pneumo

M5: Docker image, Railway deploy ADR and image smoke test (ENG-267) #22

Merged

demi.dev 1w ago

targeting main from stoopkidddd/eng-267-m5-docker-image-railwayjson-and-image-smoke-test

Merged opened by and assigned to

demi.dev 1w ago targeting main from stoopkidddd/eng-267-m5-docker-image-railwayjson-and-image-smoke-test

Linear: https://linear.app/engineereddev/issue/ENG-267

What's in this PR:

Dockerfile: multi-stage; pnpm 11.28.10 with pnpm fetch layer caching; builds protocol, crypto, web and server; production deps via pnpm deploy; runtime node:26-slim as non-root pneumo (uid 10001), WEB_DIST_DIR set, EXPOSE 8080, CMD ["node","apps/server/dist/index.js"]. ARG VERSION sets the OCI version label and the version /api/health reports. .dockerignore keeps node_modules, Android, .git, test outputs and .env* (except .env.example) out.
docker-compose.yml: the server profile builds the image and runs it against minio/minio-init.
mise run e2e:docker (scripts/e2e/docker.ts): builds the image, starts a harness MinIO, runs the container on a private Docker network in direct then proxy mode, runs the dev-client smoke beside it, and asserts health version, web assets, non-root uid, SIGTERM shutdown and image size.
check:infra runs hadolint on the Dockerfile (scripts/ci/check.sh, pinned in mise.toml).
ADR 0001 (lead decision): no railway.json. Railway Config as Code is deprecated: new services can't opt in, and existing files stop being read on 2026-12-01. Deploy settings move to ENG-271 (Railway IaC) and ENG-273 (template). Recommended overlap 0 s, draining 15 s. SPEC §4, §16.3 and §18.2 updated.
ADR 0002 (lead commit): the image cap is raised from 250 MB to 300 MB uncompressed. The spec-mandated node:26-slim base alone is 272 MB; Pneumo adds about 19 MB. SPEC §16.1 updated.

Verify (lead, rebased on e5f90f2):

mise run check:infra → exit 0 (hadolint, shellcheck, biome, workflow validation)
mise run e2e:docker → exit 0. Image 290,785,357 bytes. Direct and proxy: health version, web assets, non-root uid 10001, smoke 7/7 PASS each (authentication, pair+claim, note, link, 2 MB file, push_deleted, revoke 4401), SIGTERM shutdown.
docker images: pneumo:eng-267 291MB; no containers left.
Implementer also checked the Compose server profile (MinIO init, /api/health, web dist).

Deviations:

No railway.json (ADR 0001); the image cap is 300 MB (ADR 0002).
pnpm deploy needs registry metadata access (offline deploy fails on optional fsevents metadata); the pnpm fetch cache and the frozen offline install remain.

🤖 Generated with Claude Code

Commits 3

M5: Preserve the package version in default Docker builds (ENG-267)

+41 -12

M5: Raise the image size cap to 300 MB (ENG-267)

+30 -2

M5: Add container image and Docker smoke (ENG-267)

+341 -17

Comments

v0 1w ago

demi.dev 1w ago

pneumo-reviewer: round 1, REQUEST_CHANGES #

The image, .dockerignore, the Compose server profile, e2e:docker and the hadolint wiring all work. e2e:docker passed in both modes, hadolint is clean in mise (2.15.1) and in the Nixery amd64 image (2.14.0), and a live run of the Compose server profile came up healthy. ADR 0001, ADR 0002 and the SPEC edits are accurate and consistent (the deprecation wording and the 2026-12-01 cutoff were checked against docs.railway.com). One FIX blocks the merge.

Findings #

FIX (often 6 × bad 3 = 18), Dockerfile:2 (used at :14, :20): ARG VERSION=0.1.0 plus an unconditional npm pkg set … version="$VERSION" replaces the package.json version in every build. Railway passes build args only from matching service variables, so the owner's service, staging and every template deploy would report /api/health version 0.1.0 (and label 0.1.0) after the first bump. This breaks §16.8 and post-release version checks. Fix: an empty default, and set the version only when one is given; add a check that a build without the arg reports the package.json version.
FILE (4), scripts/e2e/docker.ts:66-205: Ctrl-C leaves the server containers, network and build tag behind (MinIO's SIGINT handler exits before finally).
FILE (3), docker.ts:15-21, 167-186: a smoke failure prints only "docker run failed"; the safe FAIL <step> lines are discarded.
FILE (2), docker.ts:12: the hard-coded tag pneumo:eng-267 is shared by concurrent runs.
FILE (1), apps/server/package.json:17: @pneumo/protocol is a runtime dependency although bundled, so ~0.9 MB of TS sources lands in the image.
FILE (1): stale railway.json mentions in CLAUDE.md:38 and .tangled/workflows/infra.yml:18.

FILE items are recorded on ENG-275; code left alone.

Spec checks #

§16.1 OK except the version handling: multi-stage, pnpm via npm (no corepack in Node 26), pnpm deploy --legacy --prod, node:26-slim, uid 10001 with root-owned read-only app files, /app holds only server dist/node_modules/package.json and web dist, EXPOSE 8080, exact CMD, 290,785,357 bytes (≤ 300 MB per ADR 0002). §16.2 OK (fake values only, verified live). §16.3, §18.2, §4 OK per ADR 0001 (overlap 0 s / draining 15 s is sound). §12.3 OK. §16.8: deviation, finding 1.

DoD checks #

hadolint and shellcheck clean; workflows valid; docker build succeeds. .dockerignore excludes .env* (keeping .env.example), Android files, .claude and .git. Layer caching works (fetch/install CACHED on rerun). Both upload modes are really exercised: the smoke runs on the private network, so presigned http://minio:9000 URLs resolve.

Tests run by the reviewer: mise run check:infra exit 0; hadolint (mise and Nixery) exit 0; mise run e2e:docker exit 0; Compose server profile live (health 200, / 200, uid 10001), then torn down. → omp round 2 for finding 1.

v1 1w ago

demi.dev 1w ago

Review round 2 (head 0b3e649): APPROVE #

Round 1 FIX resolved (Dockerfile version default):

ARG VERSION has no default, and npm pkg set runs only when VERSION is non-empty. Builds without the arg (Railway, the template, Compose) report the apps/server/package.json version from /api/health; the OCI label is empty.
The new e2e step builds a second image without the arg and asserts the empty label and the package.json version. The old ARG VERSION=0.1.0 would fail it.
ADR 0001 wording matches the behaviour. No SPEC change needed.

Round 1 FILE findings 2–6 are unchanged on purpose (filed separately). No new findings.

Verify (reviewer and lead): hadolint exit 0, mise run check:infra exit 0, mise run e2e:docker exit 0 (direct and proxy smoke 7/7 each, default-build version check, size 290.79 MB ≤ 300 MB, cleanup). CI 3/3 on 0b3e649.

Versions

pneumo-reviewer: round 1, REQUEST_CHANGES #

Findings #

Spec checks #

DoD checks #

Review round 2 (head 0b3e649): APPROVE #