run-rubygems-org

Run rubygems.org

A Rails 8 app (internal name: gemcutter) that needs Postgres, OpenSearch, and Memcached reachable on 127.0.0.1:5432/9200/11211. smoke.sh probes those ports and works whether the backing services are running in Docker, via native installs, or are already running on the host; everything else (Ruby 4, Puma, headless Chrome) runs on the host.

The agent path is ./.agents/skills/run-rubygems-org/smoke.sh — it brings services up, boots Rails on :3000 if it isn't already, curls four key endpoints, and, when a Chromium binary is available, writes three PNG screenshots to tmp/run-skill/.

All paths in this doc are relative to the repo root.

Canonical location: .agents/skills/run-rubygems-org/ — the cross-tool skill convention picked up by Codex CLI, OpenCode, and Gemini CLI. Claude Code reads it through the symlink at .claude/skills/run-rubygems-org/.

If smoke.sh fails, see "When this skill fails" below before improvising — LEARNINGS.md likely has your fix.

Prerequisites (one-time)

You need Ruby 4 (via mise/asdf from .ruby-version), and Postgres + OpenSearch + Memcached reachable on 127.0.0.1:5432/9200/11211. smoke.sh doesn't care how they're running — Docker, brew services, apt, or already-on-the-host — it just probes the ports.

If you're on a fresh machine, pick whichever path you prefer:

# Path A: Docker (one command for all three backing services)
mise install                  # ruby 4.0.5 per .ruby-version
docker compose up -d db cache search
bin/setup                     # bundle, db:prepare, db:seed, playwright install

# Path B: Native (macOS via Homebrew)
mise install
brew install postgresql@14 memcached
brew services start postgresql@14
brew services start memcached
# OpenSearch isn't in Homebrew; run it via docker even in a "native" setup:
docker run -d -p 9200:9200 \
  -e discovery.type=single-node -e DISABLE_SECURITY_PLUGIN=true \
  -e 'OPENSEARCH_JAVA_OPTS=-Xms512m -Xmx512m' \
  opensearchproject/opensearch:2.13.0
bin/setup

See CONTRIBUTING.md > Development Setup for the Linux/apt variant. bin/setup runs bin/playwright install --with-deps chromium; smoke.sh shells out to bin/playwright screenshot, which uses that same bundled Chromium — no separate browser needed.

Two traps on fresh machines:

• mise install alone doesn't put Ruby on PATH in non-interactive shells. Activate it first (eval "$(mise activate bash --shims)") or prefix commands with mise exec --; otherwise bin/setup runs against the system Ruby and bundler aborts with a version mismatch.
• Native Postgres on Linux needs trust auth. config/database.yml expects different passwords for the same postgres role in dev (devpassword) and test (testpassword), which only works when the server doesn't check passwords. Docker handles this via POSTGRES_HOST_AUTH_METHOD=trust and Homebrew trusts local connections by default, but apt's Postgres defaults to scram-sha-256 — set the local/127.0.0.1 entries in pg_hba.conf to trust and restart, or bin/rails db:prepare fails with password authentication failed for user "postgres".

Run (agent path — preferred)

./.agents/skills/run-rubygems-org/smoke.sh

That single command:

1. Probes Postgres (:5432), Memcached (:11211), and OpenSearch (:9200). If a service is already listening (native brew/apt install, or some other running container), it's left alone. If something is down and docker compose + docker-compose.yml are both available, it falls back to docker compose up -d <service>. If neither path works, it fails with the exact install commands to run.
2. If nothing is on :3000, launches bundle exec rails server -p 3000 -b 127.0.0.1 under nohup (pid → tmp/run-skill/rails.pid, log → tmp/run-skill/rails.log). Otherwise reuses the existing server.
3. Polls GET / until it returns 200 (timeout 60s — cold start is ~3s here, ~10–15s on first boot after a code change due to Tailwind compile).
4. Curls four endpoints and fails loudly on any non-200 or missing-substring:
   • / (homepage HTML)
   • /api/v1/gems/rubygem0.json (gem JSON API)
   • /versions (compact_index — what bundle and gem fetch)
   • /gems/rubygem0 (gem detail page)
5. Screenshots three pages with headless Chrome at 1280px wide → tmp/run-skill/{home,gem,signin}.png.

rubygem0 exists because db:seed creates it; if you've wiped the DB, re-run bin/rails db:seed.

To stop a server smoke.sh started:

kill "$(cat tmp/run-skill/rails.pid)"

Run (human path)

bin/rails s                  # foreground on :3000

Opens nothing — just listens. Useless headless; for an agent driving the app, always use smoke.sh so you get the screenshots and HTTP assertions.

Direct invocation (poking internal code)

Most PRs in this repo touch a model, job, policy, or controller — not the HTML surface. For those, skip smoke.sh and call the code directly. All of these run against the dev DB and respect any local changes without restarting a server:

# One-liner: import + call + observe (matches AGENTS.md examples)
bin/rails runner 'p Rubygem.find_by(name: "rubygem0")&.versions&.count'

# REPL — for exploratory work or anything multi-step
bin/rails c

# Run one test file (fastest feedback loop for a model/job change)
bin/rails test test/models/rubygem_test.rb
bin/rails test test/models/rubygem_test.rb:42      # one line
bin/rails test -n /pattern/                         # name pattern

# Enqueue / run a job inline against the dev DB
bin/rails runner 'YourJob.new.perform(Rubygem.last.id)'

If a job/policy depends on the gem index or OpenSearch being current, run the indexer first:

bundle exec rake gemcutter:index:update             # compact_index / specs.*.gz
bundle exec rake searchkick:reindex CLASS=Rubygem   # OpenSearch

When this skill fails (learn-and-teach loop)

This skill keeps a running LEARNINGS.md so each failure becomes a one-time cost, not a recurring one. When smoke.sh fails with something not already in the Troubleshooting table:

1. Read LEARNINGS.md first — Ctrl-F your error string; someone may have already solved it.
2. Read tmp/run-skill/diagnostic.txt — smoke.sh writes a structured environment snapshot (ports, docker state, rails log tail, ruby/bundle versions) every time it fails.
3. Fix it. Capture the exact command(s) that worked.
4. Add an entry to LEARNINGS.md using the template at the top of that file. Don't skip this even if the fix is "obvious" to you — it won't be obvious to the next agent in a fresh container at 3am.
5. Graduate stable fixes. If the failure mode is likely to recur (not a one-off environment quirk):
   • Add a row to the Troubleshooting table below (symptom → fix), and/or
   • Update smoke.sh to auto-detect or auto-recover (e.g. add a probe, extend the docker fallback)
   • Then mark the LEARNINGS.md entry Status: graduated with a link to your commit/PR.

This is the skill teaching itself. LEARNINGS.md is the why behind every SKILL.md / smoke.sh change — keep it even after graduation.

Tests

bin/rails test                                  # unit/integration
bin/rails test test/models/rubygem_test.rb:42   # single test
DB_HOST=db bin/rails test                       # ONLY when Postgres isn't on localhost (e.g. dev containers) — see Gotchas
bin/rails test:system                           # Playwright + Chrome
bin/ci                                          # full CI suite (lint + brakeman + tests)

Gotchas

• .env.local is excluded in the test environment. dotenv loads .env.local for development but not for test, so a DB_HOST set there is invisible to tests. config/database.yml defaults DB_HOST to localhost (TCP) for both dev and test, so a standard local setup needs no DB_HOST at all — but when Postgres lives on another host (e.g. db in a dev container), pass it inline: DB_HOST=db bin/rails test.
• There is no /up endpoint. Rails 8's default health check is not wired up; use GET / (homepage) as the readiness probe. /up returns 404.
• First boot prints ~50 lines of Datadog APM/CI-Visibility noise before "Listening on …". This is normal — dd-trace-rb autoloads in development. Don't mistake it for failure.
• OpenSearch may report status: yellow. That's expected for a single-node cluster; the smoke script only requires _cluster/health to respond, not be green.
• Seeded gems are rubygem0/rubygem1/rubygem2, not real gems. /api/v1/gems/rails.json returns "This rubygem could not be found." — use rubygem0 in any smoke check.
• Tailwind runs in a separate process (bin/rails tailwindcss:watch is auto-spawned in dev). It writes CSS asynchronously after boot; if the homepage looks unstyled in a screenshot, give it 2–3 more seconds.
• bin/setup is destructive of dev DB schema. It runs db:prepare which will create+migrate. To load real anonymized data, use script/load-pg-dump (see AGENTS.md).
• Screenshots go through bin/playwright, not a hand-found Chromium. bin/playwright is the Node CLI pinned to the playwright-ruby-client gem version, and playwright screenshot resolves its own bundled browser — so we don't care where Chromium lives on disk. If you blow away the cache, bin/playwright install --with-deps chromium rehydrates it.

Troubleshooting