k6-cloud-investigate-test

k6 Cloud test investigator

Structured 9-step workflow for investigating a Grafana Cloud k6 test or run.

This skill is workflow-only. All API calls go through gcx api against the plugin proxy, using v6 for REST and v5 for metrics. For the underlying mechanics (gcx auth, path conventions, endpoint discovery, log queries, script editing, threshold semantics, gotchas), see the k6-manage skill — every step below references it.

The content unique to this skill is:

• the ordered investigation flow (Steps 1-9 below)
• the date-alignment check ("last 7 days" ≠ "last 7 runs")
• the 3-layer pass/fail determination (result vs status vs per-check checks query)
• a worked example with realistic numbers (references/worked-example.md)

Core principles

1. Read before write. Always GET the script before any PUT. gcx k6 load-tests update-script is a write that replaces the live script with whatever file you pass — running it "just to see the URL it hits" has cost users their production scripts. If you need to learn URLs, run any non-mutating command with -vvv --log-http-payload instead.
2. Paginate when enumerating runs. The /test_runs endpoint caps at 1000 rows and gcx k6 runs list --limit 0 does not auto-follow @nextLink. Use the gcx api loop documented in k6-manage §3.
3. Verify date framing. When the user says "last 7 days", "this week", "recent runs" — confirm the most-recent run's created actually falls in that window. Surface the gap if not.
4. check() doesn't fail runs; only thresholds do. And thresholds with zero observations are reported as ✓ pass. See "Threshold semantics" below for the full deep dive; the per-check checks metric query in Step 5 catches both cases.

Prerequisites

gcx installed and authenticated against the user's stack. See k6-manage §1. Verify with:

gcx --context <stack> config check    # expect "✔ Connectivity: online"

Investigation workflow

Step 1: Identify the test and target run(s)

From the user's URL:

• /a/k6-app/tests/<id> → load test (parent of many runs)
• /a/k6-app/runs/<id> → a specific run

To go run → test: fetch the run via gcx api, see k6-manage §2 for the path-shape rules:

gcx --context <stack

and read .test_id from the response. To list runs for a test, see Step 3.

Step 2: Fetch test metadata and script

For the script, follow the GET half of the safe-edit recipe in k6-manage §5 — save a backup if you'll be editing later.

Two script endpoints exist, and the difference matters for investigation. k6-manage §5 documents both: the current load-test script and the per-run snapshot that was actually executed. They drift apart whenever the script is edited after a run. Whenever the question involves "what changed", "why did this run fail", or you're examining a run more than a few days old, also fetch the run-bundled snapshot(s) via k6-manage §5's run-script endpoint and diff against the current load-test script (or against another run's snapshot). The current load-test script is the wrong artifact to reason about a past run.

Step 3: List runs WITH pagination

Use the gcx api + @nextLink loop pattern documented in k6-manage §3 against /cloud/v6/load_tests/<test_id>/test_runs. After collecting all_runs:

Report to user: total run count, date range, latest run date. If "latest run" is more than a day old, call it out — they may believe the schedule is firing when it isn't.

Step 4: Verify date alignment with user's intent

If the user asked for "last 7 days" / "this week" / "recent": filter by date range, not by row count — "7 most recent runs" could span a day or a year depending on how often the test runs.

If len(last7) == 0: surface this to the user immediately. Don't proceed with stale data.

Step 5: Determine pass/fail status

For each run examine three independent layers:

For a run that the user thinks is "failing" but reports result: passed: check the third layer. Common pattern: every iteration's check() returns false but the run still "passes" because no threshold is defined on checks. See "Threshold semantics" below for the full deep dive (zero-observation trap, abortOnFail cloud delay, operator support).

Query the per-check breakdown via v5 (see k6-manage/references/metrics.md §7 for query_aggregate_k6 shape):

Each result entry in the response carries a check label with the check name and a single value: the success ratio (0.0–1.0). For raw success/fail counts, also query increase_nz by (check) (successes) and increase_z by (check) (failures).

Step 6: Fetch metrics for a run

Use the v5 metrics endpoints documented in k6-manage/references/metrics.md. Typical workflow:

Picking the right query method per metric type matters — see the "Query methods" tables in k6-manage/references/metrics.md. For browser tests, the URL/status breakdown comes from labels + label/{name}/values (metrics.md §4/§5), not from a separate tags endpoint.

Step 7: Fetch logs for a run

Follow the Loki recipe in k6-manage §4 (gcx api against /api/plugins/k6-app/resources/logs/..., {test_run_id="<id>"} selector, mandatory X-K6TestRun-Id header, scope start/end to run.created/run.ended). Save the response to /tmp/run_<id>_logs.json and summarise from the file as described there.

Step 8: (If asked) edit the script safely

Use the full safe-edit recipe in k6-manage §5 (GET → backup → edit → k6 inspect → 1-iter smoke → PUT with Content-Type: application/octet-stream → sha256 verify).

When the edit involves thresholds, also read "Threshold semantics" below for the zero-observation trap, the abortOnFail cloud delay, and operator support (==/!= work despite docs).

Step 9: Report

Pick a template based on the question shape.

For single-run investigations ("what's going on with run X"):

For multi-run comparisons ("what's different between passing and failing runs", "did this change cause the failures"):

The table makes side-by-side anomalies obvious (e.g. a column that's identical across passing and failing rules out that dimension as the cause; a column that flips between them is your candidate).

If the user asked for raw data dumps, also save them to /tmp/k6inv/run<id>/ and tell them the paths.

Threshold semantics

Thresholds, not check() calls, are what determine a run's result. The behaviour has several non-obvious corners that mislead first-time investigators, so they're collected here. Step 5 ("Determine pass/fail status") and Step 8 ("edit the script safely") both depend on this section.

What result means

check() calls do not affect result directly. They emit observations into the built-in checks rate metric, which a threshold may then evaluate. If no threshold references checks, failing checks are invisible to the run-level result.

The "zero observations = pass" trap

k6 reports a threshold as ✓ pass when the underlying metric has zero samples — even if the threshold expression would otherwise evaluate to false:

This bites when the script throws before reaching the relevant check() call. Force a check observation in a catch block to make the threshold see something:

Then add 'checks{check:script completed without exception}': ['rate==1.0'] to thresholds. This catches both navigation failures AND later-stage exceptions.

abortOnFail cloud delay

When k6 runs in the cloud, thresholds are evaluated every 60 seconds. The abortOnFail feature may be delayed by up to 60 seconds.

For runs shorter than 60 s, abortOnFail may not fire before iterations complete naturally. The threshold is still evaluated at end-of-run and result flips to failed — abort just doesn't save execution time.

Operators

Docs show <, >, <=, >=. == and != also work even though they aren't in the docs. For rate metrics (value 0.00–1.00), rate==1.0 means "every observation was non-zero" and rate==0 means "all observations were zero."

Workflow-specific gotchas

For the canonical gotcha list (auth expiry, doubled cloud/cloud/, 415 on script PUT, Loki missing X-K6TestRun-Id, script PUT not bumping updated, gcx k6 runs list --limit 0 not following @nextLink), see k6-manage §7. The entries below are specific to this investigation workflow:

Reference

• references/worked-example.md — a synthetic investigation with realistic findings, useful as a template for the Step 9 report.