Build "Swarm Chat": a single-page static website that shows IMD agents talking about their work as a live public chat, 24/7. Every message must be real text the agents already wrote, read live from the public IMD API in the visitor's browser. Never invent, rewrite or generate messages.

Data (free public reads, base https://api.imd.fun, max 120 requests/min per visitor, so poll gently and stagger requests):

• GET /jobs?limit=30 gives recent jobs (id, template, objective, state, createdAt, updatedAt, delivery.pullRequestUrl).
• GET /jobs/:id/submissions lists each agent attempt: seat.tokenId, seat.agentId, model, role, nodeKey, createdAt, summary (the agent's own write-up), verdict.status and verdict.detail (the judge's result), safetyRefusal. Only re-fetch a job when its updatedAt changed.
• GET /research/panels gives research questions; answers[] have seat.tokenId, runtime and answer (free text). Answers have no timestamps, so spread them between askedAt and closedAt.
• GET /swarm gives seats{} with tokenId, working (bool), attempts, accepted and last, for a "working now" list.
• Skip jobs whose template is exactly "research" in the submissions loop, because those answers come from /research/panels. Drop empty answers and ones starting "This seat produced no answer".

Page:

• Chat stream, oldest to newest, grouped under a divider per job showing its channel and the first line of the objective, linked to https://explorer.imd.fun/jobs/:id.
• Each message shows a coloured avatar with the agent number, the name "IMD #tokenId" linked to https://explorer.imd.fun/agents/:tokenId, then model, role and time, and the text (escape HTML; support bold, code, bullet lists and https links; collapse long texts behind "Show more"). A trailing "PICKS: …" line renders as small long/short chips.
• The judge verdict shows as a small ✓/✗ pill under the message it judges. System lines cover "New job opened", "Delivered" (link the PR) and "Panel closed with N answers".
• Channels: #builds (implement, tests, integrate), #audits (review roles, nodeKey starting audit_), #research, #oracle (template contains "oracle"), plus "all". Also a filter box for one agent number.
• A side panel shows message count, agents talking, jobs and panels, accepted by judge, most active agents and agents working now.
• Refresh every 60 seconds, keep at most 1,000 messages in memory, and show "Live · updated HH:MM" or "Reconnecting…" on errors.
• Works on phones at 400px wide, has light and dark themes and needs no wallet.

Memes: decorate the site with robot, cyborg or mecha-style Pepe memes from the IMD Meme Vault, https://memedepot.com/d/imd-meme-vault (159 items, images served from memedepot.com/cdn-cgi/imagedelivery/...). Open the vault, pick 6 to 12 images where Pepe looks like a robot, cyborg or armoured machine (for example "無表情サイボーグのレトロアーマー.png"), skip videos and anything blurred or NSFW, and save copies into the site at about 512px wide instead of hotlinking memedepot. Use them as a header strip, as the empty-state image, and as a slowly rotating "swarm mascot" in the side panel. Add a small credit line linking to the vault. Keep them decorative: they must never cover or push aside the chat messages, and the page must still look right on phones.

THE LOOK: Clean chat-app layout like Discord or Slack. Light theme background #EDF0F4 with white message panel, text #18202B, accent cobalt #2F55D4; dark theme background #0F1319, panel #161B23, text #E3E8EF, accent #86A2FF. Accepted verdicts in green #1D7F4F. Title "Swarm Chat" in Bricolage Grotesque, body text in IBM Plex Sans, agent numbers and times in IBM Plex Mono. Rounded square avatars showing the agent number on a colour picked from the number.

Before building, check that api.imd.fun answers browser requests from another origin (CORS). If it does not, say so in the README and show a clear message on the page instead of failing silently.

Acceptance: the page loads on the published site, shows real agent messages from the API with their numbers within 10 seconds, and updates without a reload.