feat(P3): mind/mouth split — separate voice model for the final reply (seam, default off)

The mind (chat backend/model) decides, reasons, and runs tools → a draft; the mouth
re-voices that draft in her character. Default: no mouth configured → the mind's
draft IS the reply, bit-for-bit the old behavior (and old streaming path untouched).

- config: MOUTH_BACKEND / MOUTH_MODEL. The slot for an eventual fine-tuned voice.
- chat: _mind_loop (tool/generation loop, non-stream, returns draft + tools_run),
  _voice_pass / mind.voice_messages (re-voice the draft, keep every fact/number),
  _mouth_target (active only when configured AND != mind). respond + respond_stream
  branch: mouth off = stream the mind directly (unchanged); mouth on = mind decides
  + runs tools, then the mouth streams the re-voiced reply. Falls back to the draft
  on any mouth failure (chat never breaks).
- Key payoff: the mouth needs no tool support (the mind handles tools), so it can be
  a non-tool character model (Dolphin / Claude / fine-tune). Makes the fine-tune
  easy: teach a small model to *sound* like Lyra, not to be smart.
- tests: mouth target on/off, voice_messages shape, voice_pass revoice+fallback.
  Suite 96 green, ruff clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.env.example

@@ -49,3 +49,5 @@ PING_AUTO_SALIENCE=0.8   # a thought this salient auto-pings even without an exp
 PING_COOLDOWN_MIN=60     # min minutes between AUTO pings (explicit reach-outs bypass)
 DIGEST_HOUR=18           # local hour to send her daily "what I've been thinking" digest
 CHAT_DELIBERATE=true   # think privately before answering substantive chat turns (false = faster, shallower)
+MOUTH_BACKEND=        # mind/mouth split: separate character/voice model for the final reply (empty = mind speaks)
+MOUTH_MODEL=

lyra/chat.py

@@ -1,9 +1,12 @@
 """The chat turn: assemble the prompt (lyra.mind) then speak + persist.
 
 `mind.assemble()` runs the society of parts (perceive → route → compose →
-deliberate) and hands back a ready message list + the active mode; `chat` runs the
-tool/generation loop (the "speak" part) and persists the exchange. Keeping speak
-here (not in mind) is deliberate — it's tangled with streaming and tool dispatch.
+deliberate) and hands back a ready message list + the active mode. Then:
+  - the MIND (the chat backend/model) runs the tool/generation loop — decide,
+    reason, run tools — and produces a draft.
+  - the MOUTH (a separate character model, if configured) re-voices that draft in
+    her own voice. Default: no mouth configured → the mind's draft IS the reply
+    (bit-for-bit the old behavior). The mouth slot is where a fine-tuned voice lands.
 """
 from __future__ import annotations
 
@@ -16,6 +19,7 @@ MAX_TOOL_ROUNDS = 5  # cap tool-call iterations per turn
 # tools when launched with --jinja; until it is, keep tools to cloud so MI50 chat
 # doesn't 500 on the tools param. Add "mi50" here once that flag is set.
 TOOL_BACKENDS = {"cloud"}
+_TANGLED = "(I got tangled using my tools there — say that again?)"
 
 
 def _resolve_model(backend: Backend, model_override: str | None, cfg) -> str:
@@ -29,15 +33,59 @@ def _resolve_model(backend: Backend, model_override: str | None, cfg) -> str:
     return model
 
 
+def _mouth_target(cfg, mind_backend: Backend, mind_model: str | None):
+    """The mouth (backend, model) if configured AND different from the mind; else None
+    (mouth == mind → no separate voice pass)."""
+    if not cfg.mouth_backend and not cfg.mouth_model:
+        return None
+    backend = cfg.mouth_backend or mind_backend
+    model = cfg.mouth_model or None
+    if backend == mind_backend and model == mind_model:
+        return None
+    return backend, model
+
+
 def _maybe_switch_mode(session_id: str, tool_name: str) -> None:
-    """Keep the chat framing aligned with the live data: opening a poker session
-    auto-flips this chat into Poker mode (next turn gets the card + full live tools).
-    Manual UI switching still overrides anytime."""
+    """Opening a poker session auto-flips this chat into Poker mode. Manual UI switching
+    still overrides anytime."""
     if tool_name == "start_session":
         memory.set_session_mode(session_id, modes.CASH.key)
         logbus.log("info", "mode auto-switch", session=session_id, mode=modes.CASH.key)
 
 
+def _mind_loop(messages, backend: Backend, model: str | None, tool_specs,
+               ctx: dict, session_id: str) -> tuple[str, list[str]]:
+    """Run the tool/generation loop on the MIND model (non-streaming). Mutates
+    `messages` with tool calls/results. Returns (draft_reply, tool_names_run)."""
+    tools_run: list[str] = []
+    reply = ""
+    for _ in range(MAX_TOOL_ROUNDS):
+        assistant_msg, tool_calls = llm.chat_call(
+            messages, backend=backend, model=model, tools=tool_specs
+        )
+        if not tool_calls:
+            reply = assistant_msg.get("content") or ""
+            break
+        messages.append(assistant_msg)
+        for tc in tool_calls:
+            result = toolkit.dispatch(tc["name"], tc["arguments"], ctx)
+            logbus.log("info", "tool call", session=session_id, tool=tc["name"], result=result[:80])
+            messages.append({"role": "tool", "tool_call_id": tc["id"], "content": result})
+            _maybe_switch_mode(session_id, tc["name"])
+            tools_run.append(tc["name"])
+    return reply, tools_run
+
+
+def _voice_pass(messages, draft: str, backend: Backend, model: str | None) -> str:
+    """Mouth: re-render the mind's draft in her voice. Falls back to the draft on failure."""
+    try:
+        out = llm.complete(mind.voice_messages(messages, draft), backend=backend, model=model)
+        return (out or "").strip() or draft
+    except Exception as exc:
+        logbus.log("error", "voice pass failed", error=str(exc)[:160])
+        return draft
+
+
 def respond(session_id: str, user_msg: str, backend: Backend = "cloud",
             model_override: str | None = None) -> str:
     """Produce Lyra's reply to a single user message and persist the exchange."""
@@ -48,28 +96,16 @@ def respond(session_id: str, user_msg: str, backend: Backend = "cloud",
 
     turn = mind.assemble(session_id, user_msg, backend, model)
     messages = turn.messages
-
-    # Tool loop (speak): offer her tools (scoped to the mode); run any she calls and
-    # feed results back until she returns a text reply.
     tool_specs = toolkit.specs(turn.mode.tools) if backend in TOOL_BACKENDS else None
     ctx = {"session_id": session_id, "backend": backend}
-    reply = ""
-    for _ in range(MAX_TOOL_ROUNDS):
-        assistant_msg, tool_calls = llm.chat_call(
-            messages, backend=backend, model=model, tools=tool_specs
-        )
-        if not tool_calls:
-            reply = assistant_msg.get("content") or ""
-            break
-        messages.append(assistant_msg)  # her tool-call request
-        for tc in tool_calls:
-            result = toolkit.dispatch(tc["name"], tc["arguments"], ctx)
-            logbus.log("info", "tool call", session=session_id, tool=tc["name"], result=result[:80])
-            messages.append({"role": "tool", "tool_call_id": tc["id"], "content": result})
-            _maybe_switch_mode(session_id, tc["name"])
+
+    reply, _ = _mind_loop(messages, backend, model, tool_specs, ctx, session_id)
+    mouth = _mouth_target(cfg, backend, model)
+    if mouth and reply:
+        reply = _voice_pass(messages, reply, *mouth)
     if not reply:
-        reply = "(I got tangled using my tools there — say that again?)"
-    logbus.log("info", "reply", session=session_id, chars=len(reply))
+        reply = _TANGLED
+    logbus.log("info", "reply", session=session_id, chars=len(reply), voiced=bool(mouth))
 
     memory.remember(session_id, "user", user_msg)
     memory.remember(session_id, "assistant", reply)
@@ -79,11 +115,8 @@ def respond(session_id: str, user_msg: str, backend: Backend = "cloud",
 
 def respond_stream(session_id: str, user_msg: str, backend: Backend = "cloud",
                    model_override: str | None = None):
-    """Streaming generator version of `respond`.
-
-    Yields ("delta", text) as content streams in, ("tool", name) when a tool runs,
-    and a final ("done", reply). Persists the exchange — same side effects as `respond`.
-    """
+    """Streaming generator version of `respond`. Yields ("delta", text), ("tool", name),
+    and a final ("done", reply). Same side effects as `respond`."""
     cfg = config.load()
     model = _resolve_model(backend, model_override, cfg)
     logbus.log("info", "chat request (stream)", session=session_id, backend=backend,
@@ -93,36 +126,57 @@ def respond_stream(session_id: str, user_msg: str, backend: Backend = "cloud",
     messages = turn.messages
     tool_specs = toolkit.specs(turn.mode.tools) if backend in TOOL_BACKENDS else None
     ctx = {"session_id": session_id, "backend": backend}
-    parts: list[str] = []
-    for _ in range(MAX_TOOL_ROUNDS):
-        assistant_msg = None
-        tool_calls = None
-        for ev, payload in llm.chat_call_stream(
-            messages, backend=backend, model=model, tools=tool_specs
-        ):
-            if ev == "delta":
-                parts.append(payload)
-                yield ("delta", payload)
-            elif ev == "message":
-                assistant_msg = payload
-            elif ev == "tool_calls":
-                tool_calls = payload
-        if not tool_calls:
-            break
-        messages.append(assistant_msg)  # her tool-call request
-        for tc in tool_calls:
-            result = toolkit.dispatch(tc["name"], tc["arguments"], ctx)
-            logbus.log("info", "tool call", session=session_id, tool=tc["name"], result=result[:80])
-            messages.append({"role": "tool", "tool_call_id": tc["id"], "content": result})
-            _maybe_switch_mode(session_id, tc["name"])
-            yield ("tool", tc["name"])
+    mouth = _mouth_target(cfg, backend, model)
 
-    reply = "".join(parts)
-    if not reply:
-        reply = "(I got tangled using my tools there — say that again?)"
-        yield ("delta", reply)
-    logbus.log("info", "reply", session=session_id, chars=len(reply))
+    if mouth is None:
+        # No separate voice: stream the mind directly (the original path, unchanged).
+        parts: list[str] = []
+        for _ in range(MAX_TOOL_ROUNDS):
+            assistant_msg = None
+            tool_calls = None
+            for ev, payload in llm.chat_call_stream(
+                messages, backend=backend, model=model, tools=tool_specs
+            ):
+                if ev == "delta":
+                    parts.append(payload)
+                    yield ("delta", payload)
+                elif ev == "message":
+                    assistant_msg = payload
+                elif ev == "tool_calls":
+                    tool_calls = payload
+            if not tool_calls:
+                break
+            messages.append(assistant_msg)
+            for tc in tool_calls:
+                result = toolkit.dispatch(tc["name"], tc["arguments"], ctx)
+                logbus.log("info", "tool call", session=session_id, tool=tc["name"], result=result[:80])
+                messages.append({"role": "tool", "tool_call_id": tc["id"], "content": result})
+                _maybe_switch_mode(session_id, tc["name"])
+                yield ("tool", tc["name"])
+        reply = "".join(parts)
+        if not reply:
+            reply = _TANGLED
+            yield ("delta", reply)
+    else:
+        # Mind decides + runs tools (non-streamed); mouth re-voices, streamed.
+        draft, tools_run = _mind_loop(messages, backend, model, tool_specs, ctx, session_id)
+        for name in tools_run:
+            yield ("tool", name)
+        parts = []
+        try:
+            for ev, payload in llm.chat_call_stream(
+                mind.voice_messages(messages, draft), backend=mouth[0], model=mouth[1], tools=None
+            ):
+                if ev == "delta":
+                    parts.append(payload)
+                    yield ("delta", payload)
+        except Exception as exc:
+            logbus.log("error", "voice stream failed", error=str(exc)[:160])
+        reply = "".join(parts).strip() or draft or _TANGLED
+        if not parts:
+            yield ("delta", reply)
 
+    logbus.log("info", "reply", session=session_id, chars=len(reply), voiced=bool(mouth))
     memory.remember(session_id, "user", user_msg)
     memory.remember(session_id, "assistant", reply)
     summary.maybe_summarize_async(session_id)

lyra/config.py

@@ -38,6 +38,11 @@ class Config:
     ping_quiet_hours: str  # local "start-end" 24h window to stay silent, e.g. "1-9"
     digest_hour: int       # local hour (0-23) to send her daily "what I've been thinking" digest
     chat_deliberate: bool  # think privately before answering substantive chat turns
+    # Mind/mouth split: the mind (the chat backend/model above) decides, reasons, and
+    # runs tools; the mouth re-voices the final reply in her character. Empty = mouth
+    # is the mind (no separate pass) — the slot for an eventual fine-tuned voice.
+    mouth_backend: str
+    mouth_model: str | None
     # External input feed (her #1: react to the world). Comma-separated RSS/Atom URLs.
     feeds: tuple[str, ...]
     feed_react_prob: float  # chance a would-be new thread reacts to a feed item instead
@@ -81,6 +86,8 @@ def load() -> Config:
         ping_quiet_hours=os.getenv("PING_QUIET_HOURS", "1-9"),
         digest_hour=int(os.getenv("DIGEST_HOUR", "18")),
         chat_deliberate=os.getenv("CHAT_DELIBERATE", "true").lower() not in ("0", "false", "no"),
+        mouth_backend=os.getenv("MOUTH_BACKEND", "").lower(),
+        mouth_model=os.getenv("MOUTH_MODEL") or None,
         feeds=_csv("LYRA_FEEDS", "https://hnrss.org/frontpage,https://www.pokernews.com/rss.php"),
         feed_react_prob=float(os.getenv("FEED_REACT_PROB", "0.5")),
     )

lyra/mind.py

@@ -290,6 +290,25 @@ def _deliberate_part(ctx: TurnContext) -> TurnContext:
 PIPELINE = (_perceive, _route, _compose, _deliberate_part)
 
 
+# --- mouth (the voice pass: re-render the mind's draft in her character) -----
+
+_VOICE_NOTE = (
+    "↑ That was you working the answer out — a draft Brian has NOT seen. Now say it to him "
+    "in your own voice: warm, direct, specific, in character, opinionated. Keep every fact, "
+    "number, name, and decision exactly as in the draft — change only the wording so it sounds "
+    "like you, not a generic assistant. No preamble, no meta, no 'here's a friendlier version' "
+    "— just your actual message to Brian."
+)
+
+
+def voice_messages(messages: list[Message], draft: str) -> list[Message]:
+    """Prompt for the mouth model: the full turn context + the mind's draft to re-voice."""
+    return messages + [
+        {"role": "assistant", "content": draft},
+        {"role": "system", "content": _VOICE_NOTE},
+    ]
+
+
 def assemble(session_id: str, user_msg: str, backend: Backend,
              model: str | None = None) -> TurnContext:
     """Run the parts over a fresh TurnContext and return it ready for `chat` to speak."""

tests/test_chat.py

@@ -61,3 +61,46 @@ def test_assemble_runs_the_pipeline(lyra, monkeypatch):
     assert turn.mode is not None                       # route ran
     assert turn.messages and turn.messages[-1]["role"] == "user"   # compose ran
     assert turn.messages[-1]["content"] == "hey what's up"
+
+
+# --- mind/mouth split (P3) ----------------------------------------------
+
+def test_mouth_target_off_by_default(monkeypatch):
+    import importlib
+    from lyra import config
+    monkeypatch.delenv("MOUTH_BACKEND", raising=False)
+    monkeypatch.delenv("MOUTH_MODEL", raising=False)
+    import lyra.chat as chat
+    importlib.reload(chat)
+    assert chat._mouth_target(config.load(), "cloud", "gpt-4o") is None  # mouth == mind
+
+
+def test_mouth_target_when_configured(monkeypatch):
+    import importlib
+    from lyra import config
+    monkeypatch.setenv("MOUTH_BACKEND", "local")
+    monkeypatch.setenv("MOUTH_MODEL", "dolphin3:8b")
+    import lyra.chat as chat
+    importlib.reload(chat)
+    assert chat._mouth_target(config.load(), "cloud", "gpt-4o") == ("local", "dolphin3:8b")
+
+
+def test_voice_messages_carries_draft_and_instruction(lyra):
+    _, mind = lyra
+    out = mind.voice_messages([{"role": "user", "content": "hi"}], "draft with FACT 42")
+    assert out[-2] == {"role": "assistant", "content": "draft with FACT 42"}
+    assert out[-1]["role"] == "system" and "your own voice" in out[-1]["content"].lower()
+
+
+def test_voice_pass_revoices_then_falls_back(lyra, monkeypatch):
+    _, mind = lyra
+    import importlib
+    import lyra.chat as chat
+    importlib.reload(chat)
+    monkeypatch.setattr(chat.llm, "complete", lambda msgs, backend=None, model=None: "voiced (FACT 42)")
+    assert chat._voice_pass([], "draft FACT 42", "local", "dolphin3:8b") == "voiced (FACT 42)"
+    # on failure it keeps the mind's draft (chat must not break)
+    def boom(*a, **k):
+        raise RuntimeError("mouth down")
+    monkeypatch.setattr(chat.llm, "complete", boom)
+    assert chat._voice_pass([], "draft FACT 42", "local", "dolphin3:8b") == "draft FACT 42"