Skip to content

The open-source AI voice studio. Clone, dictate, create.

Address CodeRabbit findings on the custom-endpoint PR

Moroya Sakamotobbe518dd5612
diff --git a/app/src/components/ServerTab/CapturesPage.tsx b/app/src/components/ServerTab/CapturesPage.tsx
index b26e3649..b56ca30d 100644
--- a/app/src/components/ServerTab/CapturesPage.tsx
+++ b/app/src/components/ServerTab/CapturesPage.tsx
@@ -143,7 +143,10 @@ export function CapturesPage() {
const toggleToTalkKeys = settings?.chord_toggle_to_talk_keys ?? defaultChordKeys('toggle');
const customLlmEndpointSaved = settings?.custom_llm_endpoint ?? '';
const customLlmModelSaved = settings?.custom_llm_model ?? '';
- const customLlmApiKeySaved = settings?.custom_llm_api_key ?? '';
+ // The server no longer returns the API key value; it reports whether one
+ // is stored via ``custom_llm_api_key_configured``. Anything sitting in
+ // React state below is a freshly-typed key that hasn't been sent yet.
+ const customLlmApiKeyConfigured = Boolean(settings?.custom_llm_api_key_configured);
const customLlmActive = Boolean(customLlmEndpointSaved && customLlmModelSaved);
const [chordEditor, setChordEditor] = useState<'push' | 'toggle' | null>(null);
@@ -153,7 +156,10 @@ export function CapturesPage() {
// re-render on every keystroke while the user types out a URL or token.
const [customLlmEndpointDraft, setCustomLlmEndpointDraft] = useState(customLlmEndpointSaved);
const [customLlmModelDraft, setCustomLlmModelDraft] = useState(customLlmModelSaved);
- const [customLlmApiKeyDraft, setCustomLlmApiKeyDraft] = useState(customLlmApiKeySaved);
+ // Write-only field: the input is blank on load even when a key is stored,
+ // and clears itself after a successful save so the raw value never lingers
+ // in React state where a browser cache or dev-tools inspection could grab it.
+ const [customLlmApiKeyDraft, setCustomLlmApiKeyDraft] = useState('');
useEffect(() => {
setCustomLlmEndpointDraft(customLlmEndpointSaved);
@@ -161,9 +167,6 @@ export function CapturesPage() {
useEffect(() => {
setCustomLlmModelDraft(customLlmModelSaved);
}, [customLlmModelSaved]);
- useEffect(() => {
- setCustomLlmApiKeyDraft(customLlmApiKeySaved);
- }, [customLlmApiKeySaved]);
const commitCustomLlmEndpoint = useCallback(() => {
const next = customLlmEndpointDraft.trim();
@@ -178,14 +181,18 @@ export function CapturesPage() {
}
}, [customLlmModelDraft, customLlmModelSaved, update]);
const commitCustomLlmApiKey = useCallback(() => {
- // The API key can legitimately contain leading/trailing punctuation for
- // some providers, so we only trim when the whole field is whitespace.
const trimmed = customLlmApiKeyDraft.trim();
- const next = trimmed ? customLlmApiKeyDraft : '';
- if (next !== customLlmApiKeySaved) {
- update({ custom_llm_api_key: next || null });
- }
- }, [customLlmApiKeyDraft, customLlmApiKeySaved, update]);
+ if (!trimmed) return;
+ // Fire-and-forget the key, then wipe it from state so nothing else in
+ // the tree can read it back. The server flag flips to `configured=true`
+ // on the next settings query.
+ update({ custom_llm_api_key: customLlmApiKeyDraft });
+ setCustomLlmApiKeyDraft('');
+ }, [customLlmApiKeyDraft, update]);
+ const clearCustomLlmApiKey = useCallback(() => {
+ setCustomLlmApiKeyDraft('');
+ update({ custom_llm_api_key: null });
+ }, [update]);
useEffect(() => {
fetch(`${serverUrl}/health/filesystem`)
@@ -501,6 +508,9 @@ export function CapturesPage() {
description={t('settings.captures.refinement.customEndpoint.description')}
>
<div className="space-y-2 max-w-lg">
+ <label htmlFor="customLlmEndpoint" className="block text-xs font-medium text-muted-foreground">
+ {t('settings.captures.refinement.customEndpoint.endpointLabel')}
+ </label>
<Input
id="customLlmEndpoint"
type="url"
@@ -511,7 +521,11 @@ export function CapturesPage() {
disabled={!autoRefine}
autoComplete="off"
spellCheck={false}
+ aria-label={t('settings.captures.refinement.customEndpoint.endpointLabel')}
/>
+ <label htmlFor="customLlmModel" className="block text-xs font-medium text-muted-foreground pt-1">
+ {t('settings.captures.refinement.customEndpoint.modelLabel')}
+ </label>
<Input
id="customLlmModel"
type="text"
@@ -522,18 +536,45 @@ export function CapturesPage() {
disabled={!autoRefine || !customLlmEndpointDraft}
autoComplete="off"
spellCheck={false}
+ aria-label={t('settings.captures.refinement.customEndpoint.modelLabel')}
/>
- <Input
- id="customLlmApiKey"
- type="password"
- placeholder={t('settings.captures.refinement.customEndpoint.apiKeyPlaceholder')}
- value={customLlmApiKeyDraft}
- onChange={(e) => setCustomLlmApiKeyDraft(e.target.value)}
- onBlur={commitCustomLlmApiKey}
- disabled={!autoRefine || !customLlmEndpointDraft}
- autoComplete="off"
- spellCheck={false}
- />
+ <label htmlFor="customLlmApiKey" className="block text-xs font-medium text-muted-foreground pt-1">
+ {t('settings.captures.refinement.customEndpoint.apiKeyLabel')}
+ {customLlmApiKeyConfigured ? (
+ <span className="ml-2 text-[10px] uppercase tracking-wide text-emerald-600">
+ {t('settings.captures.refinement.customEndpoint.apiKeyConfigured')}
+ </span>
+ ) : null}
+ </label>
+ <div className="flex gap-2">
+ <Input
+ id="customLlmApiKey"
+ type="password"
+ placeholder={
+ customLlmApiKeyConfigured
+ ? t('settings.captures.refinement.customEndpoint.apiKeyPlaceholderReplace')
+ : t('settings.captures.refinement.customEndpoint.apiKeyPlaceholder')
+ }
+ value={customLlmApiKeyDraft}
+ onChange={(e) => setCustomLlmApiKeyDraft(e.target.value)}
+ onBlur={commitCustomLlmApiKey}
+ disabled={!autoRefine || !customLlmEndpointDraft}
+ autoComplete="new-password"
+ spellCheck={false}
+ aria-label={t('settings.captures.refinement.customEndpoint.apiKeyLabel')}
+ className="flex-1"
+ />
+ {customLlmApiKeyConfigured && (
+ <button
+ type="button"
+ onClick={clearCustomLlmApiKey}
+ className="text-xs text-muted-foreground hover:text-destructive px-2 py-1 border border-border rounded-md"
+ aria-label={t('settings.captures.refinement.customEndpoint.apiKeyClear')}
+ >
+ {t('settings.captures.refinement.customEndpoint.apiKeyClear')}
+ </button>
+ )}
+ </div>
<p className="text-xs text-muted-foreground">
{customLlmActive
? t('settings.captures.refinement.customEndpoint.statusRemote', {
@@ -639,8 +680,14 @@ export function CapturesPage() {
<li className="flex gap-2.5">
<Lock className="h-4 w-4 shrink-0 mt-0.5 text-accent" />
<span className="leading-relaxed">
- <span className="text-foreground font-medium">{t('settings.captures.sidebar.local.title')}</span>{' '}
- {t('settings.captures.sidebar.local.body')}
+ <span className="text-foreground font-medium">
+ {customLlmActive
+ ? t('settings.captures.sidebar.remote.title')
+ : t('settings.captures.sidebar.local.title')}
+ </span>{' '}
+ {customLlmActive
+ ? t('settings.captures.sidebar.remote.body', { endpoint: customLlmEndpointSaved })
+ : t('settings.captures.sidebar.local.body')}
</span>
</li>
<li className="flex gap-2.5">
diff --git a/app/src/i18n/locales/en/translation.json b/app/src/i18n/locales/en/translation.json
index 3c0bab96..f8973fda 100644
--- a/app/src/i18n/locales/en/translation.json
+++ b/app/src/i18n/locales/en/translation.json
@@ -986,11 +986,17 @@
"customEndpoint": {
"title": "Custom LLM endpoint (advanced)",
"description": "Delegate refinement and personality rewrites to any OpenAI-compatible server — llama.cpp, vLLM, LM Studio, LocalAI, Ollama's shim, or the OpenAI API itself. Leave blank to use the built-in Qwen3.",
+ "endpointLabel": "Endpoint URL",
"endpointPlaceholder": "http://localhost:8090/v1",
+ "modelLabel": "Model name",
"modelPlaceholder": "Model name (e.g. qwen3-4b)",
+ "apiKeyLabel": "API key",
"apiKeyPlaceholder": "API key (optional)",
+ "apiKeyPlaceholderReplace": "Enter a new key to replace the stored one",
+ "apiKeyConfigured": "Stored",
+ "apiKeyClear": "Clear",
"statusBuiltin": "Using the built-in Qwen3 model selected above.",
- "statusRemote": "Sending every refinement to {{endpoint}}."
+ "statusRemote": "Sending every refinement and personality rewrite to {{endpoint}}."
}
},
"playback": {
@@ -1029,6 +1035,10 @@
"title": "Fully local.",
"body": "Whisper and the refinement LLM run on your hardware. No cloud, no accounts, your voice never leaves the machine."
},
+ "remote": {
+ "title": "Refinement runs remotely.",
+ "body": "Whisper still transcribes on your hardware, but refined transcripts and personality rewrites are sent to {{endpoint}}. Voice audio itself never leaves the machine."
+ },
"playAs": {
"title": "Play as any voice.",
"body": "Transcripts can be read back in any profile you've cloned."
diff --git a/app/src/i18n/locales/ja/translation.json b/app/src/i18n/locales/ja/translation.json
index 3171d585..f6640e86 100644
--- a/app/src/i18n/locales/ja/translation.json
+++ b/app/src/i18n/locales/ja/translation.json
@@ -981,11 +981,17 @@
"customEndpoint": {
"title": "カスタム LLM エンドポイント (上級)",
"description": "リファインとパーソナリティの書き換えを、任意の OpenAI 互換サーバー (llama.cpp、vLLM、LM Studio、LocalAI、Ollama の互換 API、OpenAI API 自体など) に委譲します。空欄のままなら内蔵の Qwen3 を使います。",
+ "endpointLabel": "エンドポイント URL",
"endpointPlaceholder": "http://localhost:8090/v1",
+ "modelLabel": "モデル名",
"modelPlaceholder": "モデル名 (例: qwen3-4b)",
+ "apiKeyLabel": "API キー",
"apiKeyPlaceholder": "API キー (任意)",
+ "apiKeyPlaceholderReplace": "上書きする新しいキーを入力",
+ "apiKeyConfigured": "保存済",
+ "apiKeyClear": "クリア",
"statusBuiltin": "上で選んだ内蔵 Qwen3 モデルを使用しています",
- "statusRemote": "リファインを {{endpoint}} に送信しています"
+ "statusRemote": "リファインとパーソナリティの書き換えを {{endpoint}} に送信しています"
}
},
"playback": {
@@ -1024,6 +1030,10 @@
"title": "完全にローカル。",
"body": "Whisper と整形用 LLM はあなたのハードウェア上で動作します。クラウドもアカウントも不要で、声がマシンの外に出ることはありません。"
},
+ "remote": {
+ "title": "整形処理はリモート実行中。",
+ "body": "Whisper はハードウェア上で書き起こしを続けますが、整形された文字起こしとパーソナリティ書き換えは {{endpoint}} に送られます。音声そのものはマシンから出ません。"
+ },
"playAs": {
"title": "どのボイスでも再生。",
"body": "クローンしたどのプロファイルでも文字起こしを読み上げできます。"
diff --git a/app/src/lib/api/types.ts b/app/src/lib/api/types.ts
index 0357a6a5..ba2d7df7 100644
--- a/app/src/lib/api/types.ts
+++ b/app/src/lib/api/types.ts
@@ -225,10 +225,22 @@ export interface CaptureSettings {
*/
custom_llm_endpoint: string | null;
custom_llm_model: string | null;
- custom_llm_api_key: string | null;
+ /**
+ * Whether a custom LLM API key is currently stored on the server. The raw
+ * key value never rides the response — this flag replaces it — so the
+ * settings UI can show a "Configured" indicator without letting the
+ * frontend rehydrate the secret into state or leak it to a browser cache.
+ * Writes still go through ``CaptureSettingsUpdate.custom_llm_api_key``.
+ */
+ custom_llm_api_key_configured: boolean;
}
-export type CaptureSettingsUpdate = Partial<CaptureSettings>;
+export type CaptureSettingsUpdate = Partial<
+ Omit<CaptureSettings, 'custom_llm_api_key_configured'>
+> & {
+ /** Write-only: setting this to a non-empty string stores it, ``null`` clears it. */
+ custom_llm_api_key?: string | null;
+};
/**
* One row in the dictation readiness checklist. ``model_name`` is the
diff --git a/backend/backends/openai_compat_backend.py b/backend/backends/openai_compat_backend.py
index 75f12045..b60cd457 100644
--- a/backend/backends/openai_compat_backend.py
+++ b/backend/backends/openai_compat_backend.py
@@ -185,4 +185,13 @@ def _extract_content(data: dict, url: str) -> str:
content = first.get("text")
if content is None:
raise ValueError(f"No content in response from {url}: {data!r}")
+ if not isinstance(content, str):
+ # Vision-capable and tool-use servers sometimes return the content as
+ # an array of typed parts instead of a bare string. Refinement /
+ # personality are text-only paths, so surface a clear error instead
+ # of tripping ``AttributeError`` on ``.strip()`` and burying the
+ # shape mismatch.
+ raise ValueError(
+ f"Unexpected content type {type(content).__name__} in response from {url}: {data!r}"
+ )
return content.strip()
diff --git a/backend/models.py b/backend/models.py
index fd7bbbc4..dad76967 100644
--- a/backend/models.py
+++ b/backend/models.py
@@ -2,8 +2,8 @@
Pydantic models for request/response validation.
"""
-from pydantic import BaseModel, Field
-from typing import Optional, List
+from pydantic import BaseModel, Field, model_validator
+from typing import Any, Optional, List
from datetime import datetime
from .utils.capture_chords import (
@@ -266,7 +266,36 @@ class CaptureSettingsResponse(BaseModel):
)
custom_llm_endpoint: Optional[str] = None
custom_llm_model: Optional[str] = None
- custom_llm_api_key: Optional[str] = None
+ # Provider credential — write-only. The response reports whether one is
+ # configured but never echoes the value, so the settings API can't be
+ # used to exfiltrate stored keys and the frontend can't rehydrate them
+ # into React state where a browser cache could pick them up.
+ custom_llm_api_key_configured: bool = False
+
+ @model_validator(mode="before")
+ @classmethod
+ def _mask_custom_llm_api_key(cls, data: Any) -> Any:
+ """Replace the raw ``custom_llm_api_key`` with a boolean ``_configured`` flag.
+
+ Runs before field validation so the stored credential is never
+ materialised on a Response instance, even in memory. Accepts both
+ SQLAlchemy ORM rows (``from_attributes=True`` path) and plain dicts
+ so tests that construct the model directly get the same treatment.
+ """
+ if isinstance(data, dict):
+ raw_key = data.pop("custom_llm_api_key", None)
+ data.setdefault("custom_llm_api_key_configured", bool(raw_key))
+ return data
+ # ORM row — build a snapshot dict, drop the secret, and hand the
+ # rest to Pydantic. ``from_attributes`` won't reach the raw
+ # attribute anymore because the returned dict wins.
+ columns = getattr(getattr(data, "__table__", None), "columns", None)
+ if columns is None:
+ return data
+ snapshot = {c.name: getattr(data, c.name, None) for c in columns}
+ raw_key = snapshot.pop("custom_llm_api_key", None)
+ snapshot["custom_llm_api_key_configured"] = bool(raw_key)
+ return snapshot
class Config:
from_attributes = True
diff --git a/backend/tests/test_openai_compat_integration.py b/backend/tests/test_openai_compat_integration.py
index d5ebbc58..f39fc600 100644
--- a/backend/tests/test_openai_compat_integration.py
+++ b/backend/tests/test_openai_compat_integration.py
@@ -110,26 +110,40 @@ async def test_backend_generates_non_empty_response():
async def test_backend_respects_examples_history():
- """Few-shot examples reach the endpoint as proper user/assistant turns."""
+ """Few-shot examples reach the endpoint as proper user/assistant turns.
+
+ Uses an arbitrary mapping (``ping → PONG``, ``foo → BAR``) so a passing
+ assertion actually proves the examples were serialised into the
+ request. A non-empty translation would still pass if ``examples``
+ were silently dropped; the model has to have seen the pattern to
+ return the mapped token on a novel input.
+ """
backend = OpenAICompatLLMBackend(
endpoint=_endpoint(),
model=_model(),
api_key=_api_key(),
)
reply = await backend.generate(
- prompt="hola",
+ prompt="baz",
system=(
- "You are a translation bot. Reply with the single-word English "
- "translation of the user's message and nothing else."
+ "You are a lookup function. Respond with exactly one uppercase word "
+ "matching the mapping shown in the examples and nothing else."
),
- max_tokens=16,
+ max_tokens=8,
temperature=0.0,
examples=[
- ("bonjour", "hello"),
- ("guten tag", "hello"),
+ ("ping", "PONG"),
+ ("foo", "BAR"),
],
)
- assert reply.strip(), "Endpoint returned an empty translation"
+ stripped = reply.strip().upper()
+ # The model may add punctuation; check that the mapped-shape token
+ # ("QUX" is the natural next entry after PONG/BAR) leads the reply.
+ assert stripped, "Endpoint returned an empty response to a few-shot prompt"
+ assert stripped.startswith(("QUX", "BAZ")), (
+ f"Reply {reply!r} doesn't look like it followed the example mapping — "
+ "the ``examples`` field may have been dropped in transit."
+ )
async def test_dispatch_returns_openai_compat_when_config_set():