Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

4.6 KB
Raw Blame
# Central product backend — Kubernetes image. (The sibling `Dockerfile` is the pm2/compose one;
# this file is the k8s-shaped variant and is what the k8s build should target.)
#
# ONE image, TWO roles — selected by the APP env var, so both Deployments run the same image and the
# manifests differ only by `env` + `replicas`:
#   APP=api  (N replicas) : node dist/server.js   ← default
#   APP=job  (1 replica)  : node dist/worker.js
#
# Why one process per container: src/worker.ts owns schema push, migrateHarnessRename, ensurePlans and
# the backfills, and those boot steps are NOT guarded across processes (only the recurring billing
# loop takes a Redis lease). Baking the pm2 ecosystem here would start a worker in EVERY replica and
# race them. K8s owns the replica count; a pm2 cluster inside a pod would also mis-size itself, since
# os.cpus() reports the node's cores, not the pod's cgroup CPU limit.

# --- builder: full install (postinstall = prisma generate) + esbuild → dist/ ------------------
FROM node:22-bookworm-slim AS builder
WORKDIR /app

# openssl so `prisma generate` detects the correct engine target (bookworm = openssl-3.0.x).
# Without it Prisma mis-detects openssl-1.1.x and the runtime can't load the query engine.
RUN apt-get update && apt-get install -y --no-install-recommends openssl && rm -rf /var/lib/apt/lists/*

# postinstall runs `prisma generate`, so the schema must be present first.
COPY package.json package-lock.json ./
COPY prisma ./prisma
RUN npm ci

COPY tsconfig.json build.mjs ./
COPY src ./src
RUN npm run build

# --- runner: dist/ + production node_modules only, non-root ----------------------------------
FROM node:22-bookworm-slim AS runner
WORKDIR /app

RUN apt-get update && apt-get install -y --no-install-recommends openssl && rm -rf /var/lib/apt/lists/*

RUN groupadd --system --gid 1001 nodejs \
    && useradd --system --uid 1001 --gid nodejs backend

COPY package.json package-lock.json ./
COPY prisma ./prisma
# Production deps WITHOUT lifecycle scripts: postinstall is `prisma generate`, which needs the Prisma
# CLI — a devDependency that `--omit=dev` does not install.
RUN npm ci --omit=dev --ignore-scripts && npm cache clean --force

# Take the Prisma pieces from the builder, where the full install already generated them:
#   .prisma  generated client + query engine
#   prisma   the CLI — the worker shells out to prisma/build/index.js for `db push` at boot
#            (lib/db/migrate.ts ensureSchema). Miss it and schema management SILENTLY no-ops.
#   @prisma  client runtime + engines
COPY --from=builder /app/node_modules/.prisma ./node_modules/.prisma
COPY --from=builder /app/node_modules/prisma ./node_modules/prisma
COPY --from=builder /app/node_modules/@prisma ./node_modules/@prisma

# Runtime artifacts only — no src/, tsconfig.json or build.mjs in the published image. dist/ mirrors
# src/ 1:1 (build.mjs sets bundle:false), so node_modules must ship with it.
COPY --from=builder /app/dist ./dist

ENV NODE_ENV=production
# Public tag/SHA only; /api/health uses these to verify the image actually serving traffic.
ARG BUILD_TAG
ARG BUILD_SHA
ENV BUILD_TAG=${BUILD_TAG}
ENV BUILD_SHA=${BUILD_SHA}
# Role selector. Override to `job` on the worker Deployment (replicas: 1).
ENV APP=api
# Each pod is one node process (no pm2 cluster here), so cap V8's old space explicitly: without it V8
# sizes the heap from the NODE's RAM, not the pod's cgroup limit, and a leak becomes an OOM-kill of the
# whole pod instead of a bounded GC. Set the pod's memory limit to ~1.5x this. Override per Deployment
# (the api role streams to many sockets and wants more headroom than the job role). Merge, don't clobber,
# any NODE_OPTIONS a manifest already sets.
ENV NODE_OPTIONS=--max-old-space-size=1024

USER backend

# 8085 = control API + /proxy data-plane; 8087 = public subdomain app-proxy (same process, so the
# api Service must expose BOTH). Documentation only — k8s publishes via the Service.
EXPOSE 8085 8087

# `exec` replaces the shell, so node ends up as PID 1 and receives SIGTERM directly; both entry
# points shut down on it (server.ts answers /api/ready 503 and keeps serving for DRAIN_DELAY_MS, then
# closes the listeners and DRAINS open websockets with a 1012 close + ~4s grace before quitting Redis,
# so give the api terminationGracePeriodSeconds >= DRAIN_DELAY_MS + 10s; see README.md for the probes;
# worker.ts stops the billing loop and hard-exits after 5s, so the job also needs >= 10).
CMD ["sh", "-c", "case \"$APP\" in api) exec node dist/server.js ;; job) exec node dist/worker.js ;; *) echo \"FATAL: unknown APP='$APP' (expected 'api' or 'job')\" >&2; exit 1 ;; esac"]