Persona & voice

Saga speaks as "the radio demon" — a theatrical 1930s broadcaster who *is* the house, not a neutral assistant wearing a character skin. This is the core pipeline every other feature's spoken output runs through (except the notes/reminders tools, which deliberately bypass it — see below).

How to use it

Not really a request-driven feature — it's always on. Anything that isn't caught by a more specific feature gets an in-character reply:

general chat, opinions, "why" / "what do you think" / "explain" questions

Longer, more reflective answers (the "philosopher register") come from the same model with a larger token budget, routed by router.classify() whenever a question reads as reflective rather than a reflex command.

How it works

• Character: SYSTEM_PROMPT in brain/models.py — defines the voice, tone rules (one sentence, rarely two; never breaks character; never admits to being an AI), and the required JSON response shape {text, style}.
• Style, not a tag: style is one of theatrical / direct / reserve / alert, each mapped to an XTTS voice seed in STYLE_MAP. direct is used for plain facts (time/date/weather/car status — see car-parker.md) so delivery doesn't bury the number in personality.
• Tier routing (brain/router.py): a pure heuristic, no LLM call. Reflection/explanation/big-topic keywords (or a long input with no "fast" signal) route to the "slow" tier (bigger num_predict); reflex-command keywords (turn on/off, time/date/weather, timer/remind) route "fast". Both tiers currently run the *same* model — see brain/models.py for the model-swap history and rationale.
• Response flow (SAGA_FLOW env, runtime-switchable via GET /config?flow=): simple (one utterance, first word = full latency), canned (instant pre-recorded opener clip while the real answer computes in the background), live3b (a small model generates a contextual opener line instead of a canned clip). Three opener-clip tiers (short/medium/long) roughly match expected wait time to the intent (HA control resolves fast; other tool calls take a beat; non-tool reflection can run long).
• Synthesis: brain/tts.py posts text + a style seed to the XTTS server; tts_formatter.py strips markdown/lists before anything reaches XTTS.
• Exception: the five notes/reminders tools skip this whole narration pass — see notes-and-reminders.md's VERBATIM_TOOLS note for why.

Status & known limits

• Style selection leans heavily toward theatrical in practice (observed in an A/B test when the fast model was last swapped) — worth an ear kept on variety in real use.
• live3b/opener flows exist but the live deployment has mostly run simple or canned — live3b is less exercised.