Troubleshooting

Auto-discover can't find the Fetch box

The container has to run with host networking (the example compose already does this). Fetcharr finds the box by listening for the announcement it broadcasts on your network (a protocol called SSDP), and those broadcasts don't reach across Docker's own private network, so the container has to share the host's network to hear them.
The host has to be on the same part of the network as the box; those broadcasts don't cross between subnets without extra setup.
You can always enter the box's IP and port manually in Settings instead.

It needs Plex's Preferences.xml bind-mounted into the container (PLEX_PREFS_PATH), which only works when Plex runs on the same host.
Paste the token manually instead; grab it from app.plex.tv (or Plex's own support article on finding your token). See Plex.

That's deliberate: Fetch reports the wrong file size while a recording is still going, so Fetcharr refuses to download it rather than save an incomplete file. It syncs on the next run after the recording finishes.

The download came up short of the size Fetch reported. The next sync picks up from where it stopped (using HTTP range requests), so a partial normally sorts itself out.

Symptoms: "No I_AM_ALIVE reply" or "Timed out waiting for I_AM_ALIVE handshake".

The reply comes from your Fetch box via Fetch's cloud, and a box whose cloud session has dozed off misses the ping even though it works fine on the LAN. Fetcharr pings twice (20 s) before giving up, and the first attempt usually wakes the box's session, so just retry the delete after a moment.
If it keeps failing, open the official Fetch mobile app: if the app can't see the box either, the box↔cloud link is down; restarting the box resets it. See Delete from Fetch and the deep dive.

A side-effect of host networking: Fetcharr isn't on any Docker bridge network. Reach it via the host's LAN IP and FETCHARR_PORT instead.

Ad detection is educated guessing, never perfect. Comskip's accuracy on Australian free-to-air varies a lot by channel (logo detection, silence thresholds, and break lengths all differ).
Run the show in DETECT mode first and check the break counts and minutes it reports on the Recordings tab before switching to CUT. Scans work the CPU hard: budget ~30 minutes per 75-minute recording on a home NAS.
Cuts land on the nearest keyframe, so a second or two either side of a break is normal.
To tune detection, place your own comskip.ini in the /config bind mount; it overrides the bundled Australian-tuned default. Every cut keeps a <file>.ts.orig backup for the retention window, so if a cut goes wrong you can rename the .orig back to recover it. See Ad removal.

Set TZ in your .env to your IANA timezone; the UI shows every timestamp in the container's zone, whatever device you're browsing from.