Stoica Mihai Claude Skills Smoke Test

Smoke Test

A smoke test (a.k.a. build-verification test / build-acceptance test / intake test) is a small, fast, broad-but-shallow suite that verifies a fresh build is stable enough to bother testing further. It is the release gate that protects the rest of the pipeline from wasting time on fundamentally broken code. The name comes from electrical engineering: power on the board and watch for smoke. If it smokes, stop.

This skill helps you design + scaffold one for the user's project. It is not a regression-suite generator and not a full QA strategy: smoke tests stay focused on a handful of critical happy-paths and connectivity checks.

When to invoke this skill

Use it when the user wants any of:

• A new smoke-test suite for a service that has none
• A review of an existing smoke-test suite for scope, speed, flakiness
• CI/CD wiring so smoke runs as a gate before regression
• Help picking the 5-10 checks that should make the cut

Skip it for:

• Unit testing or full regression testing (different discipline, larger scope)
• Pure performance / load testing (overlaps but lives elsewhere)
• "Fix this failing test" — that is debugging, not suite design

Core principles (compressed)

Smoke tests have a different shape from other tests. The user may not know these, so anchor every recommendation in them:

• Broad, not deep. Cover the most important user-facing paths. Skip edge cases, validation rules, error-message wording. If the smoke suite is failing on something that needs a debate, it is too detailed.
• API-first, UI as last resort. API-level checks finish in milliseconds and rarely flake. UI checks finish in seconds and flake often. Reserve UI smoke for flows that cannot be observed at the API layer (login redirects, SPA boot).
• **<2 minutes wall-clock** for the whole suite in CI. >2 min and developers will start skipping the gate; >15 min and the gate is effectively dead.
• 5–10 checks for a small/medium service, up to 20–50 for a large enterprise app. More than that is a regression suite in disguise.
• Idempotent and read-only-safe. Smoke tests run on every deploy, often in production. They must not leave artefacts behind. Write paths either use ephemeral test data with cleanup, or are guarded so they can be skipped against prod.
• Fail fast, gate hard. If smoke fails, the deploy is rejected. Subsequent test stages do not run.
• Diagnose-able failures. When the gate trips at 3 AM, the engineer on call needs the failing payload, the response, the correlation id — not just "test 4 failed".

The full research distillation is in references/principles.md. Read it when the user asks "why" about any of the above, or when you need the formal definitions of smoke vs sanity vs regression vs retest.

Workflow

Step 1 — Detect the stack

Run the bundled detector and read its JSON output:

python <skill-path>/scripts/detect_stack.py <repo-path>

It identifies the language(s), web/API framework, test runner(s), candidate entry points, discovered HTTP routes, and existing CI config. Route discovery is greedy/grep-based — treat the list as candidates the user should confirm, not ground truth.

If detection finds no language manifest (Analyzed 0 style result) or no framework and the user is asking specifically about HTTP smoke, say so plainly to the user and ask for the stack. Do not invent.

Step 2 — Propose the 5–10 checks

Before writing any code, present a short numbered list of what the smoke suite will cover. Each check should fit one line. Default shape:

1. Health/readiness endpoint returns 200 (/health, /healthz, /readyz, or framework default)
2. Authentication path mints a valid token (or rejects missing token with 401)
3. Primary read endpoint returns shape-correct data for known fixture
4. Primary write endpoint (optional, only if idempotent) creates a record and returns its id
5. Dependency liveness — at least one read query against the DB or message bus
6. Error path returns the correct status (404 for missing, 401 for unauthenticated)

Adapt this list to the routes the detector found. Cross out items that do not apply. Add domain-specific items (checkout funnel, login, etc.) when obvious from the routes table. Cap at 10.

Confirm the list with the user before generating code. The user may know that "X is owned by another team and not in our smoke contract" or "the /admin path is irrelevant" — let them prune.

Step 3 — Pick the template

Match the detected stack to a template in references/templates/:

• fastapi-pytest.md — FastAPI / Starlette + pytest + httpx (in-process TestClient)
• flask-pytest.md — Flask + pytest + Flask test_client()
• django-pytest.md — Django / DRF + pytest-django + Django Client
• express-vitest.md — Express + Vitest + supertest
• fastify-vitest.md — Fastify + Vitest + app.inject()
• go-httptest.md — Go net/http + standard testing + net/http/httptest.NewServer
• gin-httptest.md — Go + Gin + gin.Engine.ServeHTTP + httptest
• axum-tower.md — Rust / axum + tokio test + tower::ServiceExt

For frameworks not yet covered (Rails, NestJS, Spring, .NET, Actix, Rocket, …) read the closest template (HTTP-method-based ones for Rails/NestJS, axum-tower for Actix/Rocket) and adapt the imports + assertion library. Note explicitly in your response that the template is adapted rather than canonical, so the user calibrates trust.

Step 4 — Emit the scaffold

The generated file should:

• Sit in the project's conventional test location (tests/smoke/, test/smoke/, cmd/<service>/smoke_test.go, etc.)
• Be runnable as scaffolding — the test framework's hello-world must pass; assertions reference the real routes — but env-specific values (base URL, bearer token, test user id) are clearly marked TODO/os.getenv("SMOKE_BASE_URL", "http://localhost:8080") so the user can wire them in.
• Cap the suite at the count agreed in Step 2.
• Include a short comment block at the top stating: purpose (smoke, not regression), runtime budget (<2 min), how to run locally, how to run in CI.
• Avoid hard-coded sleep(n) — use the framework's polling/wait primitives (httpx retries, vitest vi.waitFor, Go eventually helpers).

After writing the file, show the user the runnable command. Example: "pytest tests/smoke -q -s — should run in <5 s against the in-process TestClient; set SMOKE_BASE_URL to hit a deployed environment instead."

Step 5 — Offer CI wiring (only if asked or obviously useful)

If the detector found GitHub Actions / GitLab CI and the user asked about CI, append a minimal job snippet to the recommended file path. Otherwise mention the option in one sentence and stop. Do not stuff CI config the user did not ask for.

Step 6 — Tell the user how to fail well

Close with a short reminder of the failure-debugging expectations:

• Capture stderr, response payloads, correlation ids on failure
• Treat a failing smoke run as a stop-the-line event
• Quarantine (don't delete) flaky tests, fix them within the sprint

Communicating findings

Lead with the plan (the 5-10 list), then the artefact (the scaffold file path), then runtime + caveats. Do not paste the test code into the chat — point at the file. Example shape:

Detected: FastAPI 0.110, pytest + httpx, 4 routes (/health, /users/{id}, /orders, /orders/{id}). Proposed smoke suite (6 checks):

1. GET /health returns 200 + {status: ok}
2. GET /users/1 returns alice
3. GET /users/9999 returns 404
4. POST /orders without auth returns 401
5. POST /orders with bearer + known user returns 201 + id
6. GET /orders/{id} for the order created in #5 returns the same body

Wrote tests/smoke/test_smoke.py — pytest will run in <2 s in-process. TODO markers at the top set SMOKE_BASE_URL + SMOKE_TOKEN for deployed-env runs. Caveat: I did not detect a CI config; let me know if you want a GitHub Actions snippet that runs this on every PR.

Out of scope

The research informing this skill covers smoke testing across many domains (mobile, ML, ERP, messaging). The current version of this skill focuses on web/API/software services. If the user asks for mobile or ML-pipeline smoke testing, say so plainly — the principles still apply (broad/shallow, API-first, fail fast) but the templates do not yet exist. Point them at references/principles.md for the conceptual guidance and offer to draft something stack-specific by hand.

Process metrics (Defect Escape Rate, MTTD, MTTR) are documented in references/principles.md for the user's awareness but are not emitted by this skill — they belong in a longer-running QA dashboard, not a per-build gate.

Source: Stoica-Mihai/claude-skills — distributed by TomeVault.