[sneha-20bit]

Summary

Three commits.

Fix voice notes and on-time reminders (939b355)

• Voice notes: stt.py now tries Groq Whisper (whisper-large-v3), then Gemini audio, then Sarvam, using whichever keys are set. Before, it only used Sarvam, so voice failed without a Sarvam key. The Sarvam default is now the transcription model saarika:v2.5, since saaras translates.
• Voice errors: with no voice key set, the bot says which key to add. bot.py --check shows the voice provider order.
• Reminders: the due check runs every minute instead of every 30. A reminder's first alert always arrives at the time you set, even in quiet hours and past the daily cap. Repeat nudges ("Still pending") and nudges about old ideas still wait for morning and the cap.
• Relative times: the prompt now understands "in 10 minutes" and "2 ghante mein".

Borrow plan (4561da2)

• pluto/docs/borrow-plan.md studies QwenPaw's three-layer memory and Heartbeat, and nanobot's Dream memory consolidation and how its scheduled jobs reach Telegram. It cites file paths and functions, and maps each idea onto db.py, llm.py and bot.py.
• It includes designs for a daily morning digest and a SQLite-only memory upgrade (built-in full-text search, rolling summaries, profile facts), plus what to skip for now.
• No code was copied. The ideas are re-implemented (nanobot is MIT, QwenPaw is Apache 2.0).

Sarvam script (a4702f1)

• sarvam_hello.py at the repo root makes one chat completion call to sarvam-105b-conversations with the sarvamai SDK. It reads SARVAM_API_KEY from the environment and never prints it. .env stays ignored by git.

Testing

• 99 pytest tests pass. New tests cover the voice fallback chain and request shapes, the message when no voice key is set, reminders at night, repeats waiting for morning, the cap applying only to repeats, and the one-minute check interval.
• Not run: real Groq, Gemini or Sarvam voice calls, and sarvam_hello.py itself, which hasn't been run because this session has no key.

---

Generated by Claude Code