System resources
Server total from host counters; bucket totals start from accounting deployment.
Stops active workers, health checks, probes, player streams, and fallback worker processes before restart or shutdown actions.
Reading the active V6 system map from the installed docs file.
V5 queue and resolver flow, adapted for V6 variant cataloging. The worker searches, probes, recovers, verifies, records metadata, then removes temporary media.
Shared provider lanes across season and movie completion workers. Up to four providers pull FIFO lookup work while each provider stays one-at-a-time.
Slot file: --
Real outbound subtitle-provider requests from dialogue verification. Cooldowns and local subtitle-cache hits are not counted as API requests.
Last 24 hours
Waiting for telemetry.
Season packs resolve through the V5 flow, then the best 720p, 1080p, and 4K variants are probed, verified, cataloged, and cleaned up.
Sibling workers share the queue and rotate VPN lanes 1, 2, and 3.
Search, AI checks, torrent probing, recovery, dialogue, and cleanup.
No log output yet.
Next 0 claimable targets · 0 deferred later.
Complete-series rows outrank regular queue items; fewer seasons left means a higher score. Due retries age with queued rows, and only one season per show runs at once.
Fresh successes written to DB Library, independent from noisy worker progress.
Recent completed, failed, cancelled, or deferred worker runs.
Movie completion uses the same variant catalog contract, with temporary recovery only long enough to verify and record the best magnets.
Sibling workers share the queue and rotate VPN lanes 1, 2, and 3.
Search, AI checks, torrent probing, recovery, dialogue, and cleanup.
No log output yet.
Next 0 claimable movies · 0 deferred later.
Fresh movie successes written to DB Library.
Recent completed, failed, or cancelled movie runs.
Idle dispatcher for movie or series requests. It keeps a dedicated retry ledger and only stops a request after discovered options are exhausted.
Search local title catalog.
Single always-waiting dispatcher slot for future special requests.
Waiting for the first special request.
No special request work yet.
Next 0 claimable requests - 0 deferred later.
Recent dispatch attempts, transient deferrals, and final exhaustion decisions.
History pills stay short on purpose. Use this reference to translate each code into the likely cause and next action.
This reference documents what every live panel on the Completion page means and the rules the worker follows before a season or movie is allowed into the catalog.
Queued rows become claimable when not deferred and not blocked by another active season for the same show.
Provider conveyor lanes collect candidates while each provider stays one-at-a-time.
Magnets are ranked, metadata is fetched, files are selected, and qBittorrent/libtorrent readiness is checked.
Episode coverage, dialogue, subtitles, and boundary cases must match the canonical target.
Verified media is segmented, cataloged, backed up, and tied to variant metadata.
Temporary downloads and worker ownership are removed unless the run is intentionally deferred.
Worker Slots are live sibling processes. Each slot reports its lane, stage, actual status, active target, PID, progress, working bucket, and bucket outcomes. The Flow Log is the detailed resolver trace for the selected slot. Queued Targets and Queued Movies are split into claimable rows and deferred-later rows so due work does not get hidden behind future retries.
Season queue ordering gives complete-series work priority over ordinary rows, then favors shows with fewer seasons left. Due retries age with normal queued work. The worker avoids running more than one season from the same show at the same time, which keeps cross-season qBittorrent ownership and catalog writes from colliding.
Search jobs can come from multiple source profiles. The worker records provider key, lane id, queued task count, active owner, waiting owners, heartbeat time, and parse quality so the page can distinguish a genuinely quiet provider from a busy conveyor or a blocked fetch path.
A combined file such as an Office finale can display once while covering both catalog episodes. The visible app row uses the first catalog episode number, later covered catalog rows are hidden from the normal list, and direct playback or download requests for any covered episode resolve to the same source asset.
For seasons, the gate reads conveyor_episode_numbers after mapping. That means a 2-in-1 source covering E26-E27 counts as both E26 and E27 present, while a pack that never safely covers the finale remains rejected even if its release numbering looks close.
The worker tracks first-byte timeout, selected-file progress, front-buffer readiness, no-growth windows, sparse/preallocated file parseability, qBittorrent API reachability, and lane/network mode. Recovery actions can reapply priorities, reannounce, recheck, rotate lanes, or defer when the evidence says the source is likely recoverable later.
When subtitle embedding is active, packaging requires a dialogue summary. Provider outages can mark the run deferred with a retry time instead of failed. The persisted dialogue result records which IMDb id matched, which matters for combined episode files that legitimately cover more than one catalog row.
The catalog keeps IMDb rows unchanged. Coverage metadata lives on top of package and selected-file records, so the app can show a clean E26-E27 row while old direct episode requests still resolve. Finalization state is preserved across route changes and clean stops; temporary downloads are cleaned only after cataloging or safe failure handling.
Use the Flow Log when all contenders look bad: it shows whether denial came from search quality, missing episode coverage, qBittorrent readiness, dialogue mismatch, subtitle-provider deferral, packaging failure, finalization, admin cancel, or health verification. That distinction matters because only some failures should be retried immediately.
FIFO repair queue for confirmed health issues. Slow-probe repairs replace only the affected resolution bucket, avoid the old magnet/infohash, and defer unresolved work for seven days.
Default one enabled slot. Extra siblings can be enabled; claimed jobs alternate over VPN lanes 4 and 5.
Waiting for repair work.
No repair work yet.
Next 0 claimable repairs - 0 deferred later.
Recent verified replacements, deferrals, failures, and cancel decisions.
Repair Work is the controlled replacement path for catalog variants that Health has already proven unsafe, slow, incomplete, or no longer streamable. It repairs the exact damaged bucket, keeps unrelated catalog rows intact, and now compares the replacement against the old variant before final acceptance.
Confirmed slow, failed, or damaged health evidence creates a deduped repair queue row with the old identity attached.
The worker claims FIFO work and pins child completion to the assigned repair VPN lane.
Only the affected 720p, 1080p, or 2160p bucket is rebuilt; the old magnet and infohash are blocked.
The replacement must appear in DB Library with matching kind, title, season, bucket, and a changed torrent identity.
The same streamability probe measures the old and new variants before the queue row can be marked done.
Better replacements are finalized; weak, inconclusive, or missing comparisons are deferred instead of silently taking over.
Repair is not a general completion backlog. Health has to provide the reason, the target kind, IMDb or TVDB identity, the bucket, and old torrent evidence such as magnet URI, infohash, selected files, or runtime probe details. That keeps the repair queue narrow and auditable.
Lane assignment alternates across lanes 4 and 5 so repair qBittorrent work does not collide with health probes on lanes 6-9 or the user/config-prober streamer on lane 10. The queue row records slot, process, stage, attempts, and heartbeat data so Clean Stop and Hard Cancel have visible state.
The child worker is expected to finish as a normal terminal completion run. Repair then reopens the catalog snapshot and confirms that the expected bucket exists, maps to the same title or season, and no longer points at the old identity before any success is considered.
The comparison writes compact old and replacement probe summaries into comparison_result_json. A healthy old variant can block acceptance if the replacement does not beat it by the configured margin, which prevents repair from swapping a known source for an equal or weaker one.
The history list is the audit trail: status, stage, attempts, old infohash, replacement identity, and comparison details explain why a row finished or waited. A deferred row is not ignored; it is deliberately parked until the next safe retry window.
Use Hard Cancel when a repair child is stuck at the process level. Use Clean Stop when the worker is behaving but you want it to drain safely. If a row looks stuck after child completion, check for comparing_streamability, comparison_status, and comparison_result_json in the history entry before retrying.
Live VPN lane roles, active leases, qB ownership, and libtorrent daemon workload.
The service polls user activity every 1 second. Testing starts only after no active playback, player presence, foreground app presence, or recent authenticated app activity for 300 continuous seconds.
Any active playback or fresh app/player presence cancels testing, kills active probe children, applies the best unused lane-10 config, and starts the lane-10 streamer.
It selects 3 fixtures once per local day from health/catalog rows declared healthy in the last 24h. Pinned magnets only fill gaps.
The daily set is a stable weighted sample for that local day, biased by recent speed, seeders, and peers so the same 3 are reused fairly across all configs.
For each lane-10 config, it stops the user-lane streamer, preserves port-forward data, writes the config, prepares v6lane10, then starts 3 parallel torrent probes.
Each probe prioritizes the selected file front and must reach 16 MiB of contiguous verified bytes from offset 0. Timeout is 240 seconds per config.
Probe/user lane settings are min_start_buffer_mib 16, max_startup_buffer_mib 192, urgent_window_mib 192, and lookahead_mib 8192.
A config is top-eligible only if all 3 torrents succeed against today's fixtures. Ranking is lowest worst time, then lowest mean time, then highest throughput.
At local midnight the new day starts by retesting yesterday's top 10 first. Old rows stay historical until those configs prove themselves again against today's fixtures.
After a successful sample, qB lanes 1-5 and health lanes 6-9 outside the top set are blocked from new work, drained cleanly, switched to unused top configs, verified, then unblocked.
Continuous metadata, identity, seeds, peers, and availability checks.
Continuous stream-speed checks on dedicated health lanes 6-9.
All saved health rows. Use this for inspection; queues and histories above are the operational view.
No log output yet.
This is the detailed operating reference for the Health system: how rows enter the queues, how workers claim them, how lanes are leased, how checks are judged, and exactly what each state means.
The worker maintains a local health inventory in variant_health. Each saved completion variant can be checked by the Swarm FIFO worker and the Speed Probe FIFO worker independently, which lets cheap metadata checks keep running without blocking expensive stream probes.
Health starts by syncing completion catalog variants into variant_health. The row stores title, target IDs, infohash, selected files, saved swarm numbers, resolution bucket, latest state, claim tokens, probe due time, and check counters.
The swarm path checks torrent identity and swarm viability: metadata, selected-file match, seeders, peers, availability, distributed copies, and whether the saved magnet still represents the expected media.
The probe path checks whether a variant can actually stream fast enough for its resolution bucket. It warms up, measures throughput, computes average speed over the probe, and compares that average against the required Mbps gate.
Each finished check writes a check_runs row with check type, result state, lane summary, previous values, current values, speed statistics, VPN origin, errors, and any requeue result.
Worker wakes on its loop interval, a manual Wake FIFO action, or a start/restart. The default loop interval is 300 seconds and a full pass is normalized around 24 hours.
The catalog is mirrored into variant_health so the worker has local rows to claim and can preserve health data even while the completion catalog changes.
Swarm claims use swarm claim fields; probe claims use probe claim fields. FIFO ordering prefers never-checked rows first, then the oldest checked rows, while probe rows also respect next-due timing.
The worker asks the lane manager for a lane lease owned by completion-health-worker. If no lane is available, the row is unclaimed and the slot reports a waiting state.
The worker runs the qB/torrent probe subprocess for either swarm or speed_probe, updates the slot card while running, and classifies the observed result.
The worker updates variant_health, inserts check_runs, clears claim fields, updates the slot result, and releases the lane lease even on errors.
Swarm checks are the cheap, continuous truth checks. They decide whether the magnet still has usable torrent metadata and whether the saved swarm is alive enough to keep the variant trusted.
The queue orders rows by never swarm-checked first, then oldest swarm_checked_at_utc. Each active slot owns one claim token until it finishes or returns the row to the queue.
The metadata wait has a 180 second timeout. After metadata is usable, the swarm is allowed to settle for 90 seconds before judging seeds, peers, availability, and distributed copies.
A clean dead result first becomes suspect_dead. The worker only treats it as confirmed dead when the prior state already supports that suspicion, reducing one-bad-lane false positives.
If a row was stale-dead but the swarm later appears alive with complete availability, the worker can repair the stale state back toward probe work instead of leaving the variant buried.
Probe checks answer one question: would playback stay ahead if someone pressed play? The primary judgement is buffer lead from contiguous selected-file bytes, not raw average Mbps. Mbps is still recorded as supporting telemetry and as a fallback if piece telemetry is unavailable.
| Bucket | Required fallback | Buffer-lead rule | Notes |
|---|---|---|---|
| 720p | 5.0 Mbps at 24 fps | Healthy at 180s lead; borderline at 60s lead. | Fallback bitrate converts contiguous bytes into playable seconds when duration is unknown. |
| 1080p | 10.0 Mbps at 24 fps | Healthy at 180s lead; borderline at 60s lead. | This is the normal default bucket when resolution is unknown. |
| 2160p | 35.0 Mbps at 24 fps | Healthy at 180s lead; borderline at 60s lead. | Also used for 4K and UHD labels. |
The normal probe warms up for 60 seconds, can extend warmup up to 90 seconds, then measures for 300 seconds. Slow reprobes use a longer 450 second measurement window.
buffer_lead_seconds >= 180 passes. If selected-file progress is already at least 98%, the probe can also be treated as ready because the file is effectively present.
buffer_lead_seconds >= 60 is borderline when the lead is not falling. This means playback has usable headroom, but not enough to be called healthy.
A slow result means contiguous playback lead is below 60 seconds, falling, or unavailable and the Mbps fallback also failed. Slow rows are scheduled for reprobe after 3600 seconds rather than being immediately discarded.
Slow quarantine requires confirmation from 2 distinct lanes. This prevents one bad VPN lane from deleting or replacing a variant by itself.
If verified resolution buckets are missing, the worker can nudge completion search after every 3 relevant checks so the catalog keeps trying to fill the missing bucket.
Health uses lane-manager leases with owner completion-health-worker. Swarm leases use task type health-swarm; speed probes use health-probe.
The probe policy is dedicated_lanes6_9_parallel_health. Lanes 6-9 run health swarm and speed checks through persistent libtorrent daemons with configurable parallel slots for both worker types.
Each slot card shows the claimed row, stage, lane, result state, speed, and message. A slot returning to idle means the claim was cleared and the lane lease was released.
Clean Stop asks the worker loop to stop after current work reaches a safe boundary. Running subprocesses are allowed to finish instead of being killed mid-check.
Hard Cancel is the emergency path. It marks the worker for cancellation and is expected to stop active work aggressively when the lane or subprocess is wedged.
The dashboard can add or remove sibling worker slots. Swarm and probe workers have independent slot counts and independent maximums, so one side can be scaled without changing the other.
The row passed its current health requirement. For probes, the average speed met the bucket gate or the selected file was effectively complete.
A worker slot owns the row right now. The matching claim token should clear when the check persists its result.
The variant was not a clean healthy pass, but it showed enough near-required speed to stay visible as borderline instead of immediately slow/dead.
The swarm exists but its observed seed/peer/availability evidence is weak enough that operators should treat it as risky.
The measured average speed missed the required gate. The row is retained for reprobe and may later be quarantined only after distinct-lane confirmation.
The magnet appears unusable, mismatched, or no longer represents the saved media. Confirmed dead or changed rows can trigger catalog reset and requeue work.
The worker hit an exception or probe failure that prevented a clean classification. Previous usable health can be preserved for transient errors.
The row exists in inventory but has not yet completed the relevant swarm or probe check.
| Store | Purpose | Important fields | How the dashboard uses it |
|---|---|---|---|
| variant_health | Current inventory row and queue state. | Health state, swarm state, probe state, claim tokens, next due time, speed stats, swarm stats, errors, check counts. | Inventory table, FIFO queue summaries, counts, filters, slow/borderline/dead totals. |
| check_runs | Append-only history of individual checks. | Check type, result state, previous JSON, current JSON, lane summary, VPN origin, required speed, measured speed, average speed, sample count. | Recent history panes, 24h check counts, average swarm/probe durations, debugging evidence. |
| worker slots | Live slot readout for active workers. | Slot index, state, label, resolution bucket, lane, stage, result, current speed, message. | Active slot cards for Swarm FIFO and Speed Probe FIFO. |
| worker flags | Control-plane state for start/stop behavior. | Worker PID, state, stop requested, hard cancel requested, permanent-off state. | Top status strip, On buttons, Clean Stop, Permanent Off, Hard Cancel. |
Transient error preservation is intentional. Prior useful states such as healthy, borderline_probe, slow_probe, and weak_swarm are not casually erased by a single temporary failure.
Health state, FIFO pressure, speed-probe quality, and today throughput without mixing controls into the readout.
Waiting for live Health worker status.
Waiting 0 | today 0 | forecast --
Waiting 0 | today 0 | forecast --
No forecast data loaded.
No forecast data loaded.
Swarm and speed probes stay separated so scaling, stopping, and emergency actions are easy to scan.
Metadata, identity, seed, peer, and availability checks.
Stream-speed checks, buffer lead, and slow-probe reprobes.
Active slots, FIFO pressure, and recent outcomes are separated by worker type.
Saved health rows with streamability, swarm state, lane outcome, quarantine, and requeue context.
Worker DB, flags, lanes, schedule normalization, and raw log.
No log output yet.
Full Health Worker Execution Manual with architecture, lifecycle, queue rules, states, storage, and operator guidance.
Passive dry-run pressure monitor for V6 completion downloads. It records what the future controller would pause or allow, but enforcement is off.
Waiting for storage snapshot.
Items are sorted like the eventual controller: real completion work before probes, then smallest remaining payload first.
No active decisions yet.
No pauses recommended.
Queue-less janitor for disposable completion artifacts across the bulk roots and worker-run state roots.
Sweep loop, live reference collection, and deletion decisions.
No log output yet.
Loading the disposable cleanup roots.
Last deleted items across recent sweeps.
Rules that explain how raw title metadata is interpreted before the app, dailies, completion queue, and DB Library decide what is visible or complete.
What the system does before, during, and after database mutations.
Domain checks that catch structurally valid but logically broken data.
Daily 10:00 dataset ingestion into the title catalog used by completion workers, dailies, TV search, and mobile APIs.
Snapshot, prune, write-gate, verification, and restore decisions.
Household access tokens, viewing profiles, and root-backed player cache ownership.
Create a reusable 6-digit household access token.
Reusable household access tokens and their claim state.
Profiles own watch history, active streams, prefetched files, cached movies, and cleanup timing.
Watching and prefetch slots currently attached to the shared player streamer.
| User | Slot | Title | State | Swarm | Speed | Lane | Cache |
|---|---|---|---|---|---|---|---|
| No live streams. | |||||||
Live and historical mobile playback, downloads, prefetches, successes, failures, cancellations, and cleanup outcomes.
Last 24 hours | exact events plus inferred legacy history.
Current watching, downloading, prefetching, and app/player presence by user.
| User | Type | Title | State | Progress | Speed | Swarm | Updated |
|---|---|---|---|---|---|---|---|
| No live user activity. | |||||||
Exact activity events recorded after this dashboard update.
Best-effort rows from older stream, progress, and scope tables.
Only titles currently visible in an open Android app appear here, plus explicitly labeled startup and Continue Watching session scope. Reports update in place.
This is the operating contract for User / Global Scope. The system keeps likely playback torrents peer-hot for the active Android user without intentionally downloading media bytes before Play. It is not the old global pin panel, not a byte prebuffer scheduler, and not a root-pressure control surface.
startup_prediction starts when the app foregrounds. The server predicts the first four non-empty App Layout rows before Android finishes rendering.POST /api/mobile/user-scope/visible-tiles after Compose measures visible tiles.inactive scope means the app closed, backgrounded, settings opened, or the session is otherwise no longer browsing. It clears scope immediately.daily_picks.position rows used by Android home.sort_mode, tile size, and tile ratio all affect what can be predicted.S01E01.metadata_only=true, peer_hot=true, and prebuffer_only=false.discovering means queued or resolving; metadata and peers are not ready enough yet.metadata ready means torrent metadata and file selection are known, but the alias has not proven a peer-hot runtime state.peer hot means metadata-only peer discovery is alive and maintaining swarm information.activated means a matching warmup is attached to a real active player stream and the Android app has sent fresh player_active=true presence. It should not appear for idle browsing.failed means the warmup hit a terminal error and will not be used for handoff.Daily home rows picked from verified DB Library titles for the Android app.
Configure Android home rows, category feeds, and layout controls independently from today's generated dailies list.
Tile rows drive generated Android home rows. Banner rows drive the featured slideshow. Continue Watching rows are personalized per account and only appear in Android after watch history exists.
Upload CSV files to create category pools. Dailies only choose from categories that are selected on active App Layout rows.
category_name and parent_imdb_id. Aliases also work: category_title, category, or name; and imdb_id or parent_id.
One title per row. Parent IDs must be IMDb title IDs like tt0068646. Optional columns: position, title_hint, and notes.
Example: category_name,parent_imdb_id,position,title_hint then Crime Favorites,tt0068646,1,The Godfather.
Large files are fine: the browser uploads the file directly and the server parses the cached CSV.
Configure independent Samsung TV home rows. Poster rows and banner slideshows use CSV categories; Continue Watching is personalized per account.
Stored server cache for trailer files, transparent logos, posters, banners, and Android episode stills. Downloads use the original cached files.
24/7 cache completion for verified DB Library titles over VPN lanes 1, 2, and 3.
No log output yet.
Cached trailer videos used by the Android detail hero.
Stored transparent logo assets from the persistent artwork cache.
Stored posters, banners, continue banners, and episode stills.
Everything on this page is built from verified DB Library titles, persistent server cache rows, and a worker that fills missing artwork without making the Android app wait on expensive searches.
Fetch poster, portrait banner slots, and the default banner before any logo work so Android rows have a visual fallback.
Resolve transparent title logos, normalize them to 192px height, OCR-gate them, then mirror the same clean source into full and small logo rows.
For complete series, cache stills only for concrete IMDb episode IDs and official episode numbers that belong to visible seasons.
Create 16:9 continue-watching banners from textless still frames so resume rows never depend on portrait artwork.
Download trailer media after non-trailer artwork so heavy video work does not starve posters, banners, logos, or stills.
banners → logo → episodes → continue → trailer.v6lane1, v6lane2, and v6lane3.LOGO_NORMALIZED_HEIGHT, and rejected if it is black-only.logo-ocr-h192-v4 intentionally makes old logo rows stale after logo size or OCR policy changes./var/lib/v6-server/artwork-cache/persistent/{imdb}/; trailer files live under the bulk media root.full-logo-h192-ocr-v4 and small-logo-h192-ocr-v4 force phones to refetch corrected logos.tt episode IDs.Start keeps the background worker alive; Sweep Now asks the running worker to rescan immediately.Clean Stop lets in-flight work wind down; Hard Cancel terminates now and marks running tasks interrupted.Wipe Logo PNG Cache clears logo rows only; Wipe All Artwork Cache clears image rows and files so the worker rebuilds from policy.Immortal Man for a Peaky Blinders target, then retries another candidate instead of storing a mismatch.Completion-worker triage for search, LM, magnet, dialogue, torrent, packaging, legacy finalization, and runtime failures.
Runtime knobs for completion search, subtitle providers, and VPN lanes. Completion and dialogue settings are saved to the worker runtime file so they survive dashboard restarts and are picked up by new worker runs.
The season and movie completion workers use these sources when searching for torrent candidates. Each row maps to the simulator source format: home URL, optional search URL template, and parser profile.
Used by the completion worker to verify selected media against episode/movie dialogue before cataloging and packaging.
This only changes qBittorrent payload downloads for season/movie completion. Search pages, APIs, subtitle providers, metadata fetches, health checks, and player streams keep using VPN lanes.
Keeps completion siblings low while mobile or TV users are active, then opens extra season/movie slots after the quiet window.
Lanes 1-3 run completion and artwork work. Lanes 4-5 run repair qBittorrent work. Lanes 6-9 run health libtorrent checks, and lane 10 is reserved for user playback and config proving.
Apply one AirVPN WireGuard zip to selected lanes and update forwarded torrent ports in one save.