Merge pull request #18 from techartdev/papurha/safe-fixes-and-reworks

Fix model selection, cleanup dead schema, and rework Assist agent routing

Author: techartdev

README.md

@@ -28,6 +28,7 @@ OpenClaw is a Home Assistant custom integration that connects your HA instance t
   - `openclaw.send_message`
   - `openclaw.clear_history`
   - `openclaw.invoke_tool`
+- **Integration options** for model selection and voice-specific routing
 - **Event**
   - `openclaw_message_received`
   - `openclaw_tool_invoked`
@@ -192,6 +193,13 @@ When enabled, OpenClaw tool-call responses can execute Home Assistant services.
 - **Wake word enabled**
 - **Wake word** (default: `hey openclaw`)
 - **Voice input provider** (`browser` or `assist_stt`)
+- **Voice agent ID** (optional)
+
+### Model selection
+
+- **Active Model** select entity controls the model override used for chat-card and Assist requests.
+- If no model is selected, the gateway default is used.
+- Assist conversation IDs are conservatively namespaced by agent ID so different agents do not accidentally share the same conversation context.
 
 ### Voice provider usage
 
@@ -227,11 +235,13 @@ If voice is unreliable in Brave, use Chrome/Edge for card voice input or continu
 
 Send a message to OpenClaw.
 
+> Note: file attachments are not currently supported by this service. The old `attachments` field was removed because it was accepted by the schema but never sent to the gateway.
+
 Fields:
 
 - `message` (required)
 - `session_id` (optional)
-- `attachments` (optional)
+- `agent_id` (optional)
 
 Example:
 
@@ -313,6 +323,8 @@ action:
 
 Fired when `openclaw.invoke_tool` completes.
 
+The integration also exposes native Event entities for both message-received and tool-invoked events so they can be selected directly in the automation UI.
+
 Event data includes:
 
 - `tool`

custom_components/openclaw/__init__.py

@@ -31,7 +31,6 @@
 from .api import OpenClawApiClient, OpenClawApiError
 from .const import (
     ATTR_AGENT_ID,
-    ATTR_ATTACHMENTS,
     ATTR_MESSAGE,
     ATTR_MODEL,
     ATTR_OK,
@@ -90,6 +89,7 @@
 )
 from .coordinator import OpenClawCoordinator
 from .exposure import apply_context_policy, build_exposed_entities_context
+from .helpers import extract_text_recursive
 
 _LOGGER = logging.getLogger(__name__)
 
@@ -98,6 +98,7 @@
 _VOICE_REQUEST_HEADERS = {
     "x-openclaw-source": "voice",
     "x-ha-voice": "true",
+    "x-openclaw-message-channel": "voice",
 }
 
 # Path to the chat card JS inside the integration package (custom_components/openclaw/www/)
@@ -106,7 +107,7 @@
 # URL at which the card JS is served (registered via register_static_path)
 _CARD_STATIC_URL = f"/openclaw/{_CARD_FILENAME}"
 # Versioned URL used for Lovelace resource registration to avoid stale browser cache
-_CARD_URL = f"{_CARD_STATIC_URL}?v=0.1.60"
+_CARD_URL = f"{_CARD_STATIC_URL}?v=0.1.61"
 
 OpenClawConfigEntry = ConfigEntry
 
@@ -117,7 +118,6 @@
         vol.Required(ATTR_MESSAGE): cv.string,
         vol.Optional(ATTR_SOURCE): cv.string,
         vol.Optional(ATTR_SESSION_ID): cv.string,
-        vol.Optional(ATTR_ATTACHMENTS): vol.All(cv.ensure_list, [cv.string]),
         vol.Optional(ATTR_AGENT_ID): cv.string,
     }
 )
@@ -447,13 +447,16 @@ async def handle_send_message(call: ServiceCall) -> None:
             )
             system_prompt = apply_context_policy(raw_context, max_chars, strategy)
 
+            active_model = _normalize_optional_text(options.get("active_model"))
+
             _append_chat_history(hass, session_id, "user", message)
 
             response = await client.async_send_message(
                 message=message,
                 session_id=session_id,
                 system_prompt=system_prompt,
                 agent_id=resolved_agent_id,
+                model=active_model,
                 extra_headers=extra_headers,
             )
 
@@ -469,6 +472,7 @@ async def handle_send_message(call: ServiceCall) -> None:
                         session_id=session_id,
                         system_prompt=system_prompt,
                         agent_id=resolved_agent_id,
+                        model=active_model,
                         extra_headers=extra_headers,
                     )
 
@@ -635,53 +639,6 @@ def _get_entry_options(hass: HomeAssistant, entry_data: dict[str, Any]) -> dict[
     return latest_entry.options if latest_entry else {}
 
 
-def _extract_text_recursive(value: Any, depth: int = 0) -> str | None:
-    """Recursively extract assistant text from nested response payloads."""
-    if depth > 8:
-        return None
-
-    if isinstance(value, str):
-        text = value.strip()
-        return text or None
-
-    if isinstance(value, list):
-        parts: list[str] = []
-        for item in value:
-            extracted = _extract_text_recursive(item, depth + 1)
-            if extracted:
-                parts.append(extracted)
-        if parts:
-            return "\n".join(parts)
-        return None
-
-    if isinstance(value, dict):
-        priority_keys = (
-            "output_text",
-            "text",
-            "content",
-            "message",
-            "response",
-            "answer",
-            "choices",
-            "output",
-            "delta",
-        )
-
-        for key in priority_keys:
-            if key not in value:
-                continue
-            extracted = _extract_text_recursive(value.get(key), depth + 1)
-            if extracted:
-                return extracted
-
-        for nested_value in value.values():
-            extracted = _extract_text_recursive(nested_value, depth + 1)
-            if extracted:
-                return extracted
-
-    return None
-
-
 def _summarize_tool_result(value: Any, max_len: int = 240) -> str | None:
     """Return compact string preview of tool result payload."""
     if value is None:
@@ -703,7 +660,7 @@ def _summarize_tool_result(value: Any, max_len: int = 240) -> str | None:
 
 def _extract_assistant_message(response: dict[str, Any]) -> str | None:
     """Extract assistant text from modern/legacy OpenAI-compatible responses."""
-    return _extract_text_recursive(response)
+    return extract_text_recursive(response)
 
 
 def _extract_tool_calls(response: dict[str, Any]) -> list[dict[str, Any]]:

custom_components/openclaw/api.py

@@ -104,6 +104,11 @@ def _headers(
     async def _get_session(self) -> aiohttp.ClientSession:
         """Get or create an aiohttp session."""
         if self._session is None or self._session.closed:
+            if self._session is not None and self._session.closed:
+                _LOGGER.warning(
+                    "Primary aiohttp session unavailable — creating fallback session. "
+                    "This may bypass HA connection management"
+                )
             self._session = aiohttp.ClientSession()
         return self._session
 

custom_components/openclaw/conversation.py

@@ -24,11 +24,13 @@
     ATTR_SESSION_ID,
     ATTR_TIMESTAMP,
     CONF_ASSIST_SESSION_ID,
+    CONF_AGENT_ID,
     CONF_CONTEXT_MAX_CHARS,
     CONF_CONTEXT_STRATEGY,
     CONF_INCLUDE_EXPOSED_CONTEXT,
     CONF_VOICE_AGENT_ID,
     DEFAULT_ASSIST_SESSION_ID,
+    DEFAULT_AGENT_ID,
     DEFAULT_CONTEXT_MAX_CHARS,
     DEFAULT_CONTEXT_STRATEGY,
     DEFAULT_INCLUDE_EXPOSED_CONTEXT,
@@ -38,6 +40,7 @@
 )
 from .coordinator import OpenClawCoordinator
 from .exposure import apply_context_policy, build_exposed_entities_context
+from .helpers import extract_text_recursive
 
 _LOGGER = logging.getLogger(__name__)
 
@@ -117,12 +120,20 @@ async def async_process(
         coordinator: OpenClawCoordinator = entry_data["coordinator"]
 
         message = user_input.text
-        conversation_id = self._resolve_conversation_id(user_input)
         assistant_id = "conversation"
         options = self.entry.options
         voice_agent_id = self._normalize_optional_text(
             options.get(CONF_VOICE_AGENT_ID)
         )
+        configured_agent_id = self._normalize_optional_text(
+            options.get(
+                CONF_AGENT_ID,
+                self.entry.data.get(CONF_AGENT_ID, DEFAULT_AGENT_ID),
+            )
+        )
+        resolved_agent_id = voice_agent_id or configured_agent_id
+        conversation_id = self._resolve_conversation_id(user_input, resolved_agent_id)
+        active_model = self._normalize_optional_text(options.get("active_model"))
         include_context = options.get(
             CONF_INCLUDE_EXPOSED_CONTEXT,
             DEFAULT_INCLUDE_EXPOSED_CONTEXT,
@@ -149,8 +160,9 @@ async def async_process(
                 client,
                 message,
                 conversation_id,
-                voice_agent_id,
+                resolved_agent_id,
                 system_prompt,
+                active_model,
             )
         except OpenClawApiError as err:
             _LOGGER.error("OpenClaw conversation error: %s", err)
@@ -165,8 +177,9 @@ async def async_process(
                             client,
                             message,
                             conversation_id,
-                            voice_agent_id,
+                            resolved_agent_id,
                             system_prompt,
+                            active_model,
                         )
                     except OpenClawApiError as retry_err:
                         return self._error_result(
@@ -205,8 +218,12 @@ async def async_process(
             continue_conversation=self._should_continue(full_response),
         )
 
-    def _resolve_conversation_id(self, user_input: conversation.ConversationInput) -> str:
-        """Return conversation id from HA or a stable Assist fallback session key."""
+    def _resolve_conversation_id(
+        self,
+        user_input: conversation.ConversationInput,
+        agent_id: str | None,
+    ) -> str:
+        """Return conversation id from HA with conservative agent namespacing."""
         configured_session_id = self._normalize_optional_text(
             self.entry.options.get(
                 CONF_ASSIST_SESSION_ID,
@@ -216,19 +233,25 @@ def _resolve_conversation_id(self, user_input: conversation.ConversationInput) -
         if configured_session_id:
             return configured_session_id
 
+        agent_suffix = self._normalize_optional_text(agent_id)
+
         if user_input.conversation_id:
+            if agent_suffix:
+                return f"{user_input.conversation_id}:{agent_suffix}"
             return user_input.conversation_id
 
         context = getattr(user_input, "context", None)
         user_id = getattr(context, "user_id", None)
         if user_id:
-            return f"assist_user_{user_id}"
+            base_id = f"assist_user_{user_id}"
+            return f"{base_id}:{agent_suffix}" if agent_suffix else base_id
 
         device_id = getattr(user_input, "device_id", None)
         if device_id:
-            return f"assist_device_{device_id}"
+            base_id = f"assist_device_{device_id}"
+            return f"{base_id}:{agent_suffix}" if agent_suffix else base_id
 
-        return "assist_default"
+        return f"assist_default:{agent_suffix}" if agent_suffix else "assist_default"
 
     def _normalize_optional_text(self, value: Any) -> str | None:
         """Return a stripped string or None for blank values."""
@@ -244,13 +267,14 @@ async def _get_response(
         conversation_id: str,
         agent_id: str | None = None,
         system_prompt: str | None = None,
+        model: str | None = None,
     ) -> str:
         """Get a response from OpenClaw, trying streaming first."""
-        # Try streaming (lower TTFB for voice pipeline)
         full_response = ""
         async for chunk in client.async_stream_message(
             message=message,
             session_id=conversation_id,
+            model=model,
             system_prompt=system_prompt,
             agent_id=agent_id,
             extra_headers=_VOICE_REQUEST_HEADERS,
@@ -260,62 +284,15 @@ async def _get_response(
         if full_response:
             return full_response
 
-        # Fallback to non-streaming
         response = await client.async_send_message(
             message=message,
             session_id=conversation_id,
+            model=model,
             system_prompt=system_prompt,
             agent_id=agent_id,
             extra_headers=_VOICE_REQUEST_HEADERS,
         )
-        extracted = self._extract_text_recursive(response)
-        return extracted or ""
-
-    def _extract_text_recursive(self, value: Any, depth: int = 0) -> str | None:
-        """Recursively extract assistant text from nested response payloads."""
-        if depth > 8:
-            return None
-
-        if isinstance(value, str):
-            text = value.strip()
-            return text or None
-
-        if isinstance(value, list):
-            parts: list[str] = []
-            for item in value:
-                extracted = self._extract_text_recursive(item, depth + 1)
-                if extracted:
-                    parts.append(extracted)
-            if parts:
-                return "\n".join(parts)
-            return None
-
-        if isinstance(value, dict):
-            priority_keys = (
-                "output_text",
-                "text",
-                "content",
-                "message",
-                "response",
-                "answer",
-                "choices",
-                "output",
-                "delta",
-            )
-
-            for key in priority_keys:
-                if key not in value:
-                    continue
-                extracted = self._extract_text_recursive(value.get(key), depth + 1)
-                if extracted:
-                    return extracted
-
-            for nested_value in value.values():
-                extracted = self._extract_text_recursive(nested_value, depth + 1)
-                if extracted:
-                    return extracted
-
-        return None
+        return extract_text_recursive(response) or ""
 
     @staticmethod
     def _should_continue(response: str) -> bool: