harn

harn — Harness Engineering Plugin for Claude Code

Deterministic guardrails, quality gates, and progressive disclosure for AI coding agents.

Why Harness Engineering?

The underlying AI model matters less than the system built around it. A well-harnessed agent produces consistent, safe, high-quality output regardless of the model version. Read the foundations.

Quick Start

claude plugins marketplace add sliday/claude-plugins
claude plugins install harn

Then in any project:

/harn:init        # scaffold your harness

Commands

Command	Description
/harn:init	Scaffold a complete harness — AGENTS.md, hooks, guardrails
/harn:guard	Generate or update security guardrail hooks (PreToolUse)
/harn:gate	Generate or update quality gate Stop hook
/harn:analyze	Audit your project against the CAR framework
/harn:agents-md	Build or rebuild AGENTS.md with progressive disclosure
/harn:format	Generate PostToolUse auto-formatter hook
/harn:trace	Generate PostToolUse trace logging hook for observability
/harn:checkpoint	Generate resume artifacts (LEARNED.md, CHECKPOINT.json)
/harn:mcp-check	Generate MCP server health check hook
/harn:context	Monitor context window usage and apply backpressure

What /harn:init Creates

• AGENTS.md — Agent behavior spec with progressive disclosure
• .claude/settings.json — Hook configuration
• .claude/hooks/security-guard.sh — PreToolUse guardrail
• .claude/hooks/quality-gate.sh — Stop hook for type/lint/test checks
• .claude/hooks/auto-format.sh — PostToolUse formatter (optional)
• LEARNED.md — Persistent gotchas and decisions log

CAR Framework

Every agentic system has three layers:

• Control — Constraints and specifications (AGENTS.md, linters, tests, architectural rules)
• Agency — Tools and interfaces (MCP servers, sub-agents, browser access)
• Runtime — Execution and recovery (hooks, compaction, retries, rollback)

Run /harn:analyze to score your project across all three. Learn more at harn.app.

Dependencies

• jq — JSON processing (used by hooks)
• python3 >= 3.6 — Script execution
• Language-specific toolchain for quality gates (e.g., tsc, cargo, go vet, ruff)

Troubleshooting

If a hook is blocking legitimate work or the agent is stuck in a loop, see references/escape-hatches.md.

Knowledge Base

Curated articles on harness engineering — security, context, evals, specs, and tools: harn.app/kb.

License

MIT