Skip to content

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.

fix(reasoning): resolve empty output for Qwen3/GPT-OSS and custom endpoint misrouting

Gabriel Stein18d800f4b8e2
diff --git a/src/config/constants.ts b/src/config/constants.ts
index 8af7a47b7..710b58ca1 100644
--- a/src/config/constants.ts
+++ b/src/config/constants.ts
@@ -88,7 +88,7 @@ export const MODEL_CONSTRAINTS = {
// Token Limits
export const TOKEN_LIMITS = {
- MIN_TOKENS: 100,
+ MIN_TOKENS: 512,
MAX_TOKENS: 2048,
MIN_TOKENS_ANTHROPIC: 100,
MAX_TOKENS_ANTHROPIC: 4096,
diff --git a/src/models/ModelRegistry.ts b/src/models/ModelRegistry.ts
index 967d78269..724a4e449 100644
--- a/src/models/ModelRegistry.ts
+++ b/src/models/ModelRegistry.ts
@@ -1,5 +1,5 @@
import modelDataRaw from "./modelRegistryData.json";
-import { isCloudReasoningMode } from "../stores/settingsStore";
+import { isCloudReasoningMode, getSettings } from "../stores/settingsStore";
export interface ModelDefinition {
id: string;
@@ -259,6 +259,7 @@ export function getModelProvider(modelId: string): string {
const model = getAllReasoningModels().find((m) => m.value === modelId);
if (!model) {
+ if (getSettings().reasoningProvider === "custom") return "custom";
if (modelId.includes("claude")) return "anthropic";
if (modelId.includes("gemini") && !modelId.includes("gemma")) return "gemini";
if ((modelId.includes("gpt-4") || modelId.includes("gpt-5")) && !modelId.includes("gpt-oss"))
diff --git a/src/services/ReasoningService.ts b/src/services/ReasoningService.ts
index 4e9c53d3a..2fd68e13c 100644
--- a/src/services/ReasoningService.ts
+++ b/src/services/ReasoningService.ts
@@ -441,6 +441,9 @@ class ReasoningService extends BaseReasoningService {
case "openwhispr":
result = await this.processWithOpenWhispr(text, model, agentName, config);
break;
+ case "custom":
+ result = await this.processWithOpenAI(text, trimmedModel, agentName, config);
+ break;
default:
throw new Error(`Unsupported reasoning provider: ${provider}`);
}
diff --git a/src/services/localReasoningBridge.js b/src/services/localReasoningBridge.js
index 12b19d624..84a5bdf51 100644
--- a/src/services/localReasoningBridge.js
+++ b/src/services/localReasoningBridge.js
@@ -48,17 +48,21 @@ class LocalReasoningService {
});
const result = await modelManager.runInference(modelId, text, inferenceConfig);
+ const cleanResult = result
+ .replace(/<think>[\s\S]*?<\/think>/g, "")
+ .replace(/<think>[\s\S]*$/, "")
+ .trim();
const processingTime = Date.now() - startTime;
debugLogger.logReasoning("LOCAL_BRIDGE_SUCCESS", {
modelId,
processingTimeMs: processingTime,
- resultLength: result.length,
- resultPreview: result.substring(0, 100) + (result.length > 100 ? "..." : ""),
+ resultLength: cleanResult.length,
+ resultPreview: cleanResult.substring(0, 100) + (cleanResult.length > 100 ? "..." : ""),
});
- return result;
+ return cleanResult;
} catch (error) {
const processingTime = Date.now() - startTime;
@@ -75,7 +79,7 @@ class LocalReasoningService {
}
}
- calculateMaxTokens(textLength, minTokens = 100, maxTokens = 2048, multiplier = 2) {
+ calculateMaxTokens(textLength, minTokens = 512, maxTokens = 2048, multiplier = 2) {
return Math.max(minTokens, Math.min(textLength * multiplier, maxTokens));
}
}