Install
docker pull ghcr.io/corveck79/mycelium:0.6.2
Added
- spore-nfs and spore-smb: read-only NFSv3 and SMB2/3 servers exposing the virtual library as real files, backed by the existing
/spore-stream/<token> endpoint (no new materialization logic). Server-side tricks to block Direct Play on Shield/Android TV (stub channel count, forced PGS subtitle burn-in) stopped working reliably - Android's local-network fast path bypasses the profile negotiation Linux/desktop clients respect, turning the fake stub into a black screen instead of a transcode. With real size/bytes served, Direct Play becomes correct instead of catastrophic, on every client. Both protocols share a 3-window read-ahead LRU, background prefetch, resolved-CDN-URL caching, self-healing on a dead cached URL, and a token-bucket rate limiter with retry/backoff against TorBox/CDN 429s. spore-nfs later merged into the main image (one container instead of two).
Fixed
- MKV/non-MP4 playback via
/stream (Jellyfin's path, and any other client not building a moov-first cache) could be redirected straight to a CDN URL that had gone dead earlier than its 23h in-memory cache TTL, with no validation - Jellyfin/ffmpeg would follow the dead link into a TorBox error page and fail with FFmpegException: FFmpeg exited with code 8 / ffprobe failed. Now HEAD-checked before redirecting, with an automatic re-resolve on a dead link.
play_count/last_played were written to SQLite on every single byte-range request during playback, not just on play start - under concurrent playback these writes serialized against each other for no benefit. Debounced to once per token per 60s; the CDN liveness check above is similarly cached for 120s so repeated seeks in one session don't each pay a fresh CDN round trip.- Stale
JELLYFIN_API_KEY on deploy could leave refresh_library(), merge_duplicate_versions(), refresh_missing_images() and continue-watching sync silently failing with 401 - not a code fix, but worth a mention since it was masking as playback flakiness.
