CuriosityEngine

Faculty cannot make students care about compilers. A leaderboard can.

CuriosityEngine is a Discord bot that turns every CSE subject channel into a daily competitive arena. Each morning it researches a syllabus-aligned problem, posts it to the right channel, tracks who solved it first, decays streaks for the idle, and renders a live leaderboard. It runs on Google Cloud Run at $0 idle cost: the container only exists while a request is being served.

---

The Problem

Faculty cannot generate interest in subjects. Passive lectures, no peer pressure, no feedback loop. Students disengage weeks into the semester and never recover.

How It Works

Once a day, Cloud Scheduler sends a single authenticated request to /cron/daily. The service wakes up (in ~1 second), runs the daily loop, and shuts back down:

1. Research: Vertex AI (Gemini) reads the official syllabus and generates one problem per subject channel.
2. Post: Problems go to Discord via HTTP interactions, no always-on gateway required.
3. Score: When students respond, Ed25519-verified webhook calls update Firestore in real time.
4. Decay: Streaks for non-participants drop. The leaderboard is public and ruthless.
5. Self-maintain: A second agent researches current engineering standards and opens a GitHub Pull Request improving the project. It never pushes to main.

Architecture

Component	Why it costs $0 at rest
Cloud Run (min-instances=0)	Billed per 100ms of CPU, not per hour. Zero requests = zero cost.
Firestore (serverless)	Billed per read/write operation, not per GB provisioned.
Secret Manager	Two secrets, two accesses per day. Fits free tier.
Vertex AI (Gemini)	Billed per token. One daily run costs pennies.
Cloud Scheduler	One job, one invocation per day. Free tier covers it.
Container build	Built locally with ko (Go toolchain, no Docker daemon). Nothing billable runs.

The only meaningful metered cost is Vertex AI token usage during the daily run, typically a few cents per day.

Public surfaces

POST /interactions   Discord HTTP-interaction webhook (Ed25519 verified)
POST /cron/daily     Daily trigger from Cloud Scheduler (OIDC verified)
GET  /               Server-rendered leaderboard
GET  /health         Liveness probe

Self-maintaining agent safety model

The self-update agent runs inside the same Cloud Run instance. It can read source files and open GitHub Pull Requests. It cannot:

• Push to main directly (branch protection enforced at GitHub).
• Edit its own guardrails (the files governing PR scope are in the PR diff, visible to reviewers).
• Access production secrets (Secret Manager IAM binds only to the runtime service account, which never runs locally or in CI).

Every self-update PR is a normal GitHub PR: reviewable, rejectable, and auditable.

Security and IAM

Two service accounts, minimal roles, no API keys anywhere:

Account	Role	Purpose
curiosity-runtime@...	roles/aiplatform.user, roles/datastore.user, secretAccessor (scoped to 2 secrets)	Runs the Cloud Run service
curiosity-scheduler@...	roles/run.invoker	Allows Cloud Scheduler to call /cron/daily only

Vertex AI is authenticated via Workload Identity (ADC), not an API key. Discord interactions are verified with Ed25519 signatures. Scheduler calls are verified with Google-signed OIDC tokens.

Local Development

# 1. Copy the example env file and fill in your values
cp .env.example .env

# 2. Authenticate with GCP (uses Application Default Credentials)
gcloud auth application-default login

# 3. Run the service
go run ./cmd/server

The service starts on PORT (default 8080). For Discord interactions locally, use a tool like ngrok to expose a public HTTPS URL and point your Discord app's Interactions Endpoint URL at it.

Deploy

The image is built locally with ko (the Go toolchain assembles the container, no Docker daemon and no Cloud Build, so nothing billable runs). Install ko once:

go install github.com/google/ko@latest

Then provision and deploy everything with the idempotent deploy script:

bash deploy/deploy.sh

It builds the image, pushes it to Artifact Registry, and deploys the Cloud Run service. Re-running it ships a new revision. GitHub Actions runs CI (vet, build, test, CodeQL) on every push and pull request; deployment is a deliberate, local step.

The script is idempotent: safe to rerun on an existing deployment.

Manual One-Time Steps

These cannot be automated because they involve external accounts:

1. Create a Discord application at discord.com/developers. Note the Application ID, Public Key, and Bot Token.
2. Populate Secret Manager: create two secrets in your GCP project:
   • curiosity-discord: a JSON object with keys app_id, public_key, bot_token, admin_user_id.
   • curiosity-github: a GitHub Personal Access Token with repo scope (used by the self-update agent to open PRs).
3. Add the Cloudflare DNS record: after Cloud Run assigns a domain, add a CNAME:

   curiosityengine.dmj.one  CNAME  <cloud-run-domain>  (proxied)
   

   Then map the custom domain in the Cloud Run console.
4. Set the Discord Interactions Endpoint URL to https://curiosityengine.dmj.one/interactions.

---

Built on Go 1.24, deployed on Google Cloud Run, proxied through Cloudflare.