config.example.yaml

# Saymo config template. Copy this file to `config.yaml` (gitignored) and fill in
# your own values. All secrets use `${ENV_VAR}` — keep them in your shell or a
# .env, never commit keys here.

user:
  # Your display name in prompts and turn-detection.
  name: "Your Name"
  # Any short forms / nicknames that colleagues use on the call.
  name_variants:
    - "Your Name"
  # Optional — used by prompts that accept {user_role}, {team_name}, {tech_stack}.
  role: ""          # e.g. "Senior Engineer"
  team: ""          # e.g. "Platform"
  tech_stack: ""    # e.g. "Python, Kubernetes, Postgres"
  language: ru      # `ru` or `en`

audio:
  capture_device: "BlackHole 16ch"
  playback_device: "BlackHole 2ch"
  monitor_device: ""          # e.g. "External Headphones" — to hear yourself
  recording_device: ""        # mic used for voice-sample recording
  sample_rate: 16000
  channels: 1
  chunk_size: 1024
  # --- Mic input chain ---
  # Run `saymo mic-check` to have Saymo measure your room + mic and print
  # ready-to-paste values for these fields. Defaults below are no-op, so an
  # un-calibrated install behaves identically to before the calibration
  # feature was added.
  input_gain_db: 0.0          # +3..+12 = louder; -3..-12 = quieter; 0 = off
  noise_gate_db: -120.0       # frames below this RMS (dB) go silent; -120 disables
  highpass_cutoff_hz: 0.0     # cuts rumble/hum; try 80 Hz for close-talk mics; 0 disables
  noise_reduction: false      # spectral denoise via `noisereduce` (soft dep, pip-installed)
  noise_reduction_strength: 0.75  # 0..1 — higher = more aggressive, more artefacts

stt:
  engine: whisper             # `whisper` (local) or `deepgram` (cloud)
  whisper:
    model_size: small
    device: cpu
    compute_type: int8
  deepgram:
    api_key: ${DEEPGRAM_API_KEY}
    model: nova-3
    language: ru

diarization:
  enabled: false              # optional; disabled keeps base install lightweight
  engine: disabled            # `disabled` or `pyannote`
  model: pyannote/speaker-diarization-3.1
  device: cpu                 # cpu | mps | cuda, depending on backend support
  min_speakers: 1
  max_speakers: 4
  auth_token_env: SAYMO_DIARIZATION_TOKEN  # env var name, not the token value

meeting_memory:
  enabled: true
  retain_transcripts: true     # false keeps metadata but blanks transcript text
  base_dir: ""                 # defaults to ~/.saymo/trigger_samples/
  default_window_seconds: 8.0
  summary_max_items: 5

analysis:
  llm_provider: ollama        # `ollama` (local, default) or `anthropic` (cloud)
  anthropic:
    api_key: ${ANTHROPIC_API_KEY}
    model: claude-sonnet-4-20250514
  turn_detection:
    keyword_trigger: true
    llm_trigger: true
    llm_interval_seconds: 5
    confidence_threshold: 0.8

tts:
  # Voice-cloning engines (primary — speak in the user's voice):
  #   `coqui_clone`  — XTTS v2, zero-shot or fine-tuned
  #   `qwen3_clone`  — Qwen3-TTS 12Hz with LoRA adapter (Apple Silicon)
  # Fallback engines (do NOT sound like the user):
  #   `piper`        — ~60 MB download, offline
  #   `macos_say`    — built-in, zero install
  engine: coqui_clone
  # Optional override for the live auto-mode / Q&A path. When set, _auto()
  # uses this engine for spontaneous replies while `engine` above keeps
  # handling slower prepare-time synthesis. Typical split: prepare with
  # coqui_clone (quality), realtime with qwen3_clone (latency). Empty
  # means "use `engine` for both".
  realtime_engine: ""
  piper:
    model_path: ""
  macos_say:
    voice: Milena

obsidian:
  vault_path: ""              # e.g. "~/Documents/Obsidian Vault"
  subfolder: ""
  date_format: "%Y-%m-%d"

ollama:
  url: http://localhost:11434
  model: qwen2.5-coder:7b

jira:
  url: ""                       # e.g. https://your-company.atlassian.net
  token: ${JIRA_TOKEN}
  project_key: ""               # e.g. "ABC" — scopes JQL queries to one project
  user_query: "assignee = currentUser() AND updated >= -1d ORDER BY updated DESC"
  worklog_query: "worklogAuthor = currentUser() AND worklogDate >= -10d"
  max_results: 15
  team_members: {}              # Optional: {jira_username: Display Name} for team scrum report

speech:
  style: concise
  language: ru
  source: obsidian            # `obsidian` | `jira` | `confluence` | `file`
  composer: ollama            # `ollama` | `anthropic`

safety:
  require_confirmation: true          # require a second trigger mention before auto-speaking
  confirmation_timeout_seconds: 6.0   # seconds to wait for that second mention
  hotkey_speak: "<cmd>+<shift>+s"      # force prepared playback in auto-mode
  hotkey_stop: "<cmd>+<shift>+x"
  hotkey_toggle: "<cmd>+<shift>+m"
  hotkey_takeover: "<cmd>+<shift>+u"   # manual answer mode: pause auto + best-effort mic switch
  hotkey_approved_speak: "<cmd>+<shift>+a"
  hotkey_skip: "<cmd>+<shift>+k"
  hotkey_capture_marker: "<cmd>+<shift>+c"
  hotkey_status: "<cmd>+<shift>+i"
  hotkey_hard_disable: "<cmd>+<shift>+h"
  max_speech_duration: 120

live:
  # Runtime tuning for `saymo auto`. Meeting profiles can override these under
  # `meetings.<name>.live` without editing source.
  chunk_seconds: 4.0                 # audio window sent to Whisper
  overlap_seconds: 2.0               # overlap between adjacent windows
  read_timeout_seconds: 3.0          # wait for the next captured chunk
  trigger_cooldown_seconds: 45.0     # suppress duplicate name hits after answer
  silence_rms_threshold: 0.001       # skip quiet windows before Whisper
  pre_speak_delay_seconds: 2.0       # pause after trigger before playback

# Meeting profiles — each one picks a source plugin, a call provider, and the
# trigger phrases used by `saymo auto`. Add as many as you need.
meetings:
  personal:
    description: "Personal update — first-person brief"
    provider: glip
    team: false
    source: obsidian
    trigger_phrases:
      - "Your Name"
    # Optional addressing overrides for this profile. Patterns are regular
    # expressions; use {trigger} where the matched trigger phrase should appear.
    # addressing:
    #   handoff_patterns:
    #     - "\\bпусть\\s+{trigger}\\s+ответит\\b"
    #   mention_only_patterns:
    #     - "\\bэскалировали\\s+{trigger}\\b"
    # Optional live tuning just for this profile:
    # live:
    #   chunk_seconds: 3.0
    #   overlap_seconds: 1.0
    #   trigger_cooldown_seconds: 25.0
  team:
    description: "Team update — you + teammates"
    provider: glip
    team: true
    source: obsidian
    trigger_phrases:
      - "Your Name"
      - "your team"

# Prompt templates — overrides for the defaults in saymo/speech/ollama_composer.py.
# Placeholders are resolved at runtime; see module docstrings for the full list.
prompts:
  # standup_ru: |
  #   Ты — помощник для ежедневных встреч. Составь отчёт на русском...
  #   Заметки за вчера ({yesterday_date}):
  #   ---
  #   {yesterday_notes}
  #   ---
  #   Заметки на сегодня ({today_date}):
  #   ---
  #   {today_notes}
  #   ---
  # standup_en: |
  #   You are a speaking assistant. Compose a brief verbal update in English...
  # team_scrum_ru: |
  #   Ты — помощник для командных митингов...
  # autonomous_daily_ru: |
  #   Ты — автономный голосовой ассистент...
  # qa_system_ru: |
  #   Ты — {user_name}, {user_role}...
  # expand_ru: |
  #   Тебе дали заметки. Перепиши в устную речь...

# Project-specific vocabulary — merged on top of the default abbreviation map
# in saymo/tts/text_normalizer.py. Use this instead of editing source.
vocabulary:
  abbreviations: {}           # e.g. { MYAPI: "май-эй-пи-ай", K8S: "кубернетес" }
  # Run `saymo trigger-check -p personal --mic` and add exact Whisper
  # mis-hearings with `saymo trigger-setup -p personal --heard "..."`.
  # It is safe to paste a whole transcript; Saymo saves only the name variant.
  fuzzy_expansions: {}        # e.g. { Alex: ["Alex", "Алекс", "Саша", "Саня"] }

# Tier-A response cache — pre-synthesised short answers for common stand-up
# follow-up questions. Lets real-time Q&A work on CPU-only machines: at
# runtime `_auto()` looks up a matching cached answer and plays it in ~100 ms
# instead of synthesising live. Defaults live in
# `saymo/analysis/response_cache.py::DEFAULT_RESPONSE_LIBRARY`. Entries under
# `library` extend or override those defaults — add your own intents or
# replace the wording of existing ones without touching source.
responses:
  enabled: true
  confidence_threshold: 0.6   # 0..1 — lower = more willing to match loose wording
  cache_dir: ""               # empty → ~/.saymo/audio_cache/responses/
  # When cache misses a match, try synthesising a live answer via Ollama
  # + TTS. Off by default because it's slow (several seconds) and can
  # make auto-mode feel sluggish. Turn on only if you often get questions
  # that aren't covered by cached variants.
  live_fallback: false
  # When true, runs a small Ollama classifier BEFORE the keyword match.
  # Catches rephrasings the keyword matcher misses (e.g. "когда сдашь"
  # → eta_generic). Adds ~1s latency; turn off if CPU-constrained.
  intent_classifier: false
  # Use `saymo trigger-check -p personal --text "Your Name, what is the status?"`
  # to preview trigger/addressing/question/cache decisions before a call.
  library: {}
    # Example — uncomment to override the default generic status answer:
    # status_generic:
    #   triggers:
    #     - "как дела"
    #     - "что по задаче"
    #   variants:
    #     - "В работе, детали скину в чат после митинга."
    #   description: "Generic status, my wording"
    #
    # Example — add a new intent that is not in the default library:
    # release_today:
    #   triggers:
    #     - "когда релиз"
    #     - "когда выкатка"
    #   variants:
    #     - "Планирую выкатить сегодня после обеда."
    #     - "Релиз уезжает на стейдж до конца дня."
    #   description: "Project-specific release question"