Mms
@keeb/mmsv2026.10.25.2
Security Notice
This extension includes AI agent skills that can modify AI assistant behavior. Review the skill files before installing.
01README
Media management system — content discovery, deduplication, download protocol handling, LLM-powered file organization, and subtitle extraction
02Release Notes
Organizer scene route places new shows deterministically (shows//sN) unless an existing dir looks like the same show; filler words no longer count in similar-dir matching; Ollama only health-checked when the LLM is needed; organizer test accepts scene parse items.
03Behaviors
@keeb/mms/sourcev2026.10.25.1mms_source.ts
fn search(query: string, provider: enum, resolution?: string, preferGroup?: string, preferCodec?: string, nyaaUser?: string, nyaaQuery?: string, eztvUrl?: string, eztvImdbId?: string, slug?: string, onlyEpisodes?: array, newznabUrl?: string, newznabApiKey?: string, newznabCat?: string)
Search a provider for content. Writes a single `episodes` resource
| Argument | Type | Description |
|---|---|---|
| query | string | Show name to search for |
| provider | enum | Source provider |
| resolution? | string | Resolution filter |
| preferGroup? | string | Preferred release group (e.g. MeGusta) |
| preferCodec? | string | Preferred codec (e.g. HEVC) |
| nyaaUser? | string | Nyaa uploader filter |
| nyaaQuery? | string | Override search query |
| eztvUrl? | string | Custom EZTV RSS URL |
| eztvImdbId? | string | IMDb ID (tt0239195): search EZTV's archival API instead of the rolling RSS feed |
| slug? | string | Canonical catalog slug, stamped on every item (dedup checks disk at /<slug>/s<N>) |
| onlyEpisodes? | array | Keep only these episodes, e.g. dedup's missing items; anything else is dropped |
| newznabUrl? | string | Newznab API base URL |
| newznabApiKey? | string | Newznab API key |
| newznabCat? | string | Newznab category ID |
fn search_configured()
Search all configured shows. Writes a single `episodes` resource
Resources
episodes(infinite)— All items from the most recent invocation. Single instance named
@keeb/mms/downloaderv2026.10.25.1mms_downloader.ts
Global Arguments
| Argument | Type | Description |
|---|---|---|
| transmissionUser? | string | Transmission RPC username |
fn blocklist_update()
Have Transmission re-download its peer blocklist. Optionally set the blocklist URL and enable it first.
fn add(uri: string)
Add a download to the appropriate backend (Transmission for torrents, SABnzbd for usenet)
| Argument | Type | Description |
|---|---|---|
| uri | string | Magnet URI or NZB URL |
fn add_nzb_url()
Construct an NZB URL from a vault-resolved template + a hash
fn add_from_source()
Read a source model's `episodes` resource and add every item in its episodes array as a torrent. Fan-out alternative to workflow-level forEach.
fn status()
Query all backends for current download status (factory: one resource per download)
fn remove()
Remove specific torrents from Transmission by name filter
fn clean()
Remove completed downloads from backends
Resources
download(infinite)— Tracked download state
add_batch(infinite)— Per-invocation summary of an add call — added/failed items and counts. Aggregate shape; does not track per-torrent lifecycle (see download spec for that).
blocklist(infinite)— Result of a blocklist_update call — URL, enabled flag, and rule counts before/after
@keeb/mms/dedupv2026.10.25.1mms_dedup.ts
fn filter()
Read episode data from a source model, check local directories, output only new episodes. Parses raw titles with Ollama when show/episode fields are missing.
Resources
episodes(infinite)— All actionable episodes from the most recent invocation. Single
@keeb/mms/organizerv2026.10.25.1mms_organizer.ts
fn process()
Pop pending jobs from MongoDB queue, classify with LLM, and move files to organized locations
fn pending(maxJobs?: number, retryHours?: number, name?: string)
List the video files plan would classify next (same job selection), so a scene-rules parse can run on them first. Writes pending-<name> with names[]. Moves nothing.
| Argument | Type | Description |
|---|---|---|
| maxJobs? | number | |
| retryHours? | number | |
| name? | string |
fn plan(maxJobs?: number, retryHours?: number, scene?: array, name?: string)
Classify pending jobs without moving anything: destination, route and gate context per file, plus the Jev batch for LLM-placed files. Jobs a previous apply left in staging are re-planned as soon as the organizer's rules change, else every retryHours. Writes plan-<name>; apply executes it.
| Argument | Type | Description |
|---|---|---|
| maxJobs? | number | Maximum pending jobs to plan |
| retryHours? | number | Re-plan a job left in staging after this long even if the rules haven't changed |
| scene? | array | @keeb/scenerules parse items for the pending files (see pending); enables the deterministic scene route |
| name? | string | Suffix for the plan-<name> resource |
fn apply(name?: string, judgeModel?: string, threshold?: number)
Execute plan-<name>: move each job's files into the catalog only when every file is a confident placement. A split risk, a Jev season_fits score below threshold, or no score leaves the job pending in staging (held_at/held_reason/held_rules), to be re-planned when the rules change. Writes job-<id> per job.
| Argument | Type | Description |
|---|---|---|
| name? | string | Plan to apply |
| judgeModel? | string | Model whose evaluation-<key> resources hold the season_fits scores |
| threshold? | number | Hold LLM-placed files whose season_fits P(yes) is below this |
fn evaluate(cases: array, models: array, llmOnly?: boolean, structured?: boolean, catalogHints?: boolean, scene?: array, name?: string)
Benchmark Ollama models on labeled filenames through the real classification pipeline; writes eval-<name> with accuracy, latency and misses per model. Moves nothing.
| Argument | Type | Description |
|---|---|---|
| cases | array | Filenames with the directory each belongs in |
| models | array | Ollama models to compare, e.g. ['qwen3:14b', 'gemma4:26b'] |
| llmOnly? | boolean | Send every case to the LLM (skip special cases and the fast path), to compare models rather than the pipeline |
| structured? | boolean | Constrain stage 1 with Ollama structured outputs (JSON Schema), as production does |
| catalogHints? | boolean | Tell stage 2 which existing catalog dirs look like the show, as production does |
| scene? | array | @keeb/scenerules parse items for the case filenames; enables the scene route (ignored with llmOnly) |
| name? | string | Suffix for the eval-<name> resource |
fn test(filename: string, scene?: array)
Dry-run classification: parse a filename and show where it would be organized without moving anything
| Argument | Type | Description |
|---|---|---|
| filename | string | Filename to classify |
| scene? | array | @keeb/scenerules parse items (as plan takes them); enables the scene route |
fn write_playlists()
Write an .m3u playlist of every video file, in episode order, into a catalog directory (and optionally its subdirectories)
fn write_playlists_recent()
Regenerate .m3u playlists for every directory touched by recent organizer jobs
Resources
job(infinite)— Processed media job result
playlist_run(infinite)— Playlist (.m3u) generation run summary
pending(infinite)— Video files the next plan will classify (input for a scene-rules parse)
plan(infinite)— Where each pending job would go, with review-gate context and the Jev batch to score it. Written by plan, read by apply.
eval(infinite)— Classification benchmark: accuracy and latency per Ollama model over labeled filenames
@keeb/mms/subtitlesv2026.10.25.1mms_subtitles.ts
fn extract(path: string)
Extract an embedded subtitle track from an MKV to a sibling .srt file
| Argument | Type | Description |
|---|---|---|
| path | string | Path to the input MKV file |
fn extract_recent()
Batch-extract subtitles from MKVs produced by recent organizer jobs
fn extract_dir(paths: array, recursive?: boolean)
Batch-extract subtitles from every MKV under catalog directories (backfill)
| Argument | Type | Description |
|---|---|---|
| paths | array | Catalog directories to walk |
| recursive? | boolean | Walk subdirectories too |
Resources
extraction(infinite)— Subtitle extraction result (single-file extract)
run(infinite)— Batch subtitle-extraction run summary
@keeb/subsplease-schedulev2026.04.11.1subsplease_schedule.ts
fn fetch(tz?: string)
Fetch the SubsPlease weekly schedule, one resource per scheduled show
| Argument | Type | Description |
|---|---|---|
| tz? | string | Timezone for schedule times |
Resources
show(infinite)— A show on the SubsPlease weekly schedule
@keeb/tmdb-lookupv2026.07.03.1tmdb_lookup.ts
Global Arguments
| Argument | Type | Description |
|---|---|---|
| apiKey | string | TMDB API key (v3) |
fn search(query: string, limit?: number)
Search TMDB for a TV show by name, fetches full details including series status
| Argument | Type | Description |
|---|---|---|
| query | string | TV show name to search for |
| limit? | number | Max results to return |
fn seasons(tmdbId: number)
List a TV show's seasons with cumulative absolute episode ranges — for computing episode→season breaks
| Argument | Type | Description |
|---|---|---|
| tmdbId | number | TMDB TV id |
fn searchMovie(query: string, year?: number, limit?: number)
Search TMDB for a movie by title, fetches full details (release date, runtime, vote, etc.)
| Argument | Type | Description |
|---|---|---|
| query | string | Movie title to search for |
| year? | number | Optional release-year hint to disambiguate |
| limit? | number | Max results to return |
Resources
show(infinite)— TV show info from TMDB
movie(infinite)— Movie info from TMDB
@keeb/rottentomatoesv2026.10.25.3rottentomatoes.ts
Global Arguments
| Argument | Type | Description |
|---|---|---|
| tmdbApiKey | string | TMDB API key (v3) |
| headless? | boolean | Run Playwright headless. |
| nodeModulesDir? | string | node_modules directory that already has playwright installed. When unset, the model installs a pinned playwright into $XDG_CACHE_HOME/swamp-mms/scraper (default ~/.cache) on first use, so no consumer-repo node_modules is needed. |
fn browse(sort: enum, loadMoreClicks?: number)
Scrape a Rotten Tomatoes TV browse list, enrich each show via TMDB, and write one resource per show keyed by TMDB ID.
| Argument | Type | Description |
|---|---|---|
| sort | enum | Which Rotten Tomatoes TV browse list to scrape |
| loadMoreClicks? | number | How many times to click Load More before extracting (0 = first page only) |
fn browse_movies(list?: enum, sort?: enum, loadMoreClicks?: number)
Scrape a Rotten Tomatoes movie browse shelf, enrich each movie via TMDB, and write one resource per movie keyed by TMDB movie ID. Default shelf is at_home (streaming/VOD).
| Argument | Type | Description |
|---|---|---|
| list? | enum | Which Rotten Tomatoes movie shelf to scrape. at_home = streaming/VOD (watchable now), in_theaters = theatrical, coming_soon = unreleased |
| sort? | enum | Sort order within the shelf. 'newest' is the only one RT honours reliably — the others often degrade to an alphabetical listing. |
| loadMoreClicks? | number | How many times to click Load More before extracting (0 = first page only) |
Resources
show(infinite)— A single TV show: Rotten Tomatoes scores + TMDB metadata. Spec is the TMDB ID (or rt-<slug> when no match).
movie(infinite)— A single movie: Rotten Tomatoes scores + TMDB metadata. Spec is the TMDB movie ID (or rt-<slug> when no match). Kept separate from `show` because TMDB movie and TV IDs share numeric values but name different titles.
@keeb/mms/tv-recommenderv2026.09.25.1tv_recommender.ts
Global Arguments
| Argument | Type | Description |
|---|---|---|
| personaPath? | string | Default persona markdown path. Relative paths resolve under the consumer's repo root. If unset, the bundled `personas/host.md` is used. |
fn evaluate(browseSort: enum, limit?: number, minScore?: number, catalogModel?: string)
Score rottentomatoes shows against the host persona's lanes using structured TMDB fields only, and write the ranked `recommendations` resource. No LLM, no API key, one step.
| Argument | Type | Description |
|---|---|---|
| browseSort | enum | Which RT browse list to evaluate (matches rottentomatoes browseSort tag). |
| limit? | number | Maximum number of shows to evaluate, taken by descending TMDB popularity. |
| minScore? | number | Blended RT score a lane-matching show must clear to earn Watch. Below this it is in-lane but under the bar. |
| catalogModel? | string | Model name to read the existing catalogue from, for the persona's 'new season of something I already watch' auto-include. Set empty to disable the lookup. |
fn collect(browseSort: enum, limit?: number, personaPath?: string, bundledPersona?: string)
LEGACY LLM PATH — superseded by `evaluate`. Read rottentomatoes shows for a given browseSort and the persona markdown; write a single `pack` resource the workflow forEach can drive Claude with. Requires an @keeb/anthropic/claude instance named host-evaluator.
| Argument | Type | Description |
|---|---|---|
| browseSort | enum | Which RT browse list to evaluate (matches rottentomatoes browseSort tag). |
| limit? | number | Maximum number of shows to package for evaluation. |
| personaPath? | string | Persona markdown path for this run. Overrides the model's personaPath global arg. Relative paths resolve under the consumer's repo root. |
| bundledPersona? | string | Name of a persona bundled with @keeb/mms (file in `personas/`, without the .md suffix). Use 'host' for the default. Wins over personaPath only when no path is given. |
fn aggregate(collectVersion?: number)
LEGACY LLM PATH — superseded by `evaluate`. Join the per-show Claude verdicts with the original RT metadata into a single ranked recommendation resource for the report.
| Argument | Type | Description |
|---|---|---|
| collectVersion? | number | Version of tv_recommender's `pack` resource to aggregate against. Defaults to latest. |
fn pack_typesafe(browseSort: enum, limit?: number, personaPath?: string, catalogModel?: string)
Package the current RT browse for @swamp/typesafe-ai.choice_batch: persona + watch question as instructions, one {name, state} item per show. Catalogue shows are auto-included and left out of items.
| Argument | Type | Description |
|---|---|---|
| browseSort | enum | Which RT browse list to evaluate (matches rottentomatoes browseSort tag). |
| limit? | number | Maximum number of shows to package, taken by descending TMDB popularity. |
| personaPath? | string | Persona markdown path for this run. Overrides the model's personaPath global arg. Relative paths resolve under the consumer's repo root. |
| catalogModel? | string | Model name to read the existing catalogue from. Shows already in it are auto-included and not sent for evaluation. Set empty to disable. |
fn aggregate_typesafe(evaluatorModel?: string, minScore?: number)
Join the evaluator's evaluation-<name> rubric answers onto the latest typesafe-pack, run them through the persona rules, and write the ranked `recommendations-typesafe` instance for the report.
| Argument | Type | Description |
|---|---|---|
| evaluatorModel? | string | Name of the @swamp/typesafe-ai instance whose ask_batch answered the pack's items. |
| minScore? | number | Blended RT score an in-lane show must clear to earn Watch — the same bar as evaluate. |
fn pack_catalogue(library?: string, catalogModel?: string, personaPath?: string)
Package every series in a catalogue library as ask_batch items with the persona rubric, for calibrate to measure the rubric against shows the group already chose. Writes `catalogue-pack`.
| Argument | Type | Description |
|---|---|---|
| library? | string | Catalogue library display name whose series become the calibration set — the same value passed to the catalogue model's `series` method. |
| catalogModel? | string | Model name holding the catalogue's `series` resources. |
| personaPath? | string | Persona markdown path whose rubric is being calibrated. Overrides the model's personaPath global arg. |
fn calibrate(evaluatorModel?: string)
Score the evaluator's rubric answers for the catalogue-pack with the persona rules (lane fit only, no RT bar) and write `rubric-calibration`: recall on shows the group already chose, per-statement stats, and the misses to tune against.
| Argument | Type | Description |
|---|---|---|
| evaluatorModel? | string | Name of the @swamp/typesafe-ai instance whose ask_batch answered the catalogue-pack. |
Resources
typesafe-pack(infinite)— Current RT browse as @swamp/typesafe-ai.choice_batch input: persona instructions, yes/no criteria, one item per show.
pack(infinite)— Per-show prompt material: persona text plus per-show JSON, ready for a forEach Claude call.
recommendations(infinite)— Aggregated verdicts: one row per show with parsed Watch/Skip + reasoning.
rubric-calibration(infinite)— How the persona rubric + lane rules judge the existing catalogue: recall, per-statement stats, and misses. Versions are the rubric optimizer's history.
04Reports
@keeb/mms/add-summarymethod
add_summary.ts
Summarize an add / add_from_source invocation — lists torrents/NZBs accepted and items that failed to add
downloaderadd
@keeb/mms/dedup-summarymethod
dedup_summary.ts
Summarize dedup filter results: what's new vs what was skipped
discoverydedup
@keeb/mms/discovery-summarymethod
discovery_summary.ts
Summarize discovered content from source searches
discovery
@keeb/mms/download-statusmethod
download_status.ts
Show active downloads with name, progress, and state
status
@keeb/mms/tv-recommendationsmethod
tv_recommendations.ts
Render the host's Watch/Skip verdicts for a tv_recommender.evaluate, aggregate_typesafe (or legacy aggregate) run — sorted by score within Watch, then Skip.
recommendationshost
05Skills
mms1 file
06Previous Versions
2026.10.25.1
Modified 3 behaviors
2026.10.23.1
Modified 3 behaviors
2026.07.18.1
2026.07.03.1
subtitles: extract_recent now writes a single run rollup resource per batch (window/cutoff, jobs considered, extracted/skipped/failed counts, plus per-file results and captured failures) instead of one extraction per .srt. Per-file failures no longer abort the batch.
organizer: One Piece special-case — 'One Piece - NNNN' files route to anime/completed/one-piece/s23 (current Elbaph arc) so Jellyfin resolves the correct arc name and art.
tmdb-lookup: new seasons method lists a show's seasons with cumulative absolute episode ranges, for computing episode->season breaks.
Modified 2 behaviors
2026.05.22.2
Added 1 behaviors. Added 1 reports
2026.05.22.1
Added 1 behaviors
2026.05.18.1
Add extract_recent batch method to subtitles for organize-and-clean wiring
Modified 3 behaviors
2026.05.06.2
Added 2, modified 4 behaviors. Added 1 reports
2026.04.27.1
Added 1 behaviors
2026.04.22.3
2026.04.22.2
Added 1 skills
2026.04.06.1
2026.04.02.3
2026.04.02.2
Modified 1 behaviors
2026.04.02.1
Added 1 reports
2026.03.29.1
07Stats
B
87 / 100
Downloads
147
Archive size
621.1 KB
- Has README or module doc2/2earned
- README has a code example1/1earned
- README is substantive1/1earned
- Most symbols documented0/1missing
- No slow types (deprecated)1/1earned
- Dependencies pass trust audit2/2earned
- Docs keep secrets in a vault1/1earned
- Has description1/1earned
- Platform support declared (or universal)2/2earned
- License1/2partial
no SPDX license declared
- Verified public repository2/2earned
Repository
https://github.com/keeb/swamp-mms08Platforms
09Labels