Also throttle hedge-median recompute to 0.5s, cache replace_fields format
patterns and short-circuit when no $ token, and stream HLS merge with
copyfileobj. Adds test_dash_collapse and live A/B tooling.
Segments carrying their own Range (DASH SegmentBase, HLS EXT-X-BYTERANGE) now
fail the attempt when a server ignores the Range and returns 200, instead of
writing the whole parent resource as the segment and corrupting the merge. The
resume path no longer clobbers such a slice's Range, and retries rewrite the
slice whole. All segment GETs request Accept-Encoding: identity and the rnet
path mirrors the gzip guard, so transparent CDN (de)compression can't break the
Content-Length truncation check. Adds tests pinning these plus the adaptive
start-at-cap wiring and the small-segment tail-probe gate.
by_language now applies RFC 4647 Lookup in exact mode, preferring the string-equal tag so --exact-lang separates base/regional pairs (ar/ar-EG, en/en-US) instead of collapsing them, and dedupes by id to prevent the duplicate-track-id crash. Adds a language-selection regression test.
Merge cues at the string level instead of round-tripping through
pycaption, which dropped inline tags and cue settings. STYLE and
REGION blocks are kept, duplicate boundary cues splice correctly,
and a bad timing line now skips the cue instead of crashing.
Also rank pysubs2 above SubtitleEdit for ASS to TTML/VTT since
SubtitleEdit flattens inline styling there. Documented the measured
backend fidelity table and added regression tests for the merge.
Share one best-audio selector across the episode and music paths so best/all works for songs, descriptive audio is capped to one track, and codec_priority is honored. Guard --channels and 16:9 canvas math against tracks missing channel/width metadata. Extend --best-available to video languages and full resolution misses in both normal and hybrid mode, and count only deliverable ranges when reporting missing resolutions. Rank video by resolution before bitrate, and trim group audio selections to one track per song with a warning.
BREAKING CHANGE: a missing resolution with -r HYBRID now errors without --best-available, matching non-hybrid behavior.
Warm up after half a tick and double the worker target while probes pay off, dropping to AIMD +2 on first plateau/error burst. 33% faster in shaped benchmarks; backoff semantics unchanged.
Optional ranking of forced/normal/sdh per language, mirroring audio.codec_priority; first track gets the default flag, so this also picks the default subtitle. Unset = existing behavior.
Bind sessions and download jobs to the API key (not source IP), reject path-traversal cache keys, bound decompression of peer-supplied blobs, gate the CDM device before the vault fallback, and compare API keys in constant time.
Optional Episode.air_date names daily/sports content by date instead of SxxExx (no template change). Presentation-only; `{year}` dropped only when it duplicates the date's year. Round-trips via export/import.
Allow directories.services entries to be git repo specs (URL/owner-repo/ @branch) alongside local paths: clone + TTL-pull as cache, list-order priority with a duplicate-ignored summary, a refresh-services command that force-syncs and reports the per-service diff, and a dirty-clone guard that protects local edits on the automated path. Adds binaries.Git and a redact_paths config toggle.
cdm lookups used the raw service key while decryption upper-cases its keys, so a case-mismatched cdm entry silently used the default. Add a shared ci_get helper, route cdm/decryption through resolvers, drop a redundant decryption mutation and dead downloader writes, and add a regression bench.
Derive both setup_routes and setup_swagger from one ROUTES table, and move the repeated trailing APIError->response block into an @api_handler decorator.
Delete code paths that never executed in the serve/API/remote stack:
- BridgeCancelledError: raised but never caught; replace raises with RuntimeError and drop the class.
- InputBridge.error setter and _error field: written at the handler layer, never read (session.auth_error is the surfaced channel).
- cancel_event param and _check_cancel inner fn on _perform_download: the only caller always passed None (the live job.cancel_event is untouched).
- result_with_progress monkeypatch: its status pings were redundant with the file-polled progress_sink, which is now the sole path.
- session param on _resolve_manifest_data: unused in the body.
Reconcile SessionStore's hardcoded 600s auth-state TTL into a shared AUTH_INPUT_TIMEOUT constant referenced by both the InputBridge input timeout and the store, removing the magic number without changing the longer grace window.
Update input_bridge/session_store tests for the new channels and add an assertion that auth-state entries expire at the single rule.
Move the supported interpreter range to >=3.11,<3.15, dropping the EOL 3.10 floor and unblocking installs on 3.13/3.14.
- pyproject: requires-python ">=3.10,<3.13" -> ">=3.11,<3.15"
- update_checker: replace deprecated asyncio.get_event_loop() with asyncio.to_thread() (warns on 3.12+, slated for removal)
- test_compression: simplify _run() to plain asyncio.run()
- README: bump documented Python range to 3.11 - 3.14
- uv.lock: regenerate for the new interpreter floor
Group selected videos by (resolution, range, codec) and mux each group into one MKV; only language is collapsed, so ranges/codecs/resolutions stay in separate files. Adds --merge-video flag + muxing.merge_video config (global and per-service), docs, and tests.
* Add native Music core workflow
* feat(music): add shared music core helpers, album folder template, and UX cleanup
Consolidate duplicated music-service logic into core and extend the native music workflow.
- add core/music/extract.py: shared stateless helpers (first_text, first_number, year/duration/name formatting, classify_release_kind, dedupe_track_options, build_music_from_songs) so services stop carrying their own copies
- add core/music/display.py: shared rich rendering (render_track_panel, render_album_header, render_artwork_preview) with TrackRow / MusicHeaderInfo data holders
- export the new helpers from core/music/__init__.py
- add dedicated `albums` folder template kind (output_template.folder.albums) resolving albums -> songs -> "{artist} - {album} ({year})"; whitelist music template variables (album_artist, track_total, disc_total, release_type, genre, explicit, isrc, upc, label)
- fix song filename crash: config.get_output_template(...) did not exist; use config.output_template.get("songs") with a sane default
- strip emojis from music output (renderer, dl.py music branch) to match unshackle UX; remove dead MusicSongPlan import and music_icon logic
- document the albums folder key in unshackle-example.yaml
- add tests for extract, display, and the folder template
---------
Co-authored-by: MrMovies-Dev <MrMovies-Dev@users.noreply.github.com>
Smooth Streaming manifests carry no range attributes, so every ISM video track was labelled SDR even when the stream is HDR10/HLG/DV, breaking range-based selection (-r HDR10 / -r DV) for ISM services.
- ism_init: walk the full HEVC SPS (incl. scaling-list and st_ref_pic_set skippers) to read the VUI colour triple (colour_primaries, transfer_characteristics, matrix_coeffs); expose parse_codec_private_data_colour() keyed by FourCC. No unshackle imports added.
- ism: new ISM.get_video_range() maps the CICP triple via Video.Range.from_cicp (PQ -> HDR10, HLG -> HLG, BT.709/absent -> SDR); DVHE/DVH1 FourCCs map straight to DV since DV bitstreams signal Unspecified (2,2,2) in the VUI. to_tracks() now sets range_ on every video track. Soft-fails to SDR on malformed data.
- ism: accept RnetSession in download_track() so TLS-impersonated sessions pass the type check.
- tests: real PQ/HLG/BT.709 (x265-minted) and Dolby Vision (live-manifest, DoViProfile=stn, out-of-order SPS,PPS,VPS NALs) CodecPrivateData samples; byte-level VUI assertions in test_ism_init and manifest->Range characterization in new test_ism_range.
ISM (Smooth Streaming) tracks raw-concatenate moof+mdat fragments with no ftyp/moov, so shaka-packager/mp4decrypt fail with PARSER_FAILURE (exit 2) on decrypt. The init box was previously built by n_m3u8dl_re, removed in the downloader consolidation.
Add ism_init.py, a dependency-free byte-level MP4 init-segment synthesizer that rebuilds ftyp+moov from the manifest CodecPrivateData, ported from yt-dlp's write_piff_header and N_m3u8DL-RE's MSSMoovProcessor with full codec parity:
- AVC (H264/AVC1/DAVC), with SPS/PPS picked by NAL type rather than position and NALUnitLengthField honored
- HEVC (HVC1/HEV1) with chroma format and bit depths parsed from the de-emulated SPS via exp-Golomb so 10-bit/HDR signals correctly, and profile/tier/level lifted from the SPS PTL
- Dolby Vision (DVHE/DVH1) as hvcC with a dvh1 sample entry
- AAC (AACL/AACH) with the AudioSpecificConfig synthesized from SamplingRate/Channels when the manifest omits CodecPrivateData
- EC-3 with a real dec3 box extracted from the WAVEFORMATEXTENSIBLE CodecPrivateData (Dolby GUID located by search, not fixed offset)
- TTML subtitles as stpp/sthd/subt, wired for fragmented-TTML tracks
CENC wrapping (encv/enca + sinf/tenc with default_KID) covers encrypted tracks: the per-sample IV size is derived from the fragment senc/saiz (PIFF override flag, payload arithmetic, saiz fallback) instead of assuming 8, and the constant-IV tenc form is supported. Read the track_ID from the first fragment's tfhd so the moov matches and the muxer does not drop samples.
Wire ISM.download_track to prepend the synthesized init before merging; unsupported codecs soft-fail to raw concatenation with a warning. Harden against real-world inputs: 2-letter/uppercase manifest language tags normalize to ISO-639-2 (und fallback), >65535 Hz sample rates no longer overflow the 16.16 field, truncated tfhd returns None, struct.error joins the soft-fail handler, and the emulation-prevention scan no longer over-strips consecutive escapes.
Add regression tests (37) covering box structure, every supported FourCC, 10-bit SPS parsing, ASC synthesis, dec3 extraction, IV-size derivation and the crash fixes. Validated structurally per codec with ffmpeg-minted fragments: shaka-packager parses synth-init+fragments with exit 0 and ffprobe reports the expected codec, including a live run against a public Smooth Streaming server.
unshackle's DASH parser only recognised Widevine and PlayReady ContentProtection, so services using W3C EME ClearKey had to fake a Widevine object and monkey-patch get_content_keys. Add a first-class ClearKeyCENC DRM type so services just implement a license callback.
- ClearKeyCENC (core/drm/clearkey_cenc.py): KID-based, no CDM/PSSH; builds the W3C JSON license request (unpadded base64url), parses the JWK Set response (dict/str/bytes), falls back to POSTing the manifest Laurl when the service returns None, decrypts via the same shaka/ mp4decrypt CENC path as Widevine
- DASH.get_drm emits ClearKeyCENC for scheme e2719d58-...; KID from own or sibling mp4protection cenc:default_KID, Laurl across dashif/legacy/ bare namespaces
- track.download dispatches prepare_drm for ClearKeyCENC; dl.prepare_drm gains a clearkey branch (cache/vault lookup, license-failure tolerated when content_keys pre-populated, vault push, export)
- Service.get_clearkey_license base callback (default None -> Laurl); drm_from_dict reconstructs ClearKeyCENC for export/import round-trip
- EXAMPLE service + config demo the callback
- Tests: tests/core/test_clearkey_cenc.py and an export round-trip case
- Docs: DRM_CONFIG.md ClearKey section
Instrument the full download pipeline in the structured JSONL debug logger and make adding logging to new features a one-liner.
Coverage:
- DRM license request/response, content keys (incl. remote-CDM seam) and decrypt timing across Widevine/PlayReady/ClearKey
- DASH/HLS/ISM manifest fetch + parse milestones (HLS.to_tracks also covers the m3u8_parser path used by iTunes/ATV-style services)
- Per-backend vault get/add via the Vaults manager, track selection, subtitle conversion, repackage, normalize_vui, and full mkvmerge mux (command, duration, output size, warnings)
- All external tooling (ffmpeg, ffprobe, mkvmerge, mkvpropedit, dovi_tool, SubtitleEdit, ccextractor) via a unified `tool_run` op, centralised in run_step/ffprobe + log_tool_run
DX:
- Add log_event() / timed_operation() primitives (no-op when disabled); retrofit ~91 guard/timing blocks onto them
- Fix message= collision in log_drm_operation/log_vault_query/log_service_call that raised TypeError on the live decrypt path
Redaction (redact_all = redact_text -> redact_url -> redact_path):
- Collapse content/CDN/api URLs to `redacted[.ext]`
- Strip local path prefixes (install root -> <unshackle>, venv -> <venv>, home -> ~)
- Apply to every logged string so shared logs leak no URLs, paths or usernames
- Drop per-request service_call logging (manifest parse is the request seam)
services.<TAG>.dl values only applied when the key was also set in the global dl: section (equality check against config.dl missed Click's declared defaults). Gate on Click's ParameterSource instead: CLI/env > service dl > global dl > defaults, converted via each option's own type.
- record parameter sources on serve's hand-built context so client values are never clobbered
- accept range/list as natural keys for range_/list_
- harden QualityList (YAML int) and SlowDelayRange (YAML bool) converts
Replace the class-level Track.download monkeypatch with a per-job progress sink threaded through dl.result(). The API now reports a single aggregate signal instead of each track's bouncing 0-100%:
- bitrate-weighted completion so video/audio dominate subtitles
- completed_tracks/total_tracks counts and active_tracks labels (e.g. "video 2160p DV", "audio en-US 5.1")
- downloads fill 0-90%; repackaging (when needed) and a "muxing" stage carry it to 100% so post-download work is no longer frozen at 100%
- monotonic throughout (handles the download->decrypt callable reuse)
Also:
- accept "HDR10P" as the canonical API range value ("HDR10+" still works)
- declare AUTH_METHODS opt-in on the Service base
- raise typed APIError (WORKER_ERROR/DOWNLOAD_ERROR) from the worker path
- move the progress helpers to unshackle/core/api/progress.py
* feat(api): live download phase, granular progress, swallowed-failure detection, per-request CDM
- Tee Track.download progress so the job reports real percentage (not just 5/100%) and a
human-readable phase ('downloading video 1080p') via the new job.phase field.
- Detect a swallowed download-worker failure (dl.result() prints 'Download Failed' but
exits 0) and raise, so the job is marked failed instead of completed-with-no-output.
- Per-request CDM override (dl_instance.cdm_override; get_cdm prefers it) so a job can use
a specific CDM device without mutating shared config.
* feat(api): expose service capability flags + auth methods in /services
needs_auth / has_search / has_drm derived from which Service hooks are overridden, and
auth_methods inferred from what the service's authenticate() body references (cookies /
credentials), so clients can show only the relevant auth options per service.
* feat(api): per-request credential injection for downloads
Accept a 'credential' ('user:pass') job parameter and feed it into the credentials map that
dl.get_credentials() reads, so a client can authenticate a download without persisting
anything to disk. (Kept on the deployment branch; the PR branch uses the client-sent path.)
* feat(api): gate per-request CDM override behind serve.cdm_overrides
A per-request `cdm` selects a server-side device, so honour it only when allow-listed.
`serve.cdm_overrides` opts in: a list permits those device names, or `true` permits any
(single trusted client). Unset/false rejects every override with 403, so an arbitrary device
can't be selected by default.
* fix(api): redact credentials and proxy userinfo in serialized job parameters
Job parameters can carry a raw user:pass credential and a proxy URL with embedded userinfo;
mask them wherever parameters are serialized for an API response so secrets don't leak via
the job-detail endpoint. Also read skipped-subtitle / download-failure state from the dl
instance instead of scraping stdout, and drop the dead n_m3u8dl percent branch.
* feat(api): prefer explicit AUTH_METHODS class var over source inference in /services
Inferring auth methods from authenticate() source mostly returns both options because services
call super().authenticate(cookies, credential). Prefer an explicit AUTH_METHODS class var when a
service declares one, falling back to inference.
* style: use plain hyphens instead of em-dashes in comments
* feat(api): gate per-job credentials, isolate caches, scrub error fields
Address review feedback on #113:
- Gate per-request credential/credentials behind serve.allow_job_credentials (default off,
403 when not opted in), mirroring the existing serve.cdm_overrides CDM gate.
- Isolate the token cache per credential: when a per-job credential is set, namespace
config.directories.cache by its hash in the worker, so two clients on the same service
with different credentials can't share a cached session.
- Scrub the credential, its password half, and proxy userinfo out of error_message,
error_details, error_traceback and worker_stderr before they leave via the job-detail API.
- Remove the unused _execute_download_sync in-process path (would have leaked one job's
credential into the shared global config).
- Document serve.cdm_overrides and serve.allow_job_credentials in the example config.
- Add tests for both gates (403 default, allowlist pass) and the parameter/error redaction.
* fix(dl): flag download_failed when result() swallows a worker failure
dl.result() catches a download-worker exception, reports it, and returns
without re-raising so the CLI still exits cleanly. An embedding caller (the
API worker) had no way to tell the title actually failed and would report
the job completed with no output. Expose a download_failed flag, set in the
swallow path, that the worker reads after result() returns.
* feat(api): surface skipped subtitles and pass skip_subtitle_errors
Thread skip_subtitle_errors from the job into dl() so the API can opt into
non-fatal subtitle handling, and report which subtitles were skipped: store
them on the job (dl.SkippedSubtitle dicts) and include them in job details
so a client can tell the user which weren't available.
---------
Co-authored-by: Avi Cohen <avraham.coh770@gmail.com>
Replace the hardcoded conversion if/elif in Subtitle.convert with a capability-matrix backend registry (subtitle_convert.py): each backend declares the source->target pairs it supports plus a rank, and run_conversion tries them in order as a real fallback chain. conversion_method pins a backend but still falls back (pin-then-fallback).
- Detect the cross-platform SubtitleEdit 5+ CLI (seconv) and use its --flag syntax for convert, SDH stripping, and reverse-RTL
- Protect styled ASS/SSA from automatic SRT downconversion; honor an explicit --sub-format / sidecar_format
- Read segmented fVTT (wvtt) and fTTML (stpp) directly from fragmented MP4
- Improve ASS/SSA font detection: inline \fn overrides, Format-located Fontname column, @-prefix strip, case-insensitive de-dup; covers SSA too
- Update SUBTITLE_CONFIG.md, example yaml, README; add regression tests and a backend benchmark script
-r HYBRID alone muxed the HDR10/HDR10+ base layer as a standalone output because only the ingredient DV track was flagged hybrid_base_only. The inverse was also broken: HDR10/HDR10+/DV tracks never entered the standalone-deliverable pool, so -r HYBRID,HDR10P only delivered the standalone HDR10+ by accident of the first bug.
- Add Tracks.partition_hybrid_videos: ingredient ranges (HDR10/HDR10+/DV) enter the deliverable pool only when their range is explicitly requested alongside HYBRID; replaces the duplicated filter in dl.py.
- Add Tracks.flag_hybrid_ingredients: any track in the hybrid selection but not in the deliverable selection is flagged hybrid_base_only; replaces and generalises the DV-only dv_is_deliverable special case.
Addresses review feedback on the failed-subtitle handling:
- Fully annotate download_tracks_in_passes and the download_track closure (MyPy strict).
- Replace list[str] skipped_subtitles with a documented SkippedSubtitle TypedDict
(id + language + title) so a client can report which subtitle of which title was
skipped; pin the shape with a contract test.
- Clear DOWNLOAD_CANCELLED in a finally so no failed track leaves it set for later code.
- Document why the subtitle pass must stay sequential (a concurrent pass would silently
drop in-flight subtitles via the cancel event).
- Warn only when a title skipped a subtitle and produced no video/audio/subtitle (was a
loose len(title.tracks) check that ignored chapters/attachments).
- Narrow the over-broad remove() except to ValueError with a debug log.
- Add tests: final-clear on the fatal path, all-subs-skipped keeps video/audio, and
duplicate-language subtitles distinguished by id.
ast.Num/node.n deprecated since Python 3.8; NodeVisitor falls back to visit_Num with a DeprecationWarning per dispatch. Under -W error this surfaced as a misleading fps ValueError in Video.init. Replace with visit_Constant, reject non-numeric constants, and pin parse results for int/fraction/float inputs in a regression test.
A single failing subtitle track previously aborted the whole download. Add an opt-in
--skip-subtitle-errors flag: when set, a Subtitle failure is logged and the track dropped from
the mux while the video/audio still complete (Video/Audio failures stay fatal; default
behaviour is unchanged).
Done at the right layer to avoid the shared-event race: a failed track sets the process-global
DOWNLOAD_CANCELLED event, which makes other in-flight tracks early-return without raising — so a
skipped subtitle could otherwise silently truncate the video/audio that still got muxed. The
download is split into two passes (download_tracks_in_passes): the fatal tracks download
concurrently first, then the skippable subtitles run in a separate sequential pass once nothing
else is in flight, with the event reset before each and at the start of every title. Skipped
languages are recorded on the dl instance (skipped_subtitles) for callers to surface.
Adds tests for the cancel-event interaction (a failing subtitle no longer truncates the
video/audio), the good/bad subtitle mix, the flag-off fatal path, and the per-title reset.
imSp4rky