harness

Harness

Multi-agent sprint orchestration middleware for long-running application development. Two-plugin architecture: a domain-blind core handles all orchestration, and domain skill suites provide profile-specific knowledge.

6 agents: initializer, planner, generator, evaluator, coordinator, releaser. Dual-runtime: Works with both Claude Code and OpenAI Codex CLI. Two plugins: harness (core) + harness-sdlc-suite (software delivery domain skills).

Methodology: PDCA (Plan-Do-Check-Act) + two innovations:

1. Evaluator-never-edits-code -- tool-access-level role purity
2. Authenticity gate -- binary overlay catching generic/template AI output

The core harness can be used standalone with a custom profile -- no domain skill suite required. Domain suites add pre-built profiles with evaluation criteria, artifact taxonomies, and verification procedures.

References:

• https://www.anthropic.com/engineering/harness-design-long-running-apps
• https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

---

Install

Option 1: Claude Code Marketplace (Recommended)

# Installs both core harness and SDLC suite
claude plugin install harness@harness

Then reload:

/reload-plugins

Option 2: Git Clone + Local Install

# Clone into your project
git clone https://github.com/xuzhijie-ownself/harness.git plugins/harness-repo

# Install (Mac / Linux / Git Bash)
bash plugins/harness-repo/install.sh

# Install (Windows CMD)
plugins\harness-repo\install.bat

The install script copies:

• Core skill from plugins/harness/skills/harness/
• Domain skills from plugins/harness-sdlc-suite/skills/ (6 skills)
• Agents, commands, and hooks from the core plugin

Option 3: Codex CLI

If using OpenAI Codex CLI, the plugin auto-loads from the .codex-plugin/plugin.json manifest:

# Clone the repo into your project
git clone https://github.com/xuzhijie-ownself/harness.git plugins/harness-repo

# Codex detects .codex-plugin/plugin.json automatically
# Skills from both plugins are loaded via dual skill paths
codex  # start codex in the project directory

Uninstall

# Claude Code marketplace
claude plugin uninstall harness

# Local install (removes core + all domain skills)
bash plugins/harness-repo/install.sh --uninstall

Update

# Marketplace
claude plugin update harness

# Local
cd plugins/harness-repo && git pull && bash install.sh

---

Commands

Command	Purpose
/harness:start	Scaffold harness for a new project (run once)
/harness:session	Run one supervised sprint round
/harness:run	Continuous coordinator-driven loop (unattended)
/harness:reset	Checkpoint + handoff when context fills (Variant B)
/harness:release	Version bump, changelog, and git tag

---

Roles

Agent	Spawned by	Reference
initializer	/harness:start	plugins/harness/skills/harness/roles/initializer.md
planner	/harness:start	plugins/harness/skills/harness/roles/planner.md
generator	/harness:session, coordinator	plugins/harness/skills/harness/roles/generator.md
evaluator	/harness:session, coordinator	plugins/harness/skills/harness/roles/evaluator.md
coordinator	/harness:run	plugins/harness/skills/harness/roles/coordinator.md
releaser	/harness:release, coordinator	plugins/harness/skills/harness/roles/releaser.md

The harness follows an adversarial PDCA pattern: the generator produces artifacts (Do) and the evaluator grades them (Check). The generator cannot self-approve; the evaluator cannot edit product artifacts. This separation prevents the common failure mode where a model is too lenient grading its own work. Agent files are thin YAML wrappers -- all instructions live in role files as the single source of truth.

---

Authenticity Gate

Every evaluation round applies a binary pass/fail Authenticity Gate after domain criteria scoring. This catches technically-competent-but-generic output -- artifacts that score adequately on domain criteria yet show no evidence of project-specific decision-making.

Dimension	What it checks
internal_consistency	Artifacts share consistent conventions -- structure, terminology, style form a unified whole
intentionality	Evidence of project-specific decisions, not unmodified defaults or template output
craft	Technical fundamentals correct -- hierarchy, structure, naming, formatting
fitness_for_purpose	Deliverables usable by target audience without additional explanation

The gate is dual-side: generators apply a pre-implementation checklist (prevention), evaluators apply a post-grading gate (detection). Any dimension failure fails the round regardless of domain scores.

---