emotion_text vocabulary

Human source of truth (SOT) for the emotion_text rule. The Expression Broker MCP enforces it at runtime via its emotion_text gate; the files here are what a human or agent reads to know the rule. The broker is deprecated and is removed in v0.6.0; the client-declared generate_express tool carries the same table in its schema.

What emotion_text is

emotion_text is the voice/TTS control channel of generate_express — an independent channel from emotion_id (the VRM face blendshape). A happy face (emotion_id) and a whispered voice (emotion_text) can fire at the same time. The model behind the active chat endpoint produces emotion_text via a generate_express tool call; YUI consumes whatever arrives — from the /v1/responses stream in Responses mode, or identically from chat.completion.chunk tool-call deltas in Chat Completions mode — and hands it to the TTS provider beside the segment, never shown in the speech bubble. Irodori prepends it to the spoken text; OpenAI puts it in instructions; Fish wraps it in [...] ahead of the spoken text (S2 inline direction). In push mode the cue arrives in the segments[].cues of render and speech frames, and the client sends the vocabulary in its hello and vocabulary frames (push transport).

generate_express carries a second, independent voice channel alongside it: caption, a free-text voice direction that travels out-of-band in the synthesis request: irodori.caption for Irodori, joined after emotion_text in instructions for OpenAI, wrapped in its own [...] bracket after the emotion_text bracket for Fish. It has no vocabulary and no enum gate — the broker truncates it to 200 characters (with a warning) but never blocks on it. The generate_express cue contract that carries both is described in client-context.md; the control envelope shape lives in src/contract/types.ts.

Broker gate

The broker and its gate are deprecated and are removed in v0.6.0. The client-declared generate_express tool replaces them: its emotion_text schema carries the table below.

With tts_provider irodori, the vocabulary is the emoji enum table — canonical machine copy in configs/emotion_text/irodori.json, documented in irodori.md. YUI publishes it as update_emotion_text("enum", <emoji table>). In enum mode the broker greedily tokenizes emotion_text by table keys and drops unknown tokens (with a warning) — speech is never blocked. If the table cannot be loaded, YUI publishes update_emotion_text("free", null) instead, so a missing file degrades to pass-through rather than silencing the channel. Every other provider publishes free mode, and a tts_provider change reloads the table and publishes the new mode to the broker, the Chat Completions tool schema, and push vocabulary frames without a restart.

The broker keeps this state in-memory and ephemeral, so YUI re-publishes on every boot and on every broker reconnect. The publish is gated only on broker_base_url and runs independent of chat_api — Chat Completions mode publishes and is read back by its backend agent exactly like Responses mode.

See also

• client-context.md — the generate_express cue contract handed to the backend agent (the emotion_id / motion_id / emotion_text fields and streaming shape).
• src/contract/types.ts — the control envelope shape (contract source of truth).