Blog · release

Latch 0.22.0

2026-10-19

Today's mantra, voiced on your own box (#45). No migration. The voice is opt-in.

• 🌀 Today's mantra. The Handler's mantra of the day, spoken: a 3–5 minute track of it over a quiet bed, heard five times with room to say it back, one affirmation, and an opening and a close. Two a day: wake rises at the end with a bright tone, and bed settles and fades to silence. On /trance/mantra, the Trance page and the dashboard card. The Handler page's mantra gains 🌀 play it. The words stay in the Handler's library; only the ritual moved.
• Played to the end, it counts as recited. Judged on the audio the player actually produced (clamped to the wall clock, within 5 s of the end), never on the page's "completed". Once a day, and only a track of the day the sit is on. I said it and !recite still count.
• 🔴 Not the bed one in restrictive gear. Refused on the server whatever the page offered, and the refusal names the wake one.
• The voice: latch-voice, a second container (COMPOSE_PROFILES=voice). It runs Kokoro-82M on the CPU, with ffmpeg for the mix and the MP3, so Latch's own image stays ffmpeg-free. It publishes no port, sits on an internal-only network, and its model is baked in at build time: it has no internet, and Latch reaches it by name. The address (trance.voice_url) must resolve to this machine or your network, and is checked on every call. Trance now declares that egress, and the footer shows it. Measured on a 6-core pod: ~25 s per 4-minute track, ~1.4 GB peak, handed back after each render (the model runs in a child process that exits). The binaural offset survives the encoder: decoded, L 180.00 / R 186.00 Hz (LAME's joint stereo is off).
• Rendered ahead at 03:45, both slots, three tries a day 30 minutes apart, kept 14 days. A sit never waits on a model.
• 🔴 The voice is chosen by ear first. Nothing renders until you pick one of three candidate voices, each rendered as a whole wake track at the level it plays.
• Blind is loud. A failed render says so with the reason and offers read-and-recite. A render of a mantra that has since changed is called stale and isn't offered. Yesterday's file is never played as today's. Render today's again re-renders either.
• Optional: open with today's line from your AI Handler (trance.mantra_ai_line), never a masked one. Optional: the chat bot posts the wake track in your room at wake time, once a day, as m.audio (Post today's mantra audio on /consent, off by default). That puts GET /api/v1/bot/mantra in the bot token's scope, and /bot/today gains mantra_audio ({ready, due, seconds}, null while off). Paused, nothing goes. A post is not a recite.
• New settings: trance.voice_url, trance.mantra_voice, trance.mantra_wake (07:00), trance.mantra_bed (22:30), trance.mantra_minutes (4), trance.mantra_ai_line (0).