hugomossberg/Babel v2.5.6-beta

Babel v2.5.6-beta

on GitHub

What's New in v2.5.6-beta

This beta adds an optional Bazarr Grace Period and improves model selection, subtitle handling and recovery. Existing QA and publication checks remain in place.

New Features

• Bazarr Grace Period: configure days and hours under Settings → Integrations → Bazarr before automatic AI fallback. Waiting jobs survive restarts and repeated scans, and Bazarr gets another opportunity before AI translation begins. Translate now bypasses the wait without creating a duplicate job. Requires Hybrid Mode and a configured Bazarr connection; the default is 0 hours, preserving existing behavior.

Bug Fixes

• Prevent duplicate Alpine initialization and overlapping polling, show cached libraries on startup, and clean up timers and obsolete requests.
• Improve subtitle source detection for language metadata, alternate full tracks, bare sidecars and subtitle stream layouts, while retaining language and source validation.
• Remove oversized absolute font sizes introduced by ASS/SSA-to-SRT conversion while preserving timestamps and compatible formatting. Native SRT font sizing is unchanged.
• Preserve completed recovery checkpoints after AI timeouts and bound automatic retries.
• Use the supported Jellyfin authorization header for connection testing and library refresh, with clearer, safer error handling.
• Accept legitimate Italian ordinal indicators such as 10ª and 1º without accepting real cross-script contamination.
• Improve timing-offset validation when dialogue is consolidated across cues, and invalidate stale timing verdicts after repairs.

Improvements

• Refresh model lists through provider discovery APIs, with pagination, deduplication and filtering for compatible text-generation models.
• Distinguish live listings, cached results and unverified offline suggestions. Preserve saved primary, fallback and custom model IDs; unavailable models are not silently replaced.
• List installed Ollama models and add live DeepSeek discovery. Keep DeepL's translation-specific model options separate from general-purpose LLMs.
• Improve OpenAI request compatibility for structured output and reasoning-model token parameters.
• Add AI Request Timeout (Seconds) for translation, QA and recovery. 0 keeps provider defaults; slower local inference can use a higher timeout.

Security

• Keep Plex tokens out of request URLs and prevent unexpected notifier errors from exposing credentials.
• Update the affected dependency pins: certifi 2026.10.19, pyasn1 0.6.5 and urllib3 2.8.0.
• Add repository, version, release-note and source checks before image publication, and serialize release runs.

Upgrade Notes

• Back up Babel's persistent database before upgrading. Startup adds Grace Period fields while preserving existing jobs, settings and saved models. Rolling back to older code requires the matching database backup because older versions do not understand WAITING_GRACE.
• Existing users keep the current automatic behavior unless they enable Grace Period. No API keys or model selections are migrated automatically.
• Custom discovery and connection-test URLs must use HTTP(S) without embedded credentials, query strings or fragments. Use the separate API key field.

With the standard Docker Compose configuration:

docker compose pull
docker compose up -d

Versioned image: ghcr.io/hugomossberg/babel:v2.5.6-beta

Rolling beta channel: ghcr.io/hugomossberg/babel:beta

To pin this version, use the versioned image for both Babel and the updater service in the standard Compose file. This beta does not update the latest channel. The published platform is Linux amd64.

Validation and Known Limitations

• 1,909 tests passed in both native Python and isolated Docker validation, with no failures or skipped tests. Node/Alpine regression checks passed; existing Chromium evidence was retained because the final security and release changes did not alter UI code. The locally built runtime image passed startup, authentication, database migration and restart-persistence checks.
• Provider and media-server contracts use deterministic mocks. Paid inference and the original Jellyfin, slow-Ollama and media-specific reports still need real-world confirmation. ASS/SRT appearance also depends on the player.
• A verified PT-BR dialect-quality fix and multiple Gemini API keys are not included.

Full Changelog: v2.5.5-beta...v2.5.6-beta