Smoke Test

Gateway Smoke Test

Bring up the compose stack the integration suite uses (gateway + the S3-compatible object-store origin), probe the gateway's features with live HTTP requests, reconfigure the gateway between phases, root-cause every failure, and emit a structured findings report.

Prime directive: discover, don't assume. The compose file, the GNUmakefile, and docs/getting_started.md are the source of truth for services, ports, credentials, and settings. The reference files in this skill pin a starting checklist; when discovery disagrees with a pinned value, discovery wins. Never hardcode or name the origin's implementation technology — describe it as "the S3-compatible origin" and derive every detail from the compose file.

Scope

Determine what to run based on $ARGUMENTS:

Area keyword → phase map: listing/index → P1; paths → P2; signatures/styles → P3; methods/cors → P4; cache → P5; headers → P6; buckets/ipv6 → P7; creds/assume-role/sts → P8; tls → P9. Unrecognized argument → ask the user, don't guess.

Step 1: Discover the environment

All read-only. Do this in every mode.

1. Read test/docker-compose.yaml. Record: the gateway service name and host port mapping; the origin service name, image, internal/host ports, root credentials, region, and bucket name; which gateway environment: keys are bare pass-through (a bare key whose variable is unset in the invoking shell reaches the container as an empty string that overrides the image's ENV default — every one of these must be exported explicitly); which keys use ${VAR:-default} interpolation; and which documented gateway settings have no compose key at all (those need the overlay technique in Step 3).
2. Run make help (or read the GNUmakefile) to confirm current target names. The GNUmakefile is the only supported build/test interface.
3. Read the environment-variable table in docs/getting_started.md and diff it against the checklist in references/phases.md. Any setting in the docs that no phase covers gets an improvised probe and, if it misbehaves, a coverage-gap finding.
4. Find the data-seeding function in test/run_integration_tests.sh (search for the bucket name). It names the origin's CLI client command, the alias/bucket setup, and the data directory (test/data/<bucket>/). Read it — never source or execute that script; it runs the whole suite at import time and arms exit traps.
5. Preflight:
   • docker image inspect nginx-s3-gateway — if missing, run make build.
   • Compare the image's Created timestamp against the newest commit touching common/, oss/, or Dockerfile.* (git log -1 --format=%cI -- common oss Dockerfile.*). If the image is older, warn and run make build: tests against a stale image probe the baked-in code, not the working tree.
   • Check the discovered host ports. If a port is held by a container from another project, abort with a clear message. If leftover containers from this stack's compose project exist, tear them down first (Step 6 teardown command).

Step 2: Stand up and seed

Substitute discovered values throughout; the commands below reflect the compose file at the time this skill was written.

cd "$(git rev-parse --show-toplevel)"
export COMPOSE_COMPATIBILITY=true NGINX_INTERNAL_PORT=80
# Every bare pass-through key gets an explicit value — never rely on shell
# inheritance (bare key + unset var = empty string inside the container):
export AWS_SIGS_VERSION=4 ALLOW_DIRECTORY_LIST=false \
  PROVIDE_INDEX_PAGE=false APPEND_SLASH_FOR_POSSIBLE_DIRECTORY=false \
  STRIP_LEADING_DIRECTORY_PATH="" PREFIX_LEADING_DIRECTORY_PATH="" \
  STATIC_SITE_HOSTING="" PROXY_CACHE_BYPASS_NO_CACHE=false \
  PROXY_CACHE_IGNORE_HEADERS="" ACCESS_LOG_CACHE_STATUS=false \
  IPV6_ENABLED=""
docker compose -f test/docker-compose.yaml -p ngt up -d

Seed by replicating the discovered recipe: register a CLI alias for the origin using the credentials from the compose file, create the bucket (ignore-existing), copy every entry under test/data/<bucket>/ into it, and write the runtime-generated special-character object the suite's data function creates. If the CLI client is not installed on the host, run its container image attached to the stack's network (--network ngt_default).

Readiness gate: poll http://localhost:8989/health until it returns 200 (timeout ~30 s). This also exercises the gateway's health endpoint, which the fixed suite never requests.

Assert idioms (inline; do not source suite helpers):

# Status assert:
curl -s -o /dev/null -w '%{http_code}' <url>
# Log assert — capture first; piping `compose logs` into `grep -q` dies
# with SIGPIPE under pipefail exactly when the match is found:
logs="$(docker compose -f test/docker-compose.yaml -p ngt logs nginx-s3-gateway 2>&1)"
grep -qF '<expected text>' <<< "$logs"

Step 3: Reconfigure between phases

Recreate only the gateway; the origin and its seeded data stay up, and the gateway's proxy cache (container-local) resets, keeping cache phases deterministic:

<changed exports> docker compose -f test/docker-compose.yaml -p ngt \
  up -d --force-recreate nginx-s3-gateway

For settings with no compose pass-through key (for example DIRECTORY_LISTING_PATH_PREFIX, FOUR_O_FOUR_ON_EMPTY_BUCKET, HEADER_PREFIXES_TO_STRIP, HEADER_PREFIXES_ALLOWED, PROXY_CACHE_USE_STALE, CORS_ALLOW_PRIVATE_NETWORK_ACCESS, or an S3_BUCKET_NAME override), write a minimal overlay to a scratch directory (session scratchpad or mktemp -d — never the repo) and stack it with a second -f; compose merges environment: maps by key:

# <scratch>/overlay.yaml
services:
  nginx-s3-gateway:
    environment:
      FOUR_O_FOUR_ON_EMPTY_BUCKET: "true"

docker compose -f test/docker-compose.yaml -f <scratch>/overlay.yaml \
  -p ngt up -d --force-recreate nginx-s3-gateway

After every recreate: wait on /health, then check the startup banner in the gateway logs (S3 Backend Environment: block — addressing style, signature version, IPv6 listen state) to confirm the settings you intended actually took effect before probing. A probe against a misconfigured gateway produces phantom findings.

Step 4: Run the phases in scope

Read only the ## P<N> sections of references/phases.md selected by the dispatch table. Phase index:

The listed probes are the floor, not the ceiling. If a response looks anomalous — an unexpected header, a suspicious log line, a body that is almost right — chase it with follow-up probes before moving on.

Step 5: Root-cause failures

On any failed probe, read references/root-cause.md and follow it: collect evidence up the ladder, localize to a component with the symptom table, then run the classification protocol. Record a finding for every failure, but do not classify a finding as defect until it has a minimal reproduction on a fresh recreate and has been checked against documented behavior, git log, open issues, and references/regressions.md.

FINDING: <short title>
PHASE: <P0..P9 or regression id>
FILE: <suspected component path, or n/a>
EVIDENCE: <probe command; expected vs actual; log/config excerpt>
CLASSIFICATION: defect | environment | expected-behavior | coverage-gap
SEVERITY: critical | major | minor | info

Step 6: Tear down, report, draft issues

Teardown runs unconditionally — after success, failure, or abort. The --profile flag is required: the dynamic-credentials overlay adds a depends_on for a profile-gated service, and compose rejects the merged project as invalid without it:

docker compose --profile dynamic-credentials \
  -f test/docker-compose.yaml \
  -f test/docker-compose.dynamic-credentials.yaml \
  -p ngt down --volumes --remove-orphans

Also remove any CLI alias you registered and delete scratch overlays, certificates, and secret files.

Print the report in the conversation (offer to save a copy to the scratch directory only if asked; never write it into the repo):

## Smoke Test Report: <scope> — <date>

### Environment
- Gateway image: <tag/id, created <date>>; stack: compose project ngt
- Origin: <service name and image as discovered from the compose file>

### Phase results
| Phase | Env variant | Probes | Pass | Fail | Notes |
|-------|-------------|--------|------|------|-------|

### Findings
<FINDING blocks, severity-ordered>

### Summary
- Phases run: N/N; probes: N (N pass, N fail)
- Findings: N (N defect, N environment, N expected-behavior, N coverage-gap)
- Verdict: PASS | PASS WITH FINDINGS | FAIL

Then, for each finding classified defect, draft a ready-to-file GitHub issue (title; body with reproduction commands, expected vs actual behavior, root cause, suspected component file, gateway image/commit) for gh issue create --repo nginx/nginx-s3-gateway. Present the drafts and file nothing without explicit per-issue user approval.

Guardrails

• Do NOT use or mention test.sh; the GNUmakefile is the only supported build/test interface.
• Do NOT source or execute test/run_integration_tests.sh — it runs the entire suite at import time. Read it only.
• Do NOT skip teardown. The Step 6 down --volumes --remove-orphans command runs even when a phase aborts or the user interrupts.
• Do NOT pass bare compose environment keys — export an explicit value for every pass-through key, empty string only when emptiness is intended.
• Do NOT probe a stale image. After any source edit, make build before the next probe; retesting without a rebuild exercises the baked-in copy.
• Do NOT run make test-matrix or make test-matrix-plus unless the user explicitly asks; the smoke test is not the CI matrix.
• Do NOT modify repository files. Overlays, certificates, and generated secrets live in the scratch directory.
• Do NOT name the origin's implementation technology in probes, findings, or reports — use details discovered from the compose file.
• Do NOT classify a finding as defect or draft an issue without a minimal reproduction plus checks against docs/getting_started.md, git log, gh issue list --search, and references/regressions.md; never file an issue without the user's approval.
• Do NOT continue when the discovered host ports are held by containers from another project — abort and tell the user which ports are busy.