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.

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