@@ -14,7 +14,8 @@ GROQ_API_KEY=
 ZAI_API_KEY=
 # NVIDIA Nemotron (build.nvidia.com -> Get API Key; starts with nvapi-)
 NVIDIA_API_KEY=
-# Sarvam API key for voice notes (dashboard.sarvam.ai). Optional: text works without it.
+# Voice notes are transcribed by Groq Whisper, then Gemini (both use the keys above),
+# then Sarvam. Sarvam is optional (dashboard.sarvam.ai).
 SARVAM_API_KEY=
 
 # Optional
@@ -26,5 +27,6 @@ SARVAM_API_KEY=
 # ZAI_MODEL=glm-4.6
 # NVIDIA_MODEL=nvidia/nemotron-3.5-super-vl-120b-a12b
 # NVIDIA_THINKING=0          # 1 = let Nemotron reason first (smarter, much slower)
-# SARVAM_STT_URL=https://api.sarvam.ai/speech-to-text
-# SARVAM_STT_MODEL=saaras:v3
+# STT_PROVIDERS=groq,gemini,sarvam
+# GROQ_STT_MODEL=whisper-large-v3
+# SARVAM_STT_MODEL=saarika:v2.5
@@ -41,8 +41,10 @@ Also built in: only `ALLOWED_USER_ID` may use the bot (fail closed if unset); th
 assistant can mark only the user's own open items done; `/delete` removes an item,
 its nudge history and the chat messages it came from, for good; `/forget` clears the
 chat history; `/model` picks which AI provider goes first (saved per user, others
-stay as backup); `/export` returns items and conversation; no nudges 22:00–08:00 local;
-at most 5 nudges per user per 24 hours.
+stay as backup); `/export` returns items and conversation. A reminder's first alert
+always goes out at the time the user set (checked every minute), even at night and
+past the cap; repeat nudges and idea nudges wait out quiet hours (22:00–08:00) and are
+capped at 5 per user per 24 hours.
 
 ## Layout
 
@@ -56,7 +58,7 @@ at most 5 nudges per user per 24 hours.
   core saves the raw note (rule 5).
   `offline.py` — rule-based stand-in for `simulate.py --offline`. Any object with
   `respond(message, history, items, now_local)` works as the assistant.
-- `stt.py` — Sarvam speech-to-text.
+- `stt.py` — voice-to-text: Groq Whisper → Gemini → Sarvam, whichever keys exist.
 - `tests/` — pytest with a fake assistant and fake Telegram objects; no network needed.
 
 ## Working on it
 
@@ -17,7 +17,7 @@ You need Python 3.11 or newer.
 
 1. In Telegram, message **@BotFather** → `/newbot` → copy the token.
 2. Message **@userinfobot** → copy your numeric user id.
-3. Get at least one AI key: Gemini (aistudio.google.com/apikey), Groq (console.groq.com/keys), Z AI (z.ai) or NVIDIA Nemotron (build.nvidia.com → Get API Key). With more than one, Pluto falls back automatically when one is rate-limited or down. For voice, also get a Sarvam key (dashboard.sarvam.ai).
+3. Get at least one AI key: Gemini (aistudio.google.com/apikey), Groq (console.groq.com/keys), Z AI (z.ai) or NVIDIA Nemotron (build.nvidia.com → Get API Key). With more than one, Pluto falls back automatically when one is rate-limited or down. Voice notes work with your Groq or Gemini key (Sarvam is an optional extra).
 4. Run:
 
 ```bash
@@ -56,12 +56,14 @@ The tests use a fake assistant and fake Telegram objects, so they need no keys o
 
 `/list` open items · `/done 3` · `/snooze 3` (+1 day) · `/delete 3` (forever) · `/forget` (clear chat history) · `/model` (choose which AI thinks first) · `/export` (JSON download of items and chat)
 
-## How nudges work
+## How reminders and nudges work
 
-- Checked every 30 minutes, never between 22:00 and 08:00 (`QUIET_START/END` in `core.py`).
-- Overdue items: every 6 hours, up to 5 times. Ideas with no deadline: after 3 days, then every 3 days, up to 3 times.
-- Max 5 nudges per 24 hours. Overdue items go first, and items nudged least go first, so nothing is starved.
+- Checked every minute while the bot is running.
+- **A reminder you set always arrives at the time you asked for**, even at night, and doesn't count against any limit.
+- If it's still open afterwards, Pluto nudges again every 6 hours (up to 5 times). Ideas you haven't touched get a nudge after 3 days, then every 3 days (up to 3 times).
+- Those follow-up nudges wait until morning (no follow-ups 22:00–08:00, `QUIET_START/END` in `core.py`) and are capped at 5 per 24 hours.
 - Done / Snooze / Drop buttons on every nudge.
+- Your computer must be awake with `python bot.py` running for reminders to arrive. To get them when your Mac is asleep, deploy it (see below).
 
 ## Rules built in
 
@@ -78,5 +80,5 @@ The bot must be running to nudge.
 ## Known limits
 
 - Each message sends the last 100 chat messages and your open items (plus 50 recent closed ones) to the model. "What have we discussed" only reaches that far back in the chat; saved items are always included. Add search/embeddings when this gets slow or costly.
-- Check `stt.py` against Sarvam's current docs; the model name and field names are env-configurable.
+- Voice notes: Groq `whisper-large-v3` → Gemini → Sarvam `saarika:v2.5` (`STT_PROVIDERS` changes the order).
 - AI providers are tried in order Gemini → Groq → Z AI → NVIDIA Nemotron (`PLUTO_PROVIDERS` changes it, and `/model` in chat picks which one goes first; the rest stay as backup). Default models: `gemini-2.5-flash`, `llama-3.3-70b-versatile` (Groq), `glm-4.6` (Z AI), `nvidia/nemotron-3.5-super-vl-120b-a12b` (NVIDIA, with thinking off for quick replies; `NVIDIA_THINKING=1` turns it on); change `PLUTO_MODEL`, `GROQ_MODEL`, `ZAI_MODEL`, `NVIDIA_MODEL` if a provider retires one. Fallback models may phrase replies differently, but the code-side checks (what gets saved, never citing unknown items) apply to all of them.
@@ -33,7 +33,7 @@
 log = logging.getLogger("pluto")
 
 ALLOWED = os.getenv("ALLOWED_USER_ID", "").strip()  # only this Telegram user may use the bot
-NUDGE_EVERY_SECONDS = 1800
+NUDGE_EVERY_SECONDS = 60  # reminders arrive within a minute of their time
 
 
 def allowed(update: Update) -> bool:
@@ -96,11 +96,18 @@ async def on_voice(update: Update, ctx: ContextTypes.DEFAULT_TYPE) -> None:
             filename=getattr(v, "file_name", None) or "voice.ogg",
             mime=getattr(v, "mime_type", None) or "audio/ogg",
         )
+    except stt.TranscriptionFailed as e:
+        log.warning("transcription failed: %s", e)
+        if not stt.providers():
+            msg = (f"I can't listen to voice notes yet: add one of {stt.KEY_NAMES} to .env "
+                   "and restart me. Meanwhile, please type it.")
+        else:
+            msg = "I couldn't make out that voice note just now. Please try again or type it."
+        await update.message.reply_text(msg)
+        return
     except Exception:
-        log.exception("transcription failed")
-        await update.message.reply_text(
-            "I could not understand that voice note. Please type it or try again."
-        )
+        log.exception("could not download the voice note")
+        await update.message.reply_text("I couldn't download that voice note. Please try again.")
         return
     if not text:
         await update.message.reply_text("I heard silence. Try again?")
@@ -200,8 +207,8 @@ def check_config() -> None:
         raise SystemExit("ALLOWED_USER_ID must be your numeric Telegram id (ask @userinfobot).")
     if len(llm.providers()) == 1:
         log.warning("Only one AI provider is set; add another key so Pluto has a fallback.")
-    if not os.getenv("SARVAM_API_KEY"):
-        log.warning("SARVAM_API_KEY is not set: voice notes will not work, text will.")
+    if not stt.providers():
+        log.warning("No speech-to-text key (%s): voice notes won't work, text will.", stt.KEY_NAMES)
 
 
 def build_app() -> Application:
@@ -218,7 +225,7 @@ def build_app() -> Application:
     app.add_handler(MessageHandler(filters.VOICE | filters.AUDIO, on_voice))
     app.add_handler(MessageHandler(filters.TEXT & ~filters.COMMAND, on_text))
     app.add_error_handler(on_error)
-    app.job_queue.run_repeating(nudge_job, interval=NUDGE_EVERY_SECONDS, first=60)
+    app.job_queue.run_repeating(nudge_job, interval=NUDGE_EVERY_SECONDS, first=5)
     return app
 
 
@@ -229,7 +236,7 @@ def build_app() -> Application:
     ("GROQ_API_KEY", "Groq", "AI"),
     ("ZAI_API_KEY", "Z AI", "AI"),
     ("NVIDIA_API_KEY", "NVIDIA Nemotron", "AI"),
-    ("SARVAM_API_KEY", "Sarvam (voice notes)", "optional"),
+    ("SARVAM_API_KEY", "Sarvam (voice backup)", "optional"),
 ]
 
 
@@ -240,8 +247,9 @@ def key_report() -> str:
         val = os.getenv(name) or (os.getenv("Z_AI_API_KEY") if name == "ZAI_API_KEY" else None)
         mark = "✓ set    " if val else ("✗ MISSING" if kind == "required" else "– not set")
         lines.append(f"  {mark}  {name:<20} {what}" + ("  (required)" if kind == "required" else ""))
-    order = llm.describe()
-    lines.append(f"AI order: {order}")
+    lines.append(f"AI order: {llm.describe()}")
+    voice = " → ".join(stt.providers()) or f"✗ none (add {stt.KEY_NAMES})"
+    lines.append(f"Voice notes: {voice}")
     if not llm.providers():
         lines.append("✗ Set at least one AI key.")
     return "\n".join(lines)
 
@@ -384,15 +384,18 @@ def in_quiet_hours(now) -> bool:
 def due_nudges(now=None) -> list:
     """Nudges to send now, as (user_id, item_id, Reply).
 
-    The caller sends each one and then calls db.mark_nudged(item_id, user_id, now),
-    so a failed send is retried on the next run."""
+    A reminder's first alert goes out at the time the user asked for, even in
+    quiet hours. Repeats and stale-idea nudges wait for daytime and respect the
+    daily cap. The caller sends each one and then calls
+    db.mark_nudged(item_id, user_id, now), so a failed send is retried."""
     now = now or db.now_utc()
-    if in_quiet_hours(now):
-        return []
     out = []
-    for r, reason in db.pick_nudges(now):
-        if reason == "due":
+    for r, reason in db.pick_nudges(now, quiet=in_quiet_hours(now)):
+        if reason == "reminder":
             reply = Reply(f"⏰ Reminder: {r['title']} (#{r['id']})", item_buttons(r["id"]))
+        elif reason == "due":
+            reply = Reply(f"⏰ Still pending: {r['title']} (#{r['id']}) — due {fmt_due(r['due_at'])}",
+                          item_buttons(r["id"]))
         else:
             days = (now - db.from_str(r["created_at"])).days
             reply = Reply(
 
@@ -265,13 +265,17 @@ def nudges_last_24h(user_id, now=None) -> int:
         ).fetchone()[0]
 
 
-def pick_nudges(now=None):
+def pick_nudges(now=None, quiet=False):
     """Return (item_row, reason) pairs that should be nudged now.
 
-    reason is 'due' or 'stale'. Overdue items go before stale ideas; within
-    each, items nudged the fewest times go first (so the daily cap never starves
-    an item), then the oldest. The daily cap is applied per user.
-    Quiet hours are the caller's job (see core.due_nudges)."""
+    reason is:
+      'reminder' - the first alert for something due: the user asked for this time,
+                   so it always goes out, ignoring the daily cap and quiet hours;
+      'due'      - a repeat nudge for an item still open after its due time;
+      'stale'    - an idea left untouched for a while.
+    Repeats and stale ideas respect the daily cap (per user) and are held back
+    while `quiet` (quiet hours) is true. Overdue items go before stale ideas;
+    within each, items nudged the fewest times go first, then the oldest."""
     now = now or now_utc()
     now_s = to_str(now)
     with conn() as c:
@@ -281,13 +285,16 @@ def pick_nudges(now=None):
             (now_s,),
         ).fetchall()
 
-    candidates = []
+    reminders, candidates = [], []
     for r in rows:
         last = from_str(r["last_nudged_at"]) if r["last_nudged_at"] else None
         if r["due_at"]:
-            if (
-                r["due_at"] <= now_s
-                and r["nudge_count"] < DUE_MAX_NUDGES
+            if r["due_at"] > now_s:
+                continue
+            if r["nudge_count"] == 0:
+                reminders.append((r, "reminder"))
+            elif (
+                r["nudge_count"] < DUE_MAX_NUDGES
                 and (last is None or now - last >= timedelta(hours=DUE_GAP_HOURS))
             ):
                 candidates.append((0, r["nudge_count"], r["due_at"], r["id"], r, "due"))
@@ -299,9 +306,14 @@ def pick_nudges(now=None):
                 and (last is None or now - last >= timedelta(days=IDEA_GAP_DAYS))
             ):
                 candidates.append((1, r["nudge_count"], r["created_at"], r["id"], r, "stale"))
+    reminders.sort(key=lambda t: (t[0]["due_at"], t[0]["id"]))
+    if quiet:
+        return reminders
     candidates.sort(key=lambda t: t[:4])
 
-    picked, per_user = [], {}
+    picked, per_user = list(reminders), {}
+    for r, _ in reminders:
+        per_user[r["user_id"]] = per_user.get(r["user_id"], nudges_last_24h(r["user_id"], now)) + 1
     for *_, r, reason in candidates:
         uid = r["user_id"]
         if uid not in per_user: