update groq model to GPT-OSS 120B and fixed missing subtitles

app/src/main/kotlin/com/arflix/tv/data/repository/StreamRepository.kt

@@ -2978,25 +2978,30 @@ class StreamRepository @Inject constructor(
         onPendingAddons: ((List<String>) -> Unit)? = null
     ): List<Subtitle> = withContext(Dispatchers.IO) {
         val allAddons = installedAddons.first()
-        // Include:
-        // - Addons classified as AddonType.SUBTITLE (OpenSubtitles and, going forward,
-        //   any user-added pure-subtitle addon like Wizdom/Ktuvit now that addCustomAddon
-        //   classifies them correctly).
-        // - Addons classified as CUSTOM whose manifest declares a `subtitles` resource.
-        //   This covers two cases: (a) addons installed before the classification fix
-        //   landed, which are still stored as CUSTOM; (b) hybrid addons that provide
-        //   both streams and subtitles. Both should be queried for subtitles.
+        // Selection is by CAPABILITY, not by AddonType. The old gate accepted only SUBTITLE, or
+        // CUSTOM-with-a-`subtitles`-manifest, and silently dropped everything else — which made
+        // whole addons invisible depending on *how they were installed*:
+        //   - the web installer stamps type=COMMUNITY on anything that isn't subtitle-only
+        //     (web/lib/addons.ts), and COMMUNITY was never queried at all;
+        //   - a cloud/legacy payload with no cached manifest parses back as CUSTOM + manifest=null
+        //     (parseAddons defaults the type), which failed the manifest check.
+        // Either way the symptom is identical and confusing: the addon is installed and enabled,
+        // shows in the addon list, and yet contributes zero subtitles — while OpenSubtitles
+        // (hardcoded SUBTITLE) keeps working, so it looks like "only OpenSubtitles loads".
         // Fixes issue #80.
+        val speculativeAddonIds = mutableSetOf<String>()
         val subtitleAddons = allAddons.filter { addon ->
             if (!addon.isInstalled || !addon.isEnabled) return@filter false
             if (addon.type == AddonType.SUBTITLE) return@filter true
-            if (addon.type == AddonType.CUSTOM) {
-                val declaresSubtitles = addon.manifest?.resources?.any { res ->
-                    res.name.equals("subtitles", ignoreCase = true)
-                } == true
-                return@filter declaresSubtitles
+            val declared = addon.manifest?.resources.orEmpty()
+            if (declared.isNotEmpty()) {
+                return@filter declared.any { res -> res.name.equals("subtitles", ignoreCase = true) }
             }
-            false
+            // No manifest cached — capability unknown. Ask anyway (a stream-only addon just 404s
+            // or returns an empty list) rather than dropping a possibly-working subtitle provider.
+            // These are marked speculative so they don't pay the cold-start retry below.
+            speculativeAddonIds += addon.id
+            addon.type == AddonType.CUSTOM || addon.type == AddonType.COMMUNITY
         }
 
         val videoHash = stream?.behaviorHints?.videoHash?.trim().takeUnless { it.isNullOrBlank() }
@@ -3052,8 +3057,10 @@ class StreamRepository @Inject constructor(
 
                 // Aggregator endpoints (AIOStreams) often hang or 502 on the first (cold) hit and
                 // succeed once their upstream fan-out is warm — one retry recovers those.
+                // Speculative addons (queried without a manifest, capability unknown) are expected
+                // to come back empty, so they don't get to spend another 1.5s proving it.
                 var subs = attempt()
-                if (subs.isEmpty()) {
+                if (subs.isEmpty() && addon.id !in speculativeAddonIds) {
                     delay(1_500)
                     subs = attempt()
                 }

app/src/main/kotlin/com/arflix/tv/ui/screens/player/SubtitleAiModel.kt

@@ -1,6 +1,10 @@
 package com.arflix.tv.ui.screens.player
 
+// Both names are historical: the enum value is persisted in DataStore and cloud backups, so
+// renaming either one breaks restore on other devices. They mean "the Groq model" and "the Gemini
+// model" — the actual model ids live in SubtitleTranslationService.
 enum class SubtitleAiModel {
+    // Now maps to openai/gpt-oss-120b (llama-3.3-70b decommissioned August 2026).
     GROQ_LLAMA_70B,
     // Now maps to gemini-3.5-flash-lite (2.5 retired July 2026). Name kept: the enum value is
     // persisted in DataStore and cloud backups — renaming breaks restore on other devices.

app/src/main/kotlin/com/arflix/tv/ui/screens/player/SubtitleTranslationService.kt

@@ -32,8 +32,17 @@ private val BLOCKED_FINISH_REASONS = setOf(
     "PROHIBITED_CONTENT", "SAFETY", "RECITATION", "BLOCKLIST", "SPII"
 )
 
-private const val GROQ_MODEL_ID = "llama-3.3-70b-versatile"
+// llama-3.3-70b-versatile was decommissioned by Groq (August 16, 2026). gpt-oss-120b is Groq's
+// recommended replacement: ~500 tok/s (vs 280), same 30 RPM free tier, 2x the daily token budget.
+// It is a reasoning model — see GROQ_REASONING_EFFORT below.
+private const val GROQ_MODEL_ID = "openai/gpt-oss-120b"
 private const val GROQ_URL = "https://api.groq.com/openai/v1/chat/completions"
+// gpt-oss cannot disable reasoning entirely (unlike Gemini's thinkingLevel=minimal or Qwen's
+// reasoning_effort=none) — "low" is the floor. Anything higher spends seconds of thinking on what
+// is a mechanical translation, which subtitles can't wait for. "hidden" keeps the chain-of-thought
+// out of message.content so extractJsonArray() doesn't have to parse around it.
+private const val GROQ_REASONING_EFFORT = "low"
+private const val GROQ_REASONING_FORMAT = "hidden"
 // gemini-2.5-flash was retired by Google (HTTP 404 "no longer available", July 2026).
 // gemini-3.5-flash-lite: ~2x throughput (350 vs 165 tok/s) and ~3.5x cheaper than 3.5-flash,
 // positioned by Google for high-volume translation. Same v1beta API + thinkingLevel field.
@@ -172,6 +181,8 @@ class SubtitleTranslationService(
         val body = JSONObject().apply {
             put("model", GROQ_MODEL_ID)
             put("temperature", 0.1)
+            put("reasoning_effort", GROQ_REASONING_EFFORT)
+            put("reasoning_format", GROQ_REASONING_FORMAT)
             put("messages", JSONArray().apply {
                 put(JSONObject().apply {
                     put("role", "system")

app/src/main/kotlin/com/arflix/tv/ui/screens/settings/SettingsScreen.kt

@@ -4102,7 +4102,7 @@ private fun MobileSettingsSubPage(
                         title = stringResource(R.string.ai_model_title),
                         subtitle = stringResource(R.string.ai_model_desc),
                         value = when (uiState.subtitleAiModel) {
-                            com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B -> "Groq - Llama 3.3 70B"
+                            com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B -> "Groq - GPT-OSS 120B"
                             com.arflix.tv.ui.screens.player.SubtitleAiModel.GEMINI_FLASH_25 -> "Google - Gemini 3.5 Flash Lite"
                         },
                         isFocused = false,
@@ -5397,7 +5397,7 @@ private fun TvGeneralSettingsRows(
                     title = stringResource(R.string.ai_model_title),
                     subtitle = stringResource(R.string.ai_model_desc),
                     value = when (subtitleAiModel) {
-                        com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B -> "Groq - Llama 3.3 70B"
+                        com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B -> "Groq - GPT-OSS 120B"
                         com.arflix.tv.ui.screens.player.SubtitleAiModel.GEMINI_FLASH_25 -> "Google - Gemini 3.5 Flash Lite"
                     },
                     isFocused = focusedIndex == localIndex,
@@ -5863,7 +5863,7 @@ private fun GeneralSettings(
             title = stringResource(R.string.ai_model_title),
             subtitle = stringResource(R.string.ai_model_desc),
             value = when (subtitleAiModel) {
-                com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B -> "Groq – Llama 3.3 70B"
+                com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B -> "Groq – GPT-OSS 120B"
                 com.arflix.tv.ui.screens.player.SubtitleAiModel.GEMINI_FLASH_25 -> "Google – Gemini 3.5 Flash Lite"
             },
             isFocused = focusedIndex == 29,
@@ -5945,7 +5945,7 @@ private fun AiModelDialog(
 ) {
     val isMobile = LocalDeviceType.current.isTouchDevice()
     val options = listOf(
-        Triple(com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B, "Groq – Llama 3.3 70B", stringResource(R.string.ai_groq_model_note)),
+        Triple(com.arflix.tv.ui.screens.player.SubtitleAiModel.GROQ_LLAMA_70B, "Groq – GPT-OSS 120B", stringResource(R.string.ai_groq_model_note)),
         Triple(com.arflix.tv.ui.screens.player.SubtitleAiModel.GEMINI_FLASH_25, "Google – Gemini 3.5 Flash Lite", stringResource(R.string.ai_gemini_model_note))
     )
     BackHandler { onDismiss() }

docs/subtitle-auto-match.md

@@ -27,7 +27,7 @@ context.
 | Auto Find Best Subtitle Match | `subtitle_ai_find_best_match` | `false` | Auto-runs the scan on playback. **AI-independent** (works with no API key). Off by default — users opt in. Key name kept for backward compat. |
 | AI subtitle translation (master) | `subtitle_ai_enabled` | `false` | Enables AI features: translation option in menu, AI interim during scans, hearing fallback. |
 | Auto-Select AI Translation | `subtitle_ai_auto_select` | `false` | Allows AI translation to activate **automatically** (incl. as the scan's on-screen interim). |
-| Model | `subtitle_ai_model` | `GROQ_LLAMA_70B` | Groq or Gemini for batch translation. **Hearing requires Gemini.** |
+| Model | `subtitle_ai_model` | `GROQ_LLAMA_70B` | Groq (`openai/gpt-oss-120b` — the enum name is historical) or Gemini for batch translation. **Hearing requires Gemini.** |
 | API key | `subtitle_ai_api_key` (global) | — | One key used for both batch translation and Gemini Live. |
 
 All of the above (except profile-scoped language) sync via `CloudSyncRepository`

netlify-arvio-tv-site/companion/app.js

@@ -829,7 +829,7 @@ async function renderAI() {
         <label class="model-card ${model==='GROQ_LLAMA_70B'?'selected':''}">
           <input type="radio" name="ai-model" value="GROQ_LLAMA_70B" ${model==='GROQ_LLAMA_70B'?'checked':''} onchange="updateAISetting('subtitleAiModel',this.value)">
           <div>
-            <div class="model-name">⚡ Groq Llama 70B <span class="badge badge-gold" style="font-size:10px">${t('ai_recommended')}</span></div>
+            <div class="model-name">⚡ Groq GPT-OSS 120B <span class="badge badge-gold" style="font-size:10px">${t('ai_recommended')}</span></div>
             <div class="model-desc">${t('ai_groq_desc')}</div>
           </div>
         </label>

netlify-arvio-tv-site/companion/mock-preview.html

@@ -444,7 +444,7 @@ <h2 class="sub">Scrapers</h2>
     <h2 class="sub">Select AI Model</h2>
     <div style="max-width:560px">
       <div class="model-option selected">
-        <div class="model-name">Groq — Llama 3.3 70B <span class="badge green" style="margin-right:8px">Recommended</span></div>
+        <div class="model-name">Groq — GPT-OSS 120B <span class="badge green" style="margin-right:8px">Recommended</span></div>
         <div class="model-desc">Fastest · Free · Great quality for subtitles</div>
       </div>
       <div class="model-option">

web/lib/subtitleAi.ts

@@ -1,9 +1,14 @@
 // AI subtitle translation — web port of the Android app's
-// SubtitleTranslationService/Manager: same providers (Groq llama-3.3-70b /
-// Gemini 2.5 Flash), same prompt, same ⏎ line-break sentinel, batched with a
+// SubtitleTranslationService/Manager: same providers (Groq gpt-oss-120b /
+// Gemini), same prompt, same ⏎ line-break sentinel, batched with a
 // short window, cached per cue text, 5s backoff on 429.
 
-const GROQ_MODEL_ID = "llama-3.3-70b-versatile";
+// llama-3.3-70b-versatile was decommissioned by Groq (August 16, 2026); gpt-oss-120b is the
+// recommended replacement — faster, same 30 RPM free tier. It always reasons: "low" is the floor,
+// and "hidden" keeps the chain-of-thought out of message.content. Mirrors the Android service.
+const GROQ_MODEL_ID = "openai/gpt-oss-120b";
+const GROQ_REASONING_EFFORT = "low";
+const GROQ_REASONING_FORMAT = "hidden";
 const GROQ_URL = "https://api.groq.com/openai/v1/chat/completions";
 const GEMINI_MODEL_ID = "gemini-2.5-flash";
 const GEMINI_URL = `https://generativelanguage.googleapis.com/v1beta/models/${GEMINI_MODEL_ID}:generateContent`;
@@ -118,6 +123,8 @@ export class SubtitleTranslator {
       body: JSON.stringify({
         model: GROQ_MODEL_ID,
         temperature: 0.1,
+        reasoning_effort: GROQ_REASONING_EFFORT,
+        reasoning_format: GROQ_REASONING_FORMAT,
         messages: [
           { role: "system", content: systemPrompt(this.targetLanguage) },
           { role: "user", content: JSON.stringify(lines) }