xtremexq/StremioSubMaker v1.4.0 on GitHub

New Features:

AI timestamps mode (toggle in Advanced Settings): trust the active translation provider to return/repair timestamps per batch, stream partial SRTs with AI timecodes where supported (Gemini), and rebuild partials safely while throttling with new SINGLE_BATCH_* env controls.
Single-batch Translation Mode with streaming partials to Stremio, token-aware chunking, and new SINGLE_BATCH_* env knobs to throttle streaming rebuild/log cadence.
Beta Mode added to config page - enabling it creates a "Multiple Providers" option on "AI Translation API Keys" section and shows "Advanced Configs" section for changing Gemini parameters.
Multi-provider translation pipeline: OpenAI, Anthropic, XAI/Grok, DeepSeek, DeepL, Mistral, OpenRouter, and Cloudflare Workers AI providers with per-provider keys/model pickers and server-side model discovery with a new parallel translation workflow.
Secondary provider fallback option added to config page (BETA).
Main Gemini translation workflow now supports Streaming (much faster partials and batch translations). Also implemented to other providers.

Translation Engine & Providers:

Gemini defaults now use gemini-flash-latest with SSE streaming support, real token counting, smarter output token budgets, batch headers, and timestamp-aware prompts; cache keys now include prompt variants and AI-timestamp mode.
Automatic batch-level fallback to the secondary provider on MAX_TOKENS/PROHIBITED_CONTENT/HTTP errors, preserved translation error types (incl. MULTI_PROVIDER), and mismatch correction that keeps AI-returned timecodes when requested.
Token-aware splitting tightened (soft caps per chunk, recursive halving, single-batch dynamic splits), streaming partial persistence throttled/logged to cut Redis I/O, and subtitle search cache reduced to 5k entries to lower memory use.

Configuration & UI:

Config page rebuilt with a multi-provider card (main/fallback selectors), per-provider advanced parameter editors (temperature/top-P/max tokens/timeouts/retries, Anthropic thinking, DeepL formality/formatting), encrypted storage of provider keys, and manifest/README copy updated to reflect non-Gemini providers.
File translation UI/API now accept provider-specific prompt/model/parameter overrides plus sanitized advanced settings; bypass cache is auto-forced when using advanced, multi-provider, single-batch, or AI-timestamp modes to keep shared caches clean.
Validation hardened to require at least one configured AI provider, enforce distinct main vs fallback, clamp provider parameters, normalize toggles, and normalize encrypted provider keys on load/save.

New Files:

src/services/providers/openaiCompatible.js - Universal provider for OpenAI, XAI/Grok, DeepSeek, Mistral, OpenRouter, and Cloudflare Workers AI
src/services/providers/anthropic.js - Anthropic/Claude provider with extended thinking support
src/services/providers/deepl.js - DeepL provider with beta languages and formality control
src/services/translationProviderFactory.js - Factory pattern with FallbackTranslationProvider class for automatic provider switching

New Environment Variables:

SINGLE_BATCH_LOG_ENTRY_INTERVAL - Debug log checkpoint interval for single-batch streaming
SINGLE_BATCH_SRT_REBUILD_STEP_SMALL - Partial SRT rebuild step when entries = threshold
SINGLE_BATCH_SRT_REBUILD_STEP_LARGE - Partial SRT rebuild step when entries > threshold
SINGLE_BATCH_SRT_REBUILD_LARGE_THRESHOLD - Entry threshold to switch rebuild step sizes

API Endpoints Added:

/api/models/:provider - Generic model discovery for all non-Gemini providers (OpenAI, Anthropic, DeepL, XAI, DeepSeek, Mistral, OpenRouter, Cloudflare Workers)

API Enhancements:

/api/gemini-models now accepts configStr for fetching models using saved session token
/api/translate-file now accepts overrides parameter for provider-specific settings (prompt, model, parameters)

File Translation UI:

Provider-specific overrides: prompt, model, parameters
Advanced settings override panel with input clamping
Bypass cache auto-forced when using advanced/multi-provider/single-batch/AI-timestamp modes
Multiple bug fixes

Other Bug Fixes:

Fixed critical cross-user configuration contamination bug: Added aggressive cache prevention headers (Cache-Control, Pragma, Expires) to /api/get-session/:token endpoint and implemented client-side cache-busting with timestamp query parameters. This bug was causing users to randomly see other users' language settings (e.g., Croatian, Hebrew, Polish) due to HTTP caching by browsers, proxies, or CDNs.
Translation error classification retains upstream flags (429/503/MAX_TOKENS/PROHIBITED_CONTENT/INVALID_SOURCE) so user messaging and fallback routing stay accurate.
Various minor bug fixes.

This update includes the unreleased SubMaker v1.3.6 changes below:

SubMaker v1.3.6 (unreleased)

Improvements:

3-click cache resets are now rate limited based on time (defaults: 6/15m for permanent cache, 12/15m for bypass cache) and configurable via CACHE_RESET_LIMIT_TRANSLATION, CACHE_RESET_LIMIT_BYPASS, and CACHE_RESET_WINDOW_MINUTES.
Config/UI: Source-language selection cap is now configurable via MAX_SOURCE_LANGUAGES.
Config/UI: Added combined target/learn languages cap (default 6) configurable via MAX_TARGET_LANGUAGES to prevent oversized selections.
Just Fetch mode: Added a configurable cap on fetched languages (default 9) via MAX_NO_TRANSLATION_LANGUAGES, with UI enforcement and server-side validation.
Config/UI: Updated Gemini model options to use gemini-flash-latest and gemini-flash-lite-latest for the Flash defaults.
Translation engine: Each batch prompt now carries an explicit BATCH X/Y header so the model knows which chunk it is translating.
Advanced settings: New "Send timestamps to AI" toggle sends timecodes to Gemini and trusts the model to return corrected timestamps per batch using the default translation prompt.

Bug Fixes:

SRT Translation UI: Fixed authentication failure in file upload translation page using session tokens configs
File Translation API: Added support for advanced settings override, allowing UI customizations to be properly applied during translation