fix: describe local intent answers to the model instead of replaying them

With prefer_local_intents, Home Assistant puts requests its local intents
answered into the chat log as external calls of intents like
HassGetCurrentTime. The model does not have these tools, and replayed as
its own calls they make it apologise for using tools it does not have and
doubt answers it never gave. They are now passed as a developer note with
the intent, its arguments and its result.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

AGENTS.md

@@ -0,0 +1,49 @@
+# Repository instructions
+
+## Project
+
+This is a public Home Assistant custom integration (`openai_compatible`) for OpenAI-compatible API endpoints. It adds a configurable base URL to a fork of Home Assistant core's `openai_conversation` integration. Provider entries can have conversation, AI Task, speech-to-text, and text-to-speech subentries.
+
+The README covers installation, configuration, endpoint requirements, development, and releases. `docs/provider-genericity.md` audits remaining OpenAI-specific behavior and records an open design decision; it is not a scheduled plan. The vendored core revision is recorded in `custom_components/openai_compatible/UPSTREAM_COMMIT`. `hacs.json` sets the minimum Home Assistant version (2026.9.0); `manifest.json` carries the integration version. Development and CI use Python 3.14.
+
+Production Home Assistant installs this integration through HACS from this repo's GitHub releases; publishing a release and updating in HACS is how a change reaches it (owner's OK). The `hassio` repo's `custom_components/openai_compatible` is only the nightly backup's snapshot of that install; do not edit or copy code into it.
+
+## Boundaries
+
+- `origin` is a public GitHub repository. Any push to `main` publishes the code, not only a release tag. Do not push without the owner's explicit OK.
+- Changes to a live Home Assistant installation, a configured API provider or host, or a published GitHub release/HACS download need the owner's explicit OK. Read-only inspection is fine. The documented HACS and manual installation paths can change live Home Assistant behavior.
+- Keep the fork's configurable base URL and separate `openai_compatible` domain intact when syncing upstream code. Preserve config-entry migration, provider/subentry behavior, and the four entity types; tests cover these paths.
+- Do not put secrets, tokens, passwords, or private keys in any file. Do not search for credentials. Base URLs must reject embedded credentials because URLs can appear in errors and logs; API keys belong in the config flow's key field.
+- Never list or read the UNAS or `/nfs/media`.
+
+## Working rules
+
+- Install test dependencies with `pip install -r requirements_test.txt` under Python 3.14+, then run `pytest tests/ -v`. `pyproject.toml` configures pytest and Ruff, but no Ruff command is listed in CI.
+- Keep `url_util.py` separate from vendored upstream files so upstream syncing remains manageable. When changing `strings.json`, regenerate `translations/en.json` with `python scripts/generate_translations.py`; the translation test checks that they match.
+- The README documents release steps. `python scripts/release.py VERSION` requires a clean, current `main`; it changes the manifest, commits, and tags. Pushing the tag publishes through `.github/workflows/release.yaml`. Do not run the release script or push without the owner's explicit OK.
+
+## Documentation that must be kept current
+
+- Installation, provider setup, endpoint behavior, or compatibility changes: update `README.md`.
+- Changes to remaining OpenAI-specific behavior or the capability design decision: update `docs/provider-genericity.md`.
+- Changes to repo commands, boundaries, or layout: update this file; keep `CLAUDE.md` as the import stub.
+- Minimum Home Assistant version changes: update `hacs.json` and the README.
+- Upstream syncs: keep `UPSTREAM_COMMIT` accurate. User-facing string changes: update `strings.json` and regenerate `translations/en.json`.
+- Release process or version behavior changes: update the README's Releases section and the manifest or workflow as applicable.
+
+## Definition of done
+
+- Run `pytest tests/ -v` for code or translation changes. CI also runs Hassfest and the HACS validation action on pushes and pull requests; check their results when available.
+- Keep behavior and docs aligned, including the translation output and upstream revision when affected. For releases, the tag version and `manifest.json` version must match; the release workflow enforces this.
+
+## Layout
+
+- `custom_components/openai_compatible/`: integration entry point (`__init__.py`), config flow, four entity platforms, URL/model helpers, manifest, service definitions, and UI strings/translations.
+- `tests/`: ported upstream regression tests and fork-specific tests; `tests/snapshots/` holds conversation snapshots.
+- `scripts/generate_translations.py`: builds the shipped English translation. `scripts/release.py`: prepares a version commit and annotated tag.
+- `.github/workflows/`: pytest, Hassfest/HACS validation, and tag-triggered GitHub release workflows.
+- `docs/provider-genericity.md`: provider compatibility audit and open design decision.
+
+## Owner to confirm
+
+- Which API provider endpoints or hosts does the production Home Assistant installation use?

CLAUDE.md

@@ -0,0 +1,5 @@
+# Claude Code
+
+Repository instructions for all agents live in [AGENTS.md](AGENTS.md). Edit them there.
+
+@AGENTS.md

README.md

@@ -4,6 +4,28 @@
 [![validate](https://github.com/adamjs83/ha-openai-compatible/actions/workflows/validate.yaml/badge.svg)](https://github.com/adamjs83/ha-openai-compatible/actions/workflows/validate.yaml)
 [![HACS: custom](https://img.shields.io/badge/HACS-custom-41BDF5.svg)](https://hacs.xyz)
 
+> **Fork of [adamjs83/ha-openai-compatible](https://github.com/adamjs83/ha-openai-compatible) v0.2.1**
+> with fixes for running it against a proxy that serves Claude models. Drop it once
+> upstream has fixed these issues:
+>
+> - [#3](https://github.com/adamjs83/ha-openai-compatible/issues/3): a tool call with an
+>   empty argument string (tools without parameters) is treated as `{}` instead of
+>   raising `JSONDecodeError`.
+> - [#4](https://github.com/adamjs83/ha-openai-compatible/issues/4): earlier web searches
+>   are not replayed in the history; they carry no results, and the model then disowns
+>   its own answers and searches again.
+> - [#5](https://github.com/adamjs83/ha-openai-compatible/issues/5): the home location
+>   lookup for web search uses the subentry's model instead of `gpt-4o-mini` and asks
+>   as a user message (Anthropic needs one); a failed lookup no longer blocks saving
+>   the options, and empty location fields are left out of the web search tool.
+> - Requests that Home Assistant's local intents answered (`prefer_local_intents`) are
+>   described to the model in a note instead of being replayed as its own calls of
+>   intents like `HassGetCurrentTime`, which are not among its tools; replayed, they
+>   make it apologise for using tools it does not have. Not reported upstream yet.
+>
+> HACS only installs from GitHub, so this fork is installed by copying
+> `custom_components/openai_compatible` into the Home Assistant config.
+
 Home Assistant's built-in **OpenAI** integration, with one change that matters: the
 **API base URL is configurable**. Point Home Assistant's conversation agent, AI Task,
 speech-to-text and text-to-speech at any OpenAI-compatible endpoint — LiteLLM, vLLM,
@@ -144,6 +166,16 @@ know about anyway:
   string. Options are offered based on those name patterns, and the endpoint's own error
   is what you'll see if you ask for something it can't do.
 
+## Known OpenAI-specific behaviour
+
+The base URL is configurable, but much of the rest is still upstream's OpenAI-shaped
+logic: some request parameters are sent to every provider, and which options the config
+flow offers is inferred from OpenAI model-name prefixes. That means options can appear
+for endpoints that cannot honour them, and be withheld from endpoints that could.
+
+[`docs/provider-genericity.md`](docs/provider-genericity.md) has the full audit, what
+each item costs, and the design decision that has to be made first.
+
 ## Relationship to Home Assistant core
 
 `custom_components/openai_compatible/` is derived from
@@ -151,6 +183,10 @@ know about anyway:
 (commit recorded in [`UPSTREAM_COMMIT`](custom_components/openai_compatible/UPSTREAM_COMMIT)),
 with the domain renamed to `openai_compatible` and the base URL made configurable.
 
+The manifest requires `openai>=2.45.0` rather than core's exact pin: Home Assistant
+installs custom-integration requirements against its own package constraints, so an
+exact pin stops the integration loading whenever core bumps `openai`.
+
 Base-URL handling lives in its own module,
 [`url_util.py`](custom_components/openai_compatible/url_util.py), deliberately separate
 from the vendored files so re-syncing with core stays a clean merge.

custom_components/openai_compatible/__init__.py

@@ -294,7 +294,7 @@ async def async_setup_entry(hass: HomeAssistant, entry: OpenAIConfigEntry) -> bo
     _ = await hass.async_add_executor_job(client.platform_headers)
 
     try:
-        await hass.async_add_executor_job(client.with_options(timeout=10.0).models.list)
+        await client.models.list(timeout=10.0)
     except openai.AuthenticationError as err:
         raise ConfigEntryAuthFailed(err) from err
     except openai.OpenAIError as err:

custom_components/openai_compatible/config_flow.py

@@ -104,7 +104,7 @@ from .const import (
     UNSUPPORTED_PRIORITY_SERVICE_TIERS_MODELS,
     UNSUPPORTED_WEB_SEARCH_MODELS,
 )
-from .model_list import async_fetch_model_ids, model_field
+from .model_list import async_fetch_model_ids, default_model, model_field
 from .url_util import (
     CredentialsInBaseURL,
     InvalidBaseURL,
@@ -325,14 +325,35 @@ class OpenAISubentryFlowHandler(ConfigSubentryFlow):
                 }
             )
 
+        # The model belongs on the first screen. Picking one is the entire
+        # point of adding an endpoint, and behind the "recommended" checkbox
+        # it was unreachable: ticked, this step creates the subentry and the
+        # model silently stays RECOMMENDED_CHAT_MODEL -- an OpenAI name a
+        # third-party provider will not serve. entity.py reads
+        # CONF_CHAT_MODEL regardless of the recommended flag, so a model
+        # chosen here takes effect on either path.
+        model_ids = await async_fetch_model_ids(self._get_entry().runtime_data)
+        step_schema[
+            vol.Optional(CONF_CHAT_MODEL, default=default_model(model_ids))
+        ] = model_field(model_ids)
+
         step_schema[
             vol.Required(CONF_RECOMMENDED, default=options.get(CONF_RECOMMENDED, False))
         ] = bool
 
+        errors: dict[str, str] = {}
+
         if user_input is not None:
             if user_input.get(CONF_LLM_HASS_API) is None:
                 user_input.pop(CONF_LLM_HASS_API, None)
 
+            # Checked here rather than on the next step, because the
+            # recommended path creates the subentry from this one and would
+            # otherwise skip the check entirely.
+            if user_input.get(CONF_CHAT_MODEL) in UNSUPPORTED_MODELS:
+                errors[CONF_CHAT_MODEL] = "model_not_supported"
+
+            if not errors:
                 if user_input[CONF_RECOMMENDED]:
                     if self._is_new:
                         return self.async_create_entry(
@@ -355,6 +376,7 @@ class OpenAISubentryFlowHandler(ConfigSubentryFlow):
             data_schema=self.add_suggested_values_to_schema(
                 vol.Schema(step_schema), options
             ),
+            errors=errors,
         )
 
     async def async_step_additional(
@@ -365,14 +387,8 @@ class OpenAISubentryFlowHandler(ConfigSubentryFlow):
         errors: dict[str, str] = {}
 
         step_schema: VolDictType = {
-            # Upstream makes this a free-text box defaulting to an OpenAI
-            # model name. Offer what the provider actually serves instead.
-            vol.Optional(
-                CONF_CHAT_MODEL,
-                default=RECOMMENDED_CHAT_MODEL,
-            ): model_field(
-                await async_fetch_model_ids(self._get_entry().runtime_data)
-            ),
+            # CONF_CHAT_MODEL is asked on the init step, so that it is
+            # reachable without unticking "recommended" first.
             vol.Optional(
                 CONF_MAX_TOKENS,
                 default=RECOMMENDED_MAX_TOKENS,
@@ -393,8 +409,6 @@ class OpenAISubentryFlowHandler(ConfigSubentryFlow):
 
         if user_input is not None:
             options.update(user_input)
-            if user_input.get(CONF_CHAT_MODEL) in UNSUPPORTED_MODELS:
-                errors[CONF_CHAT_MODEL] = "model_not_supported"
 
             if not errors:
                 return await self.async_step_model()
@@ -688,11 +702,17 @@ class OpenAISubentryFlowHandler(ConfigSubentryFlow):
                     ): str,
                 }
             )
+            # Ask the subentry's own model: the endpoint may not serve
+            # RECOMMENDED_CHAT_MODEL (an OpenAI name). A failed lookup only
+            # costs the city/region hint, so it must not block saving.
+            try:
                 response = await client.responses.create(
-                model=RECOMMENDED_CHAT_MODEL,
+                    model=self.options.get(CONF_CHAT_MODEL, RECOMMENDED_CHAT_MODEL),
+                    # A user message, not a system one: endpoints in front of
+                    # Anthropic models reject a request without user content.
                     input=[
                         {
-                        "role": "system",
+                            "role": "user",
                             "content": "Where are the following coordinates located: "
                             f"({zone_home.attributes[EntityStateAttribute.LATITUDE]},"
                             f" {zone_home.attributes[EntityStateAttribute.LONGITUDE]})?",
@@ -713,6 +733,10 @@ class OpenAISubentryFlowHandler(ConfigSubentryFlow):
                     ),
                 )
                 location_data = location_schema(json.loads(response.output_text) or {})
+            except (openai.OpenAIError, ValueError, vol.Invalid) as err:
+                _LOGGER.warning(
+                    "Could not look up the home location for web search: %s", err
+                )
 
         if self.hass.config.country:
             location_data[CONF_WEB_SEARCH_COUNTRY] = self.hass.config.country

custom_components/openai_compatible/entity.py

@@ -22,7 +22,6 @@ from openai.types.responses import (
     ResponseFunctionToolCall,
     ResponseFunctionToolCallParam,
     ResponseFunctionWebSearch,
-    ResponseFunctionWebSearchParam,
     ResponseIncompleteEvent,
     ResponseInputFileParam,
     ResponseInputImageParam,
@@ -181,20 +180,35 @@ def _convert_content_to_param(
     """Convert any native chat message for this agent to the native format."""
     messages: ResponseInputParam = []
     reasoning_summary: list[str] = []
-    web_search_calls: dict[str, ResponseFunctionWebSearchParam] = {}
+    # Requests that Home Assistant's local intents answered without the model
+    # (prefer_local_intents) are in the chat log as calls of intents like
+    # HassGetCurrentTime, which are not among the model's tools. Replayed as
+    # the model's own calls, they make it apologise for using tools it does
+    # not have, so they are described to it as a note instead.
+    local_calls: dict[str, llm.ToolInput] = {}
 
     for content in chat_content:
         if isinstance(content, conversation.ToolResultContent):
-            if (
-                content.tool_name == "web_search_call"
-                and content.tool_call_id in web_search_calls
-            ):
-                web_search_call = web_search_calls.pop(content.tool_call_id)
-                web_search_call["status"] = content.tool_result.get(  # type: ignore[typeddict-item]
-                    "status", "completed"
+            if (tool_call := local_calls.pop(content.tool_call_id, None)) is not None:
+                messages.append(
+                    EasyInputMessageParam(
+                        type="message",
+                        role="developer",
+                        content=(
+                            "Home Assistant's local intent recognition answered "
+                            "the last user request itself, without you: it ran "
+                            f"{tool_call.tool_name} with "
+                            f"{json_dumps(tool_call.tool_args)} and got "
+                            f"{json_dumps(content.tool_result)}. The assistant "
+                            "reply after this note is its wording, not yours. "
+                            "That intent is not one of your tools."
+                        ),
                     )
-                messages.append(web_search_call)
-            else:
+                )
+            # Earlier web searches are not replayed: the chat log only keeps
+            # their status, not their results, and a search without results
+            # makes the model disown the answer it based on that search.
+            elif content.tool_name != "web_search_call":
                 messages.append(
                     FunctionCallOutput(
                         type="function_call_output",
@@ -217,18 +231,11 @@ def _convert_content_to_param(
         if isinstance(content, conversation.AssistantContent):
             if content.tool_calls:
                 for tool_call in content.tool_calls:
-                    if (
-                        tool_call.external
-                        and tool_call.tool_name == "web_search_call"
-                        and "action" in tool_call.tool_args
-                    ):
-                        web_search_calls[tool_call.id] = ResponseFunctionWebSearchParam(
-                            type="web_search_call",
-                            id=tool_call.id,
-                            action=tool_call.tool_args["action"],
-                            status="completed",
-                        )
-                    else:
+                    if tool_call.external and tool_call.tool_name == "web_search_call":
+                        continue
+                    if tool_call.external and tool_call.tool_name != "code_interpreter":
+                        local_calls[tool_call.id] = tool_call
+                        continue
                     messages.append(
                         ResponseFunctionToolCallParam(
                             type="function_call",
@@ -418,7 +425,7 @@ async def _transform_stream(  # noqa: C901 - This is complex, but better to have
                     llm.ToolInput(
                         id=current_tool_call.call_id,
                         tool_name=current_tool_call.name,
-                        tool_args=json.loads(current_tool_call.arguments),
+                        tool_args=json.loads(current_tool_call.arguments or "{}"),
                     )
                 ]
             }
@@ -578,12 +585,20 @@ class OpenAIBaseLLMEntity(Entity):
                 ),
             )
             if options.get(CONF_WEB_SEARCH_USER_LOCATION):
+                # Leave out unknown fields: some endpoints reject empty strings,
+                # and the location lookup may have found no city or region.
+                location = {
+                    key: value
+                    for key, option in (
+                        ("city", CONF_WEB_SEARCH_CITY),
+                        ("region", CONF_WEB_SEARCH_REGION),
+                        ("country", CONF_WEB_SEARCH_COUNTRY),
+                        ("timezone", CONF_WEB_SEARCH_TIMEZONE),
+                    )
+                    if (value := options.get(option))
+                }
                 web_search["user_location"] = UserLocation(
-                    type="approximate",
-                    city=options.get(CONF_WEB_SEARCH_CITY, ""),
-                    region=options.get(CONF_WEB_SEARCH_REGION, ""),
-                    country=options.get(CONF_WEB_SEARCH_COUNTRY, ""),
-                    timezone=options.get(CONF_WEB_SEARCH_TIMEZONE, ""),
+                    type="approximate", **location
                 )
             if not options.get(
                 CONF_WEB_SEARCH_INLINE_CITATIONS,

custom_components/openai_compatible/manifest.json

@@ -1,14 +1,23 @@
 {
   "domain": "openai_compatible",
   "name": "OpenAI Compatible",
-  "after_dependencies": ["assist_pipeline", "intent"],
-  "codeowners": ["@adamjs83"],
+  "after_dependencies": [
+    "assist_pipeline",
+    "intent"
+  ],
+  "codeowners": [
+    "@adamjs83"
+  ],
   "config_flow": true,
-  "dependencies": ["conversation"],
+  "dependencies": [
+    "conversation"
+  ],
   "documentation": "https://github.com/adamjs83/ha-openai-compatible",
   "integration_type": "service",
   "iot_class": "cloud_polling",
   "issue_tracker": "https://github.com/adamjs83/ha-openai-compatible/issues",
-  "requirements": ["openai==2.45.0"],
-  "version": "1.0.0"
+  "requirements": [
+    "openai>=2.45.0"
+  ],
+  "version": "0.2.1"
 }

custom_components/openai_compatible/model_list.py

@@ -22,7 +22,7 @@ from homeassistant.helpers.selector import (
     SelectSelectorMode,
 )
 
-from .const import LOGGER
+from .const import LOGGER, RECOMMENDED_CHAT_MODEL
 
 DEFAULT_TIMEOUT = 10.0
 
@@ -48,6 +48,14 @@ async def async_fetch_model_ids(
     except openai.OpenAIError as err:
         LOGGER.debug("Could not list models: %s", err)
         return []
+    except Exception:  # noqa: BLE001 - see below
+        # Deliberately broad. This runs inside a config-flow form and has a
+        # working fallback (a free-text model field), so nothing it can raise
+        # justifies taking the dialog down: a proxy may return a payload the
+        # SDK cannot parse, and the failure modes of arbitrary
+        # OpenAI-compatible servers are not enumerable in advance.
+        LOGGER.debug("Unexpected error listing models", exc_info=True)
+        return []
 
     return sorted({model.id for model in page.data if getattr(model, "id", None)})
 
@@ -71,3 +79,17 @@ def model_field(model_ids: list[str]) -> Any:
             sort=True,
         )
     )
+
+
+def default_model(model_ids: list[str]) -> str:
+    """Return the model to preselect for a new endpoint.
+
+    Upstream's default is RECOMMENDED_CHAT_MODEL, an OpenAI name. Offering it
+    to someone configuring a third-party provider preselects a model that will
+    fail on first use, so prefer it only when the provider actually serves it
+    (or when discovery told us nothing) and otherwise start from what is
+    advertised.
+    """
+    if not model_ids or RECOMMENDED_CHAT_MODEL in model_ids:
+        return RECOMMENDED_CHAT_MODEL
+    return model_ids[0]

custom_components/openai_compatible/strings.json

@@ -123,12 +123,12 @@
           "data": {
             "chat_model": "[%key:common::generic::model%]",
             "max_tokens": "Maximum tokens to return in response",
-            "store_responses": "Store requests and responses in OpenAI",
+            "store_responses": "Store requests and responses on the provider",
             "temperature": "Temperature",
             "top_p": "Top P"
           },
           "data_description": {
-            "store_responses": "If enabled, requests and responses are stored by OpenAI and visible in your OpenAI dashboard logs"
+            "store_responses": "Sets the Responses API store flag, asking the provider to keep requests and responses in its own logs. Support varies: OpenAI shows them in its dashboard, while many compatible servers ignore the flag entirely."
           },
           "title": "Additional settings"
         },

custom_components/openai_compatible/translations/en.json

@@ -59,12 +59,12 @@
           "data": {
             "chat_model": "Model",
             "max_tokens": "Maximum tokens to return in response",
-            "store_responses": "Store requests and responses in OpenAI",
+            "store_responses": "Store requests and responses on the provider",
             "temperature": "Temperature",
             "top_p": "Top P"
           },
           "data_description": {
-            "store_responses": "If enabled, requests and responses are stored by OpenAI and visible in your OpenAI dashboard logs"
+            "store_responses": "Sets the Responses API store flag, asking the provider to keep requests and responses in its own logs. Support varies: OpenAI shows them in its dashboard, while many compatible servers ignore the flag entirely."
           },
           "title": "Additional settings"
         },
@@ -123,12 +123,12 @@
           "data": {
             "chat_model": "Model",
             "max_tokens": "Maximum tokens to return in response",
-            "store_responses": "Store requests and responses in OpenAI",
+            "store_responses": "Store requests and responses on the provider",
             "temperature": "Temperature",
             "top_p": "Top P"
           },
           "data_description": {
-            "store_responses": "If enabled, requests and responses are stored by OpenAI and visible in your OpenAI dashboard logs"
+            "store_responses": "Sets the Responses API store flag, asking the provider to keep requests and responses in its own logs. Support varies: OpenAI shows them in its dashboard, while many compatible servers ignore the flag entirely."
           },
           "title": "Additional settings"
         },

docs/provider-genericity.md

@@ -0,0 +1,113 @@
+# Provider genericity: audit and roadmap
+
+This fork currently makes one thing configurable — the API base URL — and inherits
+everything else from `homeassistant/components/openai_conversation`. That is enough to
+reach an OpenAI-compatible endpoint, and not enough to behave correctly against one.
+
+This document records what is still OpenAI-specific, what each item actually costs, and
+the design decision that has to be made before most of it can be fixed. It is an audit,
+not a plan of record: nothing here is scheduled.
+
+Line references are against the tree at the time of writing; treat them as starting
+points, not gospel.
+
+## Why a proxy makes this hard to notice
+
+A gateway such as LiteLLM can be configured to drop parameters the target provider does
+not support (`drop_params`). With that on, an endpoint absorbs OpenAI-only parameters
+silently and requests succeed — so the payload problem below never surfaces, while
+options that do nothing look like options that work.
+
+That cuts both ways, and it is the single most important thing to understand about
+testing this integration:
+
+- **A dropping proxy hides finding 1.** No error is ever raised, so passing tests
+  against it say nothing about portability.
+- **A dropping proxy makes finding 2 worse.** A parameter that is silently discarded
+  turns a misconfigured option into a no-op rather than an error, with nothing in the UI
+  or the log to indicate it had no effect.
+
+A strict server (vLLM, llama.cpp, a minimal home-grown endpoint) is the honest test
+target, because it will reject what it does not implement.
+
+## 1. OpenAI-only parameters on every request
+
+`entity.py`, in the `ResponseCreateParamsStreaming` construction at line 515:
+
+| Parameter | Sent when | Line |
+| --- | --- | --- |
+| `service_tier` | always | 520 |
+| `store` | always | 521 |
+| `prompt_cache_retention: "24h"` | always, unless the model name matches `UNSUPPORTED_EXTENDED_CACHE_RETENTION_MODELS` | 563 |
+| `reasoning`, `include: ["reasoning.encrypted_content"]` | model name starts with `o` or `gpt-5` | 525-544 |
+| `text.verbosity` | model name starts with `gpt-5` | 555-557 |
+
+`service_tier`, `store` and `prompt_cache_retention` are therefore sent to every
+provider, whatever it is. A server that validates its input strictly returns 400; a
+dropping proxy discards them and the request succeeds.
+
+**Cost:** breaks strict servers outright. Invisible behind a dropping proxy.
+
+## 2. Capability inferred from OpenAI model names
+
+There are 23 `startswith(...)` model-name gates across the component
+(`config_flow.py` 11, `entity.py` 9, `__init__.py` 2, `ai_task.py` 1), plus the
+`UNSUPPORTED_*` lists in `const.py`, which enumerate OpenAI model names.
+
+Against a third-party model string, every one of those checks is meaningless — and it
+fails in both directions:
+
+- **Options offered that the provider cannot honour.** The `UNSUPPORTED_*` lists contain
+  only OpenAI names, so a name like `claude-sonnet-4-6` never matches one and the option
+  is shown. Service tier ("Controls the cost and response time") is offered for every
+  such model. Code interpreter and web search are OpenAI *hosted tools*, offered
+  unconditionally, and almost no compatible server implements either.
+- **Options withheld that the provider does support.** Reasoning effort, reasoning
+  summary and verbosity are only offered when the name starts with `o` or `gpt-5`. Point
+  an endpoint at a reasoning-capable model whose name does not look like OpenAI's and
+  there is no way to set a reasoning effort at all — the field never renders. A dropping
+  proxy cannot compensate for this one, because the parameter is never sent.
+
+**Cost:** the config UI misrepresents the endpoint in both directions. This is the
+finding to fix first: it is what a user sees, and a dropping proxy makes it silent
+rather than loud.
+
+## 3. OpenAI branding in entities and errors
+
+- `manufacturer="OpenAI"` on every device (`entity.py:492`)
+- `DEFAULT_CONVERSATION_NAME`, `DEFAULT_AI_TASK_NAME`, `DEFAULT_STT_NAME`,
+  `DEFAULT_TTS_NAME` in `const.py` — "OpenAI Conversation", "OpenAI AI Task", …
+- user-facing error text: `OpenAI response incomplete` (459), `OpenAI response failed`
+  (473), `OpenAI response error` (475), `Error talking to OpenAI` (719-720)
+
+**Cost:** cosmetic, but the errors are the ones a user reads while debugging their own
+provider, which makes them actively misleading.
+
+## 4. Stale upstream notices
+
+- The `deprecated_generate_content` and `deprecated_generate_image` repair issues state
+  the actions will be removed "in the 2026.9.0 release". That is Home Assistant core's
+  schedule for its own integration, it has already passed, and this fork has made no
+  such commitment.
+- `organization_verification_required` is an OpenAI-platform-specific repair that
+  directs the user to OpenAI's settings page — meaningless for any other provider.
+
+**Cost:** informational, but wrong, and repair issues are prominent in the UI.
+
+## The open design decision
+
+The OpenAI API has no capability discovery: `GET /v1/models` returns ids, not feature
+support. So "which options should this endpoint offer" cannot be answered by asking the
+provider, and something has to stand in for it. The candidates:
+
+1. **A per-entry capability toggle** — "this endpoint supports OpenAI-specific
+   features", defaulted on when the host is `api.openai.com`. Explicit, one decision per
+   provider, and wrong only if the user sets it wrong.
+2. **Host detection alone** — infer from the base URL. No user burden; wrong for Azure
+   or OpenAI reached through a proxy, which is a normal setup here.
+3. **Per-endpoint feature checkboxes** — let each endpoint declare what to send. Most
+   precise, most tedious, and it asks the user to know their provider's API surface.
+
+The shape that was recommended, and not agreed: a portable payload by default, with
+OpenAI extras behind option 1. Deciding this is the prerequisite for findings 1 and 2;
+findings 3 and 4 are independent and can be done at any time.

requirements_test.txt

@@ -1,6 +1,7 @@
 # Requires Python 3.14+ (Home Assistant 2026.9 floor). Not installable on Python 3.13.
 pytest-homeassistant-custom-component>=0.13.317
-openai==2.45.0
+# Core 2026.10's pin; the manifest only sets a floor.
+openai==3.10.0
 
 # The manifest declares dependencies: ["conversation"], and the ported tests exercise
 # stt/tts too; a full HA install resolves each core component's own `requirements` at

tests/conftest.py

@@ -1,6 +1,6 @@
 """Fixtures for openai_compatible tests."""
 
-from collections.abc import Generator
+from collections.abc import AsyncGenerator, Generator
 from typing import Any
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -154,13 +154,19 @@ async def mock_config_entry_with_reasoning_model(
 @pytest.fixture
 async def mock_init_component(
     hass: HomeAssistant, mock_config_entry: MockConfigEntry
-) -> None:
-    """Initialize integration."""
+) -> AsyncGenerator[None]:
+    """Initialize integration.
+
+    The patch stays active for the whole test: setup awaits GET /models, and a
+    test that changes a subentry reloads the entry, which runs setup again.
+    """
     with patch(
         "openai.resources.models.AsyncModels.list",
+        new_callable=AsyncMock,
     ):
         assert await async_setup_component(hass, DOMAIN, {})
         await hass.async_block_till_done()
+        yield
 
 
 @pytest.fixture(autouse=True)

tests/test_base_url.py

@@ -2,6 +2,11 @@
 
 from unittest.mock import AsyncMock, patch
 
+import httpx
+import openai
+import pytest
+
+from homeassistant.config_entries import ConfigEntryState
 from homeassistant.core import HomeAssistant
 from pytest_homeassistant_custom_component.common import MockConfigEntry
 
@@ -26,7 +31,7 @@ async def test_client_built_with_base_url(hass: HomeAssistant) -> None:
         "custom_components.openai_compatible.openai.AsyncOpenAI"
     ) as mock_client:
         mock_client.return_value.platform_headers = lambda: {}
-        mock_client.return_value.with_options.return_value.models.list = AsyncMock()
+        mock_client.return_value.models.list = AsyncMock()
         await hass.config_entries.async_setup(entry.entry_id)
         await hass.async_block_till_done()
 
@@ -47,7 +52,7 @@ async def test_api_key_optional(hass: HomeAssistant) -> None:
         "custom_components.openai_compatible.openai.AsyncOpenAI"
     ) as mock_client:
         mock_client.return_value.platform_headers = lambda: {}
-        mock_client.return_value.with_options.return_value.models.list = AsyncMock()
+        mock_client.return_value.models.list = AsyncMock()
         await hass.config_entries.async_setup(entry.entry_id)
         await hass.async_block_till_done()
 
@@ -73,10 +78,56 @@ async def test_migration_backfills_missing_base_url(hass: HomeAssistant) -> None
         "custom_components.openai_compatible.openai.AsyncOpenAI"
     ) as mock_client:
         mock_client.return_value.platform_headers = lambda: {}
-        mock_client.return_value.with_options.return_value.models.list = AsyncMock()
+        mock_client.return_value.models.list = AsyncMock()
         assert await hass.config_entries.async_setup(entry.entry_id)
         await hass.async_block_till_done()
 
     assert entry.data[CONF_BASE_URL] == DEFAULT_BASE_URL
     assert entry.minor_version == 8
     assert mock_client.call_args.kwargs["base_url"] == DEFAULT_BASE_URL
+
+
+@pytest.mark.parametrize(
+    ("error", "state"),
+    [
+        (
+            openai.AuthenticationError(
+                "bad key",
+                response=httpx.Response(401, request=httpx.Request("GET", "http://x")),
+                body=None,
+            ),
+            ConfigEntryState.SETUP_ERROR,
+        ),
+        (
+            openai.APIConnectionError(request=httpx.Request("GET", "http://x")),
+            ConfigEntryState.SETUP_RETRY,
+        ),
+    ],
+)
+async def test_setup_checks_connection(
+    hass: HomeAssistant, error: Exception, state: ConfigEntryState
+) -> None:
+    """Setup must actually call GET /models and act on its failure.
+
+    AsyncModels.list returns an awaitable paginator; handing the bound method to
+    the executor returned it unawaited, so no request was made and a bad key or
+    unreachable endpoint still set up.
+    """
+    entry = MockConfigEntry(
+        domain=DOMAIN,
+        data={"api_key": "sk-test", CONF_BASE_URL: "http://192.0.2.10:4000/v1"},
+        version=2,
+        minor_version=8,
+    )
+    entry.add_to_hass(hass)
+
+    with patch(
+        "custom_components.openai_compatible.openai.AsyncOpenAI"
+    ) as mock_client:
+        mock_client.return_value.platform_headers = lambda: {}
+        mock_client.return_value.models.list = AsyncMock(side_effect=error)
+        await hass.config_entries.async_setup(entry.entry_id)
+        await hass.async_block_till_done()
+
+    mock_client.return_value.models.list.assert_awaited_once_with(timeout=10.0)
+    assert entry.state is state

tests/test_config_flow.py

@@ -177,6 +177,9 @@ async def test_creating_conversation_subentry(
 
     processed_options = RECOMMENDED_CONVERSATION_OPTIONS.copy()
     processed_options[CONF_PROMPT] = processed_options[CONF_PROMPT].strip()
+    # The init step now asks for the model, so the created subentry records
+    # one explicitly instead of falling back to the default at request time.
+    processed_options[CONF_CHAT_MODEL] = RECOMMENDED_CHAT_MODEL
 
     assert result2["data"] == processed_options
 
@@ -239,23 +242,17 @@ async def test_subentry_unsupported_model(
         subentry_flow["flow_id"],
         {
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: "o1-mini",
             CONF_PROMPT: "Speak like a pirate",
             CONF_LLM_HASS_API: ["assist"],
         },
     )
     await hass.async_block_till_done()
-    assert subentry_flow["type"] is FlowResultType.FORM
-    assert subentry_flow["step_id"] == "additional"
 
-    # Configure additional step
-    subentry_flow = await hass.config_entries.subentries.async_configure(
-        subentry_flow["flow_id"],
-        {
-            CONF_CHAT_MODEL: "o1-mini",
-        },
-    )
-    await hass.async_block_till_done()
+    # The model is asked for, and so validated, on the init step -- the flow
+    # stays there and reports the error against the field.
     assert subentry_flow["type"] is FlowResultType.FORM
+    assert subentry_flow["step_id"] == "init"
     assert subentry_flow["errors"] == {"chat_model": "model_not_supported"}
 
 
@@ -295,6 +292,7 @@ async def test_subentry_reasoning_effort_list(
         subentry_flow["flow_id"],
         {
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: model,
             CONF_PROMPT: "Speak like a pirate",
             CONF_LLM_HASS_API: ["assist"],
         },
@@ -305,9 +303,7 @@ async def test_subentry_reasoning_effort_list(
     # Configure additional step
     subentry_flow = await hass.config_entries.subentries.async_configure(
         subentry_flow["flow_id"],
-        {
-            CONF_CHAT_MODEL: model,
-        },
+        {},
     )
     assert subentry_flow["type"] is FlowResultType.FORM
     assert subentry_flow["step_id"] == "model"
@@ -349,6 +345,7 @@ async def test_subentry_reasoning_summary_visibility(
         subentry_flow["flow_id"],
         {
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: model,
             CONF_PROMPT: "Speak like a pirate",
             CONF_LLM_HASS_API: ["assist"],
         },
@@ -359,9 +356,7 @@ async def test_subentry_reasoning_summary_visibility(
     # Configure additional step
     subentry_flow = await hass.config_entries.subentries.async_configure(
         subentry_flow["flow_id"],
-        {
-            CONF_CHAT_MODEL: model,
-        },
+        {},
     )
     assert subentry_flow["type"] is FlowResultType.FORM
     assert subentry_flow["step_id"] == "model"
@@ -398,6 +393,7 @@ async def test_subentry_reasoning_summary_options(
         subentry_flow["flow_id"],
         {
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: model,
             CONF_PROMPT: "Speak like a pirate",
             CONF_LLM_HASS_API: ["assist"],
         },
@@ -407,9 +403,7 @@ async def test_subentry_reasoning_summary_options(
 
     subentry_flow = await hass.config_entries.subentries.async_configure(
         subentry_flow["flow_id"],
-        {
-            CONF_CHAT_MODEL: model,
-        },
+        {},
     )
     assert subentry_flow["type"] is FlowResultType.FORM
     assert subentry_flow["step_id"] == "model"
@@ -446,6 +440,7 @@ async def test_subentry_reasoning_summary_default_sanitized_on_model_switch(
         subentry_flow["flow_id"],
         {
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: "o3",
             CONF_PROMPT: "Speak like a pirate",
             CONF_LLM_HASS_API: ["assist"],
         },
@@ -454,7 +449,7 @@ async def test_subentry_reasoning_summary_default_sanitized_on_model_switch(
 
     subentry_flow = await hass.config_entries.subentries.async_configure(
         subentry_flow["flow_id"],
-        {CONF_CHAT_MODEL: "o3"},
+        {},
     )
     assert subentry_flow["step_id"] == "model"
 
@@ -512,6 +507,7 @@ async def test_subentry_service_tier_list(
         subentry_flow["flow_id"],
         {
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: model,
             CONF_PROMPT: "Speak like a pirate",
             CONF_LLM_HASS_API: ["assist"],
         },
@@ -522,9 +518,7 @@ async def test_subentry_service_tier_list(
     # Configure additional step
     subentry_flow = await hass.config_entries.subentries.async_configure(
         subentry_flow["flow_id"],
-        {
-            CONF_CHAT_MODEL: model,
-        },
+        {},
     )
     assert subentry_flow["type"] is FlowResultType.FORM
     assert subentry_flow["step_id"] == "model"
@@ -558,6 +552,7 @@ async def test_subentry_unsupported_reasoning_effort(
         subentry_flow["flow_id"],
         {
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: "gpt-5",
             CONF_PROMPT: "Speak like a pirate",
             CONF_LLM_HASS_API: ["assist"],
         },
@@ -568,9 +563,7 @@ async def test_subentry_unsupported_reasoning_effort(
     # Configure additional step
     subentry_flow = await hass.config_entries.subentries.async_configure(
         subentry_flow["flow_id"],
-        {
-            CONF_CHAT_MODEL: "gpt-5",
-        },
+        {},
     )
     assert subentry_flow["type"] is FlowResultType.FORM
     assert subentry_flow["step_id"] == "model"
@@ -706,11 +699,11 @@ async def test_form_credentials_in_base_url(hass: HomeAssistant) -> None:
             (
                 {
                     CONF_RECOMMENDED: False,
+                    CONF_CHAT_MODEL: "o1-pro",
                     CONF_PROMPT: "Speak like a pro",
                 },
                 {
                     CONF_TEMPERATURE: 1.0,
-                    CONF_CHAT_MODEL: "o1-pro",
                     CONF_TOP_P: RECOMMENDED_TOP_P,
                     CONF_MAX_TOKENS: 10000,
                 },
@@ -740,11 +733,11 @@ async def test_form_credentials_in_base_url(hass: HomeAssistant) -> None:
             (
                 {
                     CONF_RECOMMENDED: False,
+                    CONF_CHAT_MODEL: RECOMMENDED_CHAT_MODEL,
                     CONF_PROMPT: "Speak like a pirate",
                 },
                 {
                     CONF_TEMPERATURE: 0.3,
-                    CONF_CHAT_MODEL: RECOMMENDED_CHAT_MODEL,
                     CONF_TOP_P: RECOMMENDED_TOP_P,
                     CONF_MAX_TOKENS: RECOMMENDED_MAX_TOKENS,
                 },
@@ -797,11 +790,11 @@ async def test_form_credentials_in_base_url(hass: HomeAssistant) -> None:
             (
                 {
                     CONF_RECOMMENDED: False,
+                    CONF_CHAT_MODEL: "gpt-4o",
                     CONF_PROMPT: "Speak like super Mario",
                 },
                 {
                     CONF_TEMPERATURE: 0.8,
-                    CONF_CHAT_MODEL: "gpt-4o",
                     CONF_TOP_P: 0.9,
                     CONF_MAX_TOKENS: 1000,
                     CONF_STORE_RESPONSES: True,
@@ -852,11 +845,11 @@ async def test_form_credentials_in_base_url(hass: HomeAssistant) -> None:
             (
                 {
                     CONF_RECOMMENDED: False,
+                    CONF_CHAT_MODEL: "gpt-5.6",
                     CONF_PROMPT: "Speak like a pirate",
                 },
                 {
                     CONF_TEMPERATURE: 0.8,
-                    CONF_CHAT_MODEL: "gpt-5.6",
                     CONF_TOP_P: 0.9,
                     CONF_MAX_TOKENS: 1000,
                 },
@@ -975,11 +968,11 @@ async def test_form_credentials_in_base_url(hass: HomeAssistant) -> None:
             (
                 {
                     CONF_RECOMMENDED: False,
+                    CONF_CHAT_MODEL: "o3-mini",
                     CONF_PROMPT: "Speak like a pirate",
                 },
                 {
                     CONF_TEMPERATURE: 0.8,
-                    CONF_CHAT_MODEL: "o3-mini",
                     CONF_TOP_P: 0.9,
                     CONF_MAX_TOKENS: 1000,
                 },
@@ -1020,11 +1013,11 @@ async def test_form_credentials_in_base_url(hass: HomeAssistant) -> None:
             (
                 {
                     CONF_RECOMMENDED: False,
+                    CONF_CHAT_MODEL: "gpt-4o",
                     CONF_PROMPT: "Speak like a pirate",
                 },
                 {
                     CONF_TEMPERATURE: 0.8,
-                    CONF_CHAT_MODEL: "gpt-4o",
                     CONF_TOP_P: 0.9,
                     CONF_MAX_TOKENS: 1000,
                 },
@@ -1075,11 +1068,11 @@ async def test_form_credentials_in_base_url(hass: HomeAssistant) -> None:
             (
                 {
                     CONF_RECOMMENDED: False,
+                    CONF_CHAT_MODEL: "gpt-5-pro",
                     CONF_PROMPT: "Speak like a pirate",
                 },
                 {
                     CONF_TEMPERATURE: 0.8,
-                    CONF_CHAT_MODEL: "gpt-5-pro",
                     CONF_TOP_P: 0.9,
                     CONF_MAX_TOKENS: 1000,
                 },
@@ -1155,7 +1148,13 @@ async def test_subentry_switching(
 
     assert subentry_flow["type"] is FlowResultType.ABORT
     assert subentry_flow["reason"] == "reconfigure_successful"
-    assert subentry.data == expected_options
+    # The init step always asks for a model, so every subentry records one.
+    # Cases that do not name one get the default, which is what the flow
+    # submits for them; cases that do still win, being unpacked second.
+    assert subentry.data == {
+        CONF_CHAT_MODEL: RECOMMENDED_CHAT_MODEL,
+        **expected_options,
+    }
 
 
 @pytest.mark.parametrize("store_responses", [False, True])
@@ -1183,6 +1182,7 @@ async def test_subentry_web_search_user_location(
         {
             CONF_RECOMMENDED: False,
             CONF_PROMPT: "Speak like a pirate",
+            CONF_CHAT_MODEL: RECOMMENDED_CHAT_MODEL,
         },
     )
     assert subentry_flow["type"] is FlowResultType.FORM
@@ -1193,7 +1193,6 @@ async def test_subentry_web_search_user_location(
         subentry_flow["flow_id"],
         {
             CONF_TEMPERATURE: 1.0,
-            CONF_CHAT_MODEL: RECOMMENDED_CHAT_MODEL,
             CONF_TOP_P: RECOMMENDED_TOP_P,
             CONF_MAX_TOKENS: RECOMMENDED_MAX_TOKENS,
             CONF_STORE_RESPONSES: store_responses,
@@ -1275,6 +1274,75 @@ async def test_subentry_web_search_user_location(
     }
 
 
+async def test_subentry_web_search_user_location_lookup_fails(
+    hass: HomeAssistant,
+    mock_config_entry,
+    mock_init_component,
+) -> None:
+    """Test the location lookup uses the subentry's model and may fail.
+
+    An endpoint that only serves other providers' models rejects the OpenAI
+    default; saving must still succeed with country and time zone only.
+    """
+    subentry = next(
+        sub
+        for sub in mock_config_entry.subentries.values()
+        if sub.subentry_type == "conversation"
+    )
+    subentry_flow = await mock_config_entry.start_subentry_reconfigure_flow(
+        hass, subentry.subentry_id
+    )
+    subentry_flow = await hass.config_entries.subentries.async_configure(
+        subentry_flow["flow_id"],
+        {
+            CONF_RECOMMENDED: False,
+            CONF_PROMPT: "Speak like a pirate",
+            CONF_CHAT_MODEL: "claude-sonnet-5-5",
+        },
+    )
+    subentry_flow = await hass.config_entries.subentries.async_configure(
+        subentry_flow["flow_id"],
+        {
+            CONF_TEMPERATURE: 1.0,
+            CONF_TOP_P: RECOMMENDED_TOP_P,
+            CONF_MAX_TOKENS: RECOMMENDED_MAX_TOKENS,
+            CONF_STORE_RESPONSES: False,
+        },
+    )
+    await hass.async_block_till_done()
+    assert subentry_flow["step_id"] == "model"
+
+    hass.config.country = "DE"
+    hass.config.time_zone = "Europe/Berlin"
+    hass.states.async_set("zone.home", "0", {"latitude": 50.9, "longitude": 12.1})
+    with patch(
+        "openai.resources.responses.AsyncResponses.create",
+        new_callable=AsyncMock,
+        side_effect=BadRequestError(
+            response=httpx.Response(status_code=400, request=httpx.Request("POST", "")),
+            body=None,
+            message="unknown provider for model",
+        ),
+    ) as mock_create:
+        subentry_flow = await hass.config_entries.subentries.async_configure(
+            subentry_flow["flow_id"],
+            {
+                CONF_WEB_SEARCH: True,
+                CONF_WEB_SEARCH_CONTEXT_SIZE: "medium",
+                CONF_WEB_SEARCH_USER_LOCATION: True,
+            },
+        )
+        await hass.async_block_till_done()
+    assert mock_create.call_args.kwargs["model"] == "claude-sonnet-5-5"
+    assert mock_create.call_args.kwargs["input"][0]["role"] == "user"
+    assert subentry_flow["type"] is FlowResultType.ABORT
+    assert subentry_flow["reason"] == "reconfigure_successful"
+    assert CONF_WEB_SEARCH_CITY not in subentry.data
+    assert CONF_WEB_SEARCH_REGION not in subentry.data
+    assert subentry.data[CONF_WEB_SEARCH_COUNTRY] == "DE"
+    assert subentry.data[CONF_WEB_SEARCH_TIMEZONE] == "Europe/Berlin"
+
+
 async def test_creating_ai_task_subentry(
     hass: HomeAssistant,
     mock_config_entry: MockConfigEntry,
@@ -1307,6 +1375,7 @@ async def test_creating_ai_task_subentry(
     assert result2.get("title") == "Custom AI Task"
     assert result2.get("data") == {
         CONF_RECOMMENDED: True,
+        CONF_CHAT_MODEL: RECOMMENDED_CHAT_MODEL,
     }
 
     assert (
@@ -1354,6 +1423,7 @@ async def test_creating_ai_task_subentry_additional(
         {
             "name": "Advanced AI Task",
             CONF_RECOMMENDED: False,
+            CONF_CHAT_MODEL: "gpt-4o",
         },
     )
 
@@ -1364,7 +1434,6 @@ async def test_creating_ai_task_subentry_additional(
     result3 = await hass.config_entries.subentries.async_configure(
         result["flow_id"],
         {
-            CONF_CHAT_MODEL: "gpt-4o",
             CONF_MAX_TOKENS: 200,
             CONF_STORE_RESPONSES: True,
             CONF_TEMPERATURE: 0.5,

tests/test_conversation.py

@@ -231,19 +231,6 @@ async def test_conversation_agent(
 @pytest.mark.parametrize(
     ("description", "messages"),
     [
-        (
-            "Test function call started with missing arguments",
-            (
-                *create_function_tool_call_item(
-                    id="fc_1",
-                    arguments=[],
-                    call_id="call_call_1",
-                    name="test_tool",
-                    output_index=0,
-                ),
-                *create_message_item(id="msg_A", text="Cool", output_index=1),
-            ),
-        ),
         (
             "Test invalid JSON",
             (

tests/test_entity.py

@@ -0,0 +1,184 @@
+"""Tests for the fork's chat log conversion fixes in entity.py."""
+
+from unittest.mock import MagicMock
+
+from homeassistant.components import conversation
+from homeassistant.core import Context
+from homeassistant.helpers import llm
+
+from custom_components.openai_compatible.const import (
+    CONF_WEB_SEARCH,
+    CONF_WEB_SEARCH_COUNTRY,
+    CONF_WEB_SEARCH_TIMEZONE,
+    CONF_WEB_SEARCH_USER_LOCATION,
+)
+from custom_components.openai_compatible.entity import (
+    _convert_content_to_param,
+    _transform_stream,
+)
+
+from . import create_function_tool_call_item, create_message_item
+
+
+async def _stream(events):
+    for event in events:
+        yield event
+
+
+async def test_tool_call_without_arguments() -> None:
+    """Test an empty argument string means a call without parameters.
+
+    Some OpenAI-compatible endpoints send "" instead of "{}" for tools that
+    take no parameters, e.g. the timer status intent.
+    """
+    events = create_function_tool_call_item(
+        id="fc_1",
+        arguments=[],
+        call_id="call_1",
+        name="HassTimerStatus",
+        output_index=0,
+    )
+
+    deltas = [delta async for delta in _transform_stream(MagicMock(), _stream(events))]
+
+    tool_calls = [call for delta in deltas for call in delta.get("tool_calls", [])]
+    assert len(tool_calls) == 1
+    assert tool_calls[0].tool_name == "HassTimerStatus"
+    assert tool_calls[0].tool_args == {}
+
+
+def test_web_search_not_replayed() -> None:
+    """Test earlier web searches are left out of the history.
+
+    The chat log keeps only a search's status, not its results. Replaying the
+    bare search makes the model disown the answer it based on that search.
+    """
+    content = [
+        conversation.UserContent(content="How did the match go?"),
+        conversation.AssistantContent(
+            agent_id="conversation.test",
+            tool_calls=[
+                llm.ToolInput(
+                    id="ws_1",
+                    tool_name="web_search_call",
+                    tool_args={"action": {"type": "search", "query": "match"}},
+                    external=True,
+                )
+            ],
+        ),
+        conversation.ToolResultContent(
+            agent_id="conversation.test",
+            tool_call_id="ws_1",
+            tool_name="web_search_call",
+            result=llm.ToolResult(data={"status": "completed"}),
+        ),
+        conversation.AssistantContent(
+            agent_id="conversation.test",
+            content="It ended 2 to 2.",
+            tool_calls=[
+                llm.ToolInput(id="call_1", tool_name="GetLiveContext", tool_args={})
+            ],
+        ),
+        conversation.ToolResultContent(
+            agent_id="conversation.test",
+            tool_call_id="call_1",
+            tool_name="GetLiveContext",
+            result=llm.ToolResult(data={"success": True}),
+        ),
+    ]
+
+    messages = _convert_content_to_param(content)
+
+    assert [message["type"] for message in messages] == [
+        "message",
+        "message",
+        "function_call",
+        "function_call_output",
+    ]
+    assert messages[1]["content"] == "It ended 2 to 2."
+    assert messages[2]["call_id"] == "call_1"
+
+
+async def test_web_search_location_leaves_out_empty_fields(
+    hass, mock_config_entry, mock_init_component, mock_create_stream
+) -> None:
+    """Test unknown location fields are left out of the web search tool.
+
+    A failed location lookup stores no city or region; endpoints in front of
+    Anthropic models reject empty strings there.
+    """
+    subentry = next(iter(mock_config_entry.subentries.values()))
+    hass.config_entries.async_update_subentry(
+        mock_config_entry,
+        subentry,
+        data={
+            **subentry.data,
+            CONF_WEB_SEARCH: True,
+            CONF_WEB_SEARCH_USER_LOCATION: True,
+            CONF_WEB_SEARCH_COUNTRY: "DE",
+            CONF_WEB_SEARCH_TIMEZONE: "Europe/Berlin",
+        },
+    )
+    await hass.config_entries.async_reload(mock_config_entry.entry_id)
+    mock_create_stream.return_value = [
+        create_message_item(id="msg_A", text="Hi", output_index=0)
+    ]
+
+    await conversation.async_converse(
+        hass, "hello", None, Context(), agent_id="conversation.openai_conversation"
+    )
+
+    tools = mock_create_stream.mock_calls[0][2]["tools"]
+    web_search = next(tool for tool in tools if tool["type"] == "web_search")
+    assert web_search["user_location"] == {
+        "type": "approximate",
+        "country": "DE",
+        "timezone": "Europe/Berlin",
+    }
+
+
+def test_local_intent_replayed_as_note() -> None:
+    """Test a request a local intent answered is described, not replayed as a call.
+
+    With prefer_local_intents, the chat log holds intents like
+    HassGetCurrentTime as external tool calls. They are not the model's tools,
+    and replayed as its own calls they make it apologise for using them.
+    """
+    content = [
+        conversation.UserContent(content="What time is it?"),
+        conversation.AssistantContent(
+            agent_id="conversation.test",
+            tool_calls=[
+                llm.ToolInput(
+                    id="intent_1",
+                    tool_name="HassGetCurrentTime",
+                    tool_args={},
+                    external=True,
+                )
+            ],
+        ),
+        conversation.ToolResultContent(
+            agent_id="conversation.test",
+            tool_call_id="intent_1",
+            tool_name="HassGetCurrentTime",
+            result=llm.ToolResult(data={"speech": {"plain": {"speech": "10:29"}}}),
+        ),
+        conversation.AssistantContent(
+            agent_id="conversation.test", content="It is 10:29"
+        ),
+        conversation.UserContent(content="What is your name?"),
+    ]
+
+    messages = _convert_content_to_param(content)
+
+    assert [message["type"] for message in messages] == ["message"] * 4
+    assert [message["role"] for message in messages] == [
+        "user",
+        "developer",
+        "assistant",
+        "user",
+    ]
+    note = messages[1]["content"]
+    assert "HassGetCurrentTime" in note
+    assert "10:29" in note
+    assert "not one of your tools" in note

tests/test_model_list.py

@@ -73,6 +73,20 @@ async def test_fetch_survives_an_endpoint_that_will_not_list(error: Exception) -
     assert await async_fetch_model_ids(_client(error=error)) == []
 
 
+async def test_fetch_survives_a_non_openai_error() -> None:
+    """Discovery must never break the form it runs inside.
+
+    A proxy can return a payload the SDK cannot parse, or a client can be
+    wired in a way that raises something other than an OpenAIError; either way
+    the dropdown should degrade to a text box, not take the dialog down.
+    """
+    client = MagicMock()
+    client.with_options.return_value = client
+    client.models.list = MagicMock(return_value="not awaitable")
+
+    assert await async_fetch_model_ids(client) == []
+
+
 async def test_field_is_a_dropdown_that_still_accepts_a_typed_name() -> None:
     """A proxy may serve a model it does not advertise, so allow free text."""
     field = model_field(["claude-sonnet", "gpt-4o-mini"])

tests/test_model_picker.py

@@ -47,8 +47,8 @@ async def _model_step(
 ) -> dict:
     """Return the form that asks for the model, for any endpoint type.
 
-    stt and tts are single-step flows that ask on init; conversation and
-    ai_task_data ask on the step after it.
+    All four types ask on the first screen, so that picking a model never
+    depends on noticing the "recommended" checkbox first.
     """
     with patch(DISCOVERY, return_value=models):
         result = await hass.config_entries.subentries.async_init(
@@ -57,16 +57,6 @@ async def _model_step(
         )
     assert result["type"] is FlowResultType.FORM
     assert result["step_id"] == "init"
-
-        if subentry_type in ("stt", "tts"):
-            return result
-
-        result = await hass.config_entries.subentries.async_configure(
-            result["flow_id"],
-            {CONF_NAME: "An endpoint", CONF_RECOMMENDED: False},
-        )
-        assert result["type"] is FlowResultType.FORM
-        assert result["step_id"] == "additional"
     return result
 
 
@@ -133,17 +123,14 @@ async def test_a_non_openai_model_name_survives_the_flow(
             (mock_config_entry.entry_id, "conversation"),
             context={"source": config_entries.SOURCE_USER},
         )
+        # The model is chosen on the very first screen, alongside the name.
         result = await hass.config_entries.subentries.async_configure(
             result["flow_id"],
-            {CONF_NAME: "Sonnet via LiteLLM", CONF_RECOMMENDED: False},
-        )
-        result = await hass.config_entries.subentries.async_configure(
-            result["flow_id"], {CONF_CHAT_MODEL: "claude-sonnet-4-6"}
-        )
-        # remaining steps take their defaults
-        while result["type"] is FlowResultType.FORM:
-            result = await hass.config_entries.subentries.async_configure(
-                result["flow_id"], {}
+            {
+                CONF_NAME: "Sonnet via LiteLLM",
+                CONF_CHAT_MODEL: "claude-sonnet-4-6",
+                CONF_RECOMMENDED: True,
+            },
         )
         await hass.async_block_till_done()