HelloWorldBlazor

Dark Agent Factory 🌙🏭

A lights-out, autonomous AI software factory. Drop it into any repo, give it a project description, and let it build, test, review, and deploy while you sleep.

The factory starts agnostic. Given a request like "build me an ASP.NET Blazor app", it:

1. Detects the tech stack from the request.
2. Looks up the skills it needs locally (find-skills).
3. If a skill is missing, pulls it from a registry (skill-installer) — local FS → cache → GitHub → marketplace.
4. If still missing, it creates the skill itself (skill-creator).
5. Spawns specialized subagents (planner, architect, coder, tester, reviewer, devops, security, docs, monitor) — each with the skills they need loaded into their isolated context window.
6. Runs the lights-out cycle: plan → architect → code → test → review → deploy → monitor — logging every decision as a numbered ADR.

Run artifacts land in factory/runs/<timestamp>/.

Use as a template for new projects

Option A — GitHub template repo (recommended)

1. For every new project, click "Use this template" → Create a new repository. You get a clean copy with no history.

Option B — degit / shallow clone (no GitHub needed)

npx degit your-org/agentfactory my-new-project       # clean copy, no .git
# or
git clone --depth=1 <this-repo> my-new-project
cd my-new-project; Remove-Item -Recurse -Force .git; git init

Per-project bootstrap (always 4 steps)

cd my-new-project

# 1. Reset run state (factory artifacts; don't carry across projects)
Remove-Item -Recurse -Force factory\runs\*, factory\checkpoints\* -ErrorAction SilentlyContinue

# 2. (Optional) Keep the 4 bootstrap ADRs; new project ADRs append from 0005+.

# 3. Edit the intake
notepad factory\intake.md           # name, goals, constraints, cost cap, deadline

# 4. Open the folder in Claude Code or Copilot CLI and prompt:
#       "Run the factory on the intake."
#    The agent auto-loads AGENTS.md → ORCHESTRATOR.md and starts the 11-phase loop.

Run artifacts and decisions land under factory/runs/<UTC-timestamp>/ and docs/adr/.

Repo layout

Supported runtimes

The factory targets Claude Code and GitHub Copilot CLI primarily, but AGENTS.md makes it work with any agent that follows the open spec (Cursor, Codex, etc.).

Documentation

• Agent contract
• Orchestrator run loop
• Architecture Decision Records
• E2E Blazor dry-run walkthrough

GitHub Copilot cloud-agent integration (optional)

The factory can fan work out to the GitHub Copilot cloud agent instead of running every task locally. See ADR-0005.

How it works

1. Set dispatch_mode in factory/intake.md to one of:
   • local (default) — everything runs on your machine.
   • cloud-fanout — every coder/tester task becomes a GitHub issue assigned to Copilot; the cloud agent codes it in its own sandbox and opens a PR. Local gates (security, reviewer) still verify before merge.
   • hybrid — only tasks marked parallel-safe: true go to the cloud; the rest stay local.
2. Make sure .github/workflows/copilot-setup-steps.yml is on the default branch (it ships with this template) — it provisions Python + pyyaml + the factory's validators inside the cloud agent's sandbox.
3. Enable the Copilot coding agent on the repo (Settings → Copilot → Coding agent) and ensure gh auth status works locally.
4. Kick off a run. The orchestrator delegates eligible tasks to the cloud-dispatcher subagent (.claude/agents/cloud-dispatcher.md), which uses the github-cloud-agent skill to create issues, monitor PRs, and request reviews via the gh CLI.

The cloud agent can also be requested as a PR reviewer by the local reviewer subagent (gh pr edit <n> --add-reviewer Copilot).

License

MIT — see LICENSE.