[eric] voice: aux polish endpoint (fillers+punctuation, fail-open) on the cheap tier

backend/apps/voice/polish.py

@@ -0,0 +1,82 @@
+"""Dictation cleanup: raw whisper text -> punctuated, filler-free prose via the cheap aux tier.
+
+One fast call on whatever lane the user has configured (provider-agnostic per the aux registry).
+Every failure path returns the RAW text so dictation never breaks when the aux is unreachable.
+"""
+
+from contextlib import asynccontextmanager
+from typing import AsyncIterator, Optional
+
+from pydantic import BaseModel, ConfigDict
+from typeguard import typechecked
+
+from backend.config.Apps import SubApp
+
+
+@asynccontextmanager
+async def voice_lifespan() -> AsyncIterator[None]:
+    yield
+
+
+voice = SubApp("voice", voice_lifespan)
+
+P_POLISH_SYSTEM = (
+    "You clean up raw speech-to-text dictation. Return ONLY the cleaned text, nothing else. "
+    "Fix punctuation, capitalization, and obvious homophone errors. Remove filler words (um, uh, "
+    "like when used as filler, you know) and false starts. Apply spoken formatting commands: "
+    "'new line'/'new paragraph' become real breaks, 'period'/'comma'/'question mark' become the "
+    "mark when clearly dictated as punctuation. NEVER add content, never answer questions in the "
+    "text, never translate, never wrap in quotes, never use em-dashes. Keep the speaker's words "
+    "and tone; this is transcription cleanup, not rewriting."
+)
+
+POLISH_INPUT_CAP = 8_000
+
+
+class PolishRequest(BaseModel):
+    model_config = ConfigDict(validate_assignment=True)
+
+    text: str
+    # A one-line hint about where the user is dictating (e.g. a page title), so names spell right.
+    context: Optional[str] = None
+
+
+class PolishResponse(BaseModel):
+    model_config = ConfigDict(validate_assignment=True)
+
+    text: str
+    polished: bool
+
+
+@voice.router.post("/polish")
+@typechecked
+async def polish(body: PolishRequest) -> dict:
+    raw = (body.text or "").strip()
+    if not raw:
+        return PolishResponse(text="", polished=False).model_dump()
+    try:
+        from backend.apps.agents.core.aux_llm import aux_max_tokens_for, safe_resp_text
+        from backend.apps.agents.providers.registry import resolve_aux_model
+        from backend.apps.settings.credentials import get_anthropic_client_for_model
+        from backend.apps.settings.store import load_settings
+
+        settings = load_settings()
+        aux_model, _ = await resolve_aux_model(settings, preferred_tier="haiku")
+        client = get_anthropic_client_for_model(settings, aux_model)
+        prompt = raw[:POLISH_INPUT_CAP]
+        if body.context:
+            prompt = f"[Dictating into: {body.context[:200]}]\n{prompt}"
+        resp = await client.messages.create(
+            model=aux_model,
+            max_tokens=aux_max_tokens_for(aux_model, base=1000),
+            system=P_POLISH_SYSTEM,
+            messages=[{"role": "user", "content": prompt}],
+            # Dictation is interactive; a slow aux must never hold the paste hostage.
+            timeout=6.0,
+        )
+        cleaned = safe_resp_text(resp).strip()
+        if cleaned:
+            return PolishResponse(text=cleaned, polished=True).model_dump()
+    except Exception:
+        pass
+    return PolishResponse(text=raw, polished=False).model_dump()

backend/main.py

@@ -43,6 +43,7 @@ from backend.apps.subscription.router import subscription
 from backend.apps.auth.router import auth
 from backend.apps.web.web import web
 from backend.apps.onboarding.onboarding import onboarding
+from backend.apps.voice.polish import voice
 from backend.apps.agents.proxy.anthropic_proxy import anthropic_proxy
 from backend.apps.agents.core.openai_passthrough import openai_passthrough
 from backend.apps.workflows.workflows import workflows
@@ -50,7 +51,7 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi import WebSocket, WebSocketDisconnect
 import json
 
-main_app = MainApp([health, agents, skills, tools_lib, modes, settings, mcp_registry, skill_registry, outputs, output_versions, dashboards, swarm, service, subscription, auth, web, onboarding, anthropic_proxy, workflows, openai_passthrough])
+main_app = MainApp([health, agents, skills, tools_lib, modes, settings, mcp_registry, skill_registry, outputs, output_versions, dashboards, swarm, service, subscription, auth, web, onboarding, voice, anthropic_proxy, workflows, openai_passthrough])
 app = main_app.app
 
 # Generate per-install auth token BEFORE we bind the HTTP port. By the time any request lands, the token file exists. See backend/auth.py.