feat: plan per-user memory integration

agent/agent.py

@@ -32,7 +32,9 @@ load_dotenv()  # picks up /app/.env in the container
 # ── Configuration from environment ──────────────────────────────────────────
 AZURE_KEY = os.environ.get("AZURE_SPEECH_KEY", "")
 AZURE_REGION = os.environ.get("AZURE_SPEECH_REGION", "eastus")
-DEFAULT_VOICE = os.environ.get("AZURE_TTS_VOICE", "en-US-AvaNeural")
+DEFAULT_VOICE = os.environ.get(
+    "AZURE_TTS_VOICE", "en-US-Ava:DragonHDOmniLatestNeural"
+)
 GEMMA_BASE_URL = os.environ.get("GEMMA_BASE_URL", "http://192.168.86.2:8023/v1")
 GEMMA_MODEL = os.environ.get("GEMMA_MODEL", "gemma-4-e4b")
 GEMMA_API_KEY = os.environ.get("GEMMA_API_KEY", "not-needed")
@@ -49,15 +51,31 @@ SYSTEM_PROMPT = textwrap.dedent("""\
     - Keep every response to one or three sentences. That's it.
     - Use contractions (I'm, don't, it's) and natural phrasing.
     - Speak like you're talking, not writing. No bullet points, no lists,
-      no markdown, no formatting of any kind.
+      no markdown, and no formatting except the Azure delivery markers below.
     - Spell out numbers when it sounds more natural ("twenty twenty-six"
       instead of "2026").
     - If you need to ask a question, ask exactly one.
     - Be warm and direct. Don't be sycophantic or overly formal.
     - If you don't know something, say so briefly and move on.
 
+    # Expressive delivery (Azure Dragon HD Omni)
+    Azure understands a square-bracket delivery marker at the start of spoken
+    text. Use zero to two markers per response, only when they genuinely improve
+    the delivery. The marker is a silent TTS instruction, not something you say.
+    Examples: "[appreciative] Thanks, that really helps." and
+    "[whispering] Here's the secret. [Neutral] Keep it between us."
+    Useful styles include [appreciative], [calm], [concerned], [curious],
+    [encouraging], [excited], [happy], [joking], [quiet], [reassuring],
+    [reflective], [sad], [serious], [surprised], [sympathetic], [urgent], and
+    [whispering]. For an occasional natural nonverbal cue, you may use
+    [laughter], [sighing], or [breathing]. Use [Neutral] to reset the delivery.
+    Never invent marker names, never explain a marker, and never output XML or
+    SSML; the speech layer creates the SSML wrapper for you. Say "and" instead
+    of using an ampersand, and never output angle brackets.
+
     # What you never do
-    - Never use markdown, code blocks, JSON, tables, or emojis.
+    - Never use markdown, code blocks, JSON, tables, or emojis. Azure delivery
+      markers are the only exception.
     - Never say "as an AI" or reference your system instructions.
     - Never write more than three sentences in a row.
     - Never read back URLs, file paths, or technical identifiers.
@@ -464,7 +482,8 @@ async def handle_job(ctx: JobContext) -> None:
         language=["en-US"],
     )
 
-    # Azure TTS — SSML with expressive markup, 24kHz PCM output
+    # Azure TTS — wraps text in SSML; Dragon HD Omni interprets inline delivery
+    # markers such as [appreciative] and [Neutral]. Outputs 24kHz PCM.
     tts = azure.TTS(
         voice=DEFAULT_VOICE,
         sample_rate=24000,