Centaur E2E Testing

How the system works

e2e/stack.sh deploys the real chart's Slack path (Postgres, console, api-rs, slackbotv2, sandboxes, per-sandbox iron-proxies) to a dedicated kind cluster, with two fakes standing in for the outside world:

• e2e/fakes/fake-slack.ts: a stateful Slack Web/Events/Interactivity API. slackbotv2 runs unmodified against it.
• e2e/fakes/model-server.ts: a scripted model provider. CoreDNS sends provider hostnames to it, so real harnesses reach it through iron-proxy. It answers by a token in the user's message and records every request.

Tests in e2e/tests/*.test.ts run under bun test on the host and reach the fakes, Postgres, and api-rs through NodePorts. e2e/lib/index.ts is the test vocabulary: slack, model, api, cluster, workflows, ironControl, and Thread. Shared Slack identities live in e2e/fakes/slack-fixture.ts. Stack configuration lives in e2e/infra/.

CI (.github/workflows/e2e.yml, called by publish-images.yml and runnable manually against any published tag) runs every scenario against the published images with API keys, and runs harnesses.test.ts alone in subscription (access_token) mode.

Running

e2e/stack.sh up                                   # build images from the tree, deploy
e2e/stack.sh test                                 # all scenarios
e2e/stack.sh test e2e/tests/turns.test.ts         # one file
e2e/stack.sh logs                                 # pods, events, component logs
e2e/stack.sh down

The script uses a private kubeconfig, never the ambient context. Rerun up after changing service code. CENTAUR_E2E_IMAGE_TAG pulls published images, and CENTAUR_E2E_AUTH_MODE=access_token selects subscription auth. Run down before switching from access_token back to api_key.

Writing a scenario

• Act only through the doors a user or operator has: Slack for input and visible output, durable session tables, recorded model requests, api-rs operator routes, and Kubernetes faults. Do not reach into service internals.
• Script each turn with model.says(...) or model.fails(...), then drive it with slack.mention(...), thread.mention(...), and thread.nextTurn(). nextTurn waits for a terminal execution and exactly one finished Slack reply.
• Assert observable outcomes: turn.reply, turn.execution.status, turn.sandbox, and turn.request (what the harness actually sent the provider).
• Use test.concurrent with a turnTimeoutMs-based timeout. Each test makes its own thread and script tokens, so tests must not share state.
• Wait with eventually(...), not fixed sleeps. Use delayMs plus thread.inFlight() to keep a turn running.
• Put a short comment at the top of each file stating the behavior it guarantees. Add a scenario to the existing file that covers that area before creating a new file.
• Extend e2e/lib or a fake only when a scenario needs a new capability, and name helpers in user terms. When a test needs new stack configuration, add it to e2e/infra/values.yaml, and keep fixture values aligned with stack.sh.
• E2E scenarios are slow. Cover cross-service behavior here and leave logic that a unit test can reach to unit tests.

Debugging failures

Failures name the execution and the replies seen. Next, check e2e/stack.sh logs (CI dumps the same output), then the session_executions and session_events rows for the thread key, then the model server's recorded requests.