K6

k6 Script Generation

Efficiency note: This is a short linear recipe (read example → adapt → save → validate → review). A todo list would just mirror the headings without adding value, so skip the planning overhead and execute the steps directly.

Agent-agnostic: The steps below describe capabilities, not specific tools. Where a step says "fetch a URL" or "write a file", use whatever your agent provides for that capability (e.g. a web-fetch tool, a file-write tool, or shell curl/tee).

Step 1: Pick the right example file

Read only the file that matches the user's request. Examples provide structural scaffolding — the correct scaffold, option shapes, and import patterns.

Example files live in the examples/ directory alongside this SKILL.md.

When the request matches multiple rows (e.g. "browser" + "functional test"), prefer the row whose assertion style fits the intent. If the user says "functional test", "assert", "verify", or "expect", use functional.js even if the test involves a browser — it demonstrates expect() with auto-retrying browser matchers. Use browser.js for browser load/performance tests that don't emphasize correctness assertions.

Browser test from a plain-language journey? If the user describes a browser flow in prose and the selectors aren't known yet — i.e. they must be discovered by driving the live site — defer to the k6-browser-test skill, which explores the live page to resolve real getBy* locators before writing the test. Use the browser rows above when the page structure and selectors are already known.

Step 2: Adapt the example

Use the loaded example as the starting point. Adapt it to the user's exact requirements:

• Change endpoints, VU counts, durations, thresholds

• Add or remove scenario steps

• Rename functions and variables to match the domain

• Every expression must be complete and runnable — no { ... }, // TODO, or stubs

• Match the request — don't over-build. Implement exactly what was asked. Don't add custom request tags, extra sleep() calls, additional endpoints, or options the user didn't request. Unrequested complexity lowers quality and reduces adherence to the spec.

For multi-scenario scripts (browser + HTTP, cloud): use named scenarios with exec pointing to separate exported functions.

Step 3: Fill gaps with docs (only if needed)

The example covers common patterns. Adapt from it directly. Skip this step entirely if the example provides everything you need.

Only reach for docs if:

• The user asks for an API or option not demonstrated in the example, or

• You are not confident about the exact signature, option name, or return type

When a gap exists, first establish the docs command (one-time per session).

The k6 x docs CLI renders content only when it detects a TTY. Since agents run non-interactively, wrap every call with script to allocate a pseudo-TTY and pipe the ANSI-stripped content to stdout:

# Detect OS once (macOS vs Linux have different `script` flags):
if [[ "$(uname -s)" == "Darwin" ]]; then
  DOCS_CMD="script -q /dev/null k6 x docs"
else
  DOCS_CMD="script -qc 'k6 x docs' /dev/null"
fi

# Verify it works — should print a topic list, NOT a "browse files" guide:
$DOCS_CMD 2>/dev/null | head -5

If the output still shows "k6 documentation is a directory of markdown files", the TTY wrapper isn't working. Fall back to web docs under https://grafana.com/docs/k6/latest/ — fetch pages with whatever web-fetch capability your agent has (a built-in fetch tool, or curl in a shell).

If k6 x docs fails outright (command not found, provisioning or 404 errors), read SETUP.md — it covers auto-provisioning on k6 v1.7.0+ and the manual xk6 build for older versions.

Then look up what you need:

$DOCS_CMD <path>              # e.g. javascript-api k6-http
$DOCS_CMD <path> --depth 2
$DOCS_CMD search <term>

Common CLI paths and the 2-call strategy are in docs-guidance.md.

Do not use unpkg, @types/k6, or any npm type definition URLs.

Step 4: Save

Line 1 of every script must be a generated-by comment. Get the current UTC timestamp first (the file content depends on it, so this can't be parallelized with the write):

date -u +%Y-%m-%dT%H:%M:%SZ

Then include it as line 1:

// Generated by grafana-k6 on 2026-03-25T22:02:20.203Z

Save to k6/scripts/<descriptive-name>.js. Use lowercase kebab-case filenames. If your file-write capability doesn't create parent directories automatically, mkdir -p k6/scripts first.

The script file on disk is the deliverable — always write it. Never end the task with the script shown only in chat: Steps 5–7 (validate, review, present) all require the file to exist at k6/scripts/<name>.js. If a write fails, retry it before continuing.

Step 5: Validate

Match top-down — the first matching row wins:

What counts as passing: for functional and browser scripts, validation passes only when the exit code is 0 and the summary shows no failed checks or expect() errors — a script that completes with failing assertions is not validated. k6's error output names the exact locator or assertion that failed (and what it waited for); use it to fix the selector or logic. For load tests, exit code 0 is a pass, and a pure threshold breach under load (exit code 99) is acceptable.

If validation fails: read stderr, fix the root cause, retry up to 3 attempts. After 3 failures, present the error and ask the user how to proceed (or, when running unattended, deliver the best attempt and clearly report the unresolved error).

Step 6: Best-practices review

General checks (all scripts)

Review the script against the rules below. The checklist is authoritative — only look up docs ($DOCS_CMD best-practices or https://grafana.com/docs/k6/latest/using-k6/) if you're uncertain about a specific rule.

• export const options with realistic VUs/duration. Default VUs/durations make the test meaningful out of the box.

• Define thresholds for every load test. Without thresholds the run can't fail in CI even when performance regresses, which defeats the point of running a load test. At minimum include http_req_duration and http_req_failed (or the protocol equivalent). Pure functional tests — single-iteration expect()-only scripts — can skip this.

• Include sleep() in closed-model (VU-based) load tests. sleep() represents user think time; without it, VUs hammer endpoints faster than any real user would, inflating throughput and crowding out the system under test. This applies to VU-based HTTP, WebSocket, gRPC, crypto, and extension scripts. Open-model executors (constant-arrival-rate, ramping-arrival-rate) already pace iterations to a target rate, so sleep() there is usually unnecessary — it only ties up VUs without changing the offered load. Browser scripts use page.waitForTimeout() instead; single-iteration functional tests and one-shot connectivity/demo scripts can skip it. In event-driven WebSocket scripts, put the sleep() inside the close handler (see examples/websocket.js) — sleeping in the main function body blocks the event loop before any messages arrive.

• Assert every response. Browser scripts: use expect() from k6-testing — it auto-retries against locators and replaces waitFor() + isVisible() + check() chains. If you need a metric-tracked check() inside an async browser function, the standard check from k6 works fine as long as the predicate is synchronous — await the value first, then check it. Only reach for the k6-utils wrapper (https://jslib.k6.io/k6-utils/1.5.0/index.js) when the predicate itself must await something: a bare async predicate returns a Promise, which is always truthy, so the check silently passes. HTTP/gRPC/WS scripts: use check() for metric-tracked assertions, or expect() for functional tests. Silent failures are worse than loud ones.

• Browser scripts: wrap interactions in try/finally with page.close() in finally, so pages clean up even when assertions throw.

• gRPC scripts: wrap the iteration's calls in try/finally and call client.close() in the finally block, so the connection is released even when a check throws mid-iteration.

• WebSocket scripts: wrap JSON.parse of incoming messages in try/catch — servers can send non-JSON frames, and one bad frame shouldn't kill the VU.

• No let/var at top level — use const, since module-scope state is shared across VUs and mutability there is almost always a bug.

• No deprecated imports — use k6/websockets for WebSockets; both k6/ws and k6/experimental/websockets are deprecated.

Load & breakpoint tests

• Breakpoint tests: ramp offered load, not VUs. Use an open-model executor (ramping-arrival-rate) so the request rate is independent of how slow the system gets; a closed model (ramping-vus) throttles itself as latency rises and masks the breaking point. Pair thresholds with abortOnFail: true and a short delayAbortEval so the run stops (and reports) once an SLO is crossed.

• Track SLO-relevant custom metrics for reporting: per-endpoint latency (Trend), an error Rate, and any domain counters — plus a handleSummary that surfaces p95/p99, throughput (req/s), and error rate. See examples/executors.js and examples/metrics.js.

• Hybrid protocol + browser load tests: wrap the browser flow in try/catch and record failures to a custom error metric instead of letting an expect() throw — a single flaky browser iteration must not abort a long protocol load run.

Browser scripts — recommended practices

If the script imports k6/browser, read browser-best-practices.md and apply all checks. Fix issues and re-validate.

Step 7: Present results

1. Full script with file path

2. Validation output

3. Best-practices notes (issues found, or "all checks passed")

4. Suggested run command

k6 run --vus 10 --duration 30s k6/scripts/api-load-test.js
k6 run k6/scripts/browser-test.js
k6 cloud run k6/scripts/cloud-test.js
k6 cloud run --local-execution k6/scripts/hybrid-test.js
./k6-with-faker run k6/scripts/faker-test.js
K6_BROWSER_HEADLESS=true k6 run k6/scripts/browser-test.js

Step 8: Execute

If the user confirms, run the command.