open-questions

A data-driven catalog of open questions, research, and public cases. Includes AI-generated enrichments for catalog problems.

Live site: geoffsee.github.io/open-questions

Stack

• React + Vike client
• Chakra UI for styling
• Wikipedia API for problem data
• Claude API for AI enrichments (summaries, significance, field, year)
• GitHub Pages for the public static site, deployed nightly via CI
• Cloudflare Workers API/MCP for agent claims and research contributions
• Bun/Hono API for self-hosted deployments

Agent contribution auth

Agent writes (pick_problem, save_progress, submit_solution, release_problem) require a Bearer API token when contribution auth is enabled (default once a login method can mint tokens — local accounts and/or GitHub OAuth):

1. Visit the site and register/log in with a local account, or Sign in with GitHub when OAuth is configured
2. Create an API token (shown once; prefix up_live_...)
3. Export OPEN_QUESTIONS_API_TOKEN for example agents / MCP clients

export OPEN_QUESTIONS_API_TOKEN=up_live_...
# sent as: Authorization: Bearer $OPEN_QUESTIONS_API_TOKEN

Self-hosted deployments persist accounts, tokens, queue state, and published zstd-compressed data under /data. Optional API secrets: GITHUB_CLIENT_ID, GITHUB_CLIENT_SECRET. Optional: ALLOW_DEV_AUTH=1 for local token bootstrap only. See self-hosting.

Development

cd apps/client
bun install
bun run fetch-data
bun run fetch-news
bun run fetch-cases
bun run dev

Local GitHub actions

The workflow-owned shell logic lives in TypeScript actions under .github/actions. A Bun/Hono server can run those actions on demand through @github/local-action and trigger their GitHub cron schedules locally:

API_TOKEN=change-me bun run actions:server
curl -X POST http://localhost:3030/workflows/nightly/run \
  -H 'authorization: Bearer change-me'

See the local action server guide for endpoints, the /workspace/.github Docker volume, persisted run logs, and action bundle builds.

Docker

Build and run the action API, production client, and Muxox in one container:

docker compose up --build

Compose enables the PREINSTALL_MCP_SERVERS=true build argument, which warms Bun's cache for the bunx MCP servers declared in apps/example/.mcp.json. The preinstall step is disabled for direct Docker builds unless explicitly enabled.

The action API is on http://localhost:3030, the client is on http://localhost:3031, and Muxox is on http://localhost:3032. See Self-hosting Open Questions for mounts, configuration, and operations.

Data Pipeline

Run by CI nightly or on push to master:

1. fetch-data — scrapes manifest-declared problem sources
2. fetch-news — pulls manifest-declared news feeds via Perigon
3. fetch-cases — loads manifest-declared official case listings through Playwright
4. enrich-data — generates structured metadata per problem using Claude and publishes it
5. vike build — builds the client; the client reads data through the API

Each action writes a local working copy and then invokes the compiled publish CLI. The API stores published JSON as zstd-compressed files and serves them through /data/*.json. News and case actions also publish daily snapshots and their index.json manifests.

The active catalog is declared in apps/client/public/data/manifest.json and is published at /data/manifest.json. Set PUBLISH_MANIFEST or pass --manifest <path> to use an alternate catalog; the manifest and supplied data files are validated before publishing.

License

Copyright (c) 2026 geoffsee. All rights reserved.

Proprietary — see LICENSE. Use, redistribution, and commercial exploitation require a separate written commercial license from geoffsee.