feat: piper now not a subprocess, cache system for TTS

1 files changed, 5 insertions, 1 deletions

diff --git a/AGENTS.md b/AGENTS.md
index a074a43..8352e0e 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -23,7 +23,7 @@ src/
 ├── events/mod.rs        — handle_ready, handle_voice_state_update, handle_message
 ├── openrouter/mod.rs    — OpenRouterClient (chat, chat_with_history)
 ├── tasks/mod.rs         — cron_scheduler_loop (30s polling, stub exec bodies)
-├── tts.rs               — synthesize(): Piper subprocess → Google TTS fallback
+├── tts.rs               — TtsState (piper-rs in-process, DashMap cache, pre-gen, resampler, WAV encoder)
 └── voice/mod.rs         — VoiceManager (connect, disconnect, get_call), mp3_to_input
 ```
 
@@ -35,6 +35,10 @@ src/
 - **Shared state**: `Data` struct in `main.rs` — passed as poise `user_data`
 - **Commands**: each submodule exports `pub fn commands() -> Vec<poise::Command<Data, Error>>`
 - **Temp path**: always use `crate::config::TEMP_PATH` (resolves to `/tmp/senator_sperma`)
+- **TTS**: piper-rs in-process (no subprocess). `TtsState` in `Data` holds `OnceLock<Mutex<Piper>>` + `DashMap` cache. Synthesize via `data.tts.synthesize(text)`. Pre-generates 6 event phrases per user on first voice event. No Google TTS fallback.
+
+## Build
+- `build.rs` — compiles C stubs for espeak-rs-sys missing sonic/audio_object symbols
 
 ## Stubs to Fill (see COMPARISON.md)
 - `commands/music.rs` — `/play` is placeholder