Skip to content

:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.

chore(posthog-ai): remove hands-free voice mode (#115963)

Shy Alter15a33e54189e
diff --git a/ee/api/hands_free.py b/ee/api/hands_free.py
deleted file mode 100644
index daedb498c2e0..000000000000
--- a/ee/api/hands_free.py
+++ /dev/null
@@ -1,249 +0,0 @@
-"""Backend support for Max's hands-free mode.
-
-Hands-free mode lets a user talk to Max with their hands and eyes off the screen — built
-for mobile-web (gym, walking, driving). Speech-to-text runs via ElevenLabs Scribe directly
-from the browser using single-use tokens minted here so the ElevenLabs API key never
-reaches the client. Text-to-speech is proxied through the synthesize endpoint, again to
-keep the API key server-side; mp3 bytes stream back to the browser.
-"""
-
-from typing import Any
-
-from django.conf import settings
-from django.http import StreamingHttpResponse
-
-import requests
-import structlog
-from drf_spectacular.utils import extend_schema
-from prometheus_client import Counter
-from rest_framework import serializers, status
-from rest_framework.decorators import action
-from rest_framework.exceptions import APIException
-from rest_framework.parsers import JSONParser
-from rest_framework.permissions import IsAuthenticated
-from rest_framework.request import Request
-from rest_framework.response import Response
-from rest_framework.viewsets import GenericViewSet
-
-from posthog.api.routing import TeamAndOrgViewSetMixin
-from posthog.api.streaming import streaming_response
-from posthog.permissions import PostHogFeatureFlagPermission
-from posthog.rate_limit import (
- MaxHandsFreeSynthesizeBurstRateThrottle,
- MaxHandsFreeSynthesizeSustainedRateThrottle,
- MaxHandsFreeTokenBurstRateThrottle,
- MaxHandsFreeTokenSustainedRateThrottle,
-)
-
-from products.posthog_ai.backend.models.assistant import Conversation
-
-logger = structlog.get_logger(__name__)
-
-ELEVENLABS_TIMEOUT_SECONDS = (5, 10)
-# Generous enough for any Max summary, tight enough that an attacker minting at the
-# throttle ceiling can't ramp Scribe cost into the stratosphere.
-TTS_MAX_INPUT_CHARS = 2000
-
-# One counter, partitioned by outcome — lets us track abuse signal (provider rejection
-# spikes, repeated empty tokens) and steady-state cost (ok counts ≈ Scribe spend) without
-# blowing up cardinality via per-team labels.
-HANDS_FREE_TOKEN_COUNTER = Counter(
- "max_hands_free_token_total",
- "Outcomes for ElevenLabs Scribe single-use token mints from the hands-free endpoint.",
- labelnames=["outcome"],
-)
-HANDS_FREE_SYNTHESIZE_COUNTER = Counter(
- "max_hands_free_synthesize_total",
- "Outcomes for ElevenLabs TTS proxy requests from the hands-free endpoint.",
- labelnames=["outcome"],
-)
-HANDS_FREE_SYNTHESIZE_CHARS_COUNTER = Counter(
- "max_hands_free_synthesize_chars_total",
- "Total characters submitted for ElevenLabs TTS — proxy for TTS spend.",
-)
-
-
-class HandsFreeNotConfigured(APIException):
- status_code = status.HTTP_503_SERVICE_UNAVAILABLE
- default_detail = "Hands-free mode is not configured on this PostHog instance."
- default_code = "hands_free_not_configured"
-
-
-class HandsFreeProviderError(APIException):
- status_code = status.HTTP_502_BAD_GATEWAY
- default_detail = "The hands-free provider rejected the request."
- default_code = "hands_free_provider_error"
-
-
-class SynthesizeSerializer(serializers.Serializer):
- text = serializers.CharField(
- required=True,
- max_length=TTS_MAX_INPUT_CHARS,
- help_text="The text the assistant should speak aloud.",
- )
-
-
-class HandsFreeTokenSerializer(serializers.Serializer):
- token = serializers.CharField(help_text="Single-use ElevenLabs Scribe realtime token, valid for 15 minutes.")
-
-
-def _require_api_key(counter: Counter) -> str:
- api_key = settings.ELEVENLABS_API_KEY
- if not api_key:
- counter.labels(outcome="missing_key").inc()
- logger.warning("max_hands_free_api_key_missing")
- raise HandsFreeNotConfigured()
- return api_key
-
-
-@extend_schema(extensions={"x-product": "posthog_ai"})
-class MaxHandsFreeViewSet(TeamAndOrgViewSetMixin, GenericViewSet):
- scope_object = "INTERNAL"
- # Hands-free actions are list-level and don't operate on any model — DRF's GenericViewSet
- # needs a queryset attribute for the routing scaffolding but it is never read.
- queryset = Conversation.objects.none()
- # drf-spectacular requires every viewset to declare a serializer_class for OpenAPI
- # generation. The `token` action takes no body and `synthesize` overrides this via
- # its @extend_schema decorator, so SynthesizeSerializer is purely the schema-default
- # — never instantiated by the viewset's own get_serializer().
- serializer_class = SynthesizeSerializer
- # Server-side gate on the same feature flag the frontend checks — without it an
- # authenticated project member could POST directly to /token/ or /synthesize/ and
- # rack up ElevenLabs spend even when their org doesn't have hands-free enabled.
- posthog_feature_flag = "max-hands-free"
- permission_classes = [IsAuthenticated, PostHogFeatureFlagPermission]
-
- @extend_schema(request=None, responses={200: HandsFreeTokenSerializer})
- @action(
- detail=False,
- methods=["POST"],
- url_path="token",
- throttle_classes=[MaxHandsFreeTokenBurstRateThrottle, MaxHandsFreeTokenSustainedRateThrottle],
- )
- def token(self, request: Request, *args: Any, **kwargs: Any) -> Response:
- """Mint a single-use ElevenLabs Scribe realtime token.
-
- The browser uses the token to open a WebSocket directly to ElevenLabs — audio never
- transits PostHog infrastructure. Tokens are time-bound (15 min) and single-use; the
- per-team rate limit on this endpoint caps how often a user can mint new ones.
-
- Never logs the upstream response body — provider error responses can echo PII back and
- we don't want any of that landing in structured logs.
- """
- api_key = _require_api_key(HANDS_FREE_TOKEN_COUNTER)
- try:
- upstream = requests.post(
- f"{settings.ELEVENLABS_API_BASE_URL}/v1/single-use-token/realtime_scribe",
- headers={"xi-api-key": api_key},
- timeout=ELEVENLABS_TIMEOUT_SECONDS,
- )
- except requests.RequestException:
- HANDS_FREE_TOKEN_COUNTER.labels(outcome="provider_unreachable").inc()
- logger.exception("max_hands_free_token_failed")
- raise HandsFreeProviderError("Failed to reach the hands-free provider.")
-
- if upstream.status_code != status.HTTP_200_OK:
- HANDS_FREE_TOKEN_COUNTER.labels(outcome="provider_rejected").inc()
- logger.warning("max_hands_free_token_rejected", status_code=upstream.status_code)
- raise HandsFreeProviderError(f"Hands-free provider returned {upstream.status_code}.")
-
- try:
- body = upstream.json()
- if not isinstance(body, dict):
- raise ValueError("Expected JSON object")
- token: str = body.get("token", "")
- except ValueError:
- HANDS_FREE_TOKEN_COUNTER.labels(outcome="provider_rejected").inc()
- logger.warning("max_hands_free_token_malformed_json", status_code=upstream.status_code)
- raise HandsFreeProviderError("Hands-free provider returned a malformed response.")
- if not token:
- HANDS_FREE_TOKEN_COUNTER.labels(outcome="empty_token").inc()
- logger.warning("max_hands_free_token_empty")
- raise HandsFreeProviderError("Hands-free provider returned an empty token.")
- HANDS_FREE_TOKEN_COUNTER.labels(outcome="ok").inc()
- return Response({"token": token})
-
- # Excluded on purpose: the response is streamed audio/mpeg, and Orval's shared mutator parses
- # every response as JSON. The caller in lib/api.ts uses api.createResponse to read the stream.
- @extend_schema(exclude=True)
- @action(
- detail=False,
- methods=["POST"],
- url_path="synthesize",
- parser_classes=[JSONParser],
- throttle_classes=[MaxHandsFreeSynthesizeBurstRateThrottle, MaxHandsFreeSynthesizeSustainedRateThrottle],
- )
- def synthesize(self, request: Request, *args: Any, **kwargs: Any) -> StreamingHttpResponse:
- """Proxy text-to-speech to ElevenLabs, streaming mp3 audio back to the browser.
-
- The viewset has no per-action `parser_classes` other than this one because the
- token endpoint takes no body. Putting JSONParser here keeps the rest of the
- viewset parser-free.
- """
- serializer = SynthesizeSerializer(data=request.data)
- serializer.is_valid(raise_exception=True)
- text = serializer.validated_data["text"]
-
- api_key = _require_api_key(HANDS_FREE_SYNTHESIZE_COUNTER)
- voice_id = settings.ELEVENLABS_VOICE_ID
- if not voice_id:
- HANDS_FREE_SYNTHESIZE_COUNTER.labels(outcome="missing_voice_id").inc()
- logger.warning("max_hands_free_voice_id_missing")
- raise HandsFreeNotConfigured("Voice ID is not configured.")
- try:
- upstream = requests.post(
- f"{settings.ELEVENLABS_API_BASE_URL}/v1/text-to-speech/{voice_id}",
- headers={
- "xi-api-key": api_key,
- "Accept": "audio/mpeg",
- "Content-Type": "application/json",
- },
- json={"text": text, "model_id": settings.ELEVENLABS_TTS_MODEL_ID},
- timeout=ELEVENLABS_TIMEOUT_SECONDS,
- stream=True,
- )
- except requests.RequestException:
- HANDS_FREE_SYNTHESIZE_COUNTER.labels(outcome="provider_unreachable").inc()
- logger.exception("max_hands_free_synthesize_failed")
- raise HandsFreeProviderError("Failed to reach the hands-free provider.")
-
- if upstream.status_code != status.HTTP_200_OK:
- HANDS_FREE_SYNTHESIZE_COUNTER.labels(outcome="provider_rejected").inc()
- # ElevenLabs 4xx responses are JSON describing the account/quota state — not the
- # user's input. Safe to log a snippet so devs can see "quota_exceeded" vs
- # "voice_not_found" without hitting the dashboard.
- body_preview = ""
- try:
- body_preview = upstream.text[:300].replace("\n", " ")
- except Exception:
- pass
- logger.warning(
- "max_hands_free_synthesize_rejected",
- status_code=upstream.status_code,
- body_preview=body_preview,
- )
- upstream.close()
- raise HandsFreeProviderError(f"Hands-free provider returned {upstream.status_code}.")
-
- # Anything between the upstream 200 and returning the StreamingHttpResponse can
- # raise (a counter labels() call, the StreamingHttpResponse constructor, the
- # header assignment). Without an explicit close on those paths the connection
- # only releases when the unconsumed generator is garbage collected — which is
- # not guaranteed to be prompt under load. Guard the whole block.
- try:
- HANDS_FREE_SYNTHESIZE_COUNTER.labels(outcome="ok").inc()
- HANDS_FREE_SYNTHESIZE_CHARS_COUNTER.inc(len(text))
-
- def stream_and_close() -> Any:
- try:
- yield from upstream.iter_content(chunk_size=4096)
- finally:
- upstream.close()
-
- response = streaming_response(
- stream_and_close(), content_type="audio/mpeg", headers={"Cache-Control": "no-store"}
- )
- except Exception:
- upstream.close()
- raise
- return response
diff --git a/ee/api/test/test_hands_free.py b/ee/api/test/test_hands_free.py
deleted file mode 100644
index 44ba7eae4ad8..000000000000
--- a/ee/api/test/test_hands_free.py
+++ /dev/null
@@ -1,149 +0,0 @@
-from posthog.test.base import APIBaseTest
-from unittest.mock import MagicMock, patch
-
-from django.test import override_settings
-
-
-@override_settings(
- ELEVENLABS_API_KEY="test-key",
- ELEVENLABS_API_BASE_URL="https://api.elevenlabs.example",
-)
-class TestMaxHandsFreeAPI(APIBaseTest):
- def setUp(self) -> None:
- super().setUp()
- # The viewset gates on the max-hands-free feature flag server-side; force it on
- # so tests exercise the endpoint logic rather than the entitlement check.
- self._ff_patcher = patch("posthoganalytics.feature_enabled", return_value=True)
- self._ff_patcher.start()
- self.addCleanup(self._ff_patcher.stop)
-
- def _token_url(self) -> str:
- return f"/api/environments/{self.team.id}/max_hands_free/token/"
-
- @patch("ee.api.hands_free.requests.post")
- def test_token_returns_signed_token_from_elevenlabs(self, mock_post: MagicMock) -> None:
- mock_post.return_value = MagicMock(
- status_code=200,
- json=MagicMock(return_value={"token": "single-use-token-abc"}),
- )
-
- response = self.client.post(self._token_url(), format="json")
-
- assert response.status_code == 200
- assert response.json() == {"token": "single-use-token-abc"}
- assert mock_post.call_args.args[0].endswith("/v1/single-use-token/realtime_scribe")
- assert mock_post.call_args.kwargs["headers"]["xi-api-key"] == "test-key"
-
- @override_settings(ELEVENLABS_API_KEY="")
- def test_returns_503_when_api_key_missing(self) -> None:
- response = self.client.post(self._token_url(), format="json")
- assert response.status_code == 503
-
- @patch("ee.api.hands_free.requests.post")
- def test_returns_502_when_provider_errors(self, mock_post: MagicMock) -> None:
- mock_post.return_value = MagicMock(status_code=429, text="rate limited")
- response = self.client.post(self._token_url(), format="json")
- assert response.status_code == 502
-
- @patch("ee.api.hands_free.requests.post")
- def test_returns_502_when_provider_returns_empty_token(self, mock_post: MagicMock) -> None:
- mock_post.return_value = MagicMock(status_code=200, json=MagicMock(return_value={"token": ""}))
- response = self.client.post(self._token_url(), format="json")
- assert response.status_code == 502
-
- @patch("ee.api.hands_free.requests.post")
- def test_provider_error_does_not_log_response_body(self, mock_post: MagicMock) -> None:
- upstream = MagicMock(status_code=500, text="ECHOED API KEY OR PII")
- mock_post.return_value = upstream
-
- with patch("ee.api.hands_free.logger") as mock_logger:
- response = self.client.post(self._token_url(), format="json")
-
- assert response.status_code == 502
- log_calls = mock_logger.warning.call_args_list + mock_logger.exception.call_args_list
- for call in log_calls:
- for value in list(call.args) + list(call.kwargs.values()):
- assert "ECHOED API KEY OR PII" not in str(value)
-
- def test_requires_authentication(self) -> None:
- self.client.logout()
- response = self.client.post(self._token_url(), format="json")
- assert response.status_code in (401, 403)
-
- def _synthesize_url(self) -> str:
- return f"/api/environments/{self.team.id}/max_hands_free/synthesize/"
-
- @patch("ee.api.hands_free.requests.post")
- def test_synthesize_streams_audio_from_elevenlabs(self, mock_post: MagicMock) -> None:
- mock_post.return_value = MagicMock(
- status_code=200,
- iter_content=MagicMock(return_value=iter([b"mp3-chunk-1", b"mp3-chunk-2"])),
- )
-
- response = self.client.post(self._synthesize_url(), data={"text": "hello"}, format="json")
-
- assert response.status_code == 200
- assert response["Content-Type"] == "audio/mpeg"
- assert response["Cache-Control"] == "no-store"
- assert b"".join(response.streaming_content) == b"mp3-chunk-1mp3-chunk-2" # type: ignore[attr-defined]
- mock_post.return_value.close.assert_called_once()
-
- def test_synthesize_rejects_text_over_limit(self) -> None:
- response = self.client.post(
- self._synthesize_url(),
- data={"text": "x" * 2001},
- format="json",
- )
- assert response.status_code == 400
-
- def test_synthesize_rejects_empty_text(self) -> None:
- response = self.client.post(self._synthesize_url(), data={"text": ""}, format="json")
- assert response.status_code == 400
-
- @patch("ee.api.hands_free.requests.post")
- def test_synthesize_502_when_provider_errors(self, mock_post: MagicMock) -> None:
- mock_post.return_value = MagicMock(status_code=429, text="quota exceeded")
- response = self.client.post(self._synthesize_url(), data={"text": "hello"}, format="json")
- assert response.status_code == 502
- mock_post.return_value.close.assert_called_once()
-
- @override_settings(ELEVENLABS_API_KEY="")
- def test_synthesize_503_when_api_key_missing(self) -> None:
- response = self.client.post(self._synthesize_url(), data={"text": "hello"}, format="json")
- assert response.status_code == 503
-
- def test_endpoints_403_when_feature_flag_disabled_for_org(self) -> None:
- # Stop the class-level patch so the real PostHogFeatureFlagPermission runs against
- # an org that doesn't have the flag enabled — expect 403, no upstream calls.
- self._ff_patcher.stop()
- with patch("posthoganalytics.feature_enabled", return_value=False):
- token_response = self.client.post(self._token_url(), format="json")
- synth_response = self.client.post(self._synthesize_url(), data={"text": "hello"}, format="json")
- # Restart so addCleanup doesn't double-stop.
- self._ff_patcher.start()
- assert token_response.status_code == 403
- assert synth_response.status_code == 403
-
- def test_synthesize_requires_authentication(self) -> None:
- self.client.logout()
- response = self.client.post(self._synthesize_url(), data={"text": "hello"}, format="json")
- assert response.status_code in (401, 403)
-
- @patch("posthog.api.streaming.StreamingHttpResponse")
- @patch("ee.api.hands_free.requests.post")
- def test_synthesize_closes_upstream_when_response_construction_fails(
- self, mock_post: MagicMock, mock_streaming_response: MagicMock
- ) -> None:
- # Upstream returns 200 (so close isn't called on the error-status path) but the
- # StreamingHttpResponse constructor blows up before we can return. Without the
- # try/except in synthesize the upstream connection would leak — Django would
- # surface a 500 to the client and the requests connection would only release on
- # GC of the unconsumed iter_content generator.
- upstream = MagicMock(status_code=200)
- mock_post.return_value = upstream
- mock_streaming_response.side_effect = RuntimeError("django blew up")
-
- response = self.client.post(self._synthesize_url(), data={"text": "hello"}, format="json")
-
- assert response.status_code == 500
- upstream.close.assert_called_once()
diff --git a/ee/hogai/context/context.py b/ee/hogai/context/context.py
index 164b81ad8cc7..829184e73278 100644
--- a/ee/hogai/context/context.py
+++ b/ee/hogai/context/context.py
@@ -634,46 +634,10 @@ class AssistantContextManager(AssistantContextMixin):
prompts.append(mode_prompt)
if contextual_tools := await self._get_contextual_tools_prompt():
prompts.append(ContextMessage(content=contextual_tools, id=str(uuid4())))
- if voice_prompt := self._get_voice_mode_prompt(ui_context):
- prompts.append(ContextMessage(content=voice_prompt, id=str(uuid4())))
if formatted_ui_context := await self._format_ui_context(ui_context):
prompts.append(ContextMessage(content=formatted_ui_context, id=str(uuid4())))
return self._deduplicate_context_messages(state, prompts)
- def _get_voice_mode_prompt(self, ui_context: MaxUIContext | None) -> str | None:
- """Return a voice-mode instruction reflecting the current turn's modality.
-
- Emits a tag whenever the frontend tells us explicitly whether voice mode is on
- or off — both states need to survive in conversation history so a typed turn
- that follows a spoken one cleanly overrides the earlier voice formatting rules
- (otherwise the prior <voice_mode> instruction keeps steering the model toward
- spelled-out numbers and no markdown).
- """
- if ui_context is None or ui_context.voice_mode is None:
- return None
- if ui_context.voice_mode:
- return (
- "<voice_mode>\n"
- "The user is asking via hands-free voice mode. Your response will be read "
- "aloud by text-to-speech. Write it so it sounds natural when spoken:\n"
- "- Spell out all numbers and currencies in words "
- '(e.g. "one hundred dollars from five thousand two hundred and thirty eight users", '
- 'not "$100 from 5,238 users").\n'
- "- Spell out percentages as words "
- '(e.g. "twelve point five percent", not "12.5%").\n'
- "- No markdown — no headings, no bullets, no bold, no inline code or code blocks.\n"
- "- No emoji.\n"
- "- Use plain sentences. Keep it concise — assume the user can't see the screen.\n"
- "</voice_mode>"
- )
- return (
- "<voice_mode>\n"
- "The user is no longer in hands-free voice mode for this turn. Ignore any "
- "earlier voice-mode formatting instructions in this conversation: you may use "
- "markdown, numerals, currency symbols, code blocks, and emoji as normal.\n"
- "</voice_mode>"
- )
-
async def _get_contextual_tools_prompt(self) -> str | None:
from ee.hogai.registry import get_contextual_tool_class
diff --git a/ee/hogai/context/test/test_context.py b/ee/hogai/context/test/test_context.py
index 53e3a95cbdd2..9769ad170fe1 100644
--- a/ee/hogai/context/test/test_context.py
+++ b/ee/hogai/context/test/test_context.py
@@ -597,29 +597,6 @@ class TestAssistantContextManager(BaseTest):
self.assertIsNone(result)
mock_capture_exception.assert_called()
- def test_voice_mode_prompt_on_emits_speech_formatting_rules(self):
- prompt = self.context_manager._get_voice_mode_prompt(MaxUIContext(voice_mode=True))
- assert prompt is not None
- assert "<voice_mode>" in prompt
- assert "spelled out" in prompt.lower() or "spell out" in prompt.lower()
-
- def test_voice_mode_prompt_off_emits_explicit_override(self):
- # Explicit False must emit a counter-instruction so a typed turn after a spoken
- # turn isn't still steered by the earlier <voice_mode> prompt sitting in history.
- prompt = self.context_manager._get_voice_mode_prompt(MaxUIContext(voice_mode=False))
- assert prompt is not None
- assert "<voice_mode>" in prompt
- assert "no longer" in prompt.lower() or "ignore" in prompt.lower()
-
- @parameterized.expand(
- [
- ("voice_mode field absent", MaxUIContext(insights=None)),
- ("ui_context absent", None),
- ]
- )
- def test_voice_mode_prompt_absent_when_field_not_set(self, _name: str, ui_context: MaxUIContext | None):
- assert self.context_manager._get_voice_mode_prompt(ui_context) is None
-
def test_deduplicate_context_messages(self):
"""Test that context messages are deduplicated based on existing context message content"""
# Create state with existing context messages
diff --git a/ee/urls.py b/ee/urls.py
index 94b2dcb8ea92..fe401c1cc5b5 100644
--- a/ee/urls.py
+++ b/ee/urls.py
@@ -36,7 +36,7 @@ def extend_api_router() -> None:
router as root_router,
)
- from ee.api import hands_free, max_tools
+ from ee.api import max_tools
root_router.register(r"billing", billing.BillingViewset, "billing")
organizations_router.register(
@@ -52,8 +52,6 @@ def extend_api_router() -> None:
projects_router.register(r"max_tools", max_tools.MaxToolsViewSet, "project_max_tools", ["team_id"])
- projects_router.register(r"max_hands_free", hands_free.MaxHandsFreeViewSet, "project_max_hands_free", ["team_id"])
-
# The admin interface is disabled on self-hosted instances, as its misuse can be unsafe
if settings.ADMIN_PORTAL_ENABLED:
diff --git a/frontend/package.json b/frontend/package.json
index 7dae8707f095..b1ec243e79c0 100644
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -65,7 +65,6 @@
"@dnd-kit/modifiers": "^6.0.1",
"@dnd-kit/sortable": "^7.0.2",
"@dnd-kit/utilities": "^3.2.1",
- "@elevenlabs/client": "^1.4.0",
"@floating-ui/react": "^0.27.19",
"@marsidev/react-turnstile": "^1.4.2",
"@mathjax/src": "^4.1.2",
diff --git a/frontend/snapshots.yml b/frontend/snapshots.yml
index cb91fe0a6a5e..4f086d17fe39 100644
--- a/frontend/snapshots.yml
+++ b/frontend/snapshots.yml
@@ -13894,14 +13894,6 @@ snapshots:
hash: v1.k794b7964.69a1e5eea52abebae062fbccdfcd613e24760bd0843e54c1576cecbee868cc5d.DWFG-7lnbtlTvNy2hvY4v-IliU1BjPk9VaDqkQt4Bh0
scenes-app-posthog-ai-forms--thread-with-text-and-number-form--light:
hash: v1.k794b7964.aee1550de1b97850cb1d8c27d39e40db81c28d4fe6116a13f48d45672c143710.M1lWbiiqC33CbFfL0ZdteiJH328flmdL2oPxzwPsJgw
- scenes-app-posthog-ai-hands-free-surface--speaking--dark:
- hash: v1.k794b7964.57e6cf1b2966bbbcac0c6dc1fc79dbcec30647b444819a36036bf414e5b2e620.VxF2LBHOgDjQ7K3quyqD7Lt7Jr7BaC9pAP7aaPI_cKQ
- scenes-app-posthog-ai-hands-free-surface--speaking--light:
- hash: v1.k794b7964.d8525499c1eb57dfa8f060a24aa1e27eab84613df452f32d35562496aefe72db.zCZOWI2nhLNhZRnJImG8ChEh-q4ImegCyxRl69VrFQs
- scenes-app-posthog-ai-hands-free-surface--with-error--dark:
- hash: v1.k794b7964.8845839d904fcee5d1cf4d98c2e1c16cde91b474a4d12033729ebd0cd8c6c206.Hall9WIzoGr43FyAUQ6iahlleJIiJ5UCYFgvouPy8Fc
- scenes-app-posthog-ai-hands-free-surface--with-error--light:
- hash: v1.k794b7964.988574fff8b8f14327e6234395a9200de50d0c357261037219bf56344b249369.6f9U2b-TYev-MwpxHqJjEnLrrvTDRB0fMavjcFcUn6s
scenes-app-posthog-ai-tools--changelog-only--dark:
hash: v1.k794b7964.f17aec31dd25e93672f7d76b065a6b49052c643eb6da67458011462b883f1c9c.sYkAo1RWtXpQRQNxl5SzUkO4ELuuEshq9Kdbc0H1L0E
scenes-app-posthog-ai-tools--changelog-only--light:
diff --git a/frontend/src/lib/api.ts b/frontend/src/lib/api.ts
index 91d909686027..4945fbf67ad3 100644
--- a/frontend/src/lib/api.ts
+++ b/frontend/src/lib/api.ts
@@ -1712,11 +1712,6 @@ export class ApiRequest {
return this.teamProjectDetail(teamId).addPathComponent('conversations').addPathComponent(id)
}
- // Max hands-free mode (ElevenLabs Scribe token proxy)
- public maxHandsFree(teamId?: TeamType['id']): ApiRequest {
- return this.teamProjectDetail(teamId).addPathComponent('max_hands_free')
- }
-
// Conversations (Support product)
public conversationsTickets(teamId?: TeamType['id']): ApiRequest {
return this.projectsDetail(teamId).addPathComponent('conversations').addPathComponent('tickets')
@@ -6564,23 +6559,6 @@ const api = {
},
},
- maxHandsFree: {
- async token(options?: ApiMethodOptions): Promise<{ token: string }> {
- return await api.create(
- new ApiRequest().maxHandsFree().withAction('token').assembleFullUrl(),
- undefined,
- options
- )
- },
- async synthesize(text: string, options?: ApiMethodOptions): Promise<Response> {
- return await api.createResponse(
- new ApiRequest().maxHandsFree().withAction('synthesize').assembleFullUrl(),
- { text },
- options
- )
- },
- },
-
conversations: {
async stream(
data: {
diff --git a/frontend/src/lib/components/Shortcuts/shortcuts.ts b/frontend/src/lib/components/Shortcuts/shortcuts.ts
index cc5d781cb2ec..a286b78fd211 100644
--- a/frontend/src/lib/components/Shortcuts/shortcuts.ts
+++ b/frontend/src/lib/components/Shortcuts/shortcuts.ts
@@ -36,7 +36,6 @@ export const keyBinds: Record<string, string[]> = {
tab9: [...baseModifier, '9'],
zenMode: [...baseModifier, 'z'],
newChat: ['g', 'then', 'n'],
- maxHandsFree: ['h', 'then', 'f'],
allChats: ['g', 'then', '1'],
allProducts: ['g', 'then', '2'],
theme: ['g', 'then', 't'],
diff --git a/frontend/src/lib/constants.tsx b/frontend/src/lib/constants.tsx
index c03a11759d4e..3f8076f4c945 100644
--- a/frontend/src/lib/constants.tsx
+++ b/frontend/src/lib/constants.tsx
@@ -444,7 +444,7 @@ export const FEATURE_FLAGS = {
MAX_AI_INSIGHT_SEARCH: 'max-ai-insight-search', // owner: #team-posthog-ai
MAX_BILLING_CONTEXT: 'max-billing-context', // owner: @pawel-cebula #team-billing
MAX_DEEP_RESEARCH: 'max-deep-research', // owner: @kappa90 #team-posthog-ai
- MAX_HANDS_FREE: 'max-hands-free', // owner: #team-posthog-ai
+ MAX_WEB_ANALYTICS_NUDGE: 'posthog-ai-web-analytics-nudge', // owner: @jordanm-posthog #team-web-analytics
MCP_ANALYTICS_INTENT_ROUTING: 'mcp-analytics-intent-routing', // owner: #project-mcp-analytics
MCP_ANALYTICS_LEADERBOARD_HOME: 'mcp-analytics-leaderboard-home', // owner: #team-mcp-analytics, trial of the public leaderboard layout on the MCP analytics home tab
MCP_GATEWAY: 'mcp-gateway', // owner: #team-context-mcp — gates the MCP gateway UI AND backend enforcement of built-in agent MCP access (delegated-only installs + restricted sandbox token scope); roll out by organization
diff --git a/frontend/src/queries/schema.json b/frontend/src/queries/schema.json
index 96eee38e36c8..192cc8f15a6f 100644
--- a/frontend/src/queries/schema.json
+++ b/frontend/src/queries/schema.json
@@ -38828,9 +38828,6 @@
"$ref": "#/definitions/MaxNotebookContext"
},
"type": "array"
- },
- "voice_mode": {
- "type": "boolean"
}
},
"type": "object"
diff --git a/frontend/src/scenes/max/components/HandsFreeButton.tsx b/frontend/src/scenes/max/components/HandsFreeButton.tsx
deleted file mode 100644
index 822f7985a519..000000000000
--- a/frontend/src/scenes/max/components/HandsFreeButton.tsx
+++ /dev/null
@@ -1,38 +0,0 @@
-import { useActions, useValues } from 'kea'
-
-import { IconMicrophone } from '@posthog/icons'
-import { LemonButton } from '@posthog/lemon-ui'
-
-import { Shortcut } from 'lib/components/Shortcuts/Shortcut'
-import { keyBinds } from 'lib/components/Shortcuts/shortcuts'
-import { useFeatureFlag } from 'lib/hooks/useFeatureFlag'
-
-import { handsFreeLogic } from '../handsFreeLogic'
-
-interface HandsFreeButtonProps {
- panelId?: string
-}
-
-export function HandsFreeButton({ panelId }: HandsFreeButtonProps): JSX.Element | null {
- const flagEnabled = useFeatureFlag('MAX_HANDS_FREE')
- const { status, canUseHandsFree } = useValues(handsFreeLogic({ panelId }))
- const { toggleHandsFree } = useActions(handsFreeLogic({ panelId }))
-
- if (!flagEnabled || !canUseHandsFree || status !== 'off') {
- return null
- }
-
- return (
- <Shortcut name="maxHandsFree" keybind={[keyBinds.maxHandsFree]} intent="Enter hands-free" interaction="click">
- <LemonButton
- data-attr="max-hands-free-toggle"
- size="small"
- type="tertiary"
- icon={<IconMicrophone />}
- onClick={toggleHandsFree}
- tooltip="Enter hands-free"
- aria-label="Enter hands-free"
- />
- </Shortcut>
- )
-}
diff --git a/frontend/src/scenes/max/components/HandsFreeSurface.scss b/frontend/src/scenes/max/components/HandsFreeSurface.scss
deleted file mode 100644
index 6c6dfb02d618..000000000000
--- a/frontend/src/scenes/max/components/HandsFreeSurface.scss
+++ /dev/null
@@ -1,185 +0,0 @@
-.hands-free-surface {
- display: grid;
- grid-template-rows: minmax(2.5rem, auto) auto minmax(1.5rem, auto);
- gap: 1rem;
- place-items: center center;
- width: 100%;
- min-height: 11rem;
- padding: 1.25rem 1rem;
- text-align: center;
-}
-
-.hands-free-surface__top {
- display: flex;
- flex-direction: column;
- gap: 0.25rem;
- align-items: center;
- justify-content: center;
- width: 100%;
- min-height: 2.5rem;
-}
-
-.hands-free-surface__partial {
- max-width: 32rem;
- margin: 0;
- font-size: 1rem;
- line-height: 1.4;
- color: var(--color-text-primary);
- word-wrap: break-word;
-}
-
-.hands-free-surface__hint {
- max-width: 32rem;
- margin: 0;
- font-size: 0.8125rem;
- font-style: italic;
- color: var(--color-text-secondary);
-}
-
-.hands-free-surface__error {
- margin: 0;
- font-size: 0.8125rem;
- color: var(--color-danger);
-}
-
-.hands-free-surface__mic {
- position: relative;
- display: inline-flex;
- align-items: center;
- justify-content: center;
- width: 3rem;
- height: 3rem;
- color: var(--color-text-primary);
- appearance: none;
- cursor: pointer;
- background: var(--color-bg-fill-secondary);
- border: none;
- border-radius: 50%;
- box-shadow: 0 1px 2px rgb(0 0 0 / 20%);
- transition:
- transform 120ms ease,
- background 120ms ease;
-
- svg {
- width: 1.25rem;
- height: 1.25rem;
- }
-
- &:hover {
- transform: scale(1.04);
- }
-
- &:active {
- transform: scale(0.97);
- }
-}
-
-.hands-free-surface__mic--listening {
- color: var(--color-text-on-fill);
- background: var(--color-success);
-}
-
-.hands-free-surface__mic--speaking {
- color: var(--color-text-on-fill);
- background: var(--color-ai);
-}
-
-.hands-free-surface__mic--thinking,
-.hands-free-surface__mic--starting,
-.hands-free-surface__mic--reconnecting {
- color: var(--color-text-on-fill);
- background: var(--color-warning);
-}
-
-.hands-free-surface__mic--error {
- color: var(--color-text-on-fill);
- background: var(--color-danger);
-}
-
-// Larger tap target on phones — the primary affordance of a hands-free product,
-// and gym/walking users need more than the 48px desktop default.
-@media (max-width: 480px) {
- .hands-free-surface__mic {
- width: 4rem;
- height: 4rem;
-
- svg {
- width: 1.75rem;
- height: 1.75rem;
- }
- }
-}
-
-// Pulse applied only when status === 'listening' (your turn to talk).
-.hands-free-surface__mic--pulsing {
- animation: HandsFreeSurface__micPulse 1.6s ease-in-out infinite;
-}
-
-// Spinner ring around the mic when thinking/starting/reconnecting — keeps the mic icon
-// visible so the user still knows the button is tappable to exit.
-.hands-free-surface__mic--thinking::after,
-.hands-free-surface__mic--starting::after,
-.hands-free-surface__mic--reconnecting::after {
- position: absolute;
- inset: -4px;
- content: '';
- border: 2px solid transparent;
- border-top-color: currentColor;
- border-radius: 50%;
- animation: HandsFreeSurface__micSpin 1s linear infinite;
-}
-
-@keyframes HandsFreeSurface__micSpin {
- to {
- transform: rotate(360deg);
- }
-}
-
-@keyframes HandsFreeSurface__micPulse {
- 0%,
- 100% {
- box-shadow: 0 0 0 0 rgb(0 0 0 / 0%);
- }
-
- 50% {
- box-shadow: 0 0 0 0.75rem rgb(255 255 255 / 8%);
- }
-}
-
-.hands-free-surface__bottom {
- display: inline-flex;
- gap: 0.5rem;
- align-items: center;
-}
-
-.hands-free-surface__emoji {
- display: inline-flex;
- flex-shrink: 0;
- align-items: center;
- justify-content: center;
- font-size: 1.125rem;
- line-height: 1;
-}
-
-.hands-free-surface__emoji--pulsing {
- animation: HandsFreeSurface__pulse 1.4s ease-in-out infinite;
-}
-
-@keyframes HandsFreeSurface__pulse {
- 0%,
- 100% {
- opacity: 1;
- transform: scale(1);
- }
-
- 50% {
- opacity: 0.7;
- transform: scale(1.1);
- }
-}
-
-.hands-free-surface__label {
- font-size: 0.875rem;
- font-weight: 600;
- color: var(--color-text-primary);
-}
diff --git a/frontend/src/scenes/max/components/HandsFreeSurface.stories.tsx b/frontend/src/scenes/max/components/HandsFreeSurface.stories.tsx
deleted file mode 100644
index 045caacf81bd..000000000000
--- a/frontend/src/scenes/max/components/HandsFreeSurface.stories.tsx
+++ /dev/null
@@ -1,101 +0,0 @@
-import { Meta, StoryObj } from '@storybook/react'
-import { useActions, useMountedLogic } from 'kea'
-import { useEffect } from 'react'
-
-import { HandsFreeConnection, HandsFreeStatus, handsFreeLogic } from '../handsFreeLogic'
-import { maxLogic } from '../maxLogic'
-import { HandsFreeSurface } from './HandsFreeSurface'
-
-const TAB_ID = 'hands-free-storybook'
-
-interface DriverProps {
- status: HandsFreeStatus
- connection?: HandsFreeConnection
- partialTranscript?: string
- error?: string | null
-}
-
-// Mounts maxLogic + handsFreeLogic and drives them into a target state so each story
-// captures a single visual configuration without spinning up real WebSockets or audio.
-function HandsFreeSurfaceStory({ status, connection, partialTranscript, error }: DriverProps): JSX.Element {
- useMountedLogic(maxLogic({ panelId: TAB_ID }))
- const logic = handsFreeLogic({ panelId: TAB_ID })
- useMountedLogic(logic)
- const { setStatus, setConnection, setPartialTranscript, setError } = useActions(logic)
-
- useEffect(() => {
- setStatus(status)
- if (connection) {
- setConnection(connection)
- }
- if (partialTranscript) {
- setPartialTranscript(partialTranscript)
- }
- if (error) {
- setError(error)
- }
- }, [status, connection, partialTranscript, error, setStatus, setConnection, setPartialTranscript, setError])
-
- return (
- <div className="max-w-md mx-auto p-4 bg-bg-light border rounded">
- <HandsFreeSurface panelId={TAB_ID} />
- </div>
- )
-}
-
-const meta: Meta<typeof HandsFreeSurfaceStory> = {
- title: 'Scenes-App/PostHog AI/Hands-free surface',
- component: HandsFreeSurfaceStory,
- parameters: {
- layout: 'centered',
- },
-}
-export default meta
-
-type Story = StoryObj<typeof HandsFreeSurfaceStory>
-
-// Stories whose mic ring/pulse animates are excluded from visual regression — the
-// running CSS animation makes the snapshot diff flake. Static visual states (speaking,
-// error) keep their snapshot so we still catch unintended colour/layout drift.
-const ANIMATED_SKIP_TAGS = ['test-skip']
-
-export const Starting: Story = {
- args: { status: 'starting', connection: 'connecting' },
- tags: ANIMATED_SKIP_TAGS,
-}
-
-export const Listening: Story = {
- args: { status: 'listening', connection: 'connected' },
- tags: ANIMATED_SKIP_TAGS,
-}
-
-export const ListeningWithPartialTranscript: Story = {
- args: {
- status: 'listening',
- connection: 'connected',
- partialTranscript: 'how are the daily active users tracking against',
- },
- tags: ANIMATED_SKIP_TAGS,
-}
-
-export const Thinking: Story = {
- args: { status: 'thinking', connection: 'connected' },
- tags: ANIMATED_SKIP_TAGS,
-}
-
-export const Speaking: Story = {
- args: { status: 'speaking', connection: 'connected' },
-}
-
-export const Reconnecting: Story = {
- args: { status: 'listening', connection: 'reconnecting' },
- tags: ANIMATED_SKIP_TAGS,
-}
-
-export const WithError: Story = {
- args: {
- status: 'listening',
- connection: 'connected',
- error: 'Microphone access was denied.',
- },
-}
diff --git a/frontend/src/scenes/max/components/HandsFreeSurface.tsx b/frontend/src/scenes/max/components/HandsFreeSurface.tsx
deleted file mode 100644
index be27f43008ab..000000000000
--- a/frontend/src/scenes/max/components/HandsFreeSurface.tsx
+++ /dev/null
@@ -1,137 +0,0 @@
-import './HandsFreeSurface.scss'
-
-import { useActions, useValues } from 'kea'
-
-import { IconMicrophone } from '@posthog/icons'
-
-import { keyBinds } from 'lib/components/Shortcuts/shortcuts'
-import { useShortcut } from 'lib/components/Shortcuts/useShortcut'
-import { cn } from 'lib/utils/css-classes'
-
-import { HandsFreeStatus, handsFreeLogic } from '../handsFreeLogic'
-
-interface HandsFreeSurfaceProps {
- panelId?: string
-}
-
-const STATUS_LABEL: Record<HandsFreeStatus, string> = {
- off: '',
- starting: 'Connecting',
- listening: 'Listening',
- thinking: 'Thinking',
- speaking: 'Speaking',
-}
-
-// Emoji per visual state. Reconnecting + error deliberately render no emoji so
-// the label sits alone and stays perfectly centered under the mic.
-const STATUS_EMOJI: Record<HandsFreeStatus | 'reconnecting' | 'error', string> = {
- off: '',
- starting: '🎤',
- listening: '🎤',
- thinking: '🧠',
- speaking: '🔊',
- reconnecting: '',
- error: '',
-}
-
-const STATUS_HINT: Record<HandsFreeStatus, string> = {
- off: '',
- starting: 'One moment, getting microphone ready',
- listening: 'I will send your question when you pause speaking',
- thinking: 'Working on your answer',
- speaking: 'Start talking to interrupt',
-}
-
-function HandsFreeTopline({
- panelId,
- status,
- isReconnecting,
-}: {
- panelId?: string
- status: HandsFreeStatus
- isReconnecting: boolean
-}): JSX.Element {
- const { partialTranscript, error } = useValues(handsFreeLogic({ panelId }))
- const hint = isReconnecting ? 'Reconnecting your microphone' : STATUS_HINT[status]
- return (
- <div className="hands-free-surface__top">
- {partialTranscript ? (
- <p className="hands-free-surface__partial">{partialTranscript}</p>
- ) : (
- <p className="hands-free-surface__hint">{hint}</p>
- )}
- {error && <p className="hands-free-surface__error">{error}</p>}
- </div>
- )
-}
-
-export function HandsFreeSurface({ panelId }: HandsFreeSurfaceProps): JSX.Element | null {
- const { status, connection, error } = useValues(handsFreeLogic({ panelId }))
- const { toggleHandsFree } = useActions(handsFreeLogic({ panelId }))
-
- // Register the v-then-m exit shortcut while the surface is mounted.
- // HandsFreeButton owns the same shortcut for the "enter" path; same-name
- // re-registration in shortcutLogic handles the handover cleanly.
- useShortcut({
- name: 'maxHandsFree',
- keybind: [keyBinds.maxHandsFree],
- intent: 'Exit hands-free',
- interaction: 'function',
- callback: toggleHandsFree,
- })
-
- if (status === 'off') {
- return null
- }
-
- const isListening = status === 'listening'
- const isReconnecting = connection === 'reconnecting'
- const hasError = !!error
-
- // Reconnecting / error visuals take precedence over the listening pulse so the user
- // isn't shown a green "listening" mic when the underlying connection is dead.
- const visualState: HandsFreeStatus | 'reconnecting' | 'error' = hasError
- ? 'error'
- : isReconnecting
- ? 'reconnecting'
- : status
- const label = hasError ? 'Error' : isReconnecting ? 'Reconnecting' : STATUS_LABEL[status]
- const shouldPulse = isListening && !isReconnecting && !hasError
-
- return (
- <div
- className="hands-free-surface"
- data-attr="max-hands-free-surface"
- data-status={status}
- data-connection={connection}
- >
- <HandsFreeTopline panelId={panelId} status={status} isReconnecting={isReconnecting} />
-
- <button
- type="button"
- onClick={toggleHandsFree}
- aria-label="Exit hands-free"
- data-attr="max-hands-free-exit"
- className={cn('hands-free-surface__mic', `hands-free-surface__mic--${visualState}`, {
- 'hands-free-surface__mic--pulsing': shouldPulse,
- })}
- >
- <IconMicrophone />
- </button>
-
- <div className="hands-free-surface__bottom">
- {STATUS_EMOJI[visualState] && (
- <span
- className={cn('hands-free-surface__emoji', {
- 'hands-free-surface__emoji--pulsing': shouldPulse,
- })}
- aria-hidden
- >
- {STATUS_EMOJI[visualState]}
- </span>
- )}
- <span className="hands-free-surface__label">{label}</span>
- </div>
- </div>
- )
-}
diff --git a/frontend/src/scenes/max/components/QuestionInput.tsx b/frontend/src/scenes/max/components/QuestionInput.tsx
index 5a43149829a5..86cdfe45f849 100644
--- a/frontend/src/scenes/max/components/QuestionInput.tsx
+++ b/frontend/src/scenes/max/components/QuestionInput.tsx
@@ -10,7 +10,6 @@ import { IconArrowRight, IconCheck, IconPencil, IconStopFilled, IconTrash, IconU
import { LemonButton, LemonSwitch, LemonTextArea, Spinner } from '@posthog/lemon-ui'
import { KeyboardShortcut } from 'lib/components/KeyboardShortcut/KeyboardShortcut'
-import { useFeatureFlag } from 'lib/hooks/useFeatureFlag'
import { cn } from 'lib/utils/css-classes'
import { AIConsentPopoverWrapper } from 'scenes/settings/organization/AIConsentPopoverWrapper'
import { userLogic } from 'scenes/userLogic'
@@ -19,15 +18,12 @@ import { AgentMode } from '~/queries/schema/schema-assistant-messages'
import { ConversationQueueMessage } from '~/types'
import { ContextDisplay } from '../Context'
-import { handsFreeLogic } from '../handsFreeLogic'
import { MAX_MESSAGE_LENGTH, MESSAGE_TOO_LONG, messageLength } from '../max-constants'
import { maxGlobalLogic } from '../maxGlobalLogic'
import { maxLogic } from '../maxLogic'
import { maxThreadLogic } from '../maxThreadLogic'
import { MAX_SLASH_COMMANDS } from '../slash-commands'
import { FillInHint } from './FillInHint'
-import { HandsFreeButton } from './HandsFreeButton'
-import { HandsFreeSurface } from './HandsFreeSurface'
import { SlashCommandAutocomplete } from './SlashCommandAutocomplete'
/**
@@ -153,7 +149,7 @@ export const QuestionInput = React.forwardRef<HTMLDivElement, QuestionInputProps
ref
) {
const { dataProcessingAccepted, isPhaiSandboxFlagOn } = useValues(maxGlobalLogic)
- const { question, panelId: maxPanelId, fillInHint, typingSuggestion } = useValues(maxLogic)
+ const { question, fillInHint, typingSuggestion } = useValues(maxLogic)
const { setQuestion, setFillInHint, cancelSuggestionTyping, attachFilesToNewChat } = useActions(maxLogic)
const { user } = useValues(userLogic)
const {
@@ -182,12 +178,6 @@ export const QuestionInput = React.forwardRef<HTMLDivElement, QuestionInputProps
updateQueuedMessage,
releaseSandboxPrewarm,
} = useActions(maxThreadLogic)
- const { isActive: handsFreeActive } = useValues(handsFreeLogic({ panelId: maxPanelId }))
- // Only the hands-free row needs bottom-aligned pills — it has the mic + submit pair
- // pinned to the bottom and pills sitting in normal flow look misaligned next to them.
- // Keep the legacy items-start layout when the flag is off so existing screenshots
- // (and the layout users without hands-free see) don't change.
- const handsFreeFlagEnabled = useFeatureFlag('MAX_HANDS_FREE')
// Show info banner for conversations created during impersonation (marked as internal)
const isImpersonatedInternalConversation = user?.is_impersonated && conversation?.is_internal
@@ -368,272 +358,245 @@ export const QuestionInput = React.forwardRef<HTMLDivElement, QuestionInputProps
!streamingActive && '[--input-ring-color:var(--color-ai)]'
)}
>
- {handsFreeActive ? (
- <HandsFreeSurface panelId={maxPanelId} />
- ) : (
- <SlashCommandAutocomplete
- visible={showAutocomplete}
- onClose={() => {
- setShowAutocomplete(false)
- setAutocompleteDismissed(true)
- }}
- >
- <div className="relative w-full">
- {!inputValue && (
- <div
- id="textarea-hint"
- className="text-secondary absolute top-4 left-4 text-sm pointer-events-none"
- >
- {conversation && isSharedThread ? (
- `This thread was shared with you by ${conversation.user.first_name} ${conversation.user.last_name}`
- ) : threadLoading ? (
- 'Thinking…'
- ) : isThreadVisible ? (
- placeholder || (
- <>
- Ask follow-up{' '}
- <span className="text-tertiary opacity-80 contrast-more:opacity-100">
- or / for commands
- </span>
- </>
- )
- ) : (
+ <SlashCommandAutocomplete
+ visible={showAutocomplete}
+ onClose={() => {
+ setShowAutocomplete(false)
+ setAutocompleteDismissed(true)
+ }}
+ >
+ <div className="relative w-full">
+ {!inputValue && (
+ <div
+ id="textarea-hint"
+ className="text-secondary absolute top-4 left-4 text-sm pointer-events-none"
+ >
+ {conversation && isSharedThread ? (
+ `This thread was shared with you by ${conversation.user.first_name} ${conversation.user.last_name}`
+ ) : threadLoading ? (
+ 'Thinking…'
+ ) : isThreadVisible ? (
+ placeholder || (
<>
- Ask a question{' '}
+ Ask follow-up{' '}
<span className="text-tertiary opacity-80 contrast-more:opacity-100">
or / for commands
</span>
</>
- )}
- </div>
- )}
- {/* Postfix cue after a fill-in suggestion's typed-in prefix. */}
- {showFillInHint && (
- <div className="absolute top-4 left-4 right-4 overflow-hidden pointer-events-none">
- <FillInHint text={inputValue} hint={fillInHint} />
- </div>
- )}
- <LemonTextArea
- aria-describedby={!inputValue ? 'textarea-hint' : undefined}
- id="question-input"
- data-attr="max-chat-input"
- ref={textAreaRef}
- value={isSharedThread ? '' : inputValue}
- onChange={handleQuestionChange}
- onBlur={(e) => {
- debouncedSetQuestion.flush()
- // Release any sandbox pre-warm when the user leaves the
- // input without sending. No-op on LangGraph / unwarmed
- // conversations.
- //
- // But clicking the send button blurs the textarea *before*
- // its onClick fires the send — releasing here would cancel
- // the very warm Run the send is about to consume. When focus
- // moves to the send button, skip the release and let the send
- // path consume the warm (it does so without a DELETE).
- const next = e.relatedTarget as HTMLElement | null
- if (next?.closest('[data-attr="max-send-message"]')) {
+ )
+ ) : (
+ <>
+ Ask a question{' '}
+ <span className="text-tertiary opacity-80 contrast-more:opacity-100">
+ or / for commands
+ </span>
+ </>
+ )}
+ </div>
+ )}
+ {/* Postfix cue after a fill-in suggestion's typed-in prefix. */}
+ {showFillInHint && (
+ <div className="absolute top-4 left-4 right-4 overflow-hidden pointer-events-none">
+ <FillInHint text={inputValue} hint={fillInHint} />
+ </div>
+ )}
+ <LemonTextArea
+ aria-describedby={!inputValue ? 'textarea-hint' : undefined}
+ id="question-input"
+ data-attr="max-chat-input"
+ ref={textAreaRef}
+ value={isSharedThread ? '' : inputValue}
+ onChange={handleQuestionChange}
+ onBlur={(e) => {
+ debouncedSetQuestion.flush()
+ // Release any sandbox pre-warm when the user leaves the
+ // input without sending. No-op on LangGraph / unwarmed
+ // conversations.
+ //
+ // But clicking the send button blurs the textarea *before*
+ // its onClick fires the send — releasing here would cancel
+ // the very warm Run the send is about to consume. When focus
+ // moves to the send button, skip the release and let the send
+ // path consume the warm (it does so without a DELETE).
+ const next = e.relatedTarget as HTMLElement | null
+ if (next?.closest('[data-attr="max-send-message"]')) {
+ return
+ }
+ releaseSandboxPrewarm()
+ }}
+ onPressEnter={() => {
+ if (
+ hasQuestion &&
+ !submissionDisabledReason &&
+ (!threadLoading || queueingEnabled)
+ ) {
+ onSubmit?.()
+ submit(inputValue)
+ }
+ }}
+ onKeyDown={(event) => {
+ if (
+ event.key === 'ArrowUp' &&
+ !inputValue.trim() &&
+ queuedMessages.length > 0 &&
+ !editingQueueId
+ ) {
+ const target = event.currentTarget
+ const atStart = target.selectionStart === 0 && target.selectionEnd === 0
+ const isSingleLine = target.value.split('\n').length <= 1
+ if (!atStart || !isSingleLine) {
return
}
- releaseSandboxPrewarm()
- }}
- onPressEnter={() => {
- if (
- hasQuestion &&
- !submissionDisabledReason &&
- (!threadLoading || queueingEnabled)
- ) {
- onSubmit?.()
- submit(inputValue)
- }
- }}
- onKeyDown={(event) => {
- if (
- event.key === 'ArrowUp' &&
- !inputValue.trim() &&
- queuedMessages.length > 0 &&
- !editingQueueId
- ) {
- const target = event.currentTarget
- const atStart = target.selectionStart === 0 && target.selectionEnd === 0
- const isSingleLine = target.value.split('\n').length <= 1
- if (!atStart || !isSingleLine) {
- return
- }
- const nextMessageId = queuedMessages[0]?.id
- if (!nextMessageId) {
- return
- }
- event.preventDefault()
- setEditingQueueId(nextMessageId)
+ const nextMessageId = queuedMessages[0]?.id
+ if (!nextMessageId) {
+ return
}
- }}
- disabled={inputDisabled}
- minRows={1}
- maxRows={10}
- className={cn(
- '!border-none !bg-transparent min-h-16 py-2 pl-2 resize-none',
- handsFreeFlagEnabled ? 'pr-20' : 'pr-12',
- // Hide the native caret so only the enlarged fill-in caret shows.
- showFillInHint && 'caret-transparent'
- )}
- hideFocus
- />
- </div>
- </SlashCommandAutocomplete>
- )}
-
- {!isSharedThread &&
- !handsFreeActive && (
- // When the hands-free flag is on, reserve ~80px (pr-20) so the chip
- // row doesn't wrap under the absolutely-positioned mic + send pair.
- // Without the flag the row only has send and the legacy pr-12 is
- // enough — keep it so non-flagged users see the original layout.
- <div className={cn('pb-2', handsFreeFlagEnabled ? 'pr-20' : 'pr-12')}>
- {!isThreadVisible ? (
- <div
- className={cn(
- 'flex justify-between',
- handsFreeFlagEnabled ? 'items-end flex-wrap gap-1' : 'items-start'
- )}
- >
- <div className="flex items-start gap-1 flex-1 min-w-0">
- <div className="flex-1 min-w-0">
- <ContextDisplay size={contextDisplaySize} />
- </div>
- {isPhaiSandboxFlagOn && !conversation && (
- <div className="shrink-0">
- <input
- ref={attachmentInputRef}
- type="file"
- multiple
- className="hidden"
- data-attr="max-new-chat-attach-input"
- onChange={(event) => {
- const files = Array.from(event.target.files ?? [])
- event.target.value = ''
- attachFilesToNewChat(files, inputValue)
- }}
- />
- <LemonButton
- size="xxsmall"
- type="tertiary"
- className="border"
- icon={<IconUpload />}
- tooltip="Attach files in new PostHog AI. Your draft moves with you."
- onClick={() => attachmentInputRef.current?.click()}
- data-attr="max-new-chat-attach-file"
- >
- Attach
- </LemonButton>
- </div>
- )}
- </div>
+ event.preventDefault()
+ setEditingQueueId(nextMessageId)
+ }
+ }}
+ disabled={inputDisabled}
+ minRows={1}
+ maxRows={10}
+ className={cn(
+ '!border-none !bg-transparent min-h-16 py-2 pl-2 resize-none',
+ 'pr-12',
+ // Hide the native caret so only the enlarged fill-in caret shows.
+ showFillInHint && 'caret-transparent'
+ )}
+ hideFocus
+ />
+ </div>
+ </SlashCommandAutocomplete>
- <div
- className={cn(
- 'flex mr-1',
- handsFreeFlagEnabled
- ? 'items-end gap-1'
- : 'items-start gap-1 h-full mt-1'
- )}
- >
- {topActions}
+ {!isSharedThread && (
+ <div className="pb-2 pr-12">
+ {!isThreadVisible ? (
+ <div className="flex justify-between items-start">
+ <div className="flex items-start gap-1 flex-1 min-w-0">
+ <div className="flex-1 min-w-0">
+ <ContextDisplay size={contextDisplaySize} />
</div>
- </div>
- ) : (
- <ContextDisplay size={contextDisplaySize} />
- )}
- {promptLength !== null && (
- <div
- className={cn(
- 'text-xs text-right pr-1 pt-1',
- isOverLengthLimit ? 'text-error' : 'text-secondary'
+ {isPhaiSandboxFlagOn && !conversation && (
+ <div className="shrink-0">
+ <input
+ ref={attachmentInputRef}
+ type="file"
+ multiple
+ className="hidden"
+ data-attr="max-new-chat-attach-input"
+ onChange={(event) => {
+ const files = Array.from(event.target.files ?? [])
+ event.target.value = ''
+ attachFilesToNewChat(files, inputValue)
+ }}
+ />
+ <LemonButton
+ size="xxsmall"
+ type="tertiary"
+ className="border"
+ icon={<IconUpload />}
+ tooltip="Attach files in new PostHog AI. Your draft moves with you."
+ onClick={() => attachmentInputRef.current?.click()}
+ data-attr="max-new-chat-attach-file"
+ >
+ Attach
+ </LemonButton>
+ </div>
)}
- >
- {promptLength.toLocaleString()} / {MAX_MESSAGE_LENGTH.toLocaleString()}
</div>
- )}
- </div>
- )}
+
+ <div className="flex mr-1 items-start gap-1 h-full mt-1">{topActions}</div>
+ </div>
+ ) : (
+ <ContextDisplay size={contextDisplaySize} />
+ )}
+ {promptLength !== null && (
+ <div
+ className={cn(
+ 'text-xs text-right pr-1 pt-1',
+ isOverLengthLimit ? 'text-error' : 'text-secondary'
+ )}
+ >
+ {promptLength.toLocaleString()} / {MAX_MESSAGE_LENGTH.toLocaleString()}
+ </div>
+ )}
+ </div>
+ )}
</label>
<div
className={cn(
'absolute flex items-center',
- handsFreeFlagEnabled && 'gap-1',
isSharedThread && 'hidden',
isThreadVisible ? 'bottom-[9px] right-[9px]' : 'bottom-[7px] right-[7px]'
)}
>
- <HandsFreeButton panelId={maxPanelId} />
- {!handsFreeActive && (
- <AIConsentPopoverWrapper
- placement="bottom-end"
- showArrow
- ignoreDismissal
- onApprove={() => submit(pendingPrompt || inputValue)}
- onDismiss={() => completeThreadGeneration()}
- middleware={[
- offset((state) => ({
- mainAxis: state.placement.includes('top') ? 30 : 1,
- })),
- ]}
- hidden={!threadLoading && !pendingPrompt}
- >
- <LemonButton
- data-attr={showStopButton ? 'max-stop-generation' : 'max-send-message'}
- type={(isThreadVisible && !hasQuestion) || showStopButton ? 'secondary' : 'primary'}
- onClick={() => {
- if (threadLoading) {
- if (isQueueingSubmission) {
- if (submissionDisabledReason) {
- textAreaRef?.current?.focus()
- return
- }
- submit(inputValue)
+ <AIConsentPopoverWrapper
+ placement="bottom-end"
+ showArrow
+ ignoreDismissal
+ onApprove={() => submit(pendingPrompt || inputValue)}
+ onDismiss={() => completeThreadGeneration()}
+ middleware={[
+ offset((state) => ({
+ mainAxis: state.placement.includes('top') ? 30 : 1,
+ })),
+ ]}
+ hidden={!threadLoading && !pendingPrompt}
+ >
+ <LemonButton
+ data-attr={showStopButton ? 'max-stop-generation' : 'max-send-message'}
+ type={(isThreadVisible && !hasQuestion) || showStopButton ? 'secondary' : 'primary'}
+ onClick={() => {
+ if (threadLoading) {
+ if (isQueueingSubmission) {
+ if (submissionDisabledReason) {
+ textAreaRef?.current?.focus()
return
}
- stopGeneration()
- return
- }
- if (submissionDisabledReason) {
- textAreaRef?.current?.focus()
+ submit(inputValue)
return
}
- submit(inputValue)
- }}
- tooltip={
- // If there's a disabled reason tooltip shown below, don't show a tooltip here
- // Else, show the tooltip based on the button state
- disabledReason ? undefined : showStopButton ? (
- <>
- Let's bail <KeyboardShortcut enter />
- </>
- ) : isQueueingSubmission ? (
- <>
- Queue message <KeyboardShortcut enter />
- </>
- ) : (
- <>
- Let's go! <KeyboardShortcut enter />
- </>
- )
+ stopGeneration()
+ return
}
- loading={threadLoading && !dataProcessingAccepted}
- disabledReason={disabledReason}
- className={disabledReason ? 'opacity-[0.5]' : ''}
- size="small"
- icon={
- showStopButton ? (
- <IconStopFilled />
- ) : (
- MAX_SLASH_COMMANDS.find((cmd) => cmd.name === inputValue.split(' ', 1)[0])
- ?.icon || <IconArrowRight />
- )
+ if (submissionDisabledReason) {
+ textAreaRef?.current?.focus()
+ return
}
- />
- </AIConsentPopoverWrapper>
- )}
+ submit(inputValue)
+ }}
+ tooltip={
+ // If there's a disabled reason tooltip shown below, don't show a tooltip here
+ // Else, show the tooltip based on the button state
+ disabledReason ? undefined : showStopButton ? (
+ <>
+ Let's bail <KeyboardShortcut enter />
+ </>
+ ) : isQueueingSubmission ? (
+ <>
+ Queue message <KeyboardShortcut enter />
+ </>
+ ) : (
+ <>
+ Let's go! <KeyboardShortcut enter />
+ </>
+ )
+ }
+ loading={threadLoading && !dataProcessingAccepted}
+ disabledReason={disabledReason}
+ className={disabledReason ? 'opacity-[0.5]' : ''}
+ size="small"
+ icon={
+ showStopButton ? (
+ <IconStopFilled />
+ ) : (
+ MAX_SLASH_COMMANDS.find((cmd) => cmd.name === inputValue.split(' ', 1)[0])
+ ?.icon || <IconArrowRight />
+ )
+ }
+ />
+ </AIConsentPopoverWrapper>
</div>
</div>
{/* Info banner for conversations created during impersonation (marked as internal) */}
diff --git a/frontend/src/scenes/max/handsFreeLogic.test.ts b/frontend/src/scenes/max/handsFreeLogic.test.ts
deleted file mode 100644
index ea4da5b20098..000000000000
--- a/frontend/src/scenes/max/handsFreeLogic.test.ts
+++ /dev/null
@@ -1,110 +0,0 @@
-import { expectLogic } from 'kea-test-utils'
-
-import { useMocks } from '~/mocks/jest'
-import { initKeaTests } from '~/test/init'
-
-import { classifyPartial, handsFreeLogic, type SuppressionReason } from './handsFreeLogic'
-import { maxLogic } from './maxLogic'
-import { maxMocks } from './testUtils'
-
-describe('handsFreeLogic state machine', () => {
- let logic: ReturnType<typeof handsFreeLogic.build>
- let parent: ReturnType<typeof maxLogic.build>
-
- beforeEach(() => {
- useMocks(maxMocks)
- initKeaTests()
- parent = maxLogic({ panelId: 'hands-free-test-tab' })
- parent.mount()
- logic = handsFreeLogic({ panelId: 'hands-free-test-tab' })
- logic.mount()
- })
-
- afterEach(() => {
- logic.unmount()
- parent.unmount()
- })
-
- it('starts off with idle connection and exposes isActive=false', async () => {
- await expectLogic(logic).toMatchValues({ status: 'off', connection: 'idle', isActive: false })
- })
-
- it('exitHandsFree from off keeps status off (defensive)', async () => {
- await expectLogic(logic, () => {
- logic.actions.exitHandsFree('test')
- }).toMatchValues({ status: 'off', connection: 'idle' })
- })
-
- it('setStatus("listening") then exitHandsFree resets back to off + connection idle', async () => {
- await expectLogic(logic, () => {
- logic.actions.setStatus('listening')
- logic.actions.setConnection('connected')
- }).toMatchValues({ status: 'listening', connection: 'connected', isActive: true })
-
- await expectLogic(logic, () => {
- logic.actions.exitHandsFree('test')
- }).toMatchValues({ status: 'off', connection: 'idle', isActive: false })
- })
-
- it('setError stores the error string; entering hands-free clears it', async () => {
- await expectLogic(logic, () => {
- logic.actions.setError('something broke')
- }).toMatchValues({ error: 'something broke' })
-
- await expectLogic(logic, () => {
- logic.actions.setError(null)
- }).toMatchValues({ error: null })
- })
-
- it('setPartialTranscript stores the in-flight transcript; commitTranscript clears it', async () => {
- await expectLogic(logic, () => {
- logic.actions.setPartialTranscript('how is the prod')
- }).toMatchValues({ partialTranscript: 'how is the prod' })
-
- // commit triggers askMax via the listener; thread isn't mounted here so it'll error out,
- // but the reducer for partialTranscript clears regardless
- await expectLogic(logic, () => {
- logic.actions.commitTranscript('how is the product doing today')
- }).toMatchValues({ partialTranscript: '' })
- })
-
- it('exitHandsFree clears any in-flight partial transcript', async () => {
- await expectLogic(logic, () => {
- logic.actions.setPartialTranscript('mid sentence')
- }).toMatchValues({ partialTranscript: 'mid sentence' })
-
- await expectLogic(logic, () => {
- logic.actions.exitHandsFree('test')
- }).toMatchValues({ partialTranscript: '' })
- })
-
- it('speakAssistantResponse with empty summary while off is a no-op', async () => {
- await expectLogic(logic, () => {
- logic.actions.speakAssistantResponse({ text: '', vizCount: 0 })
- }).toMatchValues({ status: 'off' })
- })
-})
-
-describe('classifyPartial barge-in heuristic', () => {
- const spoken = 'lets look at the daily active users for the last week'
- const cases: { name: string; spoken: string; partial: string; expected: SuppressionReason }[] = [
- { name: 'empty partial -> too_short', spoken, partial: '', expected: 'too_short' },
- { name: '1-char partial -> too_short', spoken, partial: 'a', expected: 'too_short' },
- { name: '3-char unrelated -> too_short', spoken, partial: 'xyz', expected: 'too_short' },
- { name: 'allowlist "stop" bypasses too_short', spoken, partial: 'stop', expected: null },
- { name: 'allowlist "no" bypasses too_short', spoken, partial: 'no', expected: null },
- { name: 'allowlist "wait" bypasses too_short', spoken, partial: 'wait', expected: null },
- { name: 'allowlist "max" bypasses too_short', spoken, partial: 'max', expected: null },
- { name: 'substring of spoken text -> substring', spoken, partial: 'daily active', expected: 'substring' },
- { name: 'unrelated longer phrase -> not suppressed', spoken, partial: 'change the chart type', expected: null },
- {
- name: 'partial matches mid-word run in spoken -> substring',
- spoken,
- partial: 'active users',
- expected: 'substring',
- },
- ]
- it.each(cases)('$name', ({ spoken, partial, expected }) => {
- expect(classifyPartial(spoken, partial)).toBe(expected)
- })
-})
diff --git a/frontend/src/scenes/max/handsFreeLogic.ts b/frontend/src/scenes/max/handsFreeLogic.ts
deleted file mode 100644
index 56eb0af9f5ef..000000000000
--- a/frontend/src/scenes/max/handsFreeLogic.ts
+++ /dev/null
@@ -1,662 +0,0 @@
-// Hands-free conversational mode for Max — STT via ElevenLabs Scribe (browser → WS direct),
-// TTS via the browser's native speechSynthesis. Built for mobile-web use (gym, walking).
-import { MakeLogicType, actions, afterMount, connect, kea, key, listeners, path, props, reducers, selectors } from 'kea'
-import posthog from 'posthog-js'
-
-import api from 'lib/api'
-import { FEATURE_FLAGS } from 'lib/constants'
-import { lemonToast } from 'lib/lemon-ui/LemonToast/LemonToast'
-
-import { AssistantSummary, buildSpokenText } from './handsFreeUtils'
-import { maxLogic } from './maxLogic'
-import { maxThreadLogic } from './maxThreadLogic'
-
-export type HandsFreeStatus = 'off' | 'starting' | 'listening' | 'thinking' | 'speaking'
-export type HandsFreeConnection = 'idle' | 'connecting' | 'connected' | 'reconnecting' | 'closed'
-
-const SCRIBE_REALTIME_MODEL_ID = 'scribe_v2_realtime'
-
-export interface HandsFreeLogicProps {
- panelId?: string // identifies the MaxLogic instance backing this panel (scene tab id or side panel)
-}
-
-interface ScribeConnection {
- close: () => void
- on: (event: string, listener: (payload: any) => void) => void
- off: (event: string, listener: (payload: any) => void) => void
-}
-
-interface ScribeNamespace {
- connect: (options: {
- token: string
- modelId: string
- commitStrategy?: string
- microphone?: { echoCancellation?: boolean; noiseSuppression?: boolean }
- }) => ScribeConnection
-}
-
-type RealtimeEventsEnum = Record<string, string>
-type CommitStrategyEnum = { MANUAL: string; VAD: string }
-
-let cachedSdkPromise: Promise<{
- Scribe: ScribeNamespace
- RealtimeEvents: RealtimeEventsEnum
- CommitStrategy: CommitStrategyEnum
-}> | null = null
-
-async function loadScribeSdk(): Promise<{
- Scribe: ScribeNamespace
- RealtimeEvents: RealtimeEventsEnum
- CommitStrategy: CommitStrategyEnum
-}> {
- if (!cachedSdkPromise) {
- cachedSdkPromise = import(/* webpackChunkName: "elevenlabs-client" */ '@elevenlabs/client')
- .then((mod) => ({
- Scribe: (mod as unknown as { Scribe: ScribeNamespace }).Scribe,
- RealtimeEvents: (mod as unknown as { RealtimeEvents: RealtimeEventsEnum }).RealtimeEvents,
- CommitStrategy: (mod as unknown as { CommitStrategy: CommitStrategyEnum }).CommitStrategy,
- }))
- .catch((err) => {
- cachedSdkPromise = null
- throw err
- })
- }
- return cachedSdkPromise
-}
-
-// Generated by kea-typegen. Update if you're an agent, ignore if you're human.
-export interface handsFreeLogicValues {
- threadLogicKey: string // maxLogic
- canUseHandsFree: boolean
- connection: HandsFreeConnection
- error: string | null
- isActive: boolean
- partialTranscript: string
- sdkAvailable: boolean | null
- status: HandsFreeStatus
-}
-
-// Generated by kea-typegen. Update if you're an agent, ignore if you're human.
-export interface handsFreeLogicActions {
- cancelSpeaking: () => {
- value: true
- }
- commitTranscript: (text: string) => {
- text: string
- }
- enterHandsFree: () => {
- value: true
- }
- exitHandsFree: (reason?: string) => {
- reason: string
- }
- interruptSpeaking: () => {
- value: true
- }
- setConnection: (connection: HandsFreeConnection) => {
- connection: HandsFreeConnection
- }
- setError: (error: string | null) => {
- error: string | null
- }
- setPartialTranscript: (text: string) => {
- text: string
- }
- setSdkAvailable: (available: boolean) => {
- available: boolean
- }
- setStatus: (status: HandsFreeStatus) => {
- status: HandsFreeStatus
- }
- speakAssistantResponse: (summary: AssistantSummary) => {
- summary: AssistantSummary
- }
- toggleHandsFree: () => {
- value: true
- }
-}
-
-// Generated by kea-typegen. Update if you're an agent, ignore if you're human.
-export interface handsFreeLogicMeta {
- key: string
- __keaTypeGenInternalSelectorTypes: {
- isActive: (status: HandsFreeStatus) => boolean
- canUseHandsFree: (sdkAvailable: boolean | null) => boolean
- }
-}
-
-export type handsFreeLogicType = MakeLogicType<
- handsFreeLogicValues,
- handsFreeLogicActions,
- HandsFreeLogicProps,
- handsFreeLogicMeta
->
-
-export const handsFreeLogic = kea<handsFreeLogicType>([
- props({} as HandsFreeLogicProps),
- key((props) => props.panelId as string),
- path((key) => ['scenes', 'max', 'handsFreeLogic', key]),
-
- connect(({ panelId }: HandsFreeLogicProps) => ({
- values: [maxLogic({ panelId }), ['threadLogicKey']],
- })),
-
- actions({
- enterHandsFree: true,
- exitHandsFree: (reason?: string) => ({ reason: reason ?? 'user' }),
- toggleHandsFree: true,
- setStatus: (status: HandsFreeStatus) => ({ status }),
- setConnection: (connection: HandsFreeConnection) => ({ connection }),
- setError: (error: string | null) => ({ error }),
- setPartialTranscript: (text: string) => ({ text }),
- commitTranscript: (text: string) => ({ text }),
- speakAssistantResponse: (summary: AssistantSummary) => ({ summary }),
- cancelSpeaking: true,
- // Stop TTS playback and go back to listening — fired only when Scribe transcribes
- // genuine user speech over Max's TTS (voice barge-in). The mic button no longer
- // tap-interrupts; tapping it always exits hands-free entirely.
- interruptSpeaking: true,
- setSdkAvailable: (available: boolean) => ({ available }),
- }),
-
- reducers({
- status: [
- 'off' as HandsFreeStatus,
- {
- setStatus: (_, { status }) => status,
- },
- ],
- connection: [
- 'idle' as HandsFreeConnection,
- {
- setConnection: (_, { connection }) => connection,
- exitHandsFree: () => 'idle' as HandsFreeConnection,
- },
- ],
- error: [
- null as string | null,
- {
- setError: (_, { error }) => error,
- enterHandsFree: () => null,
- exitHandsFree: () => null,
- },
- ],
- partialTranscript: [
- '' as string,
- {
- setPartialTranscript: (_, { text }) => text,
- commitTranscript: () => '',
- exitHandsFree: () => '',
- },
- ],
- // null = probe not yet finished, true = Scribe present, false = SDK missing or load failed
- sdkAvailable: [
- null as boolean | null,
- {
- setSdkAvailable: (_, { available }) => available,
- },
- ],
- }),
-
- selectors({
- isActive: [(s) => [s.status], (status: HandsFreeStatus) => status !== 'off'],
- // Render the mic button when the probe hasn't completed yet (null) or has succeeded (true).
- // Hide it only once we've confirmed the SDK can't be used.
- canUseHandsFree: [(s) => [s.sdkAvailable], (sdkAvailable: boolean | null) => sdkAvailable !== false],
- }),
-
- listeners(({ actions, values, cache, props }) => ({
- toggleHandsFree: () => {
- if (values.status === 'off') {
- actions.enterHandsFree()
- } else {
- actions.exitHandsFree()
- }
- },
-
- enterHandsFree: async () => {
- if (values.status !== 'off') {
- return
- }
- actions.setStatus('starting')
- actions.setConnection('connecting')
-
- // Per-session analytics counters, captured at exit so a single PostHog query can
- // build session-length and turn-count distributions without re-aggregating events.
- cache.sessionStartedAt = Date.now()
- cache.userTurnCount = 0
- cache.assistantTurnCount = 0
- cache.interruptedCount = 0
-
- // iOS Safari requires the first audio playback (HTMLAudioElement.play()) to be
- // initiated from a user-gesture handler. The mic-button click IS that gesture, so
- // we prime a silent Audio element here. Subsequent .play() calls (driven by Max
- // responses arriving over the network) then succeed without re-prompting.
- try {
- const primer = new Audio()
- primer.muted = true
- primer.src =
- 'data:audio/mpeg;base64,SUQzBAAAAAAAI1RTU0UAAAAPAAADTGF2ZjU4LjI5LjEwMAAAAAAAAAAAAAAA//tQxAADB8AhSmxhhgAAAAAAAAAAAExBTUUzLjEwMACqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqq//tSxAADB8AhSmxhhgAAAAAAAAAAAExBTUUzLjEwMKqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqq'
- void primer.play().catch(() => undefined)
- } catch {
- // best-effort
- }
-
- let sdk
- try {
- sdk = await loadScribeSdk()
- } catch (err) {
- posthog.captureException(err)
- actions.setError('Could not load the hands-free SDK.')
- lemonToast.error('Hands-free could not load. Please try again.')
- actions.exitHandsFree('sdk_load_failed')
- return
- }
- if (values.status === 'off') {
- return
- }
-
- const { Scribe, RealtimeEvents, CommitStrategy } = sdk
- if (!Scribe || typeof Scribe.connect !== 'function') {
- // Should be unreachable — the mount-time probe should have already hidden the
- // button when the SDK doesn't expose Scribe. If we get here, capture so we can
- // see it in PostHog and bail without confusing the user.
- posthog.captureException(new Error('handsFree: Scribe namespace missing from @elevenlabs/client'))
- actions.setSdkAvailable(false)
- actions.exitHandsFree('sdk_missing_scribe')
- return
- }
-
- cache.reconnectAttempts = 0
-
- // Mints a fresh token, opens a Scribe connection, and wires event handlers.
- // Called for the initial session and on reconnect after an unexpected close.
- // Returns true on success, false when the caller should give up.
- const establishConnection = async (isReconnect: boolean): Promise<boolean> => {
- let token: string
- try {
- const response = await api.maxHandsFree.token()
- token = response.token
- } catch (err) {
- posthog.captureException(err)
- if (!isReconnect) {
- actions.setError('Could not start hands-free session.')
- lemonToast.error('Hands-free could not start. Please try again.')
- }
- actions.exitHandsFree(isReconnect ? 'reconnect_token_failed' : 'token_failed')
- return false
- }
- if (values.status === 'off') {
- return false
- }
- let connection: ScribeConnection
- try {
- connection = Scribe.connect({
- token,
- modelId: SCRIBE_REALTIME_MODEL_ID,
- // SDK defaults to MANUAL — we need VAD so silence on the mic auto-commits
- // the speech segment and triggers askMax. Without this it isn't hands-free.
- commitStrategy: CommitStrategy?.VAD ?? 'vad',
- microphone: { echoCancellation: true, noiseSuppression: true },
- })
- } catch (err) {
- posthog.captureException(err)
- const message = (err as Error)?.message ?? ''
- const isPermissionError =
- /permission|notallowed|denied/i.test(message) ||
- (err as { name?: string })?.name === 'NotAllowedError'
- if (isPermissionError) {
- actions.setError('Microphone access was denied.')
- lemonToast.error('Microphone access was denied. Hands-free disabled.')
- actions.exitHandsFree('mic_permission_denied')
- } else if (isReconnect) {
- actions.exitHandsFree('reconnect_failed')
- } else {
- actions.setError(`Hands-free connection failed: ${message || 'unknown error'}`)
- lemonToast.error('Hands-free failed to start. See console for details.')
- actions.exitHandsFree('connection_failed')
- }
- return false
- }
- cache.connection = connection
- wireConnectionEvents(connection)
- return true
- }
-
- const onOpen = (): void => actions.setConnection('connected')
- const onSessionStarted = (): void => {
- const wasStarting = values.status === 'starting'
- const wasReconnecting = cache.isReconnecting === true
- cache.isReconnecting = false
- if (wasStarting || wasReconnecting) {
- actions.setStatus('listening')
- }
- if (wasStarting) {
- posthog.capture('max hands-free entered')
- }
- if (wasReconnecting) {
- posthog.capture('max hands-free reconnected', {
- attempts: cache.reconnectAttempts ?? 1,
- })
- }
- }
- // While 'speaking', the Scribe mic stream stays open (we can't cheaply pause it)
- // and the speakers' TTS audio bleeds back into the mic — Scribe transcribes its
- // own voice. To support real barge-in (user talks over Max -> Max shuts up and
- // listens) without self-triggering, we compare incoming partials to what Max is
- // currently saying. If the partial is a substring of the spoken text, it's
- // self-transcription bleed; otherwise the user is talking over Max.
- const onPartial = (payload: any): void => {
- const text: string = payload?.text ?? payload?.transcript ?? ''
- if (values.status === 'speaking') {
- const partialLower = normaliseForBargeInMatch(text)
- const spokenLower: string = cache.spokenTextLower ?? ''
- const reason = classifyPartial(spokenLower, partialLower)
- if (reason) {
- posthog.capture('max hands-free partial suppressed', {
- reason,
- partial_chars: partialLower.length,
- spoken_chars: spokenLower.length,
- phase: 'partial',
- })
- return
- }
- // User is talking over Max — barge in: stop TTS, flip to listening, then
- // let the partial-transcript reducer below pick up the user's words.
- actions.interruptSpeaking()
- }
- if (values.status === 'listening') {
- actions.setPartialTranscript(text)
- }
- }
- const onCommitted = (payload: any): void => {
- const text: string = (payload?.text ?? payload?.transcript ?? '').trim()
- if (!text) {
- return
- }
- if (values.status === 'speaking') {
- const committedLower = normaliseForBargeInMatch(text)
- const spokenLower: string = cache.spokenTextLower ?? ''
- const reason = classifyPartial(spokenLower, committedLower)
- if (reason) {
- posthog.capture('max hands-free partial suppressed', {
- reason,
- partial_chars: committedLower.length,
- spoken_chars: spokenLower.length,
- phase: 'committed',
- })
- return
- }
- actions.interruptSpeaking()
- }
- if (values.status !== 'listening') {
- return
- }
- actions.commitTranscript(text)
- }
- const onClose = (): void => {
- if (values.status === 'off') {
- return
- }
- // Single retry-with-fresh-token on unexpected close — mobile network blips
- // (the exact gym/walking use case this targets) should not tear down the
- // session. Beyond one attempt we give up to avoid a reconnect storm.
- const attempts = (cache.reconnectAttempts ?? 0) + 1
- cache.reconnectAttempts = attempts
- if (attempts > 1) {
- actions.setConnection('closed')
- actions.exitHandsFree('connection_closed')
- return
- }
- actions.setConnection('reconnecting')
- cache.isReconnecting = true
- posthog.capture('max hands-free reconnect attempted', { attempt: attempts })
- cache.connection = undefined
- void establishConnection(true)
- }
- const onError = (payload: any): void => {
- posthog.captureException(new Error(`scribe error: ${JSON.stringify(payload ?? {})}`))
- actions.setError('Hands-free connection error.')
- actions.exitHandsFree('scribe_error')
- }
-
- const wireConnectionEvents = (connection: ScribeConnection): void => {
- connection.on(RealtimeEvents.OPEN, onOpen)
- connection.on(RealtimeEvents.SESSION_STARTED, onSessionStarted)
- connection.on(RealtimeEvents.PARTIAL_TRANSCRIPT, onPartial)
- connection.on(RealtimeEvents.COMMITTED_TRANSCRIPT, onCommitted)
- connection.on(RealtimeEvents.CLOSE, onClose)
- connection.on(RealtimeEvents.ERROR, onError)
- }
-
- await establishConnection(false)
- },
-
- exitHandsFree: ({ reason }) => {
- posthog.capture('max hands-free exited', {
- reason,
- duration_ms: cache.sessionStartedAt ? Date.now() - cache.sessionStartedAt : null,
- user_turns: cache.userTurnCount ?? 0,
- assistant_turns: cache.assistantTurnCount ?? 0,
- interruptions: cache.interruptedCount ?? 0,
- })
- actions.cancelSpeaking()
- const connection = cache.connection as ScribeConnection | undefined
- if (connection) {
- try {
- connection.close()
- } catch {
- // best-effort
- }
- cache.connection = undefined
- }
- actions.setStatus('off')
- },
-
- commitTranscript: ({ text }) => {
- const threadKey = values.threadLogicKey
- const threadLogic = threadKey
- ? maxThreadLogic.findMounted({
- panelId: props.panelId,
- conversationId: threadKey,
- })
- : null
- if (!threadLogic) {
- actions.setError('Could not find an active conversation to send the message to.')
- actions.exitHandsFree('no_thread')
- return
- }
- actions.setStatus('thinking')
- cache.userTurnCount = (cache.userTurnCount ?? 0) + 1
- posthog.capture('max hands-free user spoke', { transcript_chars: text.length })
- threadLogic.actions.askMax(text)
- },
-
- speakAssistantResponse: async ({ summary }) => {
- if (values.status === 'off') {
- return
- }
- // guard re-entry — if two assistant turns complete back-to-back, tear down the
- // previous playback (revokes blob URL, aborts inflight request) before queuing a
- // new one. Otherwise the previous audio leaks and overlaps.
- teardownSpeaking(cache)
-
- const spokenText = buildSpokenText(summary)
- if (!spokenText) {
- actions.setStatus('listening')
- return
- }
- actions.setStatus('speaking')
-
- // Store what we're saying so onPartial can match incoming transcripts against it
- // and tell self-transcription bleed apart from real user barge-in.
- cache.spokenTextLower = normaliseForBargeInMatch(spokenText)
-
- const controller = new AbortController()
- cache.speakAbortController = controller
-
- try {
- const response = await api.maxHandsFree.synthesize(spokenText, { signal: controller.signal })
- if (!response.ok) {
- throw new Error(`TTS returned ${response.status}`)
- }
- const audioBlob = await response.blob()
- if (controller.signal.aborted || values.status !== 'speaking') {
- return
- }
- const url = URL.createObjectURL(audioBlob)
- const audioElement = new Audio(url)
- cache.speakAudioElement = audioElement
- cache.speakAudioUrl = url
-
- const finished = new Promise<void>((resolve, reject) => {
- audioElement.onended = () => resolve()
- audioElement.onerror = () => reject(new Error('audio playback failed'))
- })
-
- await audioElement.play()
- // Capture once playback has actually started — distinguishes "Max attempted to
- // speak" from "Max successfully voiced a response." Counted per session via
- // assistant_turns on the exit event.
- cache.assistantTurnCount = (cache.assistantTurnCount ?? 0) + 1
- posthog.capture('max hands-free assistant spoke', {
- text_chars: spokenText.length,
- viz_count: summary.vizCount,
- })
- await finished
- } catch (err) {
- if ((err as { name?: string }).name !== 'AbortError') {
- posthog.captureException(err)
- }
- } finally {
- if (cache.speakAbortController === controller) {
- teardownSpeaking(cache)
- if (values.status === 'speaking') {
- actions.setStatus('listening')
- }
- }
- }
- },
-
- cancelSpeaking: () => {
- teardownSpeaking(cache)
- },
-
- interruptSpeaking: () => {
- // Only meaningful while we're actually playing audio — guard against stray
- // fires from elsewhere accidentally jolting us through state transitions.
- if (values.status !== 'speaking') {
- return
- }
- teardownSpeaking(cache)
- actions.setStatus('listening')
- cache.interruptedCount = (cache.interruptedCount ?? 0) + 1
- posthog.capture('max hands-free tts interrupted')
- },
- })),
-
- afterMount(({ actions, cache }) => {
- // Probe the SDK once on mount so the mic button can decide whether to render itself.
- // We capture failures to PostHog instead of surfacing dev-facing errors to end users.
- // Skip the probe entirely when the feature flag is off — the logic mounts for every
- // Max session, so an unconditional dynamic import would download the SDK chunk for
- // every user even though only flagged users can ever use it.
- if (!posthog.isFeatureEnabled(FEATURE_FLAGS.MAX_HANDS_FREE)) {
- actions.setSdkAvailable(false)
- return
- }
- cache.disposables.add(() => {
- let cancelled = false
- void loadScribeSdk()
- .then((sdk) => {
- if (cancelled) {
- return
- }
- const available = !!sdk?.Scribe && typeof sdk.Scribe.connect === 'function'
- if (!available) {
- posthog.captureException(
- new Error('handsFree: @elevenlabs/client loaded but Scribe is missing')
- )
- }
- actions.setSdkAvailable(available)
- })
- .catch((err) => {
- if (cancelled) {
- return
- }
- posthog.captureException(err)
- actions.setSdkAvailable(false)
- })
- return () => {
- cancelled = true
- }
- }, 'sdkProbe')
- }),
-])
-
-// Normalises a string to lowercase letters and digits separated by single spaces. Used to
-// match an incoming Scribe partial against the TTS content for self-transcription detection.
-function normaliseForBargeInMatch(text: string): string {
- return text
- .toLowerCase()
- .replace(/[^a-z0-9]+/g, ' ')
- .trim()
-}
-
-// Short words a gym/walking user is likely to shout to stop Max mid-sentence. These
-// always trigger barge-in even though they'd otherwise fall under the too-short or
-// substring suppression — a wrong barge-in is recoverable (user just keeps talking),
-// a swallowed "stop" is not.
-const BARGE_IN_TRIGGER_WORDS: ReadonlySet<string> = new Set(['stop', 'wait', 'cancel', 'no', 'max', 'shh', 'hey'])
-
-export type SuppressionReason = 'too_short' | 'substring' | null
-
-// Returns null when the partial should be treated as real user speech (barge in), or a
-// reason string when it should be suppressed as TTS bleed. Callers should emit the
-// reason to PostHog so we can measure misfire rate before tuning the heuristic.
-export function classifyPartial(spokenLower: string, partialLower: string): SuppressionReason {
- if (!partialLower) {
- return 'too_short'
- }
- if (BARGE_IN_TRIGGER_WORDS.has(partialLower)) {
- return null
- }
- if (partialLower.length < 4) {
- return 'too_short'
- }
- if (spokenLower.includes(partialLower)) {
- return 'substring'
- }
- return null
-}
-
-function teardownSpeaking(cache: Record<string, any>): void {
- const controller = cache.speakAbortController as AbortController | undefined
- if (controller) {
- try {
- controller.abort()
- } catch {
- // best-effort
- }
- cache.speakAbortController = undefined
- }
- const audio = cache.speakAudioElement as HTMLAudioElement | undefined
- if (audio) {
- try {
- audio.pause()
- audio.src = ''
- } catch {
- // best-effort
- }
- cache.speakAudioElement = undefined
- }
- if (cache.speakAudioUrl) {
- try {
- URL.revokeObjectURL(cache.speakAudioUrl)
- } catch {
- // best-effort
- }
- cache.speakAudioUrl = undefined
- }
- cache.spokenTextLower = undefined
-}
diff --git a/frontend/src/scenes/max/handsFreeUtils.test.ts b/frontend/src/scenes/max/handsFreeUtils.test.ts
deleted file mode 100644
index 995c028b302f..000000000000
--- a/frontend/src/scenes/max/handsFreeUtils.test.ts
+++ /dev/null
@@ -1,120 +0,0 @@
-import { AssistantMessageType } from '~/queries/schema/schema-assistant-messages'
-
-import { buildSpokenText, stripMarkdown, summariseAssistantThread } from './handsFreeUtils'
-
-type StubMessage = Parameters<typeof summariseAssistantThread>[0][number]
-
-const human = (content = 'hi'): StubMessage => ({ type: AssistantMessageType.Human, content })
-const ai = (content: string): StubMessage => ({ type: AssistantMessageType.Assistant, content })
-const viz = (): StubMessage => ({ type: AssistantMessageType.Visualization })
-const multiViz = (): StubMessage => ({ type: AssistantMessageType.MultiVisualization })
-const artifact = (): StubMessage => ({ type: AssistantMessageType.Artifact })
-
-describe('handsFreeUtils', () => {
- describe('summariseAssistantThread()', () => {
- it('returns the assistant prose message and no viz suffix when none present', () => {
- const result = summariseAssistantThread([human(), ai('here is the answer')])
- expect(result).toEqual({ text: 'here is the answer', vizCount: 0 })
- })
-
- it.each([
- ['Visualization', [ai('see chart'), viz()], 1],
- ['MultiVisualization', [ai('see charts'), multiViz()], 1],
- ['Artifact', [ai('see notebook'), artifact()], 1],
- ['mix of viz types', [ai('see charts'), viz(), multiViz(), artifact()], 3],
- ] as const)('counts %s blocks since the last human message', (_label, tail, expectedCount) => {
- const result = summariseAssistantThread([human(), ...tail])
- expect(result.vizCount).toBe(expectedCount)
- })
-
- it('only counts messages produced after the most recent human message', () => {
- const result = summariseAssistantThread([
- human('first ask'),
- ai('old answer'),
- viz(),
- human('second ask'),
- ai('new answer'),
- viz(),
- ])
- expect(result).toEqual({ text: 'new answer', vizCount: 1 })
- })
-
- it('returns the LATEST prose message when the assistant emits several in one turn', () => {
- const result = summariseAssistantThread([
- human(),
- ai('first part of the answer'),
- ai('second part of the answer'),
- ai('final summary'),
- ])
- expect(result).toEqual({ text: 'final summary', vizCount: 0 })
- })
-
- it('returns empty text when only viz messages followed the human turn', () => {
- const result = summariseAssistantThread([human(), viz(), multiViz()])
- expect(result).toEqual({ text: '', vizCount: 2 })
- })
-
- it('returns empty text and zero count when thread has no assistant turn', () => {
- const result = summariseAssistantThread([human()])
- expect(result).toEqual({ text: '', vizCount: 0 })
- })
-
- it('handles a thread that never had a human turn (e.g. agent-initiated)', () => {
- const result = summariseAssistantThread([ai('hello there'), viz()])
- expect(result).toEqual({ text: 'hello there', vizCount: 1 })
- })
- })
-
- describe('buildSpokenText()', () => {
- it.each([
- [{ text: 'just words', vizCount: 0 }, 'just words'],
- [{ text: 'words', vizCount: 1 }, "words I've added 1 chart to the chat."],
- [{ text: 'words', vizCount: 3 }, "words I've added 3 charts to the chat."],
- [{ text: '', vizCount: 2 }, "I've added 2 charts to the chat."],
- [{ text: '', vizCount: 0 }, ''],
- ])('builds %p -> %p', (summary, expected) => {
- expect(buildSpokenText(summary)).toBe(expected)
- })
- })
-
- describe('stripMarkdown()', () => {
- it.each([
- ['# Heading', 'Heading'],
- ['**bold text**', 'bold text'],
- ['_italic_', 'italic'],
- ['*emphasised*', 'emphasised'],
- ['~~struck~~', 'struck'],
- ['`inline code`', 'inline code'],
- ['1. first item\n2. second item', 'first item second item'],
- ['- bullet\n- bullet', 'bullet bullet'],
- ['[link text](https://example.com)', 'link text'],
- ['![alt](https://example.com/x.png) caption', ' caption'],
- ['> quoted line', 'quoted line'],
- ])('strips %p -> %p', (input, expected) => {
- expect(stripMarkdown(input)).toBe(expected)
- })
-
- it('leaves identifiers like foo_bar_baz alone (bold underscore guard)', () => {
- expect(stripMarkdown('the var is foo_bar_baz here')).toBe('the var is foo_bar_baz here')
- })
-
- it('omits fenced code blocks rather than reading them verbatim', () => {
- const input = 'before\n```ts\nconst x = 1\n```\nafter'
- expect(stripMarkdown(input)).toContain('code block omitted')
- expect(stripMarkdown(input)).not.toContain('const x')
- })
-
- it('strips table separator rows and pipes', () => {
- const input = '| col a | col b |\n| --- | --- |\n| one | two |'
- const result = stripMarkdown(input)
- expect(result).not.toContain('---')
- expect(result).not.toContain('|')
- expect(result).toContain('col a')
- expect(result).toContain('one')
- })
-
- it('collapses runs of whitespace so TTS does not over-pause', () => {
- expect(stripMarkdown('one\n\n\ntwo')).toBe('one two')
- })
- })
-})
diff --git a/frontend/src/scenes/max/handsFreeUtils.ts b/frontend/src/scenes/max/handsFreeUtils.ts
deleted file mode 100644
index ba42e32dbf7b..000000000000
--- a/frontend/src/scenes/max/handsFreeUtils.ts
+++ /dev/null
@@ -1,85 +0,0 @@
-import { AssistantMessageType } from '~/queries/schema/schema-assistant-messages'
-
-const VIZ_TYPES = new Set<AssistantMessageType>([
- AssistantMessageType.Visualization,
- AssistantMessageType.MultiVisualization,
- AssistantMessageType.Artifact,
-])
-
-export interface AssistantSummary {
- text: string
- vizCount: number
-}
-
-export function summariseAssistantThread(
- messages: { type: AssistantMessageType; content?: unknown }[]
-): AssistantSummary {
- let humanCutoff = -1
- for (let i = messages.length - 1; i >= 0; i--) {
- if (messages[i].type === AssistantMessageType.Human) {
- humanCutoff = i
- break
- }
- }
-
- let latestProse = ''
- let vizCount = 0
- for (let i = humanCutoff + 1; i < messages.length; i++) {
- const msg = messages[i]
- if (VIZ_TYPES.has(msg.type)) {
- vizCount += 1
- continue
- }
- if (msg.type === AssistantMessageType.Assistant && typeof msg.content === 'string' && msg.content) {
- latestProse = msg.content
- }
- }
- return { text: latestProse, vizCount }
-}
-
-export function buildSpokenText(summary: AssistantSummary): string {
- const stripped = stripMarkdown(summary.text).trim()
- if (summary.vizCount === 0) {
- return stripped
- }
- const suffix =
- summary.vizCount === 1
- ? "I've added 1 chart to the chat."
- : `I've added ${summary.vizCount} charts to the chat.`
- return stripped ? `${stripped} ${suffix}` : suffix
-}
-
-export function stripMarkdown(text: string): string {
- return (
- text
- // fenced code blocks - replaced wholesale; never spoken
- .replace(/```[\s\S]*?```/g, ' (code block omitted) ')
- // inline code
- .replace(/`([^`]+)`/g, '$1')
- // images
- .replace(/!\[[^\]]*\]\([^)]+\)/g, '')
- // links - keep link text
- .replace(/\[([^\]]+)\]\([^)]+\)/g, '$1')
- // tables - drop row-separator lines (| --- |) and pipe characters elsewhere
- .replace(/^\s*\|?\s*:?-{2,}:?(\s*\|\s*:?-{2,}:?)*\s*\|?\s*$/gm, '')
- .replace(/\|/g, ' ')
- // headings
- .replace(/^#+\s*/gm, '')
- // bold - **foo** or __foo__ - run twice to safely handle nested markers
- .replace(/\*\*([^*]+?)\*\*/g, '$1')
- .replace(/__([^_]+?)__/g, '$1')
- // italic - *foo* or _foo_ - but only when the marker is at a word boundary so
- // identifiers like foo_bar_baz aren't mangled
- .replace(/(^|[^A-Za-z0-9_])[*_]([^*_\n]+?)[*_](?=$|[^A-Za-z0-9_])/g, '$1$2')
- // strikethrough
- .replace(/~~([^~]+?)~~/g, '$1')
- // blockquotes
- .replace(/^\s*>\s?/gm, '')
- // bulleted lists
- .replace(/^\s*[-*+]\s+/gm, '')
- // ordered lists
- .replace(/^\s*\d+[.)]\s+/gm, '')
- // collapse runs of whitespace - TTS reads runs of "\n\n" as awkward pauses
- .replace(/\s+/g, ' ')
- )
-}
diff --git a/frontend/src/scenes/max/maxThreadLogic.tsx b/frontend/src/scenes/max/maxThreadLogic.tsx
index 18e4611c9f64..c8877bfdc918 100644
--- a/frontend/src/scenes/max/maxThreadLogic.tsx
+++ b/frontend/src/scenes/max/maxThreadLogic.tsx
@@ -83,8 +83,6 @@ import { isPiTaskRuntime } from 'products/posthog_ai/frontend/types/taskTypes'
import type { PermissionRequestRecord } from '../../../../products/posthog_ai/frontend/types/streamTypes'
import type { FeatureFlagsSet } from '../../lib/logic/featureFlagLogic'
import type { BillingType } from '../../types'
-import { handsFreeLogic } from './handsFreeLogic'
-import { summariseAssistantThread } from './handsFreeUtils'
import {
EnhancedToolCall,
MESSAGE_TOO_LONG,
@@ -2058,16 +2056,9 @@ export const maxThreadLogic = kea<maxThreadLogicType>([
})
}
const contextualTools = Object.fromEntries(values.tools.map((tool) => [tool.identifier, tool.context]))
- // Always send voice_mode as an explicit boolean when handsFreeLogic is mounted,
- // not just when active. Otherwise a typed turn following a spoken one inherits
- // the earlier <voice_mode> system instruction from conversation history and
- // keeps formatting for speech (no markdown, spelled-out numbers).
- const handsFree = handsFreeLogic.findMounted({ panelId: props.panelId })
- const voiceMode = handsFree ? { voice_mode: handsFree.values.isActive } : undefined
- const mergedUiContext =
- uiContext || voiceMode
- ? { ...values.compiledContext, ...uiContext, ...voiceMode }
- : values.compiledContext || undefined
+ const mergedUiContext = uiContext
+ ? { ...values.compiledContext, ...uiContext }
+ : values.compiledContext || undefined
const billingContext =
values.billingContext && values.featureFlags[FEATURE_FLAGS.MAX_BILLING_CONTEXT]
? values.billingContext
@@ -2252,11 +2243,6 @@ export const maxThreadLogic = kea<maxThreadLogicType>([
completeThreadGeneration: () => {
actions.executePendingClientToolCall()
- const handsFree = handsFreeLogic.findMounted({ panelId: props.panelId })
- if (handsFree?.values.isActive) {
- handsFree.actions.speakAssistantResponse(summariseAssistantThread(values.threadRaw))
- }
-
// Update the conversation history to include the new conversation
actions.loadConversationHistory({ doNotUpdateCurrentThread: true })
diff --git a/frontend/src/scenes/max/maxTypes.ts b/frontend/src/scenes/max/maxTypes.ts
index 8a78a1c201f1..4ad3854f6392 100644
--- a/frontend/src/scenes/max/maxTypes.ts
+++ b/frontend/src/scenes/max/maxTypes.ts
@@ -82,10 +82,6 @@ export interface MaxUIContext {
evaluations?: MaxEvaluationContext[]
notebooks?: MaxNotebookContext[]
form_answers?: Record<string, string> // question_id -> answer for create_form tool responses
- // Request modality: true when the user is asking via hands-free voice mode. Backend
- // appends a voice-formatting instruction to the prompt so the response is suitable
- // for TTS (numbers spelled out, no markdown, etc).
- voice_mode?: boolean
}
// Taxonomic filter options
diff --git a/frontend/src/scenes/project-homepage/ai-first/HomepageInput.tsx b/frontend/src/scenes/project-homepage/ai-first/HomepageInput.tsx
index 92c5f427fc5e..e36839fb46c6 100644
--- a/frontend/src/scenes/project-homepage/ai-first/HomepageInput.tsx
+++ b/frontend/src/scenes/project-homepage/ai-first/HomepageInput.tsx
@@ -2,11 +2,10 @@ import { BindLogic, useActions, useAsyncActions, useValues } from 'kea'
import posthog from 'posthog-js'
import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
-import { IconArrowRight, IconClock, IconInfo, IconMicrophone, IconPin, IconSparkles } from '@posthog/icons'
+import { IconArrowRight, IconClock, IconInfo, IconPin, IconSparkles } from '@posthog/icons'
import { LemonButton, LemonSkeleton, Tooltip } from '@posthog/lemon-ui'
import { Search } from 'lib/components/Search/Search'
-import { useFeatureFlag } from 'lib/hooks/useFeatureFlag'
import { Link } from 'lib/lemon-ui/Link'
import { ButtonPrimitive } from 'lib/ui/Button/ButtonPrimitives'
import { Label } from 'lib/ui/Label/Label'
@@ -17,7 +16,6 @@ import { FillInHint } from 'scenes/max/components/FillInHint'
import { SidebarQuestionInput } from 'scenes/max/components/SidebarQuestionInput'
import { SuggestionCard } from 'scenes/max/components/SuggestionCard'
import { SUGGESTION_CARDS_HEIGHT_PX, COLORFUL_ICONS, TopicBadges } from 'scenes/max/components/TopicBadges'
-import { handsFreeLogic } from 'scenes/max/handsFreeLogic'
import { Intro } from 'scenes/max/Intro'
import { maxGlobalLogic } from 'scenes/max/maxGlobalLogic'
import { maxLogic } from 'scenes/max/maxLogic'
@@ -43,13 +41,9 @@ import { SUGGESTIONS_LIMIT } from './homepageSuggestions'
function IdleInput(): JSX.Element {
const { query, fillInHint } = useValues(aiFirstHomepageLogic)
- const { setQuery, submitQuery, enterAiMode, startHandsFreeChat, setFillInHint } = useActions(aiFirstHomepageLogic)
- const { dataProcessingAccepted } = useValues(maxGlobalLogic)
- const handsFreeFlag = useFeatureFlag('MAX_HANDS_FREE')
- const { canUseHandsFree } = useValues(handsFreeLogic({ panelId: HOMEPAGE_TAB_ID }))
+ const { setQuery, submitQuery, enterAiMode, setFillInHint } = useActions(aiFirstHomepageLogic)
const inputRef = useRef<HTMLTextAreaElement>(null)
- const handsFreeAvailable = handsFreeFlag && canUseHandsFree && dataProcessingAccepted
// A fill-in suggestion typed its prefix in and is waiting for the user to complete it.
const showFillInHint = !!fillInHint
@@ -163,18 +157,6 @@ function IdleInput(): JSX.Element {
>
<span className="text-xxs">Tab to search</span>
</ButtonPrimitive>
- {handsFreeAvailable && (
- <Tooltip title="Start a new chat in hands-free">
- <ButtonPrimitive
- iconOnly
- data-attr="homepage-hands-free"
- className="shrink-0"
- onClick={startHandsFreeChat}
- >
- <IconMicrophone className="size-4" />
- </ButtonPrimitive>
- </Tooltip>
- )}
<Tooltip title={!query.trim() ? 'Try asking a question' : undefined}>
<ButtonPrimitive
onClick={() => {
diff --git a/frontend/src/scenes/project-homepage/ai-first/aiFirstHomepageLogic.ts b/frontend/src/scenes/project-homepage/ai-first/aiFirstHomepageLogic.ts
index 03af7ef62783..6a7b20c17335 100644
--- a/frontend/src/scenes/project-homepage/ai-first/aiFirstHomepageLogic.ts
+++ b/frontend/src/scenes/project-homepage/ai-first/aiFirstHomepageLogic.ts
@@ -7,7 +7,6 @@ import type { DashboardBasicApi } from '@posthog/products-dashboards/frontend/ge
import { tabUiStateLogic } from 'lib/logic/tabUiStateLogic'
import { navigateToHref } from 'lib/utils/navigateToHref'
-import { handsFreeLogic } from 'scenes/max/handsFreeLogic'
import { type PhaiViewMode, maxGlobalLogic } from 'scenes/max/maxGlobalLogic'
import { maxLogic, parseCommandString } from 'scenes/max/maxLogic'
import { HOMEPAGE_SUGGESTION_TOPICS } from 'scenes/max/suggestionTopics'
@@ -151,9 +150,6 @@ export interface aiFirstHomepageLogicValues {
// Generated by kea-typegen. Update if you're an agent, ignore if you're human.
export interface aiFirstHomepageLogicActions {
- enterHandsFree: () => {
- value: true
- } // handsFreeLogic
openConversation: (conversationId: string) => {
conversationId: string
} // maxLogic
@@ -222,9 +218,6 @@ export interface aiFirstHomepageLogicActions {
setSelectedTopic: (key: string | null) => {
key: string | null
}
- startHandsFreeChat: () => {
- value: true
- }
submitQuery: (mode: 'ai' | 'search') => {
mode: 'ai' | 'search'
}
@@ -297,8 +290,6 @@ export const aiFirstHomepageLogic = kea<aiFirstHomepageLogicType>([
actions: [
maxLogic({ panelId: HOMEPAGE_TAB_ID }),
['openConversation', 'startNewConversation', 'setQuestion'],
- handsFreeLogic({ panelId: HOMEPAGE_TAB_ID }),
- ['enterHandsFree'],
sceneLogic,
['setHomepage'],
tabUiStateLogic,
@@ -308,7 +299,6 @@ export const aiFirstHomepageLogic = kea<aiFirstHomepageLogicType>([
actions({
submitQuery: (mode: 'search' | 'ai') => ({ mode }),
- startHandsFreeChat: true,
enterAiMode: (trigger: string) => ({ trigger }),
setQuery: (query: string) => ({ query }),
setAnimationPhase: (phase: AnimationPhase) => ({ phase }),
@@ -635,14 +625,6 @@ export const aiFirstHomepageLogic = kea<aiFirstHomepageLogicType>([
await breakpoint(200)
actions.setAnimationPhase('content')
},
- startHandsFreeChat: () => {
- posthog.capture('homepage hands-free started')
- // Clear any in-progress conversation so hands-free always starts a fresh thread,
- // then flip into AI mode (mounting maxThreadLogic) before Scribe starts listening.
- actions.startNewConversation()
- actions.submitQuery('ai')
- actions.enterHandsFree()
- },
enterAiMode: async ({ trigger }, breakpoint) => {
// Animate into AI mode without starting a conversation
await breakpoint(300)
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index ad4ef7dde2b2..789443f6f233 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -599,9 +599,6 @@ importers:
'@dnd-kit/utilities':
specifier: ^3.2.1
version: 3.2.1([email protected])
- '@elevenlabs/client':
- specifier: ^1.4.0
- version: 1.7.0(@types/[email protected])
'@floating-ui/react':
specifier: ^0.27.19
@@ -7253,9 +7250,6 @@ packages:
'@braintree/[email protected]':
resolution: {integrity: sha512-jigsZK+sMF/cuiB7sERuo9V7N9jx+dhmHHnQyDSVdpZwVutaBu7WvNYqMDLSgFgfB30n452TP3vjDAvFC973mA==}
- '@bufbuild/[email protected]':
- resolution: {integrity: sha512-wJ8ReQbHxsAfXhrf9ixl0aYbZorRuOWpBNzm8pL8ftmSxQx/wnJD5Eg861NwJU/czy2VXFIebCeZnZrI9rktIQ==}
-
'@bufbuild/[email protected]':
resolution: {integrity: sha512-i5GE2Dk5ekdlK1TR7SugY4LWRrKSfb5T1Qn4unpIMbfxoeGKERKQ59HG3iYewacGD10SR7UzevfPnh6my4tNmQ==}
@@ -7691,12 +7685,6 @@ packages:
resolution: {integrity: sha512-3DkkN3FKJ++BQy8PE3LONoVSVBONG7hld9D+VqBp1Dcfw2PGSnWyg20NET5cJkiygWK2y2H1ra8GWk3yixzuuA==}
engines: {node: '>=18.18.2'}
- '@elevenlabs/[email protected]':
- resolution: {integrity: sha512-uLlR2GHM3lBAVZ+0zyACBYQ3hTJ5kcRh3H6fj/EMy8+N6pO43ru7hJhorMQbda19eY2sHwivjD5Vys0O0vR1Og==}
-
- '@elevenlabs/[email protected]':
- resolution: {integrity: sha512-4FdEv9YYR1EHl1KJwUFNqJcm9e1zuHi92Ky3wQgCM4UkzCiWNA24SyhLGchbXxv/Hc4w4arqBObQQ6SMj3jsKA==}
-
'@emnapi/[email protected]':
resolution: {integrity: sha512-l9Oo58x0HOP5znGzVhYW9U3e5wVuA4LAZU2AGezTmkhO1CgQRFDhDg4nneHsu/t3WniXg9QrG2nIXL/ZS8ln8Q==}
@@ -9137,12 +9125,6 @@ packages:
resolution: {integrity: sha512-pu0K1jCIdnQ12aWNaAVU5bzi7Bd1w54J3ECgANPmYLtQKP0HBj2cE/5coBD66MT10xbtIuUr7tg0Shbsvk0mDA==}
- '@livekit/[email protected]':
- resolution: {integrity: sha512-EsshAucklmpuUAfkABPxJNhzj9v2sG7JuzFDL4ML1oJQSV14sqrpTYnsaOudMAw9yOaW53NU3QQTlUQoRs4czw==}
-
- '@livekit/[email protected]':
- resolution: {integrity: sha512-Q+l57E7w/xxOBFVWzdX5rkAZO7ffyF+rlDzNUYq2SU114+5aTyCq+PK4unaEVDNd4952Af7wteKr3sOgasGuaA==}
-
resolution: {integrity: sha512-KPDeVScZgA1oq0CiPBcOa3kHIqU+pTOwRFDIhxvmf8CTNvqdZQYp5cCKW0bUk69VygB2PuTiINFWbY78aR2pQw==}
cpu: [arm64]
@@ -13465,9 +13447,6 @@ packages:
resolution: {integrity: sha512-eOIHzCUSH7SMfonMG1LsC2f8vxBFtho6NGBznK41R84YzPuvSBzrhEps33IsQiOW9+VL6NQ9DbjQJznk/S4uRA==}
- '@types/[email protected]':
- resolution: {integrity: sha512-mUMZLK3NvwRLcAAT9qmcK+9p7tpU2FHdDsntR3YI4+GY88XrgG4XiE7u1Q2LAN2/FZOz/tdMDC3GQCR4T8nFuw==}
-
resolution: {integrity: sha512-odiGr/9/qMqjcBOe5UhcNLOFHSYmKFOyr+bJ/Xu3Qp4k1pNPAlNLUVNNLcLfjQI7+W7ObX58EdD3H+3p3voOvA==}
@@ -18810,11 +18789,6 @@ packages:
resolution: {integrity: sha512-iyAZCeyD+c1gPyE9qpFu8af0Y+MRtmKOncdGoA2S5EY8iFq99dmmvkNnHiWo+pj0s7yH7l3KPIgee77tKpXPWQ==}
engines: {node: '>=18.0.0'}
- resolution: {integrity: sha512-aolY1XDAtx0nHKBNm29W9OhzBnSz1CP5kq3phvRhFfi1NbvMXs8tcACjAkZTnIKgihkp+BiJScZZ3tZv0Gz8sA==}
- peerDependencies:
- '@types/dom-mediacapture-record': ^1
-
resolution: {integrity: sha512-9bMdFfc80S+vSldBmG3HOuLVHnxRdNTlpzR6QDnzqCQtCzGUEAGTzBKYMeIM+I/sU4oZfgbcbS7X7F65/z/oxQ==}
hasBin: true
@@ -18919,10 +18893,6 @@ packages:
resolution: {integrity: sha512-9ie8ItPR6tjY5uYJh8K/Zrv/RMZ5VOlOWvtZdEHYSTFKZfIBPQa9tOAEeAWhd+AnIneLJ22w5fjOYtoutpWq5w==}
engines: {node: '>=18'}
- resolution: {integrity: sha512-HgMmCqIJSAKqo68l0rS2AanEWfkxaZ5wNiEFb5ggm08lDs9Xl2KxBlX3PTcaD2chBM1gXAYf491/M2Rv8Jwayg==}
- engines: {node: '>= 0.6.0'}
-
resolution: {integrity: sha512-BFRuQUqc7x2NWxfJBCyUrN8iYUYznzL9JROmRz1gZ6KlOIgmoD+njPVbb+VNn2nGMKggMsK79iUNErillsrx7w==}
@@ -21829,13 +21799,6 @@ packages:
resolution: {integrity: sha512-49oNpRjWRvnU8NyGVmUaYG4jtTkNonFZI86MmGRDqBphEK2EXT9gdEUoQPZhuBM8yWHxCWbobltqYO5M4XrUvQ==}
- resolution: {integrity: sha512-KrOH82c/W+GYQ0LHqtr3caRpM3ITglq3ljGUIb8LTki7ByacJZ9z+piSGiwZDsRyhQbYBOBJgr2k6X4BZXi3Kw==}
- hasBin: true
-
- resolution: {integrity: sha512-xZocWwfyp4hkbN4hLWxMjmv2Q8aNa9MhmOZ7L9aCZPT+dZsgRr6wZRrSYE3HTdyk/2pZKPSgqI7ns7Een1xMSA==}
-
resolution: {integrity: sha512-AckIIV90rPDcBcglUwXPF3kg0P0qmPsPXAj6BBEENQE1p5yA1xfmDJzfi1Tappj37Pv2mVbKpL3Z1T+Nn7k1Qw==}
@@ -23048,9 +23011,6 @@ packages:
resolution: {integrity: sha512-3KS2b+kL7fsuk/eJZ7EQdnEmQoaho/r6KUef7hxvltNA5DR8NAUM+8wJMbJyZ4G9/7i3v5zPBIMN5aybAh2/Jg==}
engines: {node: '>= 0.4'}
- resolution: {integrity: sha512-g/KzbYKbH5C2vPkaXGu8DJlHrGKHLsM25Zg9WuC9pMGfuvT+X25tZQWo5fK1BjBm8+UrVE9LDCvaY0CQk+fXDA==}
-
resolution: {integrity: sha512-uzR+FzI8qrUEIu96oaeBJmd9E7CFEiQ3goA5qCVgc4s5llSubcfGHq9yUstZx/k4s9dXHVKsE35YWoFyvEqEHA==}
@@ -23661,10 +23621,6 @@ packages:
webpack-cli:
optional: true
- resolution: {integrity: sha512-U9vjByy/sK2OMXu5mmfuZFKTMIUQe34c0JXRO+oDrxJTsntdYT2iIFwYMOV7HhMTuktcZLGf2W1N/OcSf9ssWg==}
- engines: {node: '>=6.0.0', npm: '>=3.10.0'}
-
resolution: {integrity: sha512-AMcUeyXAhbACL8S2hqqdqOLqvJ8ylmIbNwUIqQujRSouf4+eUFaXbG6F1Rbu+srlJMmxQWsiU7mOJi0nMBfM1g==}
@@ -26343,8 +26299,6 @@ snapshots:
'@braintree/[email protected]': {}
- '@bufbuild/[email protected]': {}
-
'@bufbuild/[email protected]': {}
'@bufbuild/[email protected]': {}
@@ -26779,15 +26733,6 @@ snapshots:
- bufferutil
- utf-8-validate
- '@elevenlabs/[email protected](@types/[email protected])':
- dependencies:
- '@elevenlabs/types': 0.13.0
- livekit-client: 2.19.0(@types/[email protected])
- transitivePeerDependencies:
- - '@types/dom-mediacapture-record'
-
- '@elevenlabs/[email protected]': {}
-
'@emnapi/[email protected]':
dependencies:
'@emnapi/wasi-threads': 1.2.2
@@ -28399,12 +28344,6 @@ snapshots:
dependencies:
'@lezer/common': 1.2.3
- '@livekit/[email protected]': {}
-
- '@livekit/[email protected]':
- dependencies:
- '@bufbuild/protobuf': 1.10.1
-
optional: true
@@ -33347,8 +33286,6 @@ snapshots:
'@types/[email protected]': {}
- '@types/[email protected]': {}
-
dependencies:
'@types/trusted-types': 2.0.4
@@ -40514,19 +40451,6 @@ snapshots:
rfdc: 1.4.1
wrap-ansi: 9.0.2
- dependencies:
- '@livekit/mutex': 1.1.1
- '@livekit/protocol': 1.45.8
- '@types/dom-mediacapture-record': 1.0.22
- events: 3.3.0
- jose: 6.2.3
- loglevel: 1.9.2
- sdp-transform: 2.15.0
- tslib: 2.8.1
- typed-emitter: 2.1.0
- webrtc-adapter: 9.0.5
-
dependencies:
msgpackr: 1.11.2
@@ -40623,8 +40547,6 @@ snapshots:
strip-ansi: 7.2.0
wrap-ansi: 9.0.2
-
@@ -44371,10 +44293,6 @@ snapshots:
dependencies:
compute-scroll-into-view: 3.1.1
-
-
@@ -45815,10 +45733,6 @@ snapshots:
possible-typed-array-names: 1.0.0
reflect.getprototypeof: 1.0.10
- optionalDependencies:
- rxjs: 7.8.1
-
@@ -46619,10 +46533,6 @@ snapshots:
- postcss
- uglify-js
- dependencies:
- sdp: 3.2.2
-
diff --git a/posthog/rate_limit.py b/posthog/rate_limit.py
index 04ad23442a53..3572c535aaa6 100644
--- a/posthog/rate_limit.py
+++ b/posthog/rate_limit.py
@@ -836,45 +836,6 @@ class AIResearchBurstRateThrottle(_AIThrottleBase):
action_name = "ai research burst rate limited"
-# Hands-free throttles are split per-action so a burst on /token/ (e.g. a flapping
-# mobile network triggering the reconnect path) can't lock the user out of /synthesize/
-# and vice versa. Each action's burst/sustained bucket caps a different cost surface:
-# token mints bound Scribe (STT) spend; synthesize calls bound ElevenLabs TTS spend.
-class MaxHandsFreeTokenBurstRateThrottle(_AIThrottleBase):
- # 10/min covers normal reconnects (network blip, tab refocus) and stays well below
- # any plausible legitimate flow. Dev gets a far larger ceiling for the dev loop.
- scope = "max_hands_free_token_burst"
- rate = "1000/minute" if settings.DEBUG else "10/minute"
- action_name = "max hands-free token burst rate limited"
-
-
-class MaxHandsFreeTokenSustainedRateThrottle(_AIThrottleBase):
- # Each token grants ~15 min of Scribe usage. A 2-hour gym session is ~8 tokens at
- # minimum plus reconnects, call it 15. 60/day covers several long sessions a day
- # with headroom and bounds worst-case Scribe spend per compromised account.
- scope = "max_hands_free_token_sustained"
- rate = "10000/day" if settings.DEBUG else "60/day"
- action_name = "max hands-free token sustained rate limited"
-
-
-class MaxHandsFreeSynthesizeBurstRateThrottle(_AIThrottleBase):
- # TTS calls fire once per assistant turn so the burst rate has to keep pace with
- # rapid back-and-forth. 20/min still well above any natural conversation cadence
- # but tight enough to slow a runaway client looping on synthesize.
- scope = "max_hands_free_synthesize_burst"
- rate = "1000/minute" if settings.DEBUG else "20/minute"
- action_name = "max hands-free synthesize burst rate limited"
-
-
-class MaxHandsFreeSynthesizeSustainedRateThrottle(_AIThrottleBase):
- # Bounds TTS spend per account per day. 150 calls * 2000 chars max ~= 300k chars
- # ceiling/day — comfortable for power users, hard cap on a compromised account's
- # daily ElevenLabs bill until the broader char-budget kill switch lands.
- scope = "max_hands_free_synthesize_sustained"
- rate = "10000/day" if settings.DEBUG else "150/day"
- action_name = "max hands-free synthesize sustained rate limited"
-
-
class AIResearchSustainedRateThrottle(_AIThrottleBase):
scope = "ai_research_sustained"
rate = "10/day"
diff --git a/posthog/schema.py b/posthog/schema.py
index b1130c85171f..4c5f5440c94e 100644
--- a/posthog/schema.py
+++ b/posthog/schema.py
@@ -33880,7 +33880,6 @@ class MaxUIContext(BaseModel):
form_answers: dict[str, str] | None = None
insights: list[MaxInsightContext] | None = None
notebooks: list[MaxNotebookContext] | None = None
- voice_mode: bool | None = None
class QueryRequest(BaseModel):
diff --git a/posthog/settings/integrations.py b/posthog/settings/integrations.py
index f68bf99450f1..4c0a438fa999 100644
--- a/posthog/settings/integrations.py
+++ b/posthog/settings/integrations.py
@@ -186,20 +186,6 @@ WORKOS_RADAR_ENABLED = get_from_env("WORKOS_RADAR_ENABLED", False, type_cast=str
CLOUDFLARE_TURNSTILE_SECRET_KEY = get_from_env("CLOUDFLARE_TURNSTILE_SECRET_KEY", "")
CLOUDFLARE_TURNSTILE_SITE_KEY = get_from_env("CLOUDFLARE_TURNSTILE_SITE_KEY", "")
-# ElevenLabs (Max hands-free mode)
-# STT goes browser ↔ ElevenLabs over a single-use Scribe WebSocket token (backend just mints).
-# TTS goes browser → PostHog → ElevenLabs → audio stream (backend proxies the key to ElevenLabs).
-ELEVENLABS_API_KEY = get_from_env("ELEVENLABS_API_KEY", "")
-ELEVENLABS_API_BASE_URL = get_from_env("ELEVENLABS_API_BASE_URL", "https://api.elevenlabs.io")
-# Rachel is ElevenLabs' default voice — neutral, clear at gym pace. Override if you want a
-# different feel without redeploying.
-ELEVENLABS_VOICE_ID = get_from_env("ELEVENLABS_VOICE_ID", "21m00Tcm4TlvDq8ikWAM")
-# Turbo v2.5 has ~300ms TTFB latency, sits in the free tier, and sounds clean enough for
-# gym-pace narration. Flash v2.5 is marginally faster but requires Creator ($22/mo) or above
-# on the ElevenLabs side, so devs running on free quota hit a 402 from the TTS proxy.
-# Override at the env level if you're on a paid tier and want the extra polish.
-ELEVENLABS_TTS_MODEL_ID = get_from_env("ELEVENLABS_TTS_MODEL_ID", "eleven_turbo_v2_5")
-
# PandaDoc (for legal documents: BAA/DPA). One template per document variant.
# Each call needs the matching template id, so we keep them as separate env vars —
# rotating one (e.g., when Legal updates the DPA copy) doesn't touch the others.
diff --git a/products/posthog_ai/frontend/generated/api.schemas.ts b/products/posthog_ai/frontend/generated/api.schemas.ts
index 066e9d65463a..b2260c24cb5e 100644
--- a/products/posthog_ai/frontend/generated/api.schemas.ts
+++ b/products/posthog_ai/frontend/generated/api.schemas.ts
@@ -618,11 +618,6 @@ export interface PatchedMaxCoreMemoryApi {
scraping_status?: CoreMemoryScrapingStatusEnumApi | BlankEnumApi | null
}
-export interface HandsFreeTokenApi {
- /** Single-use ElevenLabs Scribe realtime token, valid for 15 minutes. */
- token: string
-}
-
export interface JsonValueApi {}
/**
diff --git a/products/posthog_ai/frontend/generated/api.ts b/products/posthog_ai/frontend/generated/api.ts
index 8613e2f70541..245f332c3892 100644
--- a/products/posthog_ai/frontend/generated/api.ts
+++ b/products/posthog_ai/frontend/generated/api.ts
@@ -14,7 +14,6 @@ import type {
CoreMemoryListParams,
DocsSearchRequestApi,
DocsSearchResponseApi,
- HandsFreeTokenApi,
MCPToolRequestApi,
MCPToolResponseApi,
MaxCoreMemoryApi,
@@ -380,27 +379,6 @@ export const coreMemoryPartialUpdate = async (
})
}
-export const getMaxHandsFreeTokenCreateUrl = (projectId: string) => {
- return `/api/projects/${projectId}/max_hands_free/token/`
-}
-
-/**
- * Mint a single-use ElevenLabs Scribe realtime token.
- *
- * The browser uses the token to open a WebSocket directly to ElevenLabs — audio never
- * transits PostHog infrastructure. Tokens are time-bound (15 min) and single-use; the
- * per-team rate limit on this endpoint caps how often a user can mint new ones.
- *
- * Never logs the upstream response body — provider error responses can echo PII back and
- * we don't want any of that landing in structured logs.
- */
-export const maxHandsFreeTokenCreate = async (projectId: string, options?: RequestInit): Promise<HandsFreeTokenApi> => {
- return apiMutator<HandsFreeTokenApi>(getMaxHandsFreeTokenCreateUrl(projectId), {
- ...options,
- method: 'POST',
- })
-}
-
export const getMcpToolsCreateUrl = (projectId: string, toolName: string) => {
return `/api/projects/${projectId}/mcp_tools/${toolName}/`
}
diff --git a/services/mcp/src/api/generated.ts b/services/mcp/src/api/generated.ts
index ad53ec6cf637..67977571ab6e 100644
--- a/services/mcp/src/api/generated.ts
+++ b/services/mcp/src/api/generated.ts
@@ -55035,11 +55035,6 @@ export namespace Schemas {
on_overflow?: OnOverflowEnum;
}
- export interface HandsFreeToken {
- /** Single-use ElevenLabs Scribe realtime token, valid for 15 minutes. */
- token: string;
- }
-
/**
* 409 body when a guarded write was based on a stale head.
*/