Agent Harness

agent-harness

Project templates designed for letting an AI agent — Claude Code, Codex, or anything similar — write the code. One philosophy, multiple stacks:

Soft rules in natural language don't reliably survive a long agent session. The only reliable way to enforce a rule on an agent is to make it stop the build: a failing diagnostic, a failing test, a rejected commit. The agent reads the error and self-corrects.

Every template ships guardrails wired into the toolchain at error severity, plus a CLAUDE.md documenting the lifecycle (issue → branch → commit → PR) and four git/Claude hooks enforcing it.

Harnesses

Stack	Where	Templates	Extras
.NET 10	dotnet/	cli, etl-api	15 Roslyn analyzers (CI0001–CI0015), 5 architecture-test fixtures
Expo / React Native	expo/	app	Strict TS + custom ESLint rules + dependency-cruiser + coverage gates, RevenueCat baked in, staged iOS App Store submission workflow driven by Claude skills (SUBMISSION.md state machine)

.NET

git clone https://github.com/ryan75195/dotnet-agent-harness
dotnet new install .\dotnet-agent-harness\dotnet
dotnet new cli -n MyTool
cd MyTool
.\setup.ps1

Full docs: dotnet/README.md

Expo

git clone https://github.com/ryan75195/dotnet-agent-harness
.\dotnet-agent-harness\expo\new-app.ps1 -Name MyApp
cd MyApp
.\setup.ps1

What you get:

• Guardrails at error severity: no comments (local/no-comments with inline-disable comments made inert), one exported component per file, 60-line functions, no any, strict tsconfig, dependency-cruiser layer rules (lib ← components ← app, features isolated), 80% coverage on lib/features, a test file required for every module, and a pre-commit hook that runs the lot.
• Submission harness: SUBMISSION.md is a staged state machine from prerequisites to App Review. The template's CLAUDE.md routes each stage to a Claude skill — submission-doctor (local readiness audit), asc-setup (Playwright-drives App Store Connect: app record, subscription products, metadata), revenuecat-setup (Playwright-drives RevenueCat: entitlements, offerings, API keys), build-and-submit (EAS build/submit + attaching IAP to the version). Skills are idempotent and resumable; secrets never enter the repo.
• RevenueCat baked in: src/lib/purchases/ wraps configuration and a useSubscription hook; a paywall screen renders the current offering.

Claude Code plugin

The repo doubles as a Claude Code plugin marketplace. The agent-harness plugin (in plugin/) ships:

• Scaffolding skills — new-dotnet-cli, new-dotnet-etl-api, new-expo-app — which scaffold from the templates above, stamp the project with .harness.json provenance, and verify the guardrails pass.
• harness-update — pulls the latest template changes into a previously-scaffolded project: PowerShell scripts classify each changed harness-owned file (clean → copied, customized → flagged), the agent merges the flagged ones, and the project's own gate must pass before the stamp bumps.
• Feedback capture — the templates' git hooks log every pre-commit gate failure (which gate, output tail, staged diff) to ~/.agent-harness/feedback/, link the eventual fix commit, and the plugin prompts the agent to annotate each event. The harness-report skill later clusters those events into digest issues on this repo — raw material for turning recurring mistakes into new guardrails.

Install:

/plugin marketplace add ryan75195/dotnet-agent-harness
/plugin install agent-harness@agent-harness

Local development: claude --plugin-dir .\plugin. A local checkout can be pinned for offline scaffolds/updates in ~/.agent-harness/config.json: { "repoPath": "C:/path/to/dotnet-agent-harness" }.

CI

.github/workflows/template-ci.yml scaffolds, builds, and validates every template on each push: dotnet build + dotnet test for the .NET templates; typecheck + lint + dependency rules + tests + seeded-violation checks for the Expo template.

Development

.\dotnet\template-tests\scaffold-and-build.ps1 cli
.\dotnet\template-tests\scaffold-and-build.ps1 etl-api
.\expo\template-tests\scaffold-and-validate.ps1