Update README.md

updated tool call handling
updated light color handling
updated searxng handling
- all 3 should work now

README.md

@@ -49,7 +49,7 @@
 ## Quick Start
 - Create a new groqd entry.
 - Set your Groq API key.
-- Pick a model (default: `meta-llama/llama-4-maverick-17b-128e-instruct`).
+- Pick a model (default: `meta-llama/llama-4-scout-17b-16e-instruct`).
 - Set a personality prompt if you want tone/style changes.
 
 ## Options
@@ -62,13 +62,6 @@ Open **groqd → Options** to configure:
 - Auto-fetch URLs in user messages
 - Time format (12h / 24h)
 
-## Recommended Prompt Additions
-To make tone warmer and lyrical:
-
-"""
-Be warm, playful, and a bit lyrical. If the user quotes lyrics, continue one line then respond empathetically.
-"""
-
 ## Notes
 - TTS/STT are handled by Home Assistant and other integrations.
 - Web search is optional and only used if enabled.

custom_components/groqd/__init__.py

@@ -2,7 +2,8 @@
 
 from __future__ import annotations
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
+from typing import Any
 
 import groq
 
@@ -12,7 +13,7 @@ from homeassistant.core import HomeAssistant
 
 from .const import DOMAIN
 
-PLATFORMS: list[Platform] = [Platform.CONVERSATION]
+PLATFORMS: list[Platform] = [Platform.CONVERSATION, Platform.BUTTON]
 
 
 @dataclass
@@ -20,6 +21,7 @@ class GroqdRuntimeData:
     """Runtime data for groqd."""
 
     client: groq.AsyncClient
+    conversation_agent: Any = field(default=None)
 
 
 type GroqdConfigEntry = ConfigEntry[GroqdRuntimeData]

custom_components/groqd/button.py

@@ -0,0 +1,43 @@
+"""Button platform for groqd — provides a 'Clear session memory' button."""
+
+from __future__ import annotations
+
+from homeassistant.components.button import ButtonEntity
+from homeassistant.core import HomeAssistant
+from homeassistant.helpers.entity_platform import AddEntitiesCallback
+from homeassistant.helpers import device_registry as dr
+
+from . import GroqdConfigEntry
+from .const import DOMAIN, LOGGER
+
+
+async def async_setup_entry(
+    hass: HomeAssistant,
+    config_entry: GroqdConfigEntry,
+    async_add_entities: AddEntitiesCallback,
+) -> None:
+    """Set up groqd button entities."""
+    async_add_entities([GroqdClearMemoryButton(config_entry)])
+
+
+class GroqdClearMemoryButton(ButtonEntity):
+    """Button that wipes all persisted conversation history for this entry."""
+
+    _attr_has_entity_name = True
+    _attr_translation_key = "clear_memory"
+
+    def __init__(self, entry: GroqdConfigEntry) -> None:
+        self._entry = entry
+        self._attr_unique_id = f"{entry.entry_id}_clear_memory"
+        self._attr_device_info = dr.DeviceInfo(
+            identifiers={(DOMAIN, entry.entry_id)},
+        )
+
+    async def async_press(self) -> None:
+        """Handle button press — delegate to the conversation agent."""
+        agent = self._entry.runtime_data.conversation_agent
+        if agent is not None:
+            await agent.async_clear_memory()
+            LOGGER.info("groqd: clear memory button pressed for entry %s", self._entry.entry_id)
+        else:
+            LOGGER.warning("groqd: clear memory pressed but conversation agent not ready")

custom_components/groqd/const.py

@@ -30,7 +30,7 @@ CONF_SEARXNG_LANGUAGE = "searxng_language"
 CONF_AUTO_FETCH_URLS = "auto_fetch_urls"
 CONF_TIME_FORMAT = "time_format"
 
-DEFAULT_CHAT_MODEL = "meta-llama/llama-4-maverick-17b-128e-instruct"
+DEFAULT_CHAT_MODEL = "meta-llama/llama-4-scout-17b-16e-instruct"
 DEFAULT_CONTEXT_MESSAGES = 20
 DEFAULT_MAX_TOKENS = 4096
 DEFAULT_TEMPERATURE = 1.0

custom_components/groqd/conversation.py

@@ -95,17 +95,153 @@ async def async_setup_entry(
     async_add_entities([agent])
 
 
+def _make_nullable(schema: Any) -> Any:
+    """Wrap a schema so it also accepts JSON null (Python None).
+
+    Models often emit ``null`` for optional parameters they don't need rather
+    than omitting the key entirely.  Making optional properties nullable in the
+    schema sent to Groq prevents server-side validation failures for those
+    ``null`` values.  ``_coerce_args_for_schema`` then strips the nulls before
+    forwarding to Home Assistant, which does not accept null slot values.
+    """
+    if not isinstance(schema, dict):
+        return schema
+    if schema.get("type") == "null":
+        return schema  # already null
+    if "anyOf" in schema:
+        options = schema["anyOf"]
+        if any(isinstance(o, dict) and o.get("type") == "null" for o in options):
+            return schema  # null already present in union
+        return {**schema, "anyOf": [*options, {"type": "null"}]}
+    if "type" in schema:
+        return {"anyOf": [schema, {"type": "null"}]}
+    return schema
+
+
+def _is_slot_wrapper(schema: dict) -> bool:
+    """Return True if schema looks like an HA intent slot wrapper.
+
+    HA represents intent slot values as ``{"value": X}`` (sometimes also
+    with a ``"text"`` sibling key).  Detecting this pattern lets us relax
+    the schema so Groq models can emit the bare value directly (e.g. "red"
+    instead of {"value": "red"}).
+    """
+    if schema.get("type") != "object":
+        return False
+    props = schema.get("properties", {})
+    return "value" in props and len(props) <= 2
+
+
+def _relax_schema_arrays(schema: Any) -> Any:
+    """Recursively relax schemas so Groq's server-side validation accepts
+    common model output patterns that differ from the strict HA schema.
+
+    Two relaxations are applied:
+
+    * ``type: array`` → ``anyOf: [array, string]``
+      Models often emit ``"light"`` instead of ``["light"]``.
+
+    * HA slot-wrapper objects ``{type: object, properties: {value: X}}``
+      → ``anyOf: [object, X]``
+      Models often emit ``"red"`` instead of ``{"value": "red"}``.
+
+    After Groq accepts the output, ``_coerce_args_for_schema`` fixes the
+    args back to what HA expects before forwarding the tool call.
+    """
+    if not isinstance(schema, dict):
+        return schema
+    if schema.get("type") == "object":
+        props = schema.get("properties", {})
+        required = set(schema.get("required", []))
+        is_wrapper = _is_slot_wrapper(schema)
+        # Capture the raw value schema BEFORE recursing (used in anyOf below)
+        raw_value_schema = props.get("value") if is_wrapper else None
+        result = dict(schema)
+        new_props: dict[str, Any] = {}
+        for k, v in props.items():
+            relaxed = _relax_schema_arrays(v)
+            # Optional properties must also accept null so the model can emit
+            # null instead of omitting the key and still pass Groq validation.
+            if k not in required:
+                relaxed = _make_nullable(relaxed)
+            new_props[k] = relaxed
+        result["properties"] = new_props
+        if is_wrapper and raw_value_schema is not None:
+            # Allow either the full {value: X} object OR a bare value of the
+            # same base type — but WITHOUT any enum constraints.  Models often
+            # produce values outside the declared enum (e.g. XY colour strings
+            # instead of named colours); stripping the enum here lets Groq
+            # accept the output so we can pass it to HA and let HA report the
+            # proper error if the value is truly invalid.
+            raw_type = raw_value_schema.get("type", "string")
+            loose_alt: dict[str, Any] = {"type": raw_type} if raw_type else {"type": "string"}
+            return {"anyOf": [result, loose_alt]}
+        return result
+    if schema.get("type") == "array":
+        return {"anyOf": [schema, {"type": "string"}]}
+    return schema
+
+
+def _coerce_args_for_schema(args: Any, schema: Any) -> Any:
+    """Recursively coerce args so they match what HA expects.
+
+    Handles the two common model output mismatches:
+
+    * ``"light"`` (string) where schema expects ``["light"]`` (array).
+    * ``"red"`` (string) where schema expects ``{"value": "red"}``
+      (HA intent slot wrapper object).
+    """
+    if not isinstance(schema, dict):
+        return args
+    schema_type = schema.get("type")
+    # Array: wrap bare string in a list
+    if schema_type == "array":
+        if isinstance(args, str):
+            return [args]
+        return args
+    if schema_type == "object":
+        props = schema.get("properties", {})
+        # Models sometimes double-encode an object as a JSON string, e.g.
+        # "color": "{\"xy\":[0.64,0.33]}" instead of a real dict.
+        # Try to JSON-decode such strings before further coercion.
+        if isinstance(args, str):
+            try:
+                decoded = json.loads(args)
+                if isinstance(decoded, dict):
+                    args = decoded
+            except (json.JSONDecodeError, ValueError):
+                pass
+        # Slot wrapper: model sent bare value instead of {"value": X}
+        if "value" in props and not isinstance(args, dict):
+            return {"value": args}
+        if isinstance(args, dict):
+            result = dict(args)
+            for key, prop_schema in props.items():
+                if key in result and isinstance(prop_schema, dict):
+                    result[key] = _coerce_args_for_schema(result[key], prop_schema)
+            # Strip None values — null means "omit" for optional HA parameters.
+            # HA's voluptuous validators don't accept None for optional slots;
+            # they simply expect the key to be absent.
+            return {k: v for k, v in result.items() if v is not None}
+    return args
+
+
 def _format_tool(
     tool: llm.Tool, custom_serializer: Callable[[Any], Any] | None
-) -> ChatCompletionToolParam:
-    """Format tool specification."""
+) -> tuple[ChatCompletionToolParam, dict]:
+    """Format tool specification.
+
+    Returns the Groq-compatible tool param (with relaxed array schemas) and
+    the original schema dict (for use in argument coercion before calling HA).
+    """
+    original_schema = convert(tool.parameters, custom_serializer=custom_serializer)
     tool_spec = FunctionDefinition(
         name=tool.name,
-        parameters=convert(tool.parameters, custom_serializer=custom_serializer),
+        parameters=_relax_schema_arrays(original_schema),
     )
     if tool.description:
         tool_spec["description"] = tool.description
-    return ChatCompletionToolParam(type="function", function=tool_spec)
+    return ChatCompletionToolParam(type="function", function=tool_spec), original_schema
 
 
 def _parse_tool_choice(value: str) -> Any:
@@ -198,12 +334,29 @@ def _extract_urls(text: str, limit: int = 3) -> list[str]:
             break
     return cleaned
 
+def _normalize_searxng_url(url: str) -> str:
+    """Return the bare SearXNG base URL, stripping any /search path or OpenSearch suffix.
+
+    Users sometimes copy the OpenSearch template URL (e.g.
+    ``https://searxng.example.com/search=%s``) instead of just the base
+    URL.  We strip the ``/search`` path component and anything after it so
+    the code can always safely append ``/search`` with its own query params.
+    A subpath mount is preserved (e.g. ``https://example.com/searxng``).
+    """
+    url = url.strip()
+    # Strip OpenSearch-style /search=%s (and variants like /search?q=%s)
+    url = re.sub(r"/search[=?].*$", "", url, flags=re.IGNORECASE)
+    # Strip a trailing bare /search as well
+    url = re.sub(r"/search/?$", "", url, flags=re.IGNORECASE)
+    return url.rstrip("/")
+
+
 async def _run_searxng(
     hass: HomeAssistant,
     options: dict[str, Any],
     tool_args: dict[str, Any],
 ) -> dict[str, Any]:
-    base_url = options.get(CONF_SEARXNG_URL, DEFAULT_SEARXNG_URL).rstrip("/")
+    base_url = _normalize_searxng_url(options.get(CONF_SEARXNG_URL, DEFAULT_SEARXNG_URL))
     params = {
         "q": tool_args.get("query", ""),
         "format": "json",
@@ -269,14 +422,25 @@ class GroqdConversationEntity(
         data = await self._store.async_load() or {}
         self._persisted_history = data.get("history", {})
         self._memory_index = data.get("memory_index", {})
+        self.entry.runtime_data.conversation_agent = self
         self.entry.async_on_unload(
             self.entry.add_update_listener(self._async_entry_update_listener)
         )
 
     async def async_will_remove_from_hass(self) -> None:
+        self.entry.runtime_data.conversation_agent = None
         conversation.async_unset_agent(self.hass, self.entry)
         await super().async_will_remove_from_hass()
 
+    async def async_clear_memory(self) -> None:
+        """Clear all conversation history and persisted memory."""
+        self.history.clear()
+        self._persisted_history.clear()
+        self._memory_index.clear()
+        if self._store:
+            await self._store.async_save({"history": {}, "memory_index": {}})
+        LOGGER.info("groqd: memory cleared for entry %s", self.entry.entry_id)
+
     async def async_process(
         self, user_input: conversation.ConversationInput
     ) -> conversation.ConversationResult:
@@ -284,6 +448,7 @@ class GroqdConversationEntity(
         intent_response = intent.IntentResponse(language=user_input.language)
         llm_api: llm.APIInstance | None = None
         tools: list[ChatCompletionToolParam] | None = None
+        tool_schemas: dict[str, dict] = {}
         user_name: str | None = None
         llm_context = llm.LLMContext(
             platform=DOMAIN,
@@ -309,7 +474,11 @@ class GroqdConversationEntity(
                 return conversation.ConversationResult(
                     response=intent_response, conversation_id=user_input.conversation_id
                 )
-            tools = [_format_tool(tool, llm_api.custom_serializer) for tool in llm_api.tools]
+            _formatted = [_format_tool(tool, llm_api.custom_serializer) for tool in llm_api.tools]
+            tools = [f[0] for f in _formatted]
+            tool_schemas = {
+                tool.name: f[1] for tool, f in zip(llm_api.tools, _formatted)
+            }
         if options.get(CONF_SEARXNG_ENABLED, DEFAULT_SEARXNG_ENABLED):
             if tools is None:
                 tools = []
@@ -406,7 +575,7 @@ class GroqdConversationEntity(
             formatted_time = now.strftime("%-I:%M %p")
         prompt_parts.append(
             f"Current local time: {formatted_time}. "
-            "Always express time exactly in this numeric format; do not spell out numbers."
+            "When reading times aloud, use natural spoken format (e.g., 'three thirty-four' not 'three thousand thirty-four')."
         )
 
         if options.get(CONF_AUTO_FETCH_URLS, DEFAULT_AUTO_FETCH_URLS):
@@ -547,6 +716,12 @@ class GroqdConversationEntity(
                     except Exception as err:
                         tool_response = {"error": type(err).__name__, "error_text": str(err)}
                 elif llm_api:
+                    # Coerce any string values back to arrays where the HA tool
+                    # schema requires them (Groq models sometimes emit a bare
+                    # string instead of a one-element array, e.g. "light" vs
+                    # ["light"] for the domain parameter).
+                    if tool_name in tool_schemas:
+                        tool_args = _coerce_args_for_schema(tool_args, tool_schemas[tool_name])
                     tool_input = llm.ToolInput(
                         tool_name=tool_name,
                         tool_args=tool_args,

custom_components/groqd/manifest.json

@@ -1,7 +1,7 @@
 {
   "domain": "groqd",
   "name": "groqd",
-  "version": "0.1.0",
+  "version": "0.1.1",
   "config_flow": true,
   "documentation": "https://github.com/sudoxreboot/groqd",
   "issue_tracker": "https://github.com/sudoxreboot/groqd/issues",

custom_components/groqd/translations/en.json

@@ -1,4 +1,11 @@
 {
+  "entity": {
+    "button": {
+      "clear_memory": {
+        "name": "Clear session memory"
+      }
+    }
+  },
   "config": {
     "step": {
       "user": {