shadok-ai 0.13.7 → 0.13.9

package/README.md

@@ -140,6 +140,16 @@ Enter to skip; you can add it later from the web UI).
 - **Engine room** — the raw TUI screen, live, with clickable keys, for anything
   the chat can't express. With tmux there is also an **experimental real
   terminal** (xterm.js over the pane's byte stream) when a snapshot isn't enough.
+- **Voice notes on Telegram** — send a voice message to an agent's topic and it
+  is transcribed locally, then acted on. What was understood is echoed back
+  first (`🎙 « … »`), so a misheard instruction is visible before it runs —
+  and the agent is told the words were **spoken**, so it asks about an odd
+  file name or branch instead of acting on a mishearing.
+  Transcription is installed **on first use** — a static ffmpeg, whisper.cpp and
+  a 181 MB model — and the topic says so rather than going quiet for minutes.
+  `SHADOK_WHISPER_MODEL` picks a bigger one if your recordings need it.
+  Set `SHADOK_WHISPER_DIR` to a directory shared between containers and one copy
+  serves them all.
 - **Files an agent hands you** — when an agent's result IS a file (a report, an
   export, a screenshot, an archive) it arrives as a download card in the chat,
   inline if it is an image, and as an upload in the Telegram topic. It goes

package/context/pilot-prompt.md

@@ -23,6 +23,12 @@ Claude Code sessions in parallel. A human pilots you from a browser chat
   Avoid very wide previews in the options.
 - The user may be on a phone: conclusion first, compact responses, no long
   tables where a list would do.
+- **A message marked `vocal`** in its context header is the **transcription**
+  of something spoken, not text the human typed. Transcription mishears
+  homophones, proper nouns and technical terms. Read the intention rather than
+  the letter — and when something load-bearing looks garbled (a file name, a
+  branch, a command, a number), **ask instead of guessing**: acting on a
+  misheard instruction costs far more than one clarifying question.
 - **Handing over a file.** The person reading you is in a browser or on
   Telegram — **not on this machine**. A local path is a string they cannot open,
   so when your result IS a file (a report, an export, a screenshot, an archive)

package/dist/server.js

@@ -3363,7 +3363,14 @@ wss.on("connection", (ws, req) => {
                         // speaking", so they get no header.
                         let submitText = text;
                         if (origin === "web" || origin === "telegram" || origin === "cli") {
-                            submitText = markPromptMeta(text, promptMetaHeader(origin, new Date(), author, defaultTimeZone()));
+                            // A transcribed voice note says so IN the platform field rather than
+                            // through a different `origin`: the gate just above matches the origin
+                            // EXACTLY, so an origin of "telegram vocal" would drop the header
+                            // entirely — losing the time and the sender to gain one word.
+                            // `parsePromptMeta` reads the platform as the first field and the
+                            // sender as everything past the time, so this widens neither.
+                            const platform = msg.voice === true ? `${origin} vocal` : origin;
+                            submitText = markPromptMeta(text, promptMetaHeader(platform, new Date(), author, defaultTimeZone()));
                         }
                         // Push the ledger DELTA ahead of the message: rows changed since this
                         // agent last saw the ledger, so it learns what siblings resolved /