Ultra E2e Test MCP Server

ultra-e2e-test

The end-to-end test for TEMPLATE1: a project generated with every feature, then released and updated through each template release, so that init, template-update, releases and MCP publishing are proven on a real repository rather than only in the template's own CI.

Getting started

You need:

• Node 24 (.node-version), for the scripts and every Node module.
• Go 1.26 (services/api-go/go.mod), and golangci-lint for the complete local check; without it, verify.mjs reports golangci-lint as skipped, and CI still runs it.
• Python 3.13 or newer and uv (services/api-py/.python-version), which installs the rest.
• The GitHub CLI, only to apply repository settings with configure-github.mjs.

Each toolchain version is pinned once, in the file named beside it above, so there is no .tool-versions to keep in step; mise and asdf can read those files directly.

1. Install the dependencies of every module present:

   node scripts/setup.mjs
   

2. Run the checks every CI job runs on the code:

   node scripts/verify.mjs             # everything
   node scripts/verify.mjs <module>    # the chassis plus the named modules only
   

   It runs each module's own checks and tests, and the API contract against every task service present. A check it cannot run is reported as skipped, never as passed. A few checks run only in CI, because they need a container engine, a network service or a CI-only tool: actionlint and zizmor on the workflows, npm audit signatures, building and starting each container image, and the report-only scans in security.yml.

3. Start what you are working on. Each module's entry under What's here names the command that starts it.

What's here

• scripts/ — setup.mjs installs every module, verify.mjs runs every module's checks, check-hygiene.mjs guards the repository's shape, check-docs.mjs its documentation, check-contract.mjs holds every task service to the one API contract in scripts/contract/, and configure-github.mjs applies the repository settings (squash merging, the required verify check, security features).
• services/api-go/ — Go task API in Clean Architecture layers. go run ./cmd/api there serves it on port 8080. README
• services/api-ts/ — TypeScript task API with a pure domain core. npm start there serves it on port 8080. README
• services/api-py/ — Python task API, same routes and layers. uv run --directory src python -m api_py.main there serves it on port 8080. README
• services/mcp-server/ — MCP server exposing the task API to an AI assistant. npm start there serves it over stdio, calling the task API at TASK_API_URL; the README shows how to register it with a client. README
• apps/web/ — React single-page app, organised by feature. npm run dev there serves it at http://localhost:5173, with /api passed to a task service on port 8080. README
• packages/ts-library/ — TypeScript library published to npm. npm run verify there builds it and checks the package consumers would install. README
• architecture/ — LikeC4 model of the system. npm run dev there previews every view. README
• docs/ — the documentation index and the rules for keeping it true; docs/adr/ holds the architecture decision records.
• .claude/skills/ — step-by-step procedures coding agents follow for recurring tasks. .github/prompts/ holds Copilot prompt files that wrap one of them; AGENTS.md holds the rules all of them follow.
• .github/ — workflows, issue forms, pull request template, Dependabot and code owners.

Working with a coding agent

AGENTS.md holds the rules every coding agent follows here; CLAUDE.md, GEMINI.md and Copilot's instructions point to it rather than repeating it. For a recurring task, such as recording a decision or taking a template update, the matching procedure in .claude/skills/ is the one to follow. Ask the agent to run node scripts/verify.mjs and to say what it ran before it reports a change as done.

Continuous integration

• verify.yml — on every pull request, every push to main, and in a merge queue: repository hygiene, chassis tests, actionlint, a security audit of the workflows with zizmor, and one job per module — each service job also runs the API contract and starts the service's container image to prove it answers — all feeding the aggregate verify job, which is the only required check (ADR-0002).
• pr-title.yml — pull request titles follow Conventional Commits.
• copilot-setup-steps.yml — the environment GitHub's Copilot coding agent prepares before it works here: every toolchain the selected features need, then node scripts/setup.mjs. It runs on its own only when it changes.
• security.yml — report-only scans that fail only when a scan could not run: gitleaks over new commits and weekly over history, npm audit for every npm lockfile, and a Trivy scan of every container image the repository builds, for fixable high and critical vulnerabilities in its operating-system and language packages.
• security.yml, Go — govulncheck, reporting only vulnerabilities in code the Go service actually calls.
• security.yml, Python — pip-audit over the Python service's lockfile.
• codeql.yml — CodeQL analysis; enable it by setting the repository variable CODEQL_ENABLED=true (needs a public repository or GitHub Advanced Security).
• scorecard.yml — OpenSSF Scorecard: an outside measurement of the practices this repository claims, published and uploaded to code scanning; enable it with SCORECARD_ENABLED=true on a public repository. Some checks measure the project rather than the workflows, and a new or single-maintainer repository scores low on them: Code-Review and Branch-Protection while pull requests merge without a second person's review, Maintained for its first 90 days, SAST until CodeQL has run on recent pull requests, and CII-Best-Practices until the project registers for the badge.
• mcp-publish.yml — after a release, pushes the MCP server's image, built for amd64 and arm64, to GitHub Container Registry and its server.json to the MCP Registry, tokenlessly; enable it with MCP_PUBLISH_ENABLED=true (how).
• release.yml — release-please on main, off until RELEASE_ENABLED=true, which configure-github.mjs sets. Releases start at 0.1.0. GitHub holds the checks of a pull request opened by github-actions[bot] until someone approves them, so releasing is: open the release pull request, Approve workflows to run, wait for verify, merge. A RELEASE_PLEASE_TOKEN secret holding a GitHub App or personal token removes that step.
• Library publishing — with release selected, each release publishes packages/ts-library to npm with provenance once NPM_PUBLISH_ENABLED=true is set and npm trusts the workflow (how).
• architecture.yml — publishes the architecture model to GitHub Pages once PAGES_ENABLED=true is set.

Dependabot proposes grouped updates weekly for every ecosystem present, SHA-pinned actions included.

Taking template updates

This project was generated from ultra-e2e-test, and CHANGELOG.md records the release it came from. When a later release fixes something you want, scripts/template-update.mjs brings the change in: it regenerates the project as the old and the new release would have made it, with this project's name and features, and applies the difference as a three-way merge. What you changed yourself is kept, and a conflict is left to resolve like any merge conflict.

node scripts/template-update.mjs --to vX.Y.Z --dry-run   # what would change
node scripts/template-update.mjs --to vX.Y.Z             # apply, then review, verify and commit

Updates only move forward: a release older than the one the project is on is refused.

The update-from-template skill walks an agent through the whole procedure.

Contributing and security

CONTRIBUTING.md describes the workflow, SECURITY.md how to report a vulnerability privately, and CODE_OF_CONDUCT.md the expected conduct. Guidance for coding agents is in AGENTS.md.

License

MIT