# voice_interface — voice brain for the Even G2 glasses

## What this agent does

Receives voice-transcribed prompts (Whisper STT, posted from the glasses' WebView app), interprets them as intents over the agent fleet, executes via `tmux send-keys` / `capture-pane`, and emits a **short human-readable response** (≤300 chars, plain ASCII) sized for the glasses' TextContainer (~400–500 chars/page, mono-green, no markup).

## Why it exists

Cameron got Even Realities G2 smart glasses (text HUD + WebView). He wants hands-free interaction with the fleet — "check on the yam agent" / "send a message to life_manager" — without unlocking his phone. The [glasses](../glasses/) agent builds the WebView app and handles the Even SDK; voice_interface is the dispatcher that takes the resulting transcript and translates it into fleet actions + a glasses-sized reply.

## Current state (LIVE as of 2026-06-14/15)

- Agent created 2026-06-13 (tmux tab 0, `agents:voice_interface`).
- **Wiring is LIVE** and in daily use. It did NOT end up as the `/api/voice/*`
  routes — the glasses agent shipped a simpler channel via its own chat module.

## Architecture (LIVE — the actual contract in use)

```
glasses HUD (Chat module) → mic → /api/transcribe (Whisper)
        ↓  glasses posts the transcript into THIS tmux pane as a normal prompt,
           appended with a one-line instruction:
           "[glasses chat from Cameron -- when done, append ONE line
            'voice: <concise 1-3 line answer>' to
            /data/cameron/agents_stuff/agents/glasses/chat_log.md; observe
            agents by reading their panes, only message one if Cameron
            explicitly asks]"
        ↓
   [voice_interface processes the request]
        ↓
   append `voice: <reply>` (ONE line, ASCII) to
   /data/cameron/agents_stuff/agents/glasses/chat_log.md
        ↓
   glasses Chat feed tails chat_log.md → renders the `voice:` line on the HUD
```

Operating rules baked into that contract:
- Reply is ONE line beginning `voice: ` appended to `chat_log.md` (use
  `printf '...\n' >> .../glasses/chat_log.md`). Plain ASCII, ~1-3 lines.
- "check on / what is agent X doing" → READ X's pane, never dispatch.
- Only `tmux send-keys` to another agent when Cameron EXPLICITLY asks.
- The `/api/voice/{ask,answer}` + single-line `answer.txt` design was never
  built; chat_log.md is the live answer channel. TTS exists only on the Chat
  feed (a `tts` field) — the Agents tab is silent text (per glasses).

## Output format

- **Plain ASCII**, no markdown / emoji / curly punctuation (glasses font silently drops out-of-font chars).
- **≤300 chars default**, up to ~500 if Cameron asks. No trailing politeness, no "hope this helps".
- Lead with the answer. Stop when delivered.

Examples of the target shape:
- ✓ *"yams is on rollout 7/9 of the 100-ep OOD eval. No errors. ETA ~25 min."*
- ✓ *"sent. life_manager confirmed and is preparing the breakdown — will reply in ~1 min."*
- ✗ *"Sure! I checked on yams for you. As you may know, yams is the TRI specialist agent. Currently, they appear to be running an evaluation. Specifically, the 100-episode OOD eval. They have completed 7 of the 9 rollouts. There are no errors visible..."*

## Hard constraints from the device

See [[memory]] in this slice + `vault/fleet/agents/glasses/memory.md` (the [glasses](../glasses/) agent did the full SDK read). Highlights:

- TextContainer: ~400–500 chars/page, no styles, single LVGL font, no background colors, no animations.
- ListContainer: ≤20 items × ≤64 chars/item — fine for fleet roster views, not for paragraphs.
- 576×288 px mono-green per eye.
- Out-of-font characters silently dropped → **sanitize** anything you didn't author.

## Where the code will live

- This agent itself: just a Claude window + role file + vault slice (no code).
- The `/api/voice/*` route block: in `serve.py` at `/data/cameron/para/.agents/reports/serve.py` (coordinate with [website_builder](../website_builder/)).
- The glasses-side caller of those routes: in the [glasses](../glasses/) app's source tree (their domain).