no credits
• No popup? Ensure a notification daemon is running (notify-send needs a daemon)
• Cancel button doesn't stop audio? Fixed: the poll loop now uses [ -s ] (non-empty check) instead of [ -f ] (exists). The action file is created as empty by the shell redirect; it's only filled when the user clicks Abbrechen.
• pw-play doesn't work — it can't decode MP3; talk uses mpv which handles all formats