Snail.MCP.E2E¶

Task environments for .NET apps: run the app with its services, call its API, read logs, run tests

An agent that changed an app can check that the app works, not only that it compiles. The server raises the task's environment — the app built from the working copy, beside the services it needs — and lets the agent call it, read what it printed and run the project's end-to-end tests against it. Built for Snail.Swarm.Agent; each environment is a Nomad job.

Its tools: docs/reference/tools, written from the server itself. Built on Snail.Sdk.

The manifest¶

A project says what its environment holds in .snail/environment.json:

{
  "app": {
    "project": "src/Shop.Api/Shop.Api.csproj",
    "port": 8080,
    "health": "/health",
    "env": { "Mongo__Url": "mongodb://{mongo}:27017", "Rabbit__Host": "{rabbit}" }
  },
  "services": {
    "mongo": { "image": "mongo:8.0", "port": 27017 },
    "rabbit": { "image": "rabbitmq:4-alpine", "port": 5672 }
  },
  "tests": { "project": "tests/Shop.E2E/Shop.E2E.csproj", "env": { "APP_URL": "{app}" } }
}

{name} in a setting becomes that service's host inside the environment, {app} the app's address. Port and health default to 8080 and /health; comments and trailing commas are allowed. Snail.Swarm.Agent keeps a whole sample in tests/Samples/Shop.

How an environment runs¶

• The session's address names the working copy, /mcp?root=/workspaces/<task>/step-N; the task is the copy's first folder, so every step of a task reaches the same environment, and e2e_up replaces it with the current files. A task's id is a DNS label of at most 37 characters; a part's host is e2e-<task>--<part>.
• The environment is the Nomad job e2e-<task>: one task per service from its image, and the app built and run with the .NET SDK image. Before them the server's own task s3l_prepare copies the task's working copy and packages into the allocation; the app sees neither the other tasks' copies nor the shared package volume. After them s3l_keep writes the task's packages back to its own folder of that volume, so the next raise restores from disk.
• Everything sits on one internal Docker network. The way out is the server's proxy, which lets through NuGet's hosts only, with limits of its own: 256 connections, 10 s for a request head, 2 minutes of silence.
• e2e_up and e2e_test run as jobs: the call answers within its wait (60 s unless asked otherwise) with the result, or with the job's job://e2e/… resource, whose text holds the result once it ends. A raise or a test run that could not happen — no manifest, no Nomad, stopped by e2e_down — answers ok: false with its error, hint and type, flagged isError, and its job ends failed with the same words. One environment does one thing at a time: the same call while it runs — a test run with the same filter — waits on the same job, another waits its turn, e2e_down cancels them. A job of one task is unknown to the sessions of another; the last twenty ended jobs of each task stay readable.
• The tests run as the batch job e2e-<task>--tests in the same network, with APP_URL pointing at the app. A run ends on its own time limit, and one a restart left going is stopped at the next start.
• An environment nobody called for 60 minutes is stopped, unless a raise or test run is at work on it.
• The manifest is read only from a file of the copy itself: a link, or anything but a regular file, is refused.
• The server is native (Native AOT) on runtime-deps: it answers its health check about 40 ms after it starts and holds about 40 MB, against 130 ms and 100 MB framework-dependent. It listens on every interface by design — in its container, and under launchd too, since the platform's workers reach it from elsewhere — and refuses a request a browser makes (one with Origin) with 403. Every client is given a session, so one subscribed to a raise or a test run is told of its end.

Running it¶

docker build -t s3l-e2e-mcp:local .

In Snail.Swarm.Agent the server stands beside the MCP kit: SWARM_ENVIRONMENTS=1 sh deploy/stand.sh up (deploy/compose.environments.yml).