Validation Coverage

This document explains how harness-starter-kit is validated. The tests prove that the kit can install baseline files, keep documentation references healthy, and run local drift checks. They do not prove that harness adoption reduces agent mistakes; measure that separately with docs/evaluation.md.

Local Validation

Run these checks before changing starter-kit templates, installer behavior, drift scripts, command documents, or README structure. On macOS/Linux, use python3 instead of python when python is unavailable:

python -m unittest discover -s tests
python -m py_compile scripts/apply_harness.py scripts/check_agent_skills_package.py scripts/check_docs_drift.py scripts/check_structure.py scripts/check_encoding_hygiene.py scripts/check_effectiveness_plan.py scripts/check_failure_memory.py scripts/check_decision_memory.py scripts/harness_doctor.py
python scripts/check_agent_skills_package.py
python scripts/check_docs_drift.py
python scripts/check_structure.py
python scripts/check_encoding_hygiene.py
python scripts/check_effectiveness_plan.py
python scripts/check_failure_memory.py
python scripts/check_decision_memory.py
python scripts/harness_doctor.py --target .

Profile README consistency is covered by tests/test_profile_consistency.py, which runs as part of the full unittest suite. Run it directly when adding or changing profile guidance:

python -m unittest tests.test_profile_consistency

Agent Skills package consistency is covered by tests/test_agent_skills_package.py and scripts/check_agent_skills_package.py. Run them directly when changing agent-skills/, Codex plugin metadata, Claude Code plugin metadata, skill guidance, or the skills installation docs:

python -m unittest tests.test_agent_skills_package
python scripts/check_agent_skills_package.py

When Claude Code is installed locally, also run its plugin validator:

claude plugin validate agent-skills

Fixture Smoke Tests

Automated fixture smoke tests cover harness installation for:

• Node.js / TypeScript
• Next.js
• Django
• FastAPI
• Flask
• React
• Spring Boot
• Android / Kotlin Gradle
• Vue
• Go
• Rust

These tests verify that the installer preserves existing files, writes expected profile snippets, and produces runnable generic drift checks.

Manual And Opt-In E2E Checks

Additional end-to-end adoption checks have been run manually against:

• a Node.js ES module project using node --test, repeated installer runs, the TypeScript profile check_harness.py, and intentional drift failures
• a FastAPI project using pytest, mypy, generated drift checks, and the FastAPI profile check_harness.py

FastAPI E2E coverage is available as an opt-in automated test because it creates a virtual environment and installs dependencies:

$env:RUN_FASTAPI_E2E = "1"
python -m unittest tests.test_fastapi_profile_e2e

In GitHub Actions, run the Harness Check workflow manually and enable run_fastapi_e2e to execute the dependency-installing test.

Lifecycle Pilots

Pilot lifecycle tests have validated prompt-first adoption from blank repositories into minimal Django and Next.js projects. These tests verified generic-first adoption, later stack-specific profile absorption, filled measurement plans, and runnable local checks. The Next.js pilot also verified post-adoption cleanup and Git hygiene after removing the local kit clone.

See docs/examples/lifecycle-pilot-results.md for the detailed pilot summary.

Live Adoption Target

harnessworks/harness_starter_kit_django is the active dogfooding repository for this kit. It is used to validate prompt-first adoption, Django profile absorption, /harness update, failure memory, and effectiveness measurement workflows in a real target repository.

This is operational evidence, not proof that harness adoption reduces agent mistakes by itself. Record comparable task outcomes in effectiveness reports.

Dogfood Benchmarks

Additional dogfood reports record harnessed-only target outcomes without claiming effectiveness improvement:

• TodayBus harnessed-only dogfood benchmark for a Next.js public-data target
• Harness ERP Spring/Maven dogfood benchmark for a Spring Boot backend and vanilla frontend target with initial, backend-follow-up, and frontend-follow-up comparable task groups plus non-comparable CI and design-baseline maintenance evidence

Use the dogfood evidence adoption checklist before adding another target as validation or effectiveness evidence.

Example Reports

Use these examples when checking whether a target adoption report is complete:

• examples/node-adoption-report.md
• examples/nextjs-adoption-report.md
• examples/django-adoption-report.md
• examples/flask-adoption-report.md
• examples/spring-adoption-report.md
• examples/node-effectiveness-report.md

What This Does Not Prove

Fixture tests, smoke tests, and lifecycle pilots validate installation behavior, runnable checks, and measurement readiness. They do not prove that harness adoption reduces repeated agent mistakes, CI failure rates, or human rework.

Use docs/evaluation.md and docs/templates/effectiveness-report.md to record comparable before/after or harnessed-only observations.