Publish to ghcr.io and simplify setup to compose + .env

[dean1850]

Installing jellylook meant cloning the repo and building the image locally, and the setup instructions were split between the README and a wall of # comments in .env.example — including a hashed-out Tautulli block users had to un-comment by hand.

A GitHub Actions workflow now builds and publishes a multi-arch image to ghcr.io/dean1850/jellylook, so installing is: download two files, fill in keys, docker compose up -d. Every comment is gone from .env.example and the explanations live in README tables instead.

Changes

• .github/workflows/docker-publish.yml — builds an amd64 image, boots it, and checks /health, the UI, /api/status and /api/settings, asserts the container's own HEALTHCHECK reports healthy, and asserts startup fails fast on missing keys. Only then does it publish linux/amd64 + linux/arm64. Tags: latest on main, semver on v* tags, short sha, and the branch name.
• docker-compose.yml pulls the published image. The local build path moves to docker-compose.build.yml for development. The ./data bind mount is unchanged, so existing installs keep their SQLite database.
• Dockerfile — OCI labels (links the package to this repo), tzdata so TZ actually affects log timestamps, a HEALTHCHECK, and DATA_DIR. Still runs as root, matching the deliberate revert in Revert Dockerfile to run as root #2.
• requirements.txt pinned so a given tag is reproducible.
• .env.example — all comments and the hashed-out Tautulli block removed. All three history sources are now listed uncommented; HISTORY_SOURCE alone decides which are required, so nothing needs commenting in or out.
• README.md — no-clone quick start, update instructions, a "building from source" section, and Configuration tables covering every variable, its default and where to get each key.
• .dockerignore added.
• RECS_PER_SCAN bug fix — see below.

Fixing RECS_PER_SCAN

Testing this migration surfaced a pre-existing bug: RECS_PER_SCAN in .env never did anything. init_db() seeded recs_per_scan=60 into app_settings, and _recs_per_scan() reads that table before falling back to .env, so the env var could not take effect on any install past its first run.

recs_per_scan is no longer seeded. A missing row means "use .env"; a row is written only when the user saves one in the Settings panel, so a deliberate panel choice still wins. Two supporting details:

• Saving the panel at the .env value clears the row rather than storing it. The panel POSTs recs_per_scan on every save, so without this, saving any setting once would silently pin the batch size and make a later .env edit look ignored all over again.
• GET /api/settings returns the effective value, so the panel shows the number a scan would really use instead of a hard-coded 60.

Existing databases already carry the seeded row, so a one-time migration drops it. A new schema_meta table records that the migration ran, so a value the user picks later is never deleted on a subsequent start.

Verification

Local end-to-end — 90/90 checks pass. The real FastAPI app was booted under uvicorn against stub Jellystat / Jellyfin / Tautulli / TMDb / OMDb / Seerr / LLM servers and driven over HTTP:

• full scan in both jellystat and tautulli modes, end to end
• already-watched titles excluded, TMDb-unresolvable titles dropped, post-enrichment dedupe
• library ownership flagging via Jellyfin provider ids and via Plex title keys through Tautulli
• metadata cache proven: a second scan made zero TMDb and zero OMDb calls
• paging, sorting (match/imdb/year), type filters, the film alias, and a rejected bad filter
• settings validation: clamping, 400 on a bad enum, no partial application, unknown keys ignored
• every Seerr route including 409-as-already-requested and upstream-500-as-502
• fail-fast startup: missing key exits non-zero and names the variable; bad LLM_PROVIDER rejected
• RECS_PER_SCAN=13 yields a 13-title batch; a saved override of 20 wins; saving 13 again returns control to .env; a simulated pre-migration database picks up .env while keeping a later user choice across restarts
• .env.example parses as a real .env, its values coerce to the right types, and an unfilled copy reports exactly the five keys the README says are required

In CI on this branch every smoke-test step passed and the multi-arch push succeeded. The published manifest is a clean OCI index:

linux/amd64  sha256:71765e51eab3d11d…
linux/arm64  sha256:687001046e60c638…

Confirmed pullable anonymously, so no registry login is needed. No unknown/unknown platform entries (provenance: false).

I could not build the image on my own machine — this environment's egress policy blocks Docker Hub's blob CDN — so CI is the authority on the image itself. Everything above the image layer was verified locally.

One correction testing forced

I had written a README warning that trailing # comments in .env break startup. I tested it: both Docker Compose and pydantic-settings strip them. The warning was wrong and was removed. Likewise, a note claiming forks' packages land private was removed — packages inherit the repo's visibility, verified with an anonymous pull.

Note

The workflow also triggers on claude/** branches, which is how the publish path was proven before merging. It means future branches publish a branch-tagged image (handy for testing, easy to drop — it's one line under on.push.branches). The claude-github-docker-ghcr-migration-x1x6pk tag can be deleted from the package once this merges.