API Chaos Runner Skill for Claude Code

Overview

Most API tests run against perfect networks: <1ms latency, no packet loss, infinite bandwidth, deterministic ordering. Real production isn't like that. Network chaos testing drives the existing tests under controlled network impairment - the team discovers which retry / circuit-breaker / timeout patterns actually hold up before real customers find out.

The canonical open-source primitive is Toxiproxy - Shopify's "TCP proxy to simulate network and system conditions for chaos and resiliency testing" (toxiproxy-readme (opens in new window)). The pattern: sit Toxiproxy between client and upstream; manipulate toxics (latency, timeout, bandwidth, etc.) during test execution.

This skill is build-an-X - the workflow chains the team's existing API tests (Postman / Karate / RestAssured / Tavern / Schemathesis) through a Toxiproxy-managed connection and orchestrates a per-scenario chaos matrix.

When to use

The API has documented resilience requirements (retry on 5xx, circuit-break on 3 consecutive failures, timeout at 5s, etc.) and the team needs to verify them, not just document them.
A new external dependency just got added; the team wants to pressure-test fallback behavior.
An incident postmortem identified a "we should have caught this in testing" item; chaos coverage is the prevention mechanism for that class.
The team has integration tests already and wants to multiply their signal value via fault injection.

If the team is just starting API testing and has no resilience patterns to verify, this skill is overkill - start with happy-path coverage via postman-collections or the language-native equivalents first.

How to use

Confirm documented resilience requirements exist (retry, circuit-breaker, timeout, fallback). If none, start with happy-path coverage instead of chaos.
Pick the chaos primitive - Toxiproxy for per-API chaos in CI (Step 1).
Build the chaos matrix: for each existing test scenario, list the toxics to inject and the expected behavior per condition (Step 2).
Wire Toxiproxy between the app and its upstream and register a proxy per dependency (references/toxiproxy-wiring.md).
Run the matrix, always including a no-toxic control row; capture one JUnit XML per scenario (Step 3).
Aggregate results into a resilience matrix and report which assertions break under which toxic (Step 4).
Inject only toxics that map to a documented expectation; run the full matrix nightly, the control row per PR.

Step 1 - Pick the chaos primitive

Tool	Layer	Best for
Toxiproxy	TCP proxy	Per-connection latency / bandwidth / timeout / drop. Most precise.
Pumba	Docker container	Container-level chaos (kill, pause, network).
Gremlin (commercial)	Multi-platform	Production-grade chaos with audit / approval flow.
LitmusChaos	Kubernetes operator	Cloud-native; experiments declared as CRDs.
`tc qdisc` (Linux native)	Network interface	Lowest level; most setup; CI-friendly only with `--cap-add NET_ADMIN`.

Default recommendation: Toxiproxy for per-API chaos in CI. The others fit when the team is already in those ecosystems (Docker-Compose-heavy projects, Kubernetes-first projects).

Step 2 - Define the chaos matrix

For each existing API test scenario, define a matrix of conditions to run it under:

Scenario	Toxic	Expected behavior
Order create (POST /orders)	None (control)	201 in <500ms
Order create	latency=1000ms	201 in <2s (within timeout budget)
Order create	latency=10000ms	504 with retry-after, OR client gives up
Order create	bandwidth=10kbps	201 (eventual) OR 408 timeout
Order create	reset_peer	502 with retry attempted
Order create	timeout	504; circuit-breaker opens after 3rd

Per toxiproxy-readme (opens in new window), the canonical toxic types include latency, down (forced failure), bandwidth, slow_close, timeout, slicer, limit_data, reset_peer.

The matrix is the load-bearing artifact: what the team expects under each condition is what differentiates resilience verification from "did the test pass?" The Expected column drives the assertions.

Step 3 - Wire Toxiproxy and run the matrix

Sit Toxiproxy between the app and its upstream, register a proxy per dependency, and add/remove toxics around each test run. Point the app at Toxiproxy's listen ports; it forwards to the real upstream when no toxic is active. A matrix runner loops the scenarios, adding one toxic per row and producing one JUnit XML per scenario to aggregate in the report stage.

Full docker-compose setup, control-API proxy registration, toxiproxy-cli toxic commands, and the runner shell script: references/toxiproxy-wiring.md.

Step 4 - Report what broke under what

A successful chaos run produces a resilience matrix report:

## API Chaos Matrix - verdict: REVIEW

| Scenario         | Control | Latency 1s | Bandwidth 10k | Timeout 5s | Reset peer |
|------------------|:-------:|:----------:|:-------------:|:----------:|:----------:|
| POST /orders     |    ✅   |     ✅     |       ✅      |     ✅     |     ❌     |
| GET /orders/:id  |    ✅   |     ✅     |       ❌      |     ✅     |     ✅     |
| DELETE /orders/:id |  ✅   |     ✅     |       ✅      |     ✅     |     ✅     |

### Failures

| Test | Toxic       | Expected                                          | Actual |
|------|-------------|---------------------------------------------------|--------|
| POST /orders | reset_peer | 502 + retry attempted; second attempt succeeds | 502; no retry observed in client logs |
| GET /orders/:id | bandwidth=10k | 200 in <30s | 408 timeout at 10s |

A green matrix isn't the goal - finding where resilience is missing is the goal. A failure under a chaos scenario is a feature request, not a bug in the test.

Choosing what to inject

Match toxics to documented resilience requirements:

Resilience pattern documented	Toxic to inject
Retry on 5xx	`down` (forced 5xx) or `reset_peer`
Timeout after Nms	`latency=N+500` (force the timeout)
Circuit-breaker after 3 failures	`down` for ≥3 requests
Fallback to cache when upstream unreachable	`down` indefinitely
Bulkhead under load	`bandwidth=very-low`
Slow-loris client	`slow_close` on response

Run only the toxics that map to a documented expectation; running every toxic against every endpoint is noise.

Worked example

The orders service documents "retry once on connection reset." The team already has a Postman collection covering POST /orders.

Wire Toxiproxy in front of orders-api-real and register the orders-api proxy (Step 3).
Matrix rows: control (no toxic) and reset_peer. Expected for reset_peer: 502, retry attempted, second attempt succeeds.
Run the matrix. Control passes - 201 in <500ms.
Under reset_peer the run returns 502 with no retry in the client logs, so POST /orders reset_peer is marked failing in the matrix.

Verdict: REVIEW. The documented retry policy is not actually implemented. The output is a fix request against the client (add the retry), not a bug in the test - exactly the resilience gap the matrix exists to surface.

Anti-patterns

Anti-pattern	Why it fails	Fix
Chaos in production	Real users observe; oncall pages.	CI / staging only. Production chaos requires the team's full chaos engineering practice (Gremlin / Litmus + approval flow).
Per-PR chaos matrix	Adds 10+ minutes; team disables.	Nightly chaos runs; PR runs only the control row.
Asserting "chaos must not break anything"	Every system has a breaking point; the test trivially fails.	Assert specific resilience behavior under specific conditions; document the breaking point as accepted.
Using `down` for everything	`down` forces 5xx; doesn't model real-world latency / bandwidth.	Mix `latency`, `bandwidth`, `timeout`, `reset_peer` for realistic mixes.
Skipping the control row	Without control, the matrix can't distinguish chaos failures from test bugs.	Always run a no-toxic scenario as the baseline.

Limitations

Toxiproxy is TCP-level. UDP flaws (DNS resolver weirdness, QUIC) need different tooling.
Doesn't model partial failures within a single connection. Toxiproxy treats each connection uniformly; for "retry on the 3rd request after 2 successes" patterns, layer a counting middleware.
Per-test-suite isolation. When tests run in parallel against the same Toxiproxy instance, they fight over the same toxic state; serialize chaos scenarios or use one Toxiproxy instance per worker.

References

toxiproxy (opens in new window) - main repo: install, control API, toxic types, language-native clients.
references/toxiproxy-wiring.md - docker-compose setup, proxy registration, toxic commands, matrix runner.
Pumba - https://github.com/alexei-led/pumba
LitmusChaos - https://litmuschaos.io/
Principles of Chaos Engineering - https://principlesofchaos.org/
postman-collections, tavern-testing, karate-testing, restassured-testing - example-based test suites that this skill drives through chaos.

Reference

Toxiproxy wiring and matrix runner

View source

Full setup for placing Toxiproxy between the client and upstream, registering proxies, adding toxics, and running the chaos matrix. Referenced from the api-chaos-runner skill's Step 3.

Setup (Docker example)

# docker-compose.test.yml
services:
  toxiproxy:
    image: ghcr.io/shopify/toxiproxy:latest
    ports:
      - 8474:8474   # control API
      - 5432:5432   # proxied DB
      - 8080:8080   # proxied API
  app:
    build: .
    environment:
      DATABASE_URL: 'postgres://user:pass@toxiproxy:5432/db'
      EXTERNAL_API_URL: 'http://toxiproxy:8080'

Per toxiproxy-readme (opens in new window), the application points at Toxiproxy's listen ports rather than the upstream. Toxiproxy forwards to the real upstream when no toxic is active.

Define proxies via the control API

# Register the upstream
curl -d '{"name":"orders-api","listen":"0.0.0.0:8080","upstream":"orders-api-real:8080"}' \
  http://toxiproxy:8474/proxies

Add a toxic during a test

Per toxiproxy-readme (opens in new window):

# 1000ms latency on every request through this proxy
toxiproxy-cli toxic add -t latency -a latency=1000 orders-api

# Bandwidth cap at 10 KB/s
toxiproxy-cli toxic add -t bandwidth -a rate=10 orders-api

# Forced timeout
toxiproxy-cli toxic add -t timeout -a timeout=5000 orders-api

# Remove all toxics
toxiproxy-cli toxic remove orders-api -n <toxic-name>

For a stateless add-test-remove cycle, the language-native client libraries (toxiproxy-python, toxiproxy-node, toxiproxy-ruby, toxiproxy-go) wrap the HTTP API.

Run the matrix

A minimal runner shell script:

#!/usr/bin/env bash
# scripts/chaos-matrix.sh
set -e

PROXY=orders-api
TEST_CMD="npx newman run collections/orders.postman_collection.json -e environments/chaos.json -r cli,junit --reporter-junit-export results-$1.xml"

run_with_toxic() {
  local label="$1"; local type="$2"; local args="$3"
  echo "=== $label ==="
  toxiproxy-cli toxic remove "$PROXY" -n latency 2>/dev/null || true
  toxiproxy-cli toxic remove "$PROXY" -n bandwidth 2>/dev/null || true
  toxiproxy-cli toxic remove "$PROXY" -n timeout 2>/dev/null || true
  if [ -n "$type" ]; then
    toxiproxy-cli toxic add -t "$type" $args "$PROXY"
  fi
  $TEST_CMD "$label" || true   # don't bail; we want the matrix
}

run_with_toxic 'control'   ''        ''
run_with_toxic 'latency-1s' latency  '-a latency=1000'
run_with_toxic 'bandwidth' bandwidth '-a rate=10'
run_with_toxic 'timeout'   timeout   '-a timeout=5000'

The matrix produces one JUnit XML per scenario. Aggregate them in the report stage.