E2e Alertmanager Test

e2e-alertmanager-test

Takes the exp_alerts fixtures out of a Prometheus promtool-style unit test file, POSTs each one to a real Alertmanager's /api/v2/alerts endpoint, and renders what the resulting notification would look like in several formats. It does not test routing/inhibition logic inside this tool itself — it exercises whatever routing your target Alertmanager instance is actually configured with, and shows you the rendered output.

When to use this skill

• Before merging an alert change, to see the actual notification content (subject line, body, Slack attachment) a human on-call would receive — not just "does it fire," but "does it read well."
• As a CI step against a throwaway/ephemeral Alertmanager container, producing a rendered-output artifact that reviewers can diff between runs, the same way a UI snapshot test is diffed. Fed a skeleton config (see below), this run is fully hermetic — see Achieving full hermeticity in CI.
• This tool's own external dependency is the Alertmanager instance, not historical data — unlike alert-hysteresis/stale-alerts-analyzer, it needs no real production history, so a fresh, disposable Alertmanager instance is sufficient and the run is fully reproducible.

Prerequisites

1. A Prometheus unit-test file with exp_alerts fixtures — typically the output of autogen-promql-tests (see that skill) with real, filled-in labels/annotations, or a hand-written promtool test file. Test cases whose exp_alerts is empty (asserting "does not fire") are silently skipped — there's nothing to render for those.
2. A running Alertmanager reachable over HTTP, defaulting to http://localhost:9093. For CI or local preview, run one disposably:

   docker run -d -p 9093:9093 prom/alertmanager
   

Achieving full hermeticity in CI

The e2e-alertmanager-test binary itself is never hermetic on its own — it always needs to POST to a real, reachable Alertmanager. But the routing config that Alertmanager instance loads is entirely under your control, and that's where every external dependency can be removed:

1. Treat the team's real alertmanager.yml — the one that governs production route: and inhibit_rules: — as the source of truth for routing/grouping/inhibition logic. That's the logic worth exercising in a test.
2. Build a skeleton config for CI that keeps that routing tree intact but overrides every receiver's connection details — smtp_smarthost, Slack api_url, PagerDuty/webhook URLs and keys — with local, no-op values. For example, point smtp_smarthost at a local SMTP catcher (e.g. MailHog/smtp4dev) instead of the real relay, so the receiver still accepts a connection and the message headers can be rendered and diffed, but no mail actually leaves the runner.
3. Start the ephemeral Alertmanager for the CI job with the skeleton config, not the real one, and point --alertmanager-url at it.

With this setup, the Alertmanager under test still runs the real routing/grouping/inhibition rules — so a routing regression still surfaces in the diff — but delivery stops at the local catcher: no real email, Slack message, or PagerDuty incident is ever created. That's what makes this tool's CI output (rendered notification content, including SMTP headers) safe to treat as a fully hermetic, deterministic artifact. Generating and maintaining the skeleton config (importing the team's routing tree while stripping receiver secrets) is on you — this repo doesn't ship tooling for that merge/override step, and --alertmanager-config (below) does not perform it either; it only enriches this tool's own rendered output text.

Keep the skeleton config's routing tree in sync with the real one, ideally generated from it rather than hand-maintained — if it drifts, a green CI run stops guaranteeing anything about production routing.

Setup

go build -o bin/e2e-alertmanager-test ./cmd/e2e-alertmanager-test
# or: go install github.com/conallob/o11y-analysis-tools/cmd/e2e-alertmanager-test@latest

Usage

e2e-alertmanager-test [options]

Typical invocations

# Plain-text email preview
e2e-alertmanager-test --tests=./alerts_test.yml --output=email

# Full HTML email rendering
e2e-alertmanager-test --tests=./alerts_test.yml --output=email-html --full

# Slack attachment JSON preview
e2e-alertmanager-test --tests=./alerts_test.yml --output=slack

# Everything, redirected to a diffable artifact for CI
e2e-alertmanager-test --tests=./alerts_test.yml --output=full > notifications.txt

Reading the output

Each test case with non-empty exp_alerts becomes one numbered "Test #N", printing the alert name and then the rendered output in the requested format(s), followed by a summary:

═══════════════════════════════════════════════════════════
Summary
═══════════════════════════════════════════════════════════
Total test cases: 3
Successful: 3
Failed: 0

Exit code 1 if any alert failed to POST successfully (e.g. Alertmanager unreachable or returned non-200), 0 otherwise. Note: "successful" here means the POST to Alertmanager succeeded, not that routing/inhibition matched any particular expectation — this tool doesn't assert routing outcomes, it only renders notification content.

Agent workflow

1. Confirm a unit-test file with real (non-TODO) exp_alerts exists — if not, point the user at autogen-promql-tests first.
2. Confirm an Alertmanager is reachable at --alertmanager-url (start a disposable one via Docker if none is running and this is a local/CI context). Never point this at a production Alertmanager — it really posts alerts, and a receiver with live connection details will really dispatch them. For CI, load the disposable instance with a skeleton config that keeps real routing/grouping/inhibition but swaps receiver connection details for local no-op ones, so the run stays hermetic even though it exercises real routing logic.
3. Run with --output=full to capture every rendering for a comprehensive review, or --output=slack/--output=email-html when a user specifically wants to review one channel's format.
4. For CI: redirect output to a file and store it as a build artifact so subsequent runs can be diffed to catch unintended notification-content regressions from template/label changes.