Skip to content

The web app

The web app is a SolidJS single-page app, built into the image and served by judgebot at http://localhost:8787. It is laid out as a conversation: your questions on one side, the judge’s answers on the other, and the question box docked at the bottom. It runs the same pipeline as the Discord bot and shows the same answer:

  • the ruling
  • its citations, in one section that shows each one’s label (a rule, a ruling, Oracle text or a prior call) and opens to every verbatim quote, each linked to its source
  • the cards the question was resolved to
  • the confidence and the CR version.

An answer naming a card that is not out yet, and that the loaded Comprehensive Rules may not cover, opens with the same note as on Discord (unreleased cards). The feed shows the note on a stored answer as it was when the answer was given.

A reply that is not an answer (busy, rate limited, a card not found, an error, the maintenance notice) is a titled notice in the judge’s message. The header links to this documentation and has a theme switch (Auto, Dark, Light). The footer names the source repository, the commit the instance was built from and the operator’s support address (JUDGE_OPERATOR_EMAIL, without which the web app does not start). A line above it gives the Comprehensive Rules release loaded and how long ago the data was last refreshed, for example Comprehensive Rules 2026-09-25 · refreshed 3 hours ago.

Answers are written in the light Markdown Discord renders (bold card names, rule numbers in backticks, occasional italics). The web app renders those three marks and nothing else.

A name that could mean several cards gets a “did you mean…?” choice. Questions asked in one browser session share history, so a follow-up works. Before the first question the page offers four examples. Choosing one fills the box but does not send it, because an answer spends the operator’s money and one of the visitor’s rate-limited questions.

Nobody is logged in there. Nothing about the visitor is stored beyond a random session id, one per browser tab, that groups their questions and their ratings. Their questions are listed on the feed without a name. The browser itself remembers the theme, the feed’s layout and whether to count web ratings (localStorage); none of them is sent to the server.

An answer in the chat, on the feed and on a call’s page has three buttons: Correct, Partially correct and Incorrect.

  • Correct is recorded at once.
  • Partially correct and Incorrect open a form under the answer, which the keyboard reaches first: the kind of problem (required), What’s wrong and The right answer (up to 1000 characters each) and a name (up to 100), all three optional. The reason goes to the operator’s feedback queue beside the question and answer; the judge never reads it.
  • Rating an answer the tab already rated asks first, under the buttons: Replace or Keep. Correct over Correct says it is already recorded.

A web rating is anonymous, so it never shapes which past answers the judge sees as examples; only Discord’s ratings do. Each rating keeps the address and browser it came from for FEEDBACK_IP_DAYS (default 90), for dealing with abuse, then loses both. Ratings are limited to 30 per address per ten minutes, apart from the questions’ limit.

Under each answer on the feed and its page, the counts are Discord’s. Include N from the web adds the web’s, a choice this browser remembers. The web’s are hidden by default because anyone can inflate them: a new tab is a new rater.

When the operator sets JUDGE_PUBLIC_FEEDBACK=true, the reasons raters gave are listed there too, with the kind of problem and their words, never who gave them; the form says so before a rater writes one. A reason the operator dismisses in the feedback queue leaves the public pages. Rating needs --api beside --web, as asking does.

/feed lists answered questions from every interface, newest first: Discord, the web app, the MCP judge tool, judge-cli and agent sessions. Each entry shows where it was asked, when, the Discord username of whoever asked (anonymous for the web, the agent tools, calls from before 2.0.0 and anyone who used /forget), the question, the answer with its citations and cards, how it was rated, and a note when its citations no longer hold.

  • Only a question asked with /judge-private (or an agent call made private) stays off it. A Discord question asked in a channel few people can see is still listed. Servers and channels are never shown, and a mention in a question shows as @… or #….
  • /feed?interface=discord (or web, mcp, cli, agent_session) lists one interface’s.
  • The search box finds answers whose question, answer or asker’s username contains what you type (/feed?q=…); @name lists one Discord user’s (/feed?user=name).
  • Compact (the default) shows one line per question, which opens in place to the whole exchange; Expanded shows every exchange in full.
  • /calls/<id> is one answer’s own page, and /users/<name> one Discord user’s public questions. A username someone gave up and another took lists only the newer holder’s questions once they ask one publicly; the earlier holder’s stay in the feed.
  • JUDGE_FEED=false removes the pages and their endpoints. JUDGE_FEED_DAYS hides calls older than that many days from all of them, without deleting anything (JUDGE_CALL_RETENTION_DAYS deletes).

Turning the feed on lists every earlier public call too, including those asked before it existed.

/dashboard counts what the feed lists: every question not asked private, within JUDGE_FEED_DAYS when that is set. Private questions are never counted, in any figure.

  • The number of questions, of Discord askers, and the share answered.
  • Questions per day by interface, as a chart (weeks above three months). Show as a table lists the same figures as text.
  • How the questions went: answered, a “which card?” question, a card not found, out of scope, declined by the model, or failed.
  • The ratings: correct, partially correct, incorrect. Discord’s are counted; Include N from the web adds the web’s, the same choice as on the feed.
  • The most-asked topics, cards and Discord askers. A card opens a feed search for it, an asker their page, and an interface its part of the feed.
  • The range: 7, 30 or 90 days, or a year (/dashboard?days=90, or ?from=YYYY-MM-DD&to=… for any range up to 366 days). The interface chips narrow every figure to one.
  • /dashboard?user=name is one Discord user’s figures, linked from their page (/users/name). Outcomes are not counted per user: a question that was not answered records no asker.
  • With JUDGE_PUBLIC_COST=true it also shows what the answers cost, in all and per question. With JUDGE_PUBLIC_MODELS=true it shows which models read and answered the questions, and narrows to one model. Both are off by default.

The dashboard is there wherever the feed is, and JUDGE_FEED=false removes both.

The web app is the --web role of judgebot. --api serves POST /api/judge, the route the web app asks its questions through, and the MCP transport is --mcp. Any of the three also serves GET /api/health and GET /api/about. The compose service runs --discord --api --web --jobs unless JUDGE_ROLES in .env says otherwise, so a scripts/deploy.sh deployment serves the web app. JUDGE_ROLES='--discord --api --jobs', for example, is a deployment with no public page.

While the instance migrates its database for a major release, a question gets a short notice and a Try again button that sends the same question once the upgrade is done.

The operator’s view of the instance is part of the same build, served by the --admin role on a port of its own (ADMIN_ADDR, default 127.0.0.1:8789), never by the public listener. It needs ADMIN_TOKEN; judge-cli admin-url prints the page’s address with the token in its #token= part. The page keeps the token for the tab and removes it from the address bar. From another machine, forward the port over SSH, or on a trusted network publish it on the host’s LAN address with a compose override (Deployment).

  • Calls: every stored call, private ones included, newest first. Filters for the interface, the asker (Discord username or id), the model, private or public, and text.
  • A call: the exchange as the feed shows it, then what is stored with it: who asked and where, the thread, every rating with its rater and reason, and the run that answered it with each billed request’s model, tokens and cost.
  • Feedback: the feedback queue, open items first when the tab opens. Every Partially correct or Incorrect rating given a reason, newest first, filtered by status, kind of problem, interface and rater. An item shows the rater’s reason, its state, the call whole, and a form that resolves it (with a note and what fixed it) or dismisses it (with a note). A rater who replaces their reason reopens it.
  • Failures: the failed calls, each opening onto the error and what the model sent.
  • Dashboard (/admin/usage): the public dashboard over every call, private ones included and counted apart, narrowed by range, interface, asker (Discord username or id) and model; then the failed calls, the spend per day and month (chat and embeddings, judgebot ingest included), the budget period, the cap and what is left (as the instance’s processes recorded them), the runs by outcome, model, interface and stage, the cost per answer, latency, the dearest runs, the runs the spend cap refused, the lowest-rated calls, the feedback queue’s counts by kind of problem and the latest data refreshes. judge-cli stats prints the same report.

Anonymous traffic is rate limited per address (API_RATE_LIMIT questions per API_RATE_WINDOW_SECS, default 4 per 5 minutes), ahead of the JUDGE_CONCURRENCY slots and the spend cap. The feed’s reads have a window of their own, 120 a minute per address, so browsing does not use up questions, and the dashboard’s another, 30 a minute. The HTTP API page has the request and reply shapes. Behind a Cloudflare Tunnel, the deployment runbook puts a Cloudflare rate-limiting rule in front of all of it; behind the bundled public site or your own proxy, API_CLIENT_IP=proxy keeps the count per visitor.

Terminal window
cargo run --release -p judgebot -- --api --web # API + the built page on localhost:8787
npm --prefix web ci
npm --prefix web run dev # Vite dev server with /api proxied to :8787
# and /admin/api to an --admin listener on :8789