home-ops: Skill for Claude Code

Tdarr libraries, flows and flow nodes

Relocated verbatim from AGENTS.md on 2026-09-01 so it loads only when this subsystem is in play. The text below is unchanged; only line breaks were inserted. AGENTS.md keeps a one-sentence pointer. Add new findings here or to the owning document, not back into AGENTS.md - see its "Maintaining this file" section for the rule.

• Tdarr's librariesToNotProcess is a Tdarr Pro feature and is a silent no-op on this unlicensed install - never use it as a scope boundary. It is stored in NodeJSONDB, rendered in the UI and returned by GET /api/v2/get-nodes, so it looks live; talos-3 carried it for Series and still transcoded a TV-Shows file on 2026-08-31. Both queue readers gate it on auth (getQueuedFiles.js: if (auth && librariesToNotProcess.length > 0); getStagedFiles.js: the same conjunction), and auth is authStatus(false) from Tdarr_Server/srcug/server/auth.js, which is a licence check that POSTs tdarrKey to tdarr.io - not a login. tdarrKey is empty here and POST /api/v2/auth-status {"data":{"saU":false}} returns false, which is the exact call the dispatcher makes. This also refutes any "the server pushes work past a node-side accept filter" theory: get-new-task is a node poll (the Server relay sending job to Node relay: <node> line is the reply), and there is no node-side accept filter at all. What does hold, with no auth check anywhere in qb/qbUtils.js + plugins/queueQueryFuncs.js, is the library-level toggles - processLibrary, processTranscodes, processHealthChecks and the per-hour schedule - which build the db NOT IN (...) clause for table1. Series is scoped out with processTranscodes: false (2026-08-31), which leaves health checks and scanning on. Proof method, and why the same file is simultaneously accepted by table4 and refused by table1: docs/tdarr-errored-remuxes.md §1.

• A Tdarr flow's customFunction node reads args.inputs.code; a node whose inputsDB key is function silently runs Tdarr's default stub, which returns outputNumber: 1 unconditionally. lib.loadDefaultValues fills the missing code with the built-in example, so the node is green, the job report even prints your real source (under "function", next to the stub under "code"), and nothing anywhere reports an error. All five customFunction nodes in movies_av1_nvenc_v1 were in this state - including the three guards that were supposed to reject a bad encode before Replace Original File - so every transcode replaced its source with zero verification, across all 6514 job reports. Confirm execution behaviourally, never by reading the flow: grep a job report for a string only your code can emit (Size check: <digit>), or check that a node returned an output number the stub cannot produce. Fixed by rekeying to code; mechanism, evidence and the still-open e_dv_bypass / br_*_bypass edges that make the DV and bitrate checks inert: docs/tdarr-errored-remuxes.md §2 and §3.6. Reviewable node sources and a read-only unit harness that runs them against the parked masters' real ffProbeData are in docs/tdarr/flow-nodes/.

• Never enable forceConform on Tdarr's Set Container node - it deletes streams rather than converting them. ffmpegCommandSetContainer/1.0.0/index.js sets stream.removed = true for mov_text/eia_608/timed_id3/data when targeting mkv, which on the parked 4K masters is 1-46 subtitle tracks destroyed per file in an irreversible in-place rewrite, with no guard in the flow noticing. The supported path is to convert: a customFunction between Set Container and Set Video Encoder that pushes ['-c:{outputIndex}', 'srt'] onto each mov_text stream's outputArgs, and marks removed only for data/bin_data and zero-dimension mjpeg cover art (the separate cause of Amelie's dimensions not set). Encoder-specific output arguments have the same shape of trap: -preset medium/-global_quality/-look_ahead are av1_qsv-only and made libsvtav1 fail at init on 88/88 CPU jobs, so cargs* now reads back the chosen encoder from the video stream's outputArgs and branches. Both proven live: docs/tdarr-errored-remuxes.md §3.