From 62169ae8fee67148d2b66a34eb942b5f3de7aa56 Mon Sep 17 00:00:00 2001
From: witbrock <witbrock@gmail.com>
Date: Sat, 12 Sep 2026 21:08:15 -0400
Subject: [PATCH 1/2] Repair allowed-model saves and metadata-backed
 transcription selection

---
 docs/engineering/speech_io.md                 |  45 +++++
 src/backend/languagemodels/llm_interface.py   |   8 +-
 src/backend/server/routes/settings_routes.py  | 169 ++++++++++++++----
 src/backend/server/routes/speech_routes.py    |   5 +-
 src/backend/services/model_audio_metadata.py  |  29 +++
 .../services/speech_transcription_service.py  |  38 +++-
 .../web/von_interface/static/js/dictation.js  |  10 +-
 .../von_interface/static/js/modelInventory.js |  92 ++++++++++
 .../von_interface/static/js/settingsPage.js   | 144 +++++++++++++--
 .../web/von_interface/static/styles.css       |   7 +
 .../von_interface/templates/settings_tab.html |   9 +
 tests/backend/test_model_audio_metadata.py    |  98 ++++++++++
 tests/backend/test_speech_transcription.py    |  60 +++++++
 tests/frontend/modelInventory.test.js         |  69 +++++++
 .../settingsModelPoolControls.test.js         |  53 +++++-
 15 files changed, 762 insertions(+), 74 deletions(-)
 create mode 100644 src/backend/services/model_audio_metadata.py
 create mode 100644 src/frontend/web/von_interface/static/js/modelInventory.js
 create mode 100644 tests/backend/test_model_audio_metadata.py
 create mode 100644 tests/frontend/modelInventory.test.js

diff --git a/docs/engineering/speech_io.md b/docs/engineering/speech_io.md
index 66b6317ff..4730db535 100644
--- a/docs/engineering/speech_io.md
+++ b/docs/engineering/speech_io.md
@@ -231,3 +231,48 @@ handoff/reconciliation, actor denial and telemetry isolation/read-back.
 
 Streaming references: [OpenAI realtime transcription](https://developers.openai.com/api/docs/guides/realtime-transcription),
 [WebRTC transport](https://developers.openai.com/api/docs/guides/voice-webrtc?api=realtime).
+
+## Recorded-transcription metadata and model settings
+
+The Settings additional-model inventory projects each provider's existing
+catalogue (OpenAI, OpenRouter, Gemini, Meta and all scanned Ollama hosts). It
+labels catalogue availability separately from the selected target's persisted
+allow list and from assigned browser, primary, server-default, RAG and recorded
+transcription roles. An allowed alternative is not automatically assigned a
+role. Provider availability means discovery succeeded with the configured key
+or host; it does not establish quota or guarantee a later inference request.
+Remote Ollama hosts remain subject to the existing remote-scan setting. Meta
+retains its existing fixed catalogue. A disabled or unreachable provider gets
+a provider-specific status and can be rechecked with Refresh provider inventory.
+
+**Allow and save** checks discovery, retains the selected primary and existing
+alternatives, saves through the canonical scoped settings API, and checks the
+returned pool before enabling use. Failed saves retain the edit and a retry
+action. The 60-second browser observation bound releases a stuck Settings
+interaction; it is not proof the server did not save. An unknown outcome asks
+for **Reload saved model pool** before retrying. Reload discards local edits in
+favour of canonical state. Saving and reloading update the selectors immediately.
+
+The represented model registry's `capabilities.audio_transcription` boolean
+(on `#V#has_model_capabilities_json`) declares compatibility with recorded file
+transcription. Exact provider/model identity or a registry model alias matches
+the entry; model-name substrings are not evidence of support. Explicit false
+can disable compatibility. The compatibility defaults for the existing OpenAI
+file adapter are `gpt-transcribe`, `gpt-4o-transcribe` and
+`gpt-4o-mini-transcribe`. Discovery still determines availability; defaults
+absent from a provider catalogue are displayed as unavailable, not invented as
+accessible models. Additional registry entries can extend the eligible list.
+Only the implemented OpenAI file-transcription transport is currently supported;
+OpenRouter audio input, general chat, realtime transcription and speech synthesis
+do not establish compatibility with that transport.
+
+The recorded-transcription preference belongs to this browser and is distinct
+from the scoped allow list. Its selector uses the actor's effective pool even
+when an administrator is editing the organisation target. Disabled, unavailable
+and stale selections remain visible but cannot be newly selected. The server
+rechecks the requested model against metadata, the effective pool and the
+configured OpenAI key, then applies the existing execution authority check before
+sending audio. A stale explicit selection fails visibly rather than silently
+switching to another model. Automatic selection retains the existing environment
+preference and compatibility ordering, then considers metadata-backed models.
+Live dictation continues to use its separate realtime transport.
diff --git a/src/backend/languagemodels/llm_interface.py b/src/backend/languagemodels/llm_interface.py
index a65ff4e41..8c379ebeb 100644
--- a/src/backend/languagemodels/llm_interface.py
+++ b/src/backend/languagemodels/llm_interface.py
@@ -2469,9 +2469,9 @@ def get_embedding(self, text: str, model: Optional[str] = None) -> List[float]:
             )
             raise RuntimeError(f"OpenAI embedding error: {str(e)}") from e
 
-    def list_models(self) -> List[str]:
-        """List available models from OpenAI (focus on GPT models) with caching/backoff."""
-        cache_key = "openai:models"
+    def list_models(self, *, include_all: bool = False) -> List[str]:
+        """Discover all model IDs for inventory, or the legacy GPT chat catalogue."""
+        cache_key = "openai:models:all" if include_all else "openai:models"
         now = time.time()
 
         # Check cache
@@ -2501,7 +2501,7 @@ def list_models(self) -> List[str]:
             models_response = self.client.models.list()
             for model in getattr(models_response, "data", []):  # type: ignore[attr-defined]
                 mid = getattr(model, "id", None)
-                if isinstance(mid, str) and "gpt" in mid:
+                if isinstance(mid, str) and (include_all or "gpt" in mid):
                     models.append(mid)
             models = sorted(models)
             logger.debug(f"Found OpenAI models (filtered): {models}")
diff --git a/src/backend/server/routes/settings_routes.py b/src/backend/server/routes/settings_routes.py
index 697c58d02..916f19bbe 100644
--- a/src/backend/server/routes/settings_routes.py
+++ b/src/backend/server/routes/settings_routes.py
@@ -962,18 +962,24 @@ def get_llm_info():
 def execution_cost_display():
     """Read viewer-scoped thresholds or replace the authenticated user's pair."""
     from ...services.execution_cost_display_service import (
-        PREDICATE, resolve_preferences, validate_preferences,
+        PREDICATE,
+        resolve_preferences,
+        validate_preferences,
     )
 
     actor = get_effective_user_concept_id()
     if not actor:
         return _jsonify_no_store({"error": "authenticated_actor_context_required"}, 401)
     effective = get_effective_context(
-        request.headers.get("X-Von-Window-Session"), dict(session), actor,
+        request.headers.get("X-Von-Window-Session"),
+        dict(session),
+        actor,
     )
     if request.method == "PUT":
         import json
-        from ...services.ontology_mutation_command_service import execute_governed_ontology_method
+        from ...services.ontology_mutation_command_service import (
+            execute_governed_ontology_method,
+        )
         from ...services.text_value_service import upsert_singleton_text_relation
 
         try:
@@ -981,19 +987,32 @@ def execution_cost_display():
         except ValueError as exc:
             return _jsonify_no_store({"error": str(exc)}, 400)
         text = json.dumps(value, sort_keys=True)
-        provenance = {"source": "execution_cost_display_preferences", "actor_concept_id": actor}
+        provenance = {
+            "source": "execution_cost_display_preferences",
+            "actor_concept_id": actor,
+        }
         result = execute_governed_ontology_method(
             method_name="upsert_singleton_text_relation",
-            arguments={"concept_id": actor, "predicate": PREDICATE, "text": text,
-                       "language": "en-NZ", "provenance": provenance},
+            arguments={
+                "concept_id": actor,
+                "predicate": PREDICATE,
+                "text": text,
+                "language": "en-NZ",
+                "provenance": provenance,
+            },
             mutate=lambda: upsert_singleton_text_relation(
-                subject_concept_id=actor, predicate=PREDICATE, text=text,
-                lang="en-NZ", provenance=provenance,
+                subject_concept_id=actor,
+                predicate=PREDICATE,
+                text=text,
+                lang="en-NZ",
+                provenance=provenance,
             ),
         )
         if result.get("success") is False:
             return _jsonify_no_store(result, 503)
-    return _jsonify_no_store(resolve_preferences(actor, effective.get("organisation_id")))
+    return _jsonify_no_store(
+        resolve_preferences(actor, effective.get("organisation_id"))
+    )
 
 
 @settings_bp.route("/llm/cost_summary", methods=["GET"])
@@ -1014,9 +1033,7 @@ def get_llm_cost_summary():
             {"error": "An authenticated actor and namespace are required."}, 401
         )
     if not provider or not model:
-        return _jsonify_no_store(
-            {"error": "provider and model are required."}, 400
-        )
+        return _jsonify_no_store({"error": "provider and model are required."}, 400)
     try:
         limit = int(request.args.get("limit") or 200)
         window_days = int(request.args.get("window_days") or 30)
@@ -2049,11 +2066,7 @@ def _settings_key_presence(env_var_name: str) -> tuple[bool, str | None]:
         if read_secret_file(os.getenv(f"{name}_FILE")):
             return True, "secret_file"
 
-    dotenv_keys = tuple(
-        key
-        for name in candidates
-        for key in (name, f"{name}_FILE")
-    )
+    dotenv_keys = tuple(key for name in candidates for key in (name, f"{name}_FILE"))
     dotenv_values = read_repo_dotenv_values(dotenv_keys)
     for name in candidates:
         if dotenv_values.get(name):
@@ -2101,7 +2114,7 @@ def get_ollama_models():
 
 
 @settings_bp.route("/models/openai", methods=["GET"])
-def get_openai_models():
+def get_openai_models(*, include_all=False):
     """API endpoint to get available OpenAI models regardless of current selection."""
     try:
         # Get the configured OpenAI API key environment variable
@@ -2109,7 +2122,11 @@ def get_openai_models():
 
         # Create a direct OpenAIClient instance to get models
         client = OpenAIClient(api_key_env_var=api_key_env_var)
-        models = client.list_models()
+        models = (
+            client.list_models(include_all=True)
+            if include_all
+            else client.list_models()
+        )
 
         if models:
             current_app.logger.info(f"Retrieved {len(models)} OpenAI models")
@@ -2255,6 +2272,66 @@ def get_meta_models():
         )
 
 
+@settings_bp.route("/models/inventory/<provider>", methods=["GET"])
+def get_model_inventory(provider):
+    """Project discovery and transport metadata without enabling any model."""
+    from ...services.model_audio_metadata import recorded_transcription_supported
+    from ...services.model_registry_service import get_model_registry_snapshot
+    from ...languagemodels.model_defaults import TRANSCRIPTION_MODELS
+
+    loaders = {
+        "openai": lambda: get_openai_models(include_all=True),
+        "openrouter": get_openrouter_models,
+        "gemini": get_gemini_models,
+        "meta": get_meta_models,
+        "ollama": get_ollama_models_from_all_hosts,
+    }
+    if provider not in loaders:
+        return _jsonify_no_store({"error": "Unsupported model provider."}, 400)
+    response = current_app.make_response(loaders[provider]())
+    payload = response.get_json()
+    available = response.status_code == 200
+    discovered = payload.get("models", []) if isinstance(payload, dict) else payload
+    discovered = discovered if available and isinstance(discovered, list) else []
+    registry_models = get_model_registry_snapshot().get("models", [])
+    entries = []
+    for value in discovered:
+        model = (
+            value.get("model") or value.get("name")
+            if isinstance(value, dict)
+            else value
+        )
+        if not isinstance(model, str) or not model:
+            continue
+        entry = {"provider": provider, "model": model, "available": True}
+        if isinstance(value, dict) and (value.get("host") or value.get("host_url")):
+            entry["host"] = value.get("host") or value.get("host_url")
+        entries.append(entry)
+    if provider == "openai":
+        for model in TRANSCRIPTION_MODELS:
+            if not any(e["model"] == model for e in entries):
+                entries.append(
+                    {"provider": provider, "model": model, "available": False}
+                )
+    for entry in entries:
+        entry["audio_transcription"] = recorded_transcription_supported(
+            provider, entry["model"], registry_models=registry_models
+        )
+    return _jsonify_no_store(
+        {
+            "provider": provider,
+            "available": available,
+            "error": (
+                None
+                if available
+                else f"{provider}: catalogue unavailable; check the provider key, host and access, then refresh."
+            ),
+            "models": entries,
+        },
+        200,
+    )
+
+
 @settings_bp.route("/model_parameters/capabilities", methods=["GET"])
 def get_model_parameter_capabilities():
     provider = str(request.args.get("provider") or "").strip().lower()
@@ -2355,10 +2432,14 @@ def first_val(v):
                 preferred_language = language_rows[0].get("text")
         # Legacy member_of_organisation describes membership, not preference.
         # Never choose the first membership as a landing context.
-        from ...services.login_organisation_preference_service import initialise_login_context
+        from ...services.login_organisation_preference_service import (
+            initialise_login_context,
+        )
+
         organisation_concept_id = (
             initialise_login_context(session)["organisation_concept_id"]
-            if session.get("user_email") and session.get("user_concept_id") else None
+            if session.get("user_email") and session.get("user_concept_id")
+            else None
         )
         return (
             jsonify(
@@ -2674,9 +2755,9 @@ def verify_gemini_api_key():
         )
     api_key = _resolve_settings_api_key(
         api_key_env_var,
-        fallback_env_vars=("GOOGLE_API_KEY",)
-        if api_key_env_var == "GEMINI_API_KEY"
-        else (),
+        fallback_env_vars=(
+            ("GOOGLE_API_KEY",) if api_key_env_var == "GEMINI_API_KEY" else ()
+        ),
     )
     if not api_key:
         return _jsonify_no_store(
@@ -2780,7 +2861,10 @@ def verify_openrouter_api_key():
     """Verify the fixed OpenRouter key source by listing its current catalogue."""
 
     payload = request.get_json(silent=True) or {}
-    if str(payload.get("api_key_env_var") or "OPENROUTER_API_KEY").strip() != "OPENROUTER_API_KEY":
+    if (
+        str(payload.get("api_key_env_var") or "OPENROUTER_API_KEY").strip()
+        != "OPENROUTER_API_KEY"
+    ):
         return _jsonify_no_store(
             {
                 "success": False,
@@ -4285,9 +4369,9 @@ def test_gemini_model():
 
     api_key = _resolve_settings_api_key(
         api_key_env_var,
-        fallback_env_vars=("GOOGLE_API_KEY",)
-        if api_key_env_var == "GEMINI_API_KEY"
-        else (),
+        fallback_env_vars=(
+            ("GOOGLE_API_KEY",) if api_key_env_var == "GEMINI_API_KEY" else ()
+        ),
     )
     if not api_key:
         return _jsonify_no_store(
@@ -4420,13 +4504,8 @@ def test_meta_model():
             400,
         )
 
-    actor_user_concept_id, actor_source = (
-        get_effective_user_concept_id_with_source()
-    )
-    if (
-        not actor_user_concept_id
-        or actor_source == LEGACY_IDENTITY_HEADER_ACTOR_SOURCE
-    ):
+    actor_user_concept_id, actor_source = get_effective_user_concept_id_with_source()
+    if not actor_user_concept_id or actor_source == LEGACY_IDENTITY_HEADER_ACTOR_SOURCE:
         return _jsonify_no_store(
             {
                 "success": False,
@@ -4657,13 +4736,25 @@ def test_openrouter_model():
         failure_kind = "unexpected_error"
         reason = "Unexpected OpenRouter model probe failure."
         if status_code == 401 or "auth" in lowered or "api key" in lowered:
-            failure_kind, reason = "authentication_error", "OpenRouter authentication failed."
+            failure_kind, reason = (
+                "authentication_error",
+                "OpenRouter authentication failed.",
+            )
         elif status_code == 403 or "permission" in lowered:
-            failure_kind, reason = "permission_denied", "OpenRouter access was denied for this model."
+            failure_kind, reason = (
+                "permission_denied",
+                "OpenRouter access was denied for this model.",
+            )
         elif status_code == 404 or "not found" in lowered:
-            failure_kind, reason = "model_not_found", "The selected OpenRouter model was not found."
+            failure_kind, reason = (
+                "model_not_found",
+                "The selected OpenRouter model was not found.",
+            )
         elif status_code == 429 or "rate limit" in lowered or "quota" in lowered:
-            failure_kind, reason = "rate_limited", "OpenRouter rate or quota limit was reached."
+            failure_kind, reason = (
+                "rate_limited",
+                "OpenRouter rate or quota limit was reached.",
+            )
         elif "connect" in lowered or "network" in lowered:
             failure_kind, reason = "connection_error", "Could not reach OpenRouter."
         current_app.logger.warning(
diff --git a/src/backend/server/routes/speech_routes.py b/src/backend/server/routes/speech_routes.py
index 57a9167f2..df547d8a0 100644
--- a/src/backend/server/routes/speech_routes.py
+++ b/src/backend/server/routes/speech_routes.py
@@ -49,7 +49,7 @@ def speech_capabilities():
     if not actor:
         return jsonify(error="authentication_required"), 401
     try:
-        data = transcription_capability(actor, organisation)
+        data = transcription_capability(actor, organisation, request.args.get("model", ""))
     except SpeechUnavailable as exc:
         data = {"available": False, "reason": str(exc)}
     except Exception:
@@ -95,6 +95,7 @@ def transcribe_speech():
             context=request.form.get("context", ""),
             vocabulary=vocabulary,
             language=request.form.get("language", ""),
+            model=request.form.get("model", ""),
         )
     except (ValueError, json.JSONDecodeError) as exc:
         return jsonify(error="invalid_audio_request", message=str(exc)), 400
@@ -108,7 +109,7 @@ def transcribe_speech():
         return (
             jsonify(
                 error="transcription_failed",
-                message="Transcription failed. Your recording is available to retry.",
+                message="OpenAI transcription failed. Check model access and the configured provider key in Settings. Your recording is available to retry.",
             ),
             502,
         )
diff --git a/src/backend/services/model_audio_metadata.py b/src/backend/services/model_audio_metadata.py
new file mode 100644
index 000000000..370a9dd7a
--- /dev/null
+++ b/src/backend/services/model_audio_metadata.py
@@ -0,0 +1,29 @@
+"""Recorded-audio transport metadata, supplemented by the represented registry.
+
+The compatibility entries describe existing transport support, not permission or
+provider availability. Registry capabilities.audio_transcription may extend or
+disable that support; only the implemented OpenAI file transport is eligible.
+"""
+
+from ..languagemodels.model_defaults import TRANSCRIPTION_MODELS
+
+
+def recorded_transcription_supported(provider, model, *, registry_models=None):
+    provider = str(provider or "").lower()
+    model = str(model or "").removeprefix(provider + ":")
+    supported = provider == "openai" and model in TRANSCRIPTION_MODELS
+    if registry_models is None:
+        from .model_registry_service import get_model_registry_snapshot
+
+        registry_models = get_model_registry_snapshot().get("models", [])
+    for entry in registry_models:
+        if entry.get("provider") != provider:
+            continue
+        names = [entry.get("model_id"), *(entry.get("model_aliases") or [])]
+        if model not in names and provider + ":" + model not in names:
+            continue
+        capability = (entry.get("capabilities") or {}).get("audio_transcription")
+        if isinstance(capability, bool):
+            supported = capability
+            break
+    return provider == "openai" and supported
diff --git a/src/backend/services/speech_transcription_service.py b/src/backend/services/speech_transcription_service.py
index c1d86f589..ee8d56e4c 100644
--- a/src/backend/services/speech_transcription_service.py
+++ b/src/backend/services/speech_transcription_service.py
@@ -12,6 +12,7 @@
 from ..languagemodels.llm_interface import assert_model_execution_allowed
 from ..languagemodels.model_defaults import TRANSCRIPTION_MODELS
 from .settings_service import get_openai_env_var, resolve_enabled_llm_settings
+from .model_audio_metadata import recorded_transcription_supported
 
 MAX_AUDIO_BYTES = 24_000_000
 AUDIO_TYPES = {
@@ -28,7 +29,9 @@ class SpeechUnavailable(Exception):
     pass
 
 
-def transcription_model(actor: str, organisation: str | None) -> str:
+def transcription_model(
+    actor: str, organisation: str | None, requested_model=""
+) -> str:
     enabled = resolve_enabled_llm_settings(
         user_concept_id=actor, org_concept_id=organisation
     )
@@ -37,20 +40,29 @@ def transcription_model(actor: str, organisation: str | None) -> str:
         for e in enabled
         if isinstance(e, dict) and e.get("provider") == "openai"
     }
-    preferred = os.environ.get("VON_TRANSCRIPTION_MODEL", "").strip()
-    candidates = (preferred,) if preferred else TRANSCRIPTION_MODELS
+    preferred = str(
+        requested_model or os.environ.get("VON_TRANSCRIPTION_MODEL", "")
+    ).strip()
+    candidates = (preferred,) if preferred else (*TRANSCRIPTION_MODELS, *sorted(models))
     for model in candidates:
-        if model in TRANSCRIPTION_MODELS and model in models:
+        if model in models and recorded_transcription_supported("openai", model):
             return model
     raise SpeechUnavailable(
-        "Enable an OpenAI transcription model in Settings: "
+        (
+            f"OpenAI {preferred} is not an enabled recorded-transcription model. "
+            if preferred
+            else ""
+        )
+        + "Enable an OpenAI transcription model in Settings: "
         + ", ".join(TRANSCRIPTION_MODELS)
         + "."
     )
 
 
-def transcription_capability(actor: str, organisation: str | None) -> dict:
-    model = transcription_model(actor, organisation)
+def transcription_capability(
+    actor: str, organisation: str | None, requested_model=""
+) -> dict:
+    model = transcription_model(actor, organisation, requested_model)
     if not os.environ.get(get_openai_env_var() or "OPENAI_API_KEY"):
         raise SpeechUnavailable("The server's configured OpenAI key is unavailable.")
     return {
@@ -84,7 +96,15 @@ def normalise_language(language):
 
 
 def transcribe_audio(
-    *, actor, organisation, audio, mime_type, context="", vocabulary=None, language=""
+    *,
+    actor,
+    organisation,
+    audio,
+    mime_type,
+    context="",
+    vocabulary=None,
+    language="",
+    model="",
 ):
     import openai
 
@@ -93,7 +113,7 @@ def transcribe_audio(
         raise ValueError("Unsupported recording format. Use MP4, WebM, MP3 or WAV.")
     if not audio or len(audio) > MAX_AUDIO_BYTES:
         raise ValueError("Recording is empty or exceeds the 24 MB upload limit.")
-    capability = transcription_capability(actor, organisation)
+    capability = transcription_capability(actor, organisation, model)
     model = capability["model"]
     assert_model_execution_allowed(
         provider="openai",
diff --git a/src/frontend/web/von_interface/static/js/dictation.js b/src/frontend/web/von_interface/static/js/dictation.js
index 38864166d..7a30e7bc1 100644
--- a/src/frontend/web/von_interface/static/js/dictation.js
+++ b/src/frontend/web/von_interface/static/js/dictation.js
@@ -139,11 +139,13 @@ export function createDictationController({ input, button, status, cancelButton,
     async function transcribe(capture) {
         if (!valid(capture)) return;
         render('transcribing', 'Transcribing with conversation vocabulary…');
+        capture.model ??= selectedTranscriptionModel();
         capture.abort = new AbortController();
         const form = new FormData();
         const mime = capture.blob.type.split(';')[0];
         form.append('audio', capture.blob, mime === 'audio/mp4' ? 'dictation.mp4' : 'dictation.webm');
         form.append('language', capture.context.language || '');
+        form.append('model', capture.model || '');
         form.append('context', capture.context.text || '');
         form.append('vocabulary', JSON.stringify(capture.context.vocabulary || []));
         try {
@@ -163,7 +165,7 @@ export function createDictationController({ input, button, status, cancelButton,
         statusDismissed = false;
         if (current) cancel();
         const context = getContext();
-        const capture = { key: context.key, context, base: input.value,
+        const capture = { key: context.key, context, base: input.value, model: selectedTranscriptionModel(),
             selection: { start: input.selectionStart, end: input.selectionEnd }, chunks: [], bytes: 0 };
         capture.completion = new Promise(resolve => { capture.resolve = resolve; });
         current = capture;
@@ -261,7 +263,8 @@ export function createDictationController({ input, button, status, cancelButton,
 
     async function refreshCapabilities() {
         try {
-            const response = await fetchImpl('/api/speech/capabilities', { credentials: 'same-origin', cache: 'no-store' });
+            const model = selectedTranscriptionModel();
+            const response = await fetchImpl('/api/speech/capabilities' + (model ? `?model=${encodeURIComponent(model)}` : ''), { credentials: 'same-origin', cache: 'no-store' });
             const data = await response.json();
             capability = response.ok ? { ...data.transcription, streaming: data.streaming } : { available: false, reason: 'Sign in to use recorded transcription.' };
         } catch (_) {
@@ -302,6 +305,7 @@ export function createDictationController({ input, button, status, cancelButton,
     root.addEventListener?.('pagehide', cancel);
     root.addEventListener?.('focus', refreshCapabilities);
     root.addEventListener?.('von-preferences-changed', refreshCapabilities);
+    root.document?.addEventListener('von:settingsChanged', refreshCapabilities);
     root.document?.addEventListener('visibilitychange', visibility);
     render('idle');
     void refreshCapabilities();
@@ -324,6 +328,8 @@ export function createDictationController({ input, button, status, cancelButton,
             retryButton.removeEventListener('click', retry); root.removeEventListener?.('pagehide', cancel);
             root.removeEventListener?.('focus', refreshCapabilities);
             root.removeEventListener?.('von-preferences-changed', refreshCapabilities);
+            root.document?.removeEventListener('von:settingsChanged', refreshCapabilities);
             root.document?.removeEventListener('visibilitychange', visibility);
         } };
 }
+import { selectedTranscriptionModel } from './modelInventory.js';
diff --git a/src/frontend/web/von_interface/static/js/modelInventory.js b/src/frontend/web/von_interface/static/js/modelInventory.js
new file mode 100644
index 000000000..cc7b133df
--- /dev/null
+++ b/src/frontend/web/von_interface/static/js/modelInventory.js
@@ -0,0 +1,92 @@
+// Discovery is evidence of availability, never model-pool authority.
+export const TRANSCRIPTION_MODEL_KEY = 'chatRecordedTranscriptionModel';
+
+export function selectedTranscriptionModel() {
+    try { return localStorage.getItem(TRANSCRIPTION_MODEL_KEY) || ''; } catch { return ''; }
+}
+
+export async function modelSettingsRequest(url, options = {}, timeoutMs = 60000) {
+    const controller = new AbortController();
+    let timer;
+    try {
+        return await Promise.race([
+            (async () => {
+                const response = await fetch(url, { ...options, signal: controller.signal });
+                const data = await response.json();
+                if (!response.ok) throw new Error(data.message || data.error || `Request failed (${response.status}).`);
+                return data;
+            })(),
+            new Promise((_, reject) => {
+                timer = setTimeout(() => {
+                    controller.abort();
+                    reject(new Error('No response within 60 seconds. The save outcome is unknown; reload the model pool before retrying.'));
+                }, timeoutMs);
+            }),
+        ]);
+    } finally {
+        clearTimeout(timer);
+    }
+}
+
+export function renderModelInventory({ inventories, enabled, effectiveEnabled = enabled, roles, allow, busy = false, ready = true }) {
+    const list = document.getElementById('additionalModelInventory');
+    const selector = document.getElementById('settingsTranscriptionModelSelect');
+    if (!list && !selector) return;
+    const same = (a, b) => a.provider === b.provider && a.model === b.model
+        && (a.host || '') === (b.host || '');
+    const models = inventories.flatMap((inventory) => inventory.models || []);
+    for (const entry of [...enabled, ...effectiveEnabled]) {
+        if (!models.some((candidate) => same(candidate, entry))) models.push({ ...entry, available: false });
+    }
+    list?.replaceChildren();
+    for (const inventory of inventories) {
+        if (inventory.error && list) {
+            const error = document.createElement('p');
+            error.textContent = inventory.error;
+            list.append(error);
+        }
+    }
+    for (const entry of models) {
+        const allowed = ready && enabled.some((candidate) => same(candidate, entry));
+        const assigned = roles.filter((role) => same(role.entry || {}, entry)).map((role) => role.label);
+        if (entry.provider === 'openai' && selectedTranscriptionModel() === entry.model) assigned.push('Recorded transcription preference');
+        const row = document.createElement('div');
+        row.className = 'settings-model-pool-entry';
+        const label = document.createElement('span');
+        label.className = 'settings-model-pool-entry-label';
+        label.textContent = `${entry.provider}: ${entry.model}${entry.host ? ` (${entry.host})` : ''} · ${entry.available ? 'Provider available' : 'Provider/model unavailable'} · ${allowed ? 'Allowed' : 'Not allowed'} · ${assigned.join(', ') || 'No role assigned'}`;
+        const button = document.createElement('button');
+        button.type = 'button';
+        button.className = 'btn-secondary';
+        button.textContent = allowed ? 'Allowed' : 'Allow and save';
+        button.disabled = busy || !ready || allowed || !entry.available;
+        button.addEventListener('click', () => allow(entry));
+        row.append(label, button);
+        list?.append(row);
+    }
+    if (selector) {
+        selector.replaceChildren(new Option('Automatic from enabled transcription models', ''));
+        const eligible = models.filter((entry) => entry.audio_transcription === true);
+        for (const entry of eligible) {
+            const allowed = ready && effectiveEnabled.some((candidate) => same(candidate, entry));
+            const option = new Option(`${entry.provider}: ${entry.model} · ${!entry.available ? 'unavailable' : allowed ? 'enabled' : 'enable in model pool'}`, entry.model);
+            option.disabled = !entry.available || !allowed;
+            selector.append(option);
+        }
+        const selected = selectedTranscriptionModel();
+        if (selected && !eligible.some((entry) => entry.model === selected)) {
+            const stale = new Option(`${selected} · unavailable or not eligible; choose another model`, selected);
+            stale.disabled = true;
+            selector.append(stale);
+        }
+        selector.value = selected;
+        selector.disabled = !ready || busy;
+        selector.onchange = () => {
+            if (selector.selectedOptions[0]?.disabled) return;
+            localStorage.setItem(TRANSCRIPTION_MODEL_KEY, selector.value);
+            window.dispatchEvent(new CustomEvent('von-preferences-changed'));
+            window.parent?.dispatchEvent(new CustomEvent('von-preferences-changed'));
+            renderModelInventory({ inventories, enabled, effectiveEnabled, roles, allow, busy, ready });
+        };
+    }
+}
diff --git a/src/frontend/web/von_interface/static/js/settingsPage.js b/src/frontend/web/von_interface/static/js/settingsPage.js
index 9aa366691..b149785e5 100644
--- a/src/frontend/web/von_interface/static/js/settingsPage.js
+++ b/src/frontend/web/von_interface/static/js/settingsPage.js
@@ -1,3 +1,4 @@
+import { modelSettingsRequest, renderModelInventory } from './modelInventory.js';
 import { profileButton } from './components/participantProfile.js';
 import { CONVERSATION_LAYOUT_KEY, CONVERSATION_LAYOUT_KEYS, CONVERSATION_TRAY_HOVER_KEY, loadConversationLayoutPreferences, normaliseConversationLayout, saveConversationLayoutPreference } from './utils/conversationLayoutPreferences.js';
 import { supportsAudioRecording } from './dictation.js';
@@ -177,6 +178,7 @@ let latestCapabilityIndexStatus = null;
 let capabilityIndexStatusUnsubscribe = null;
 let latestRagRuntimeConfiguration = null;
 let currentEnabledLlmAlternatives = [];
+let persistedTargetEnabledLlms = [];
 let currentServerDefaultLlm = null;
 let stagedScopedPrimaryLlm = null;
 let scopedModelStateReady = false;
@@ -2037,6 +2039,7 @@ function applyScopedModelSettings(settings, generation = scopedModelLoadGenerati
   currentEffectiveEnabledLlms = normaliseEnabledLlmEntries(
     settings?.effective_enabled_llms || settings?.enabled_llms,
   );
+  persistedTargetEnabledLlms = normaliseEnabledLlmEntries(settings?.enabled_llms);
   currentEnabledLlmAlternatives = normaliseEnabledLlmEntries(settings?.enabled_llms)
     .filter((entry) => !resolved || !sameLlmSlot(entry, resolved));
   explicitlyRemovedWorkflowPoolKeys = new Set();
@@ -2401,6 +2404,10 @@ function renderModelPoolTargetScopeNote() {
   note.textContent = 'Saving changes only the current user primary and pool.';
 }
 
+function isPersistedTargetModelAllowed(entry) {
+  return scopedModelStateReady && persistedTargetEnabledLlms.some((candidate) => sameLlmSlot(candidate, entry));
+}
+
 function updatePoolActionButton(buttonId, entry) {
   const button = document.getElementById(buttonId);
   if (!button) return;
@@ -2424,10 +2431,10 @@ function updatePoolActionButton(buttonId, entry) {
 
   const existing = currentEnabledLlmAlternatives.find((candidate) => sameLlmSlot(candidate, entry));
   if (existing && sameLlmEntry(existing, entry)) {
-    button.disabled = true;
-    button.textContent = isPersistedEffectiveModelAllowed(entry)
+    button.disabled = scopedModelSaveInFlight || isPersistedTargetModelAllowed(entry);
+    button.textContent = isPersistedTargetModelAllowed(entry)
       ? 'Already allowed for this scope'
-      : 'Pending allowed-model save';
+      : 'Save allowed model now';
     return;
   }
 
@@ -2522,7 +2529,82 @@ function renderWorkflowModelPool() {
   updateSelectedPremiumEligibilityControls();
 }
 
+let additionalModelInventories = [];
+let modelInventoryGeneration = 0;
+
+function renderAdditionalModelInventory() {
+  renderModelInventory({
+    inventories: additionalModelInventories,
+    enabled: persistedTargetEnabledLlms,
+    effectiveEnabled: currentEffectiveEnabledLlms,
+    roles: [
+      { entry: currentEffectiveLlm, label: 'Effective primary' },
+      { entry: currentResolvedLlm, label: 'Save target primary' },
+      { entry: browserChatPrimaryEntry(), label: 'Browser chat' },
+      { entry: currentServerDefaultLlm, label: 'Server default' },
+      { entry: latestRagRuntimeConfiguration?.llm_resolution?.effective, label: 'RAG language model' },
+      { entry: latestRagRuntimeConfiguration?.embedder_resolution?.effective, label: 'RAG embedder' },
+    ],
+    ready: scopedModelStateReady,
+    busy: scopedModelSaveInFlight,
+    allow: allowAndSaveWorkflowPoolEntry,
+  });
+}
+
+async function refreshAdditionalModelInventory() {
+  if (!document.getElementById('additionalModelInventory')) return;
+  const generation = ++modelInventoryGeneration;
+  additionalModelInventories = [];
+  renderAdditionalModelInventory();
+  await Promise.all(['openai', 'openrouter', 'gemini', 'meta', 'ollama'].map(async (provider) => {
+    let inventory;
+    try {
+      inventory = await modelSettingsRequest(`/api/settings/models/inventory/${provider}`, {
+        cache: 'no-store', headers: await buildSettingsFetchHeaders(),
+      });
+    } catch (_) {
+      inventory = { provider, models: [], error: `${provider}: model discovery unavailable. Check provider configuration and refresh the inventory.` };
+    }
+    if (generation !== modelInventoryGeneration) return;
+    additionalModelInventories.push(inventory);
+    renderAdditionalModelInventory();
+  }));
+}
+
+async function allowAndSaveWorkflowPoolEntry(entry) {
+  if (scopedModelSaveInFlight || !entry) return false;
+  const generation = scopedModelLoadGeneration;
+  scopedModelSaveInFlight = true;
+  renderModelScopeOverview();
+  setInlineStatusMessage(document.getElementById('modelPoolStatusMessage'), `Checking ${entry.provider}: ${entry.model} availability…`, null);
+  try {
+    const inventory = await modelSettingsRequest(`/api/settings/models/inventory/${entry.provider}`, {
+      cache: 'no-store', headers: await buildSettingsFetchHeaders(),
+    });
+    if (!inventory.available || !inventory.models?.some((candidate) => candidate.available && sameLlmSlot(candidate, entry))) {
+      throw new Error(inventory.error || `${entry.provider}: ${entry.model} is unavailable. Check the provider key or host and refresh its model list.`);
+    }
+    if (generation !== scopedModelLoadGeneration) return false;
+  } catch (error) {
+    if (generation === scopedModelLoadGeneration) {
+      setInlineStatusMessage(document.getElementById('modelPoolStatusMessage'), `${entry.provider}: ${error.message}`, 'error');
+    }
+    return false;
+  } finally {
+    scopedModelSaveInFlight = false;
+    renderModelScopeOverview();
+  }
+  addOrUpdateWorkflowPoolEntry(entry);
+  const saved = await persistModelScopeSettings();
+  if (!saved) {
+    const status = document.getElementById('modelPoolStatusMessage');
+    setInlineStatusMessage(status, `${entry.provider}: ${entry.model}: ${status?.textContent || 'Save failed; retry.'}`, 'error');
+  }
+  return saved;
+}
+
 function renderModelScopeOverview() {
+  renderAdditionalModelInventory();
   const browserPrimary = browserChatPrimaryEntry();
   setModelScopeSummaryText(
     'browserChatModelSummary',
@@ -2630,6 +2712,7 @@ function restoreScopedPrimary() {
 }
 
 async function persistModelScopeSettings() {
+  if (scopedModelSaveInFlight) return false;
   if (!scopedModelStateReady) {
     setInlineStatusMessage(
       document.getElementById('modelPoolStatusMessage'),
@@ -2641,6 +2724,7 @@ async function persistModelScopeSettings() {
   }
   scopedModelSaveInFlight = true;
   updateScopedModelSaveAvailability();
+  renderModelScopeOverview();
   setInlineStatusMessage(
     document.getElementById('modelPoolStatusMessage'),
     'Saving scoped primary and allowed models...',
@@ -2655,15 +2739,18 @@ async function persistModelScopeSettings() {
       'success',
     );
     void refreshSelectedOpenAiCostSummary();
+    return true;
   } catch (error) {
     setInlineStatusMessage(
       document.getElementById('modelPoolStatusMessage'),
       error?.message || 'Failed to save scoped primary and allowed models.',
       'error',
     );
+    return false;
   } finally {
     scopedModelSaveInFlight = false;
     updateScopedModelSaveAvailability();
+    renderModelScopeOverview();
   }
 }
 
@@ -3970,7 +4057,7 @@ function setupSpeechSettingsSection() {
   const sttSupported = isSpeechRecognitionSupported();
 
   if (supportNote) {
-    supportNote.textContent = `${supportsAudioRecording() ? 'Recorded dictation is supported; enable an OpenAI transcription model (gpt-transcribe, gpt-4o-transcribe or gpt-4o-mini-transcribe) in your model pool.' : 'Recorded dictation needs HTTPS and microphone recording support.'} Browser recognition: ${sttSupported ? 'available with limited context support' : 'unavailable'}. Spoken playback: ${ttsSupported ? 'available' : 'unavailable'}.`;
+    supportNote.textContent = `${supportsAudioRecording() ? 'Recorded dictation uses enabled models with file-transcription metadata and a supported provider transport. Allow a model in the inventory, then select it here. Live dictation uses a separate realtime transport.' : 'Recorded dictation needs HTTPS and microphone recording support.'} Browser recognition: ${sttSupported ? 'available with limited context support' : 'unavailable'}. Spoken playback: ${ttsSupported ? 'available' : 'unavailable'}.`;
   }
 
   if (!ttsSupported) {
@@ -4389,6 +4476,7 @@ export function __testOnly_setModelScopeState({
     );
   currentEffectiveEnabledLlms = normaliseEnabledLlmEntries(effectiveEnabledLlms);
   currentServerDefaultLlm = buildCanonicalLlmEntry(serverDefaultLlm);
+  persistedTargetEnabledLlms = normaliseEnabledLlmEntries(enabledLlms);
   currentEnabledLlmAlternatives = normaliseEnabledLlmEntries(enabledLlms)
     .filter((entry) => !currentResolvedLlm || !sameLlmSlot(entry, currentResolvedLlm));
   explicitlyRemovedWorkflowPoolKeys = new Set();
@@ -4410,6 +4498,14 @@ export function __testOnly_addOrUpdateWorkflowPoolEntry(entry) {
   return addOrUpdateWorkflowPoolEntry(entry);
 }
 
+export function __testOnly_allowAndSaveWorkflowPoolEntry(entry) {
+  return allowAndSaveWorkflowPoolEntry(entry);
+}
+
+export function __testOnly_refreshAdditionalModelInventory() {
+  return refreshAdditionalModelInventory();
+}
+
 export function __testOnly_useBrowserModelAsScopedPrimary() {
   return useBrowserModelAsScopedPrimary();
 }
@@ -5028,6 +5124,8 @@ document.addEventListener('DOMContentLoaded', async () => {
   await loadAndDisplaySettings();
   setupConversationHistorySettingsSection();
   setupSpeechSettingsSection();
+  void refreshAdditionalModelInventory();
+  document.getElementById('refreshModelInventoryButton')?.addEventListener('click', refreshAdditionalModelInventory);
   setupConceptUiSettings();
   // Load DB info
   try { await loadAndDisplayDbInfo(); } catch { }
@@ -5244,19 +5342,19 @@ document.getElementById('testOpenRouterModelButton')?.addEventListener('click',
 document.getElementById('testMetaModelButton')?.addEventListener('click', testSelectedMetaModel);
 document.getElementById('testOllamaModelButton')?.addEventListener('click', testSelectedOllamaModel);
 document.getElementById('addOpenAiToWorkflowPoolButton')?.addEventListener('click', () => {
-  addOrUpdateWorkflowPoolEntry(selectedOpenAiPoolEntry());
+  void allowAndSaveWorkflowPoolEntry(selectedOpenAiPoolEntry());
 });
 document.getElementById('addGeminiToWorkflowPoolButton')?.addEventListener('click', () => {
-  addOrUpdateWorkflowPoolEntry(selectedGeminiPoolEntry());
+  void allowAndSaveWorkflowPoolEntry(selectedGeminiPoolEntry());
 });
 document.getElementById('addOpenRouterToWorkflowPoolButton')?.addEventListener('click', () => {
-  addOrUpdateWorkflowPoolEntry(selectedOpenRouterPoolEntry());
+  void allowAndSaveWorkflowPoolEntry(selectedOpenRouterPoolEntry());
 });
 document.getElementById('addMetaToWorkflowPoolButton')?.addEventListener('click', () => {
-  addOrUpdateWorkflowPoolEntry(selectedMetaPoolEntry());
+  void allowAndSaveWorkflowPoolEntry(selectedMetaPoolEntry());
 });
 document.getElementById('addOllamaToWorkflowPoolButton')?.addEventListener('click', () => {
-  addOrUpdateWorkflowPoolEntry(selectedOllamaPoolEntry());
+  void allowAndSaveWorkflowPoolEntry(selectedOllamaPoolEntry());
 });
 document.getElementById('useBrowserModelAsScopedPrimaryButton')?.addEventListener(
   'click',
@@ -5267,6 +5365,9 @@ document.getElementById('restoreScopedPrimaryButton')?.addEventListener(
   restoreScopedPrimary,
 );
 document.getElementById('saveModelPoolButton')?.addEventListener('click', persistModelScopeSettings);
+document.getElementById('reloadModelPoolButton')?.addEventListener('click', () => {
+  if (!scopedModelSaveInFlight) void reloadScopedModelSettings();
+});
 document.getElementById('saveRuntimeModelSettingsButton')?.addEventListener('click', persistRuntimeModelSettings);
 document.getElementById('modelPoolTargetScopeSelect')?.addEventListener('change', (event) => {
   if (!actorSessionReady) {
@@ -6216,16 +6317,18 @@ async function saveAllSettings({
       settings.rag_llm = ragLlmSetting;
     }
 
-    const response = await fetch('/api/settings/', {
+    const requestOptions = {
       method: 'POST',
       headers: await buildSettingsFetchHeaders({ 'Content-Type': 'application/json' }),
       body: JSON.stringify(settings),
-    });
-    const payload = await response.json().catch(() => ({}));
-
-    if (!response.ok) {
-      const message = payload?.message || `Failed to save settings (${response.status})`;
-      throw new Error(message);
+    };
+    let payload;
+    if (includeChatModels) {
+      payload = await modelSettingsRequest('/api/settings/', requestOptions);
+    } else {
+      const response = await fetch('/api/settings/', requestOptions);
+      payload = await response.json().catch(() => ({}));
+      if (!response.ok) throw new Error(payload.message || `Failed to save settings (${response.status})`);
     }
 
     if (includeChatModels && modelScopeGenerationAtSave !== scopedModelLoadGeneration) {
@@ -6248,10 +6351,16 @@ async function saveAllSettings({
     }
 
     if (includeChatModels && payload?.resolved_llm) {
+      if (!Array.isArray(payload.enabled_llms) || settings.enabled_llms.some(
+        (entry) => !payload.enabled_llms.some((saved) => sameLlmEntry(entry, saved)),
+      )) throw new Error('Saved pool did not contain all requested models. Reload the pool and retry.');
       currentResolvedLlm = payload.resolved_llm;
+      if (selectedModelPoolTargetScope === 'user' || effectiveModelScope(currentEffectiveLlm) === 'organisation') {
+        currentEffectiveLlm = payload.resolved_llm;
+      }
       stagedScopedPrimaryLlm = buildCanonicalLlmEntry(payload.resolved_llm);
       if (Array.isArray(payload?.enabled_llms)) {
-        const persistedTargetEnabledLlms = normaliseEnabledLlmEntries(payload.enabled_llms);
+        persistedTargetEnabledLlms = normaliseEnabledLlmEntries(payload.enabled_llms);
         currentEnabledLlmAlternatives = [...persistedTargetEnabledLlms]
           .filter((entry) => !currentResolvedLlm || !sameLlmSlot(entry, currentResolvedLlm));
         if (
@@ -6325,6 +6434,7 @@ async function saveAllSettings({
         true,
       );
     }
+    if (includeChatModels) throw error;
     return false;
   }
 }
diff --git a/src/frontend/web/von_interface/static/styles.css b/src/frontend/web/von_interface/static/styles.css
index 099173d0f..0d340130d 100644
--- a/src/frontend/web/von_interface/static/styles.css
+++ b/src/frontend/web/von_interface/static/styles.css
@@ -15524,6 +15524,13 @@ body.settings-body {
     gap: 8px;
 }
 
+#additionalModelInventory {
+    display: grid;
+    gap: 8px;
+    max-height: 32rem;
+    overflow: auto;
+}
+
 .settings-model-pool-entry {
     display: flex;
     align-items: center;
diff --git a/src/frontend/web/von_interface/templates/settings_tab.html b/src/frontend/web/von_interface/templates/settings_tab.html
index 3f2012db4..ed4ae8854 100644
--- a/src/frontend/web/von_interface/templates/settings_tab.html
+++ b/src/frontend/web/von_interface/templates/settings_tab.html
@@ -233,6 +233,9 @@ <h3 class="speech-settings-title">Text-to-speech</h3>
 
                     <section class="speech-settings-section" aria-label="Dictation settings">
                         <h3 class="speech-settings-title">Dictation</h3>
+                        <label for="settingsTranscriptionModelSelect">Recorded transcription model (this browser)</label>
+                        <select id="settingsTranscriptionModelSelect" disabled></select>
+                        <p class="settings-note">Only enabled, available file-transcription models can be selected. A stale preference remains visible until you choose another model.</p>
                         <p class="speech-settings-note">Recorded audio uses conversation vocabulary. The browser controls below apply only to browser recognition.</p>
 
                         <div class="speech-settings-row">
@@ -314,11 +317,17 @@ <h3>Scoped allowed models</h3>
                         The scoped primary is always allowed. Add or remove alternatives explicitly; configuring
                         or testing a model does not allow it.
                     </p>
+                    <details>
+                        <summary>Additional models — all providers and assigned roles</summary>
+                        <button id="refreshModelInventoryButton" type="button" class="btn-secondary">Refresh provider inventory</button>
+                        <div id="additionalModelInventory" role="region" aria-label="Additional model inventory"></div>
+                    </details>
                     <div id="workflowModelPoolList" class="settings-model-pool-list"></div>
                     <div class="inline-form inline-form-tight mt-2">
                         <button id="saveModelPoolButton" class="btn-azure" type="button" disabled>
                             Save scoped primary &amp; allowed models
                         </button>
+                        <button id="reloadModelPoolButton" class="btn-secondary" type="button">Reload saved model pool</button>
                     </div>
                     <div id="modelPoolStatusMessage" class="status-message" role="status"></div>
                 </div>
diff --git a/tests/backend/test_model_audio_metadata.py b/tests/backend/test_model_audio_metadata.py
new file mode 100644
index 000000000..115ee03a3
--- /dev/null
+++ b/tests/backend/test_model_audio_metadata.py
@@ -0,0 +1,98 @@
+from flask import Flask, jsonify
+import pytest
+from src.backend.services.model_audio_metadata import recorded_transcription_supported
+from src.backend.services import model_registry_service
+from src.backend.server.routes import settings_routes
+
+
+def test_metadata_extends_and_disables_compatibility_without_guessing_from_names():
+    registry = [
+        {
+            "provider": "openai",
+            "model_id": "new-audio",
+            "capabilities": {"audio_transcription": True},
+        },
+        {
+            "provider": "openai",
+            "model_id": "gpt-transcribe",
+            "capabilities": {"audio_transcription": False},
+        },
+        {
+            "provider": "openrouter",
+            "model_id": "other-audio",
+            "capabilities": {"audio_transcription": True},
+        },
+    ]
+    assert recorded_transcription_supported(
+        "openai", "new-audio", registry_models=registry
+    )
+    assert not recorded_transcription_supported(
+        "openai", "gpt-transcribe", registry_models=registry
+    )
+    assert not recorded_transcription_supported(
+        "openai", "gpt-realtime-whisper", registry_models=registry
+    )
+    assert not recorded_transcription_supported(
+        "openai", "general-chat", registry_models=registry
+    )
+    assert not recorded_transcription_supported(
+        "openrouter", "other-audio", registry_models=registry
+    )
+    for model in ("gpt-transcribe", "gpt-4o-transcribe", "gpt-4o-mini-transcribe"):
+        assert recorded_transcription_supported("openai", model, registry_models=[])
+
+
+@pytest.mark.parametrize(
+    "provider,loader,payload",
+    [
+        ("openai", "get_openai_models", ["gpt-4o-transcribe", "chat"]),
+        ("openrouter", "get_openrouter_models", ["vendor/chat"]),
+        (
+            "ollama",
+            "get_ollama_models_from_all_hosts",
+            {"models": [{"name": "local", "host_url": "http://localhost:11434"}]},
+        ),
+        ("gemini", "get_gemini_models", ["gemini-model"]),
+        ("meta", "get_meta_models", ["meta-model"]),
+    ],
+)
+def test_inventory_preserves_provider_discovery_and_host_metadata(
+    monkeypatch, provider, loader, payload
+):
+    monkeypatch.setattr(
+        model_registry_service, "get_model_registry_snapshot", lambda: {"models": []}
+    )
+    monkeypatch.setattr(settings_routes, loader, lambda **kw: (jsonify(payload), 200))
+    app = Flask(__name__)
+    with app.test_request_context():
+        response = app.make_response(settings_routes.get_model_inventory(provider))
+        data = response.get_json()
+    assert data["available"]
+    assert all(model["provider"] == provider for model in data["models"])
+    if provider == "ollama":
+        assert data["models"][0]["host"] == "http://localhost:11434"
+    if provider == "openai":
+        models = {model["model"]: model for model in data["models"]}
+        assert models["gpt-4o-transcribe"]["audio_transcription"]
+        assert not models["gpt-transcribe"]["available"]
+        assert not models["chat"]["audio_transcription"]
+
+
+def test_disabled_provider_has_actionable_failure_and_no_available_models(monkeypatch):
+    monkeypatch.setattr(
+        model_registry_service, "get_model_registry_snapshot", lambda: {"models": []}
+    )
+    monkeypatch.setattr(
+        settings_routes,
+        "get_openai_models",
+        lambda **kw: (jsonify(error="private provider details"), 400),
+    )
+    app = Flask(__name__)
+    with app.test_request_context():
+        data = app.make_response(
+            settings_routes.get_model_inventory("openai")
+        ).get_json()
+    assert not data["available"]
+    assert "openai" in data["error"] and "check" in data["error"]
+    assert "private provider details" not in data["error"]
+    assert not any(model["available"] for model in data["models"])
diff --git a/tests/backend/test_speech_transcription.py b/tests/backend/test_speech_transcription.py
index abfca3741..a010acae2 100644
--- a/tests/backend/test_speech_transcription.py
+++ b/tests/backend/test_speech_transcription.py
@@ -10,6 +10,11 @@
 
 @pytest.fixture
 def provider(monkeypatch):
+    from src.backend.services import model_registry_service
+
+    monkeypatch.setattr(
+        model_registry_service, "get_model_registry_snapshot", lambda: {"models": []}
+    )
     monkeypatch.setattr(
         speech,
         "resolve_enabled_llm_settings",
@@ -145,3 +150,58 @@ def test_bad_recording_never_reaches_provider(provider, data, mime):
             actor="#V#a", organisation=None, audio=data, mime_type=mime
         )
     provider[0].audio.transcriptions.create.assert_not_called()
+
+
+def test_explicit_metadata_model_selection_is_checked_against_effective_pool(
+    provider, monkeypatch
+):
+    from src.backend.services import model_registry_service
+
+    monkeypatch.setattr(
+        model_registry_service,
+        "get_model_registry_snapshot",
+        lambda: {
+            "models": [
+                {
+                    "provider": "openai",
+                    "model_id": "future-transcription",
+                    "capabilities": {"audio_transcription": True},
+                }
+            ]
+        },
+    )
+    monkeypatch.setattr(
+        speech,
+        "resolve_enabled_llm_settings",
+        lambda **kw: [
+            {"provider": "openai", "model": "future-transcription"},
+            {"provider": "openai", "model": "general-chat"},
+        ],
+    )
+    assert (
+        speech.transcription_model("#V#speaker", None, "future-transcription")
+        == "future-transcription"
+    )
+    for model in ("general-chat", "removed-transcription", "gpt-4o-transcribe"):
+        with pytest.raises(speech.SpeechUnavailable, match="not an enabled"):
+            speech.transcription_model("#V#speaker", None, model)
+    client, _ = provider
+    speech.transcribe_audio(
+        actor="#V#speaker",
+        organisation=None,
+        audio=b"fixture",
+        mime_type="audio/webm",
+        model="future-transcription",
+    )
+    assert (
+        client.audio.transcriptions.create.call_args.kwargs["model"]
+        == "future-transcription"
+    )
+
+
+def test_disabled_provider_cannot_transcribe_even_with_enabled_model(
+    provider, monkeypatch
+):
+    monkeypatch.delenv("TEST_SPEECH_KEY")
+    with pytest.raises(speech.SpeechUnavailable, match="OpenAI key"):
+        speech.transcription_capability("#V#speaker", None)
diff --git a/tests/frontend/modelInventory.test.js b/tests/frontend/modelInventory.test.js
new file mode 100644
index 000000000..5d65c1278
--- /dev/null
+++ b/tests/frontend/modelInventory.test.js
@@ -0,0 +1,69 @@
+/** @jest-environment jsdom */
+import { modelSettingsRequest, renderModelInventory, TRANSCRIPTION_MODEL_KEY } from '../../src/frontend/web/von_interface/static/js/modelInventory.js';
+
+beforeEach(() => {
+    localStorage.clear();
+    document.body.innerHTML = '<div id="additionalModelInventory"></div><select id="settingsTranscriptionModelSelect"></select>';
+});
+
+const audio = { provider: 'openai', model: 'future-transcriber', available: true, audio_transcription: true };
+const chat = { provider: 'openrouter', model: 'vendor/chat', available: true };
+const local = { provider: 'ollama', model: 'local-model', host: 'http://localhost:11434', available: true };
+const inventories = [{ models: [audio, chat, local] }];
+
+test('all providers, hosts, roles and enablement are visible; only transcription metadata enables selection', () => {
+    const allow = jest.fn();
+    renderModelInventory({ inventories, enabled: [audio, local], roles: [{ entry: local, label: 'Browser chat' }], allow });
+    const rows = [...document.querySelectorAll('.settings-model-pool-entry')];
+    expect(rows).toHaveLength(3);
+    expect(rows[0].textContent).toContain('No role assigned');
+    expect(rows[1].textContent).toContain('openrouter: vendor/chat');
+    expect(rows[2].textContent).toContain('ollama: local-model (http://localhost:11434)');
+    expect(rows[2].textContent).toContain('Browser chat');
+    rows[1].querySelector('button').click();
+    expect(allow).toHaveBeenCalledWith(chat);
+    const select = document.getElementById('settingsTranscriptionModelSelect');
+    expect([...select.options].map(o => o.value)).toEqual(['', audio.model]);
+    select.value = audio.model;
+    select.dispatchEvent(new Event('change'));
+    expect(localStorage.getItem(TRANSCRIPTION_MODEL_KEY)).toBe(audio.model);
+    expect(document.getElementById('additionalModelInventory').textContent).toContain('Recorded transcription preference');
+});
+
+test('stale, disabled and unallowed selections stay visible but cannot be selected', () => {
+    localStorage.setItem(TRANSCRIPTION_MODEL_KEY, 'stale-chat');
+    renderModelInventory({ inventories, enabled: [], roles: [], allow: jest.fn() });
+    const select = document.getElementById('settingsTranscriptionModelSelect');
+    expect(select.value).toBe('stale-chat');
+    expect(select.selectedOptions[0].disabled).toBe(true);
+    expect(select.options[1].disabled).toBe(true);
+    renderModelInventory({ inventories: [{ models: [{ ...audio, available: false }] }], enabled: [audio], roles: [], allow: jest.fn() });
+    expect(select.options[1].disabled).toBe(true);
+    expect(document.querySelector('button').disabled).toBe(true);
+});
+
+test('organisation save target does not grant transcription to a user override', () => {
+    renderModelInventory({ inventories, enabled: [audio], effectiveEnabled: [local], roles: [], allow: jest.fn() });
+    expect(document.querySelector('button').disabled).toBe(true);
+    expect(document.getElementById('settingsTranscriptionModelSelect').options[1].disabled).toBe(true);
+});
+
+test('unresponsive requests terminate visibly, ignore late results, and permit retry', async () => {
+    jest.useFakeTimers();
+    let finish;
+    global.fetch = jest.fn(() => new Promise(resolve => { finish = resolve; }));
+    const request = modelSettingsRequest('/api/settings/', {}, 60000);
+    const check = expect(request).rejects.toThrow('outcome is unknown');
+    await jest.advanceTimersByTimeAsync(60000);
+    await check;
+    expect(global.fetch.mock.calls[0][1].signal.aborted).toBe(true);
+    finish({ ok: true, json: async () => ({ stale: true }) });
+    global.fetch.mockResolvedValue({ ok: true, json: async () => ({ saved: true }) });
+    expect(await modelSettingsRequest('/api/settings/')).toEqual({ saved: true });
+    jest.useRealTimers();
+});
+
+test('provider-specific errors survive the transport', async () => {
+    global.fetch = jest.fn().mockResolvedValue({ ok: false, status: 400, json: async () => ({ message: 'OpenRouter: check model access.' }) });
+    await expect(modelSettingsRequest('/api/settings/')).rejects.toThrow('OpenRouter: check model access.');
+});
diff --git a/tests/frontend/settingsModelPoolControls.test.js b/tests/frontend/settingsModelPoolControls.test.js
index 9e55b70b2..b4023ef72 100644
--- a/tests/frontend/settingsModelPoolControls.test.js
+++ b/tests/frontend/settingsModelPoolControls.test.js
@@ -5,6 +5,7 @@ const path = require('path');
 
 import {
     __testOnly_addOrUpdateWorkflowPoolEntry,
+    __testOnly_allowAndSaveWorkflowPoolEntry,
     __testOnly_applyCapabilityIndexStatusCard,
     __testOnly_applySharedRuntimeModelWriteAccess,
     __testOnly_buildPersistedLlmSelections,
@@ -146,7 +147,55 @@ describe('settings model scope and workflow pool controls', () => {
             model_parameters: { reasoning_effort: 'low' },
         })).toBe(true);
         expect(document.getElementById('addOpenAiToWorkflowPoolButton').textContent)
-            .toBe('Pending allowed-model save');
+            .toBe('Save allowed model now');
+        expect(document.getElementById('addOpenAiToWorkflowPoolButton').disabled).toBe(false);
+    });
+
+    test('Allow checks the provider, persists once, retains existing entries and enables use after read-back', async () => {
+        sessionStorage.setItem('von_current_user', JSON.stringify({ concept_id: '#V#user-a' }));
+        const primary = { provider: 'ollama', model: 'gemma4:latest', scope: 'user' };
+        const old = { provider: 'openrouter', model: 'existing' };
+        const added = { provider: 'openai', model: 'gpt-5.6-luna', model_parameters: { reasoning_effort: 'low' } };
+        __testOnly_setModelScopeState({ resolvedLlm: primary, enabledLlms: [primary, old] });
+        let stored;
+        global.fetch.mockImplementation(async (url, options) => {
+            if (url.includes('/models/inventory/')) return { ok: true, json: async () => ({ available: true, models: [{ ...added, available: true }] }) };
+            if (options?.method === 'POST') {
+                stored = JSON.parse(options.body);
+                return { ok: true, json: async () => ({ resolved_llm: primary, enabled_llms: stored.enabled_llms }) };
+            }
+            return { ok: true, json: async () => ({ resolved_llm: primary, enabled_llms: stored?.enabled_llms || [] }) };
+        });
+        expect(await __testOnly_allowAndSaveWorkflowPoolEntry(added)).toBe(true);
+        expect(stored.enabled_llms).toEqual(expect.arrayContaining([old, added]));
+        expect(stored.active_llm.model).toBe(primary.model);
+        expect(document.getElementById('testOpenAiModelButton').disabled).toBe(false);
+        expect(document.getElementById('modelPoolStatusMessage').textContent).toContain('saved');
+        expect((await __testOnly_reloadScopedModelSettings()).applied).toBe(true);
+        expect(document.getElementById('testOpenAiModelButton').disabled).toBe(false);
+        expect(stored.enabled_llms.filter(entry => entry.model === added.model)).toHaveLength(1);
+    });
+
+    test('unavailable providers do not mutate the staged pool and expose an actionable failure', async () => {
+        __testOnly_setModelScopeState({ resolvedLlm: { provider: 'ollama', model: 'primary', scope: 'user' } });
+        global.fetch.mockResolvedValue({ ok: true, json: async () => ({ available: false, error: 'OpenRouter: configure provider key.', models: [] }) });
+        expect(await __testOnly_allowAndSaveWorkflowPoolEntry({ provider: 'openrouter', model: 'missing' })).toBe(false);
+        expect(__testOnly_getScopedModelState().enabledAlternatives).toEqual([]);
+        expect(document.getElementById('modelPoolStatusMessage').textContent).toContain('configure provider key');
+        expect(global.fetch.mock.calls.some(([, options]) => options?.method === 'POST')).toBe(false);
+    });
+
+    test('failed persistence retains the edit and exposes the real error with a retry button', async () => {
+        sessionStorage.setItem('von_current_user', JSON.stringify({ concept_id: '#V#user-a' }));
+        __testOnly_setModelScopeState({ resolvedLlm: { provider: 'ollama', model: 'primary', scope: 'user' } });
+        const added = { provider: 'openai', model: 'gpt-5.6-luna', model_parameters: { reasoning_effort: 'low' } };
+        global.fetch.mockImplementation(async (url) => url.includes('/models/inventory/')
+            ? { ok: true, json: async () => ({ available: true, models: [{ ...added, available: true }] }) }
+            : { ok: false, status: 500, json: async () => ({ message: 'Canonical persistence unavailable; retry.' }) });
+        expect(await __testOnly_allowAndSaveWorkflowPoolEntry(added)).toBe(false);
+        expect(document.getElementById('modelPoolStatusMessage').textContent).toContain('Canonical persistence unavailable');
+        expect(document.getElementById('addOpenAiToWorkflowPoolButton').disabled).toBe(false);
+        expect(document.getElementById('testOpenAiModelButton').disabled).toBe(true);
     });
 
     test('fetches and renders bounded actor-scoped cost evidence without inventing zero', async () => {
@@ -446,6 +495,8 @@ describe('settings model scope and workflow pool controls', () => {
             .rejects.toThrow('finish loading');
         expect(global.fetch).not.toHaveBeenCalled();
 
+        // Allow the asynchronous window-session header preparation to complete.
+        await new Promise(resolve => setTimeout(resolve, 0));
         pending.get('/api/settings/?user_concept_id=%23V%23user-b&organisation_concept_id=%23V%23org&model_scope=user')({
             ok: true,
             json: async () => ({

From 35b05527508077e005d20a8c6e6b66ba9475d713 Mon Sep 17 00:00:00 2001
From: witbrock <witbrock@gmail.com>
Date: Sat, 12 Sep 2026 21:10:21 -0400
Subject: [PATCH 2/2] Verify inventory discovery and Settings browser recovery
 on current main

---
 src/backend/server/routes/settings_routes.py  | 107 +++++++-----------
 .../von_interface/templates/settings_tab.html |   2 +-
 tests/backend/test_model_audio_metadata.py    |  13 +++
 tests/browser/modelEnablement.cjs             |  96 ++++++++++++++++
 4 files changed, 148 insertions(+), 70 deletions(-)
 create mode 100644 tests/browser/modelEnablement.cjs

diff --git a/src/backend/server/routes/settings_routes.py b/src/backend/server/routes/settings_routes.py
index 916f19bbe..0c9f10646 100644
--- a/src/backend/server/routes/settings_routes.py
+++ b/src/backend/server/routes/settings_routes.py
@@ -962,24 +962,18 @@ def get_llm_info():
 def execution_cost_display():
     """Read viewer-scoped thresholds or replace the authenticated user's pair."""
     from ...services.execution_cost_display_service import (
-        PREDICATE,
-        resolve_preferences,
-        validate_preferences,
+        PREDICATE, resolve_preferences, validate_preferences,
     )
 
     actor = get_effective_user_concept_id()
     if not actor:
         return _jsonify_no_store({"error": "authenticated_actor_context_required"}, 401)
     effective = get_effective_context(
-        request.headers.get("X-Von-Window-Session"),
-        dict(session),
-        actor,
+        request.headers.get("X-Von-Window-Session"), dict(session), actor,
     )
     if request.method == "PUT":
         import json
-        from ...services.ontology_mutation_command_service import (
-            execute_governed_ontology_method,
-        )
+        from ...services.ontology_mutation_command_service import execute_governed_ontology_method
         from ...services.text_value_service import upsert_singleton_text_relation
 
         try:
@@ -987,32 +981,19 @@ def execution_cost_display():
         except ValueError as exc:
             return _jsonify_no_store({"error": str(exc)}, 400)
         text = json.dumps(value, sort_keys=True)
-        provenance = {
-            "source": "execution_cost_display_preferences",
-            "actor_concept_id": actor,
-        }
+        provenance = {"source": "execution_cost_display_preferences", "actor_concept_id": actor}
         result = execute_governed_ontology_method(
             method_name="upsert_singleton_text_relation",
-            arguments={
-                "concept_id": actor,
-                "predicate": PREDICATE,
-                "text": text,
-                "language": "en-NZ",
-                "provenance": provenance,
-            },
+            arguments={"concept_id": actor, "predicate": PREDICATE, "text": text,
+                       "language": "en-NZ", "provenance": provenance},
             mutate=lambda: upsert_singleton_text_relation(
-                subject_concept_id=actor,
-                predicate=PREDICATE,
-                text=text,
-                lang="en-NZ",
-                provenance=provenance,
+                subject_concept_id=actor, predicate=PREDICATE, text=text,
+                lang="en-NZ", provenance=provenance,
             ),
         )
         if result.get("success") is False:
             return _jsonify_no_store(result, 503)
-    return _jsonify_no_store(
-        resolve_preferences(actor, effective.get("organisation_id"))
-    )
+    return _jsonify_no_store(resolve_preferences(actor, effective.get("organisation_id")))
 
 
 @settings_bp.route("/llm/cost_summary", methods=["GET"])
@@ -1033,7 +1014,9 @@ def get_llm_cost_summary():
             {"error": "An authenticated actor and namespace are required."}, 401
         )
     if not provider or not model:
-        return _jsonify_no_store({"error": "provider and model are required."}, 400)
+        return _jsonify_no_store(
+            {"error": "provider and model are required."}, 400
+        )
     try:
         limit = int(request.args.get("limit") or 200)
         window_days = int(request.args.get("window_days") or 30)
@@ -2066,7 +2049,11 @@ def _settings_key_presence(env_var_name: str) -> tuple[bool, str | None]:
         if read_secret_file(os.getenv(f"{name}_FILE")):
             return True, "secret_file"
 
-    dotenv_keys = tuple(key for name in candidates for key in (name, f"{name}_FILE"))
+    dotenv_keys = tuple(
+        key
+        for name in candidates
+        for key in (name, f"{name}_FILE")
+    )
     dotenv_values = read_repo_dotenv_values(dotenv_keys)
     for name in candidates:
         if dotenv_values.get(name):
@@ -2122,11 +2109,7 @@ def get_openai_models(*, include_all=False):
 
         # Create a direct OpenAIClient instance to get models
         client = OpenAIClient(api_key_env_var=api_key_env_var)
-        models = (
-            client.list_models(include_all=True)
-            if include_all
-            else client.list_models()
-        )
+        models = client.list_models(include_all=True) if include_all else client.list_models()
 
         if models:
             current_app.logger.info(f"Retrieved {len(models)} OpenAI models")
@@ -2432,14 +2415,10 @@ def first_val(v):
                 preferred_language = language_rows[0].get("text")
         # Legacy member_of_organisation describes membership, not preference.
         # Never choose the first membership as a landing context.
-        from ...services.login_organisation_preference_service import (
-            initialise_login_context,
-        )
-
+        from ...services.login_organisation_preference_service import initialise_login_context
         organisation_concept_id = (
             initialise_login_context(session)["organisation_concept_id"]
-            if session.get("user_email") and session.get("user_concept_id")
-            else None
+            if session.get("user_email") and session.get("user_concept_id") else None
         )
         return (
             jsonify(
@@ -2755,9 +2734,9 @@ def verify_gemini_api_key():
         )
     api_key = _resolve_settings_api_key(
         api_key_env_var,
-        fallback_env_vars=(
-            ("GOOGLE_API_KEY",) if api_key_env_var == "GEMINI_API_KEY" else ()
-        ),
+        fallback_env_vars=("GOOGLE_API_KEY",)
+        if api_key_env_var == "GEMINI_API_KEY"
+        else (),
     )
     if not api_key:
         return _jsonify_no_store(
@@ -2861,10 +2840,7 @@ def verify_openrouter_api_key():
     """Verify the fixed OpenRouter key source by listing its current catalogue."""
 
     payload = request.get_json(silent=True) or {}
-    if (
-        str(payload.get("api_key_env_var") or "OPENROUTER_API_KEY").strip()
-        != "OPENROUTER_API_KEY"
-    ):
+    if str(payload.get("api_key_env_var") or "OPENROUTER_API_KEY").strip() != "OPENROUTER_API_KEY":
         return _jsonify_no_store(
             {
                 "success": False,
@@ -4369,9 +4345,9 @@ def test_gemini_model():
 
     api_key = _resolve_settings_api_key(
         api_key_env_var,
-        fallback_env_vars=(
-            ("GOOGLE_API_KEY",) if api_key_env_var == "GEMINI_API_KEY" else ()
-        ),
+        fallback_env_vars=("GOOGLE_API_KEY",)
+        if api_key_env_var == "GEMINI_API_KEY"
+        else (),
     )
     if not api_key:
         return _jsonify_no_store(
@@ -4504,8 +4480,13 @@ def test_meta_model():
             400,
         )
 
-    actor_user_concept_id, actor_source = get_effective_user_concept_id_with_source()
-    if not actor_user_concept_id or actor_source == LEGACY_IDENTITY_HEADER_ACTOR_SOURCE:
+    actor_user_concept_id, actor_source = (
+        get_effective_user_concept_id_with_source()
+    )
+    if (
+        not actor_user_concept_id
+        or actor_source == LEGACY_IDENTITY_HEADER_ACTOR_SOURCE
+    ):
         return _jsonify_no_store(
             {
                 "success": False,
@@ -4736,25 +4717,13 @@ def test_openrouter_model():
         failure_kind = "unexpected_error"
         reason = "Unexpected OpenRouter model probe failure."
         if status_code == 401 or "auth" in lowered or "api key" in lowered:
-            failure_kind, reason = (
-                "authentication_error",
-                "OpenRouter authentication failed.",
-            )
+            failure_kind, reason = "authentication_error", "OpenRouter authentication failed."
         elif status_code == 403 or "permission" in lowered:
-            failure_kind, reason = (
-                "permission_denied",
-                "OpenRouter access was denied for this model.",
-            )
+            failure_kind, reason = "permission_denied", "OpenRouter access was denied for this model."
         elif status_code == 404 or "not found" in lowered:
-            failure_kind, reason = (
-                "model_not_found",
-                "The selected OpenRouter model was not found.",
-            )
+            failure_kind, reason = "model_not_found", "The selected OpenRouter model was not found."
         elif status_code == 429 or "rate limit" in lowered or "quota" in lowered:
-            failure_kind, reason = (
-                "rate_limited",
-                "OpenRouter rate or quota limit was reached.",
-            )
+            failure_kind, reason = "rate_limited", "OpenRouter rate or quota limit was reached."
         elif "connect" in lowered or "network" in lowered:
             failure_kind, reason = "connection_error", "Could not reach OpenRouter."
         current_app.logger.warning(
diff --git a/src/frontend/web/von_interface/templates/settings_tab.html b/src/frontend/web/von_interface/templates/settings_tab.html
index 28f75d76c..ae700ffa2 100644
--- a/src/frontend/web/von_interface/templates/settings_tab.html
+++ b/src/frontend/web/von_interface/templates/settings_tab.html
@@ -234,7 +234,7 @@ <h3 class="speech-settings-title">Text-to-speech</h3>
                     <section class="speech-settings-section" aria-label="Dictation settings">
                         <h3 class="speech-settings-title">Dictation</h3>
                         <label for="settingsTranscriptionModelSelect">Recorded transcription model (this browser)</label>
-                        <select id="settingsTranscriptionModelSelect" disabled></select>
+                        <select id="settingsTranscriptionModelSelect" class="speech-settings-select" disabled></select>
                         <p class="settings-note">Only enabled, available file-transcription models can be selected. A stale preference remains visible until you choose another model.</p>
                         <p class="speech-settings-note">Recorded audio uses conversation vocabulary. The browser controls below apply only to browser recognition.</p>
 
diff --git a/tests/backend/test_model_audio_metadata.py b/tests/backend/test_model_audio_metadata.py
index 115ee03a3..740120500 100644
--- a/tests/backend/test_model_audio_metadata.py
+++ b/tests/backend/test_model_audio_metadata.py
@@ -96,3 +96,16 @@ def test_disabled_provider_has_actionable_failure_and_no_available_models(monkey
     assert "openai" in data["error"] and "check" in data["error"]
     assert "private provider details" not in data["error"]
     assert not any(model["available"] for model in data["models"])
+
+
+def test_openai_inventory_includes_non_gpt_ids_without_changing_chat_catalogue(monkeypatch):
+    from types import SimpleNamespace
+    from src.backend.languagemodels import llm_interface
+    monkeypatch.setattr(llm_interface, '_MODEL_CACHE', {})
+    monkeypatch.setattr(llm_interface, '_MODEL_FAIL_BACKOFF', {})
+    client = object.__new__(llm_interface.OpenAIClient)
+    client.client = SimpleNamespace(models=SimpleNamespace(list=lambda: SimpleNamespace(data=[
+        SimpleNamespace(id='gpt-chat'), SimpleNamespace(id='whisper-1'), SimpleNamespace(id='embedding-model')
+    ])))
+    assert client.list_models(include_all=True) == ['embedding-model', 'gpt-chat', 'whisper-1']
+    assert client.list_models() == ['gpt-chat']
diff --git a/tests/browser/modelEnablement.cjs b/tests/browser/modelEnablement.cjs
new file mode 100644
index 000000000..a2dfec234
--- /dev/null
+++ b/tests/browser/modelEnablement.cjs
@@ -0,0 +1,96 @@
+const fs = require('fs');
+const http = require('http');
+const path = require('path');
+// Actual Settings markup and modules against isolated synthetic HTTP state.
+// Run: PLAYWRIGHT_BROWSERS_PATH=/tmp/von-playwright node tests/browser/modelEnablement.cjs
+const root = process.cwd();
+fs.mkdirSync('/tmp/von-model-browser', { recursive: true });
+const { chromium } = require(path.join(root, 'node_modules/playwright'));
+let stored = { resolved_llm: { provider: 'ollama', model: 'local-model', host: 'http://localhost:11434', scope: 'user' }, enabled_llms: [{ provider: 'ollama', model: 'local-model', host: 'http://localhost:11434' }] };
+const audio = { provider: 'openai', model: 'gpt-4o-transcribe', available: true, audio_transcription: true };
+let saveCount = 0;
+let failSave = false;
+const server = http.createServer(async (req, res) => {
+  const url = new URL(req.url, 'http://localhost');
+  if (url.pathname === '/') {
+    let html = fs.readFileSync(path.join(root, 'src/frontend/web/von_interface/templates/settings_tab.html'), 'utf8');
+    html = html.replace(/\{\{ url_for\('static', filename='([^']+)'\) \}\}/g, '/static/$1');
+    html = html.replace(/<script[\s\S]*?<\/script>/g, '');
+    res.setHeader('Content-Type', 'text/html'); return res.end(html);
+  }
+  if (url.pathname.startsWith('/static/')) {
+    const file = path.join(root, 'src/frontend/web/von_interface', url.pathname);
+    res.setHeader('Content-Type', file.endsWith('.js') ? 'text/javascript' : file.endsWith('.css') ? 'text/css' : 'text/plain');
+    if (fs.existsSync(file)) return res.end(fs.readFileSync(file));
+    res.statusCode = 404; return res.end();
+  }
+  res.setHeader('Content-Type', 'application/json');
+  if (url.pathname.includes('/models/inventory/')) {
+    const provider = url.pathname.split('/').pop();
+    const models = provider === 'openai' ? [audio, { provider, model: 'general-chat', available: true }] : provider === 'ollama' ? [{ ...stored.resolved_llm, available: true }] : provider === 'openrouter' ? [{ provider, model: 'vendor/chat', available: true }] : [];
+    return res.end(JSON.stringify({ provider, available: true, models }));
+  }
+  if (url.pathname === '/api/settings/') {
+    if (req.method === 'POST') {
+      saveCount++;
+      let body = ''; for await (const chunk of req) body += chunk;
+      if (failSave) { res.statusCode = 500; return res.end(JSON.stringify({ message: 'Fixture persistence temporarily unavailable; retry.' })); }
+      const data = JSON.parse(body);
+      stored = { resolved_llm: data.active_llm, enabled_llms: data.enabled_llms };
+    }
+    return res.end(JSON.stringify(stored));
+  }
+  return res.end('{}');
+});
+(async () => {
+  await new Promise(resolve => server.listen(0, '127.0.0.1', resolve));
+  const url = `http://127.0.0.1:${server.address().port}`;
+  const browser = await chromium.launch({ headless: true });
+  const page = await browser.newPage({ viewport: { width: 1280, height: 1000 } });
+  const errors = []; page.on('pageerror', error => errors.push(error.message));
+  await page.goto(url);
+  await page.evaluate(async (stored) => {
+    sessionStorage.setItem('von_current_user', JSON.stringify({ concept_id: '#V#fixture-user' }));
+    const mod = await import('/static/js/settingsPage.js');
+    window.settingsFixture = mod;
+    mod.__testOnly_setModelScopeState({ resolvedLlm: stored.resolved_llm, enabledLlms: stored.enabled_llms, actorReady: true });
+    mod.__testOnly_initialiseSettingsConcernNavigation();
+    mod.__testOnly_renderModelScopeOverview();
+    await mod.__testOnly_refreshAdditionalModelInventory();
+  }, stored);
+  await page.locator('[data-settings-concern-tab="models"]').click();
+  await page.locator('summary').filter({ hasText: 'Additional models' }).click();
+  const row = page.locator('#additionalModelInventory .settings-model-pool-entry').filter({ hasText: 'openai: gpt-4o-transcribe' });
+  failSave = true;
+  await row.locator('button').click();
+  await page.waitForFunction(() => document.getElementById('modelPoolStatusMessage').textContent.includes('temporarily unavailable'));
+  failSave = false;
+  await row.locator('button').click();
+  await page.waitForFunction(() => document.getElementById('modelPoolStatusMessage').textContent.includes('saved for the current scope'));
+  if (saveCount !== 2 || stored.enabled_llms.length !== 2 || stored.resolved_llm.model !== 'local-model') throw Error('Save/recovery or preservation failed');
+  await page.locator('#reloadModelPoolButton').click();
+  await page.waitForFunction(() => document.getElementById('modelPoolStatusMessage').textContent.includes('Model settings loaded'));
+  const select = page.locator('#settingsTranscriptionModelSelect');
+  await page.locator('[data-settings-concern-tab="conversations"]').click();
+  await select.selectOption('gpt-4o-transcribe');
+  await select.scrollIntoViewIfNeeded();
+  await page.screenshot({ path: '/tmp/von-model-browser/transcription-desktop.png' });
+  const result = await page.evaluate(() => ({
+    selected: localStorage.getItem('chatRecordedTranscriptionModel'),
+    options: [...document.querySelector('#settingsTranscriptionModelSelect').options].map(o => ({ value: o.value, disabled: o.disabled })),
+    inventory: document.getElementById('additionalModelInventory').textContent,
+    overflow: document.documentElement.scrollWidth > innerWidth,
+  }));
+  if (result.options.some(o => o.value === 'general-chat') || result.selected !== 'gpt-4o-transcribe') throw Error('Filtering/selection failed');
+  await page.locator('[data-settings-concern-tab="models"]').click();
+  await page.screenshot({ path: '/tmp/von-model-browser/desktop.png', fullPage: true });
+  await page.setViewportSize({ width: 390, height: 844 });
+  await page.screenshot({ path: '/tmp/von-model-browser/mobile.png', fullPage: true });
+  await page.locator('[data-settings-concern-tab="conversations"]').click();
+  await select.scrollIntoViewIfNeeded();
+  await page.screenshot({ path: '/tmp/von-model-browser/transcription-mobile.png' });
+  result.mobileOverflow = await page.evaluate(() => document.documentElement.scrollWidth > innerWidth);
+  fs.writeFileSync('/tmp/von-model-browser/evidence.json', JSON.stringify({ url, saveCount, result, errors, fixture: 'Actual Settings template/modules; synthetic discovery and canonical settings HTTP responses; no provider generation or live data.' }, null, 2));
+  console.log(JSON.stringify({ saveCount, selected: result.selected, options: result.options, overflow: result.overflow, mobileOverflow: result.mobileOverflow, errors }));
+  await browser.close(); server.close();
+})().catch(error => { console.error(error); server.close(); process.exit(1); });