card-grader/CLAUDE.md

Card Grader

Self-hosted web app that estimates PSA trading card grades from photos, using vision-capable LLMs (Claude and/or GPT) as the grading engine. Runs as a single Docker container on the user's Unraid home server, exposed at hippofam.com/cards through Nginx Proxy Manager.

No framework: the whole backend is Python stdlib (http.server.ThreadingHTTPServer) plus sqlite3. Optional third-party packages are anthropic, openai, and Pillow (image processing only — never required for the server to boot).

Architecture

• app.py — HTTP routing/handlers. Reads Basic Auth username forwarded by the reverse proxy for attribution only (who graded what), never for access control — auth itself is enforced by Nginx Proxy Manager in front of the container via htpasswd. Admin-only settings are gated by ADMIN_USERNAME (CARD_GRADER_ADMIN_USER env var). Request bodies are capped (MAX_BODY_BYTES); oversized/malformed bodies get a 413 and the connection is closed rather than trusting a client-supplied length.
• vision.py — Provider abstraction over Anthropic and OpenAI vision APIs (_call_anthropic / _call_openai, dispatched from a common entrypoint). MODELS holds per-model pricing/capabilities. GRADING_SYSTEM is the large system prompt encoding PSA's actual grading rubric (centering tolerances, corner/edge/surface criteria, qualifiers — MC/OC/PD/ST/MK, half-grades, Altered Authentic) — verified against psacard.com, beckett.com, and cgccards.com, not guessed. Self-reported "high" confidence is capped to "medium" when neither centering nor edge whitening was independently measurable from the image geometry.
• cardimage.py — Pillow-based image pipeline: card boundary detection, corner/edge/surface crop generation, centering/edge-whitening/aspect-ratio measurement from pixels (not just left to the model's eye). Includes holder/toploader-aware detection (detect_card_box): a combinatorial search over candidate edge positions constrained to standard card aspect ratios, so the app can locate the actual card inside a slab/toploader rather than just detecting the holder's outline. When a holder is detected, edge-whitening measurement is refused (the holder's own inner edge corrupts that specific measurement) but centering still runs (border-width geometry, unaffected by clear plastic), and aspect-ratio measurement deliberately uses the unrefined box to avoid circularity (refinement selects for standard-ratio rectangles, so measuring the refined box's ratio would trivially always look "normal").
• store.py — SQLite (WAL mode) persistence. grades and grade_events tables, JSON columns for structured data, schema migrations via idempotent ALTER TABLE in store.init(). List/detail API responses explicitly exclude source_images_json to keep payloads small. DB path overridable via CARD_GRADER_DB_PATH.
• static/ — Vanilla JS/CSS/HTML frontend, PWA-enabled (manifest, service worker). "PSA slab" visual theme. __BASE__ is templated at request time so the app works correctly when served from a sub-path (/cards) behind the proxy.

Local development

python3 app.py

Needs ANTHROPIC_API_KEY and/or OPENAI_API_KEY in the environment to actually grade; the server itself has no required third-party deps. A local grades.db is gitignored and is scratch/test data only — it is not the production database (that lives on the server, see below).

Deployment

Target: Unraid server, container path /mnt/user/appdata/card-grader, built from Dockerfile / docker-compose.yml in this repo. Non-root container user (99:100 / nobody:users), healthcheck, log rotation, init: true. Reverse-proxied by Nginx Proxy Manager at hippofam.com/cards with per-user HTTP Basic Auth.

Current deploy mechanism (as of 2026-08-25): manual. Changes are copied to the server by hand (scp) and the container is rebuilt over SSH; the two copies of the repo (local Mac, server) are kept in sync by committing on both sides. There is currently no automated push-to-deploy path connected — see below.

Superseded approach, fully torn down: git-shell restricted deploy. A carddeploy service account (login shell git-shell, one bare repo, a post-receive hook triggering a single root-owned deploy script via one narrowly-scoped sudo rule) was built, tested, and verified working end-to-end — then explicitly torn down at the user's request in favor of Forgejo (below). Nothing from it remains: the account, sudoers rule, bare repo, /boot/config/go persistence block, and the sshd_config AllowUsers addition were all removed and verified reverted. The local git remote and SSH keypair for it were deleted too. If this pattern (restricted git-shell account + sudo-scoped deploy script) is useful as a reference for scoping other access requests later, it's a solid model — it was abandoned for tooling reasons (moving to a real git host), not because it didn't work.

In progress: Forgejo. Decision made 2026-08-25 to self-host Forgejo (soft-fork of Gitea) as the git remote and CI/CD trigger, replacing the git-shell approach entirely. Plan, run by the user themselves (not by Claude — user explicitly wants to execute server-side setup steps personally):

1. Forgejo itself runs as its own Docker container on the same Unraid box (/mnt/user/appdata/forgejo), proxied at git.hippofam.com via Nginx Proxy Manager.
2. This repo gets a forgejo remote and is pushed there (full history, not a fresh copy).
3. Forgejo Actions (instance + repo level) is enabled, and a self-hosted runner is registered at the repo level (not instance-wide) as its own container (/mnt/user/appdata/forgejo-runner), with the host Docker socket mounted in so it can rebuild the card-grader container.
4. A .forgejo/workflows/deploy.yml workflow (checkout + docker compose up -d --build on push to main) does the actual deploy.

Known tradeoff, not yet resolved: unlike the git-shell mailbox (which could only ever trigger one hardcoded script), a CI runner executes whatever a workflow YAML says — mounting the Docker socket gives it control over any container on the box, not just this one. The access boundary becomes "who can push a workflow file to this repo" rather than a sudoers rule. If tighter isolation is wanted later, a docker-socket-proxy in front of the runner (scoped to just the card-grader compose project) was proposed but not yet built — revisit if this matters more once the basic pipeline is working.

Status: walkthrough delivered to the user 2026-08-25; not yet executed. Do not assume any of this is live — confirm current state before relying on it or making claims about deploy mechanism to the user.

A prior root SSH key/access to the whole box was explicitly revoked by the user before any of the above; every access grant since then (including a one-time temporary root key used only to run the git-shell setup script, itself removed immediately after) has been narrowly scoped and time-boxed on request. Treat that as the standing expectation for any future server-access ask on this project — default to the smallest scope that does the job, time-box it, and confirm teardown/persistence explicitly rather than leaving it open-ended.

Conventions and constraints established for this project

• PSA grading rules must be verified against a real source (PSA's own site, Beckett, CGC) before being encoded into GRADING_SYSTEM — this rubric has had real accuracy bugs (e.g. a self-contradictory centering tolerance, a missing 5% front-centering leeway rule for grades 7+) found and fixed through actual verification, not assumption.
• The vision prompt's cannot_assess escape hatch is deliberately narrow: surface condition genuinely can't be separated from dust/scratches on a toploader/slab's plastic, so it keeps the escape hatch. Corner and edge geometry (sharpness, whitening, chipping) does read through clear plastic and must be judged normally — an earlier prompt version gave the model an escape hatch for corners/edges too, which caused false "cannot tell from photo" results on cards that were plainly visible through the holder. Don't reintroduce that.
• No pre-flight/"precheck" step before a paid grading call. This was tried and explicitly rejected by the user — their workflow is screenshots that can't be pulled out of the toploader, so a pre-check step doesn't fit and was reverted.
• No browser automation (Playwright or otherwise) to scrape eBay listings. Tried a plain curl-based fetch once (not Playwright) and got an immediate 403; iterating on headers/fingerprinting to get past that would be bot-detection evasion regardless of which tool performs it, and is out of scope. This feature is paused; if revisited, the legitimate path is a client-side bookmarklet/extension that uses the user's own authenticated browser session rather than a server-side fetch.
• Never trust the reverse proxy's forwarded auth header for anything beyond attribution (whose name to log against a grade) — access control is the proxy's job, not the app's.