v0.0.1-alpha.6

Operations

• Updated pinned Docker release actions to their Node 24 generations, removing
  GitHub's Node 20 deprecation annotations from the release workflow.

Container image

ghcr.io/freshlabdev/voicy@sha256:a478f9817e78393c7a459712d3f9a29fdb2941847940327719fce7ac7e699596

v0.0.1-alpha.5

Fixed

• Repeated ordinary and ephemeral panel actions now treat Telegram's
  message is not modified response as an idempotent success instead of
  retrying the update and eventually dropping it.

Container image

ghcr.io/freshlabdev/voicy@sha256:6dfd627e34a17f2c9b4d8e6b1a4009279f6b1101bd45855433dba21309d98b22

v0.0.1-alpha.4

Voicy now has one canonical name from repository folder to production schema,
reliable Telegram initialization and job accounting, Rich Markdown transcript
delivery, complete settings, and useful personal and global statistics.

Added

• /start panel with Language, Settings, Stats, Help, About, and Close.
• /language [ru|en] in DM opens the language panel or sets the language
  across the shared Core language hub.
• Settings for smart formatting, paragraphs, filler words, profanity filter,
  diarization, quote style, and metadata.
• Per-user Deepgram options: smart_format, paragraphs, filler_words,
  profanity_filter, diarize_model=latest. detect_language and
  mip_opt_out remain always-on.
• Localized speaker turns from Deepgram paragraph-level speaker metadata.
• Personal and global statistics tabs with users, transcript kinds, audio
  duration, words, Kyiv peak hour, and last detected language.
• Owner-bound deep-link recovery for a long ephemeral transcript when neither
  ephemeral Rich Markdown nor direct-message delivery is available.
• Media size and duration limits, hourly retention cleanup, and stuck-job
  health reporting.

Changed

• Removed Deepgram WebSocket code and Telegram draft delivery. Prerecorded
  POST /v1/listen is the only STT transport.
• Removed transcript-document delivery and its setting. Transcripts use Bot API
  sendRichMessage, preserving replies and topics, with safe splitting above
  32,768 characters.
• Settings callbacks carry the desired value and are safe to retry.
• Telegram startup retries deleteWebhook, getMe, and command registration
  until all succeed. /healthz remains unhealthy before initialization.
• A terminal job transition and user-stat update now commit atomically and only
  once. Failed and empty jobs never affect user-facing statistics.
• transcripts primary key is (file_id, variant) and cached speaker turns are
  stored with the text.
• The folder, stack, database schema, Core bot key, role, and container are all
  named voicy.

Migrations

• 002_user_settings.sql adds settings, cache variants, speaker turns,
  retrieval tokens, and last-use timestamps.
• Core 011_voicy.sql migrates legacy presence and language rows, renames the
  schema and role, and narrows voicy_core to the three shared API functions it
  uses.

Operations

• Updated to Go 1.26.6 and fixed reachable standard-library and x/text
  vulnerabilities.
• CI and release workflows pin actions, pin govulncheck, run full release
  verification, require tagged code to be on main, fail on missing changelog,
  and publish an immutable image digest with SBOM and provenance.
• Production Compose accepts only an explicit VOICY_IMAGE reference.

Container image

ghcr.io/freshlabdev/voicy@sha256:d6dc88deffc97640b886273b7357a379eb3902102af421063d88540c23ea5ec8

v0.0.1-alpha.3

Menus and transcripts follow the family panel contract. Long voices keep
Deepgram paragraph breaks.

Added

• Menu callbacks are owner-scoped (m:<user_id>:<action>). A foreign tap is
  ignored with a short toast.
• Close deletes the panel instead of replacing it with a checkmark.
• Empty stats show a short empty state, not zeros.
• Deepgram Listen requests paragraphs=true. The formatted reply uses that
  paragraph text when Deepgram returns it.

Changed

• /start, Help, and Stats use <b>title</b> + <i>hint</i> +
  <blockquote> body, with Back/Close on subpanels.
• Transcripts are a clean blockquote. Language, duration, and confidence are
  no longer appended.

Known Limitations

• Cached transcripts from earlier alphas stay a single paragraph until that
  file_id is transcribed again.

v0.0.1-alpha.2

File uploads now go to Deepgram prererecorded Listen instead of Live streaming.

Changed

• Transcribe sends the downloaded Telegram file to POST /v1/listen. Default
  Content-Type is audio/ogg when Telegram does not provide one.
• Live WebSocket streaming remains in the client but is no longer the file
  upload path.

Fixed

• Empty or mis-detected audio containers are less likely to fail Deepgram
  because files are sent as prererecorded OGG rather than a raw stream.

v0.0.1-alpha.1

First public alpha. A Go Telegram bot that turns a voice message or video circle
into text with Deepgram.

Added

• DM transcription: send a voice or video circle and receive the text.
• Group /v replies with a public transcript of the replied voice or circle.
• Group /vp is an ephemeral command: a private placeholder inside Telegram's
  15-second window, then editEphemeralMessageText. DM fallback if the edit
  fails. Ordinary non-ephemeral group /vp stays quiet.
• Deepgram Live streaming (nova-3, language=multi) with prererecorded
  Listen fallback. Interim text uses Telegram sendMessageDraft where the
  client accepts drafts.
• file_id cache: store Telegram file id plus transcript, never audio bytes.
  The same file is not sent to Deepgram again.
• First-party Telegram HTTP client. No third-party Bot API SDK.
• Shared core identity via core.touch. Domain tables live in
  voicetotext.*.
• User-facing stats count only successful non-empty transcripts. Empty and
  failed runs are logged only.
• /healthz, Docker Compose for local development, and a WS04 production
  compose file that joins core_net.

Operations

• Local Compose seeds a minimal core schema (deploy/core-init.sql).
• Production uses voicetotext_core on shared core-postgres
  (deploy/ws04/compose.yaml). Core registration is
  core/migrations/009_voicetotext.sql.

Known Limitations

• Alpha: live validation is limited. Group drafts are often ignored by Telegram;
  /v still sends a public final message.
• Privacy mode stays on. Groups do not auto-transcribe every voice.
• Deepgram streaming of an already-downloaded file falls back to REST Listen
  when the WebSocket path fails.

Migrations

• 001_init.sql creates transcripts, jobs, user_stats, and
  runtime_state in the voicetotext schema.