Skip to content

Notify & read-aloud

Two ways asmltr uses voice to keep you in the loop when you're away from a chat window:

  • Notify — the assistant (or a schedule) proactively reaches you with a spoken / pushed message, trying the best channel that can actually get through.
  • Notification reader — your phone reads incoming notifications aloud over your headphones as a short spoken synopsis, skipping the noise.

Both are voice-first, both respect quiet hours, and both are configurable.

Reaching you — asmltr notify

When something needs your attention out-of-band — a scheduled morning brief, "your build finished," an alert while you're away — the assistant calls:

asmltr notify "<message>" [--title "<title>"] [--force] [--silent]
Flag Effect
--title A short heading shown with the message.
--force Deliver even during quiet hours (use only when it's genuinely urgent).
--silent Skip the spoken step — deliver as text only.

A schedule whose prompt says "notify me…" or "send me a message" is exactly this command. It's also the primitive the morning brief uses.

The delivery ladder

asmltr notify doesn't just fire and hope — it walks a ladder and stops at the first step that can actually reach you:

  1. Read aloud — if a connected assistant device (the phone app or a headless control link) is present and allowed to speak, it's spoken aloud through your configured voice.
  2. Push — a push notification to the device (when a push sender is configured).
  3. Text fallback — a message to a configured channel (Telegram / Discord / email).

If a step isn't reachable or isn't configured, the ladder falls through to the next. The command reports which step delivered it (✓ notified via android) or that nothing landed — so a notification never silently vanishes into the void.

Quiet hours suppress the spoken step (so a 3 AM brief won't wake you); the message still travels the rest of the ladder as text unless you passed --force. If no quieter step is configured, it's held rather than spoken.

Configuring delivery

Set the policy in the dashboard under Settings → Notifications → Notify delivery:

  • Quiet hours — the window where the spoken step is suppressed.
  • Only read aloud over headphones — never speak over the phone's loudspeaker.
  • Text fallback — the channel + target that catches messages the spoken step can't (recommended, so quiet-hours notifications still reach you silently).

Under the hood this is stored at ~/.asmltr/notify.json and served by POST /v2/notify + /v2/notify/config; connectors are reached through the connector manager's unified send path, so there are no host-specific scripts to wire up.

Reading phone notifications aloud

The notification reader speaks a natural-language synopsis of incoming phone notifications over your headphones — like a conversational, selective version of a car's "read my messages." It's a native feature of the Android app.

How it works

  • The app watches posted notifications (you grant Android's Notification access once).
  • Each one is judged on-device by the local engine — it returns whether to speak it, a priority score, and a one-sentence synopsis. Because this runs on the local Agent SDK (not a metered/cloud API), the content of your private notifications never leaves the device for a third-party key.
  • If it clears your threshold, the app reads the synopsis aloud ("You've got a direct message on Discord — they're done with the project").
  • A burst of notifications is summarized together rather than read one by one.
  • It waits for a gap rather than talking over you — see below.

It won't talk over you

A synopsis is only ever spoken into a free ear. Before anything is read, the app asks whether this is actually a good moment:

Situation What happens
You're on a call — cellular or VoIP (WhatsApp, Meet, Discord…) Held. Nothing is spoken.
Do Not Disturb is on Held. DND means quiet, and the assistant is not exempt.
Navigation is speaking, or an alarm/other assistant is Waits it out (~20s), then holds if it's still going.
Music or a podcast is playing Spoken — but the music ducks to a murmur underneath, then comes back.
A call starts mid-sentence Playback stops immediately and the rest is held.

"Held" never means "lost". The synopsis goes into a backlog and the eyes appear in a corner of the screen wearing a "!":

  • Tap them → the whole backlog is read aloud in one go, in order, in your voice. If five things arrived during a twenty-minute call, you hear all five when you're ready.
  • Drag them to the ✕ (it fades in at the bottom of the screen while you drag) → discard the backlog unheard.
  • Fling them off any edge → same thing, faster.
  • Drag them anywhere else → they snap to the nearest side and stay there; the app remembers.

The backlog survives the app being restarted by the system, and entries older than three hours expire on their own — a synopsis from this morning is noise, not news.

The same gate covers asmltr notify's spoken step, so a proactive message from the assistant can't barge into a call either. (Do Not Disturb is not applied there: that path already has its own quiet hours, and the notify ladder decides its own timing.)

Gating & settings (app → ⚙ Notifications)

  • Enable readout on/off, and only over headphones (skip when on the speaker).
  • Quiet hours — no readout during your configured window.
  • Priority threshold — read only notifications scored at or above your bar.
  • Per-app allow/deny and a sender allow-list; ongoing/transport/foreground-service noise is filtered out automatically.
  • Verbosity — headline vs. full synopsis — and burst-summarize on/off.
  • Hold instead of interrupting — on by default. Turning it off restores barge-in behavior (still ducked, still respecting the headphones gate).
  • Respect Do Not Disturb — on by default.
  • Wait out navigation prompts and alarms — on by default. Turn it off if you'd rather hear notifications while a route is being narrated.

The settings screen also shows the live gate state ("Right now: free to speak" / "a call is in progress" / "3 held") and a Test held badge button, so the corner UI can be tried without staging a phone call.

Privacy

Notification text is sensitive by nature. The synopsis is generated locally, nothing is retained beyond what's needed to de-duplicate a burst, and readout only happens over a private audio route (headphones). Android's own consent screen gates the whole feature.