agent-harness

agent-harness

A Claude Code plugin that wraps every code change in a plan → build → verify loop.

The main agent acts as an orchestrator. Three subagents do the work:

• harness-planner — interactive brainstorm + writing-plan; produces a per-task spec and plan with explicit acceptance criteria.
• harness-builder — TDD implementation against the plan.
• harness-evaluator — skeptical verification against per-rubric checks; returns a structured verdict.

The orchestrator loops planner → builder → evaluator until the verdict is pass. Hooks enforce the contract: the main agent cannot edit code; the builder runs only on the right branch; the evaluator can only write its verdict. The build cannot exit until verified.

Installing the plugin

/plugin install agent-harness@<source>

Bootstrapping a project

In any project where you want the harness active, run:

/agent-harness-init

This creates .claude/agent-harness/ with config.toml, a project-context.md template for you to fill in, and copies of the harness scripts and check rubrics. It also adds the project-side allowlist entries to .claude/settings.json and registers .harness/ as a gitignored state directory.

The plugin's hooks fire automatically on SessionStart, PreToolUse, PostToolUse, SubagentStop, and Stop — no project-side hook wiring needed.

Disabling the harness

Per-session: /agent-harness-off and /agent-harness-on. A fresh Claude Code session re-enables automatically.

Layout

Path	Purpose
agents/	The three subagent prompts.
commands/	Slash commands (agent-harness-{init,on,off,status}).
hooks/hooks.json	Hook registrations referencing scripts via ${CLAUDE_PLUGIN_ROOT}.
scripts/lib/	Shared bash libraries (state, config, session).
scripts/harness/	Harness ops scripts the orchestrator and subagents call.
scripts/hook_*.sh	The five hook implementations.
scripts/tests/	Test suites for the scripts.
checks/*.md	Evaluation rubrics the evaluator runs against.
templates/	Per-project bootstrap templates instantiated by /agent-harness-init.

Per-project layout (after /agent-harness-init)

<project>/
├── .claude/
│   ├── agent-harness/
│   │   ├── config.toml          ← required_commands, code_paths
│   │   ├── project-context.md   ← project-specific guidance for subagents
│   │   ├── scripts/             ← copied from plugin
│   │   └── checks/              ← copied from plugin
│   └── settings.json            ← extended with allowlist entries
└── .harness/                    ← gitignored; per-session state
    └── state/<session-id>/

License

MIT — see LICENSE.